THE ARGO PROJECT: Machine vision based motion capture for tracking the trajectory of the pose of a mobile rigid body. Geoffrey Allan Liggins B. A. Sc. University of British Columbia, 1993 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENT FOR THE DEGREE OF MASTER OF APPLIED SCIENCE in THE FACULTY OF GRADUATE STUDIES DEPARTMENT OF MECHANICAL ENGINEERING We accept this thesis as conforming to the required standard THE UNIVERSITY OF BRITISH COLUMBIA April, 1998 © Geoffrey Allan Liggins, 1998 In presenting this thesis in partial fulfilment of the requirements for an advanced degree at the University of British Columbia, I agree that the Library shall make it freely availabe for reference and study. I further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the head of my department or by his or her representatives. It is understood that copying or publication of this for financial gain shall not be allowed without my written permis sion. Department of Mechanical Engineering The University of British Columbia 2324 Main Mall Vancouver, B. C. Canada V6T 1Z4 Date: Abstract The objective of the Argo Project is to develop a tool that will track in real-time the motion of unconstrained, self-propelled, model ships in seakeeping tests done in towing tanks and manoeu vring basins. To meet the unconstrained requirement, the tracking system must be non- contact and can not interfere with the operation or motion of the model ship. An additional operating re quirement is that the sensor must cover an area in excess of thirty square metres. An optical based sensor was selected as it satisfied these constraints. Tracking the motion of the model ship is achieved with a predictive, extended Kalman filter (EKF), using feature point extraction from multiple synchronized images. The EKF is used be cause it can readily integrate and filter multiple noisy data sets. As well, it can generate an estimate of the pose, namely the position and orientation, of the model ship relative to the reference frame of the test tank. While this project is focused on ship tracking there are many other applications for a system of this kind. TM The system under development makes use of the Qualisys camera and video processor hard ware that extract image feature points and return them to a host computer. The incoming image TM feature points are then fed into tracking software developed in MATLAB . The tracking software uses estimates of the image to do feature point correspondence and sorts the incoming data vector into the expected order. The sorted data vector is then used as the input vector for the EKF which computes the photogrammetric equations and computes the state vector for the pose of the mobile object being tracked. This work is being undertaken at the University of British Columbia (UBC), Deparment of Me chanical Engineering, Maritime Engineering and Naval Architecture Research Laboratory. The organizaitons that assisted in this research effort are the Centre for Cold Ocean Resources Engi neering, Intelligent Systems Group (C-CORE) and the National Research Council - Institute for Marine Dynamics (NRC-IMD). ii Table of Contents Abstract ii List of Tables vList of Figures viList of Abbreviations and Nomenclature viiAcknowledgements xiiPreface xiv Chapter 1: Introduction 1 1.1 Motivation1.2 Motion Tracking 3 1.3 Thesis Structure 4 Chapter 2: Current Motion Capture and Analysis Technology 6 2.1 Introduction 6 2.2 Selection of a Motion Capture System 6 2.2.1 Requirements for monitoring model ship motions 7 2.2.2 Evaluation of sensor systems 9 2.2.3 Imaging Technology 12 2.2.4 Markers 14 2.2.5 Commercial motion capture systems 16 2.2.6 Selected motion capture system 8 2.3 Motion Analysis Techniques 19 2.3.1 Photogramme try 20 2.3.2 Tracking 2 2.4 Related Research Projects 5 2.5 Summary 27 Chapter 3: Photogrammetry 9 3.1 Introduction3.2 Coordinate Frame Definitions 30 3.3 The Pin Hole Camera and Colinear Theory 31 3.4 Systematic Errors 34 3.5 Direct Linear Transformation (DLT) 6 3.6 Photogrammetry and Pose Extraction 37 3.7 The Correspondence Problem 41 3.8 Summary 4Chapter 4: State Estimation, Tracking and the Extended Kalman Filter 48 4.1 Introduction4.1.1 State Variables 50 4.2 Linear State Estimation and The Discrete Linear Kalman Filter 51 4.2.1 Linear stochastic system model 54.2.2 System Noise and Uncertainty 2 4.2.3 State Estimation 54 4.2.4 The Kalman Gain and the State Error Covariance 57 4.2.5 Kalman filter regulation 9 4.2.6 Kalman Filter Initialization 61 4.3 Nonlinear State Estimation and the Extended Kalman Filter (EKF) 64.3.1 Linearization 64 4.3.2 About the Extended Kalman Filter 66 4.3.3 Argo Project EKF requirements 7 iii 4.4 Design and implementation considerations 67 4.5 Summary 70 Chapter 5: Argo Project Implementation 72 5.1 Introduction5.2 Methodology 3 5.2.1 State Variables and the System Dynamics Model 75.2.2 Observable Outputs and the Observation Model 6 5.2.3 Expected Errors and Covariance Models 80 5.2.4 Data Acceptance and Rejection, and Pair Matching 82 5.2.5 Initialization of Tracker and State Error Covariance 5 5.2.6 Incoming Data and Simulated Data Generation 86 5.3 Hardware Implementation 87 5.3.1 Video Cameras and Processors 88 5.3.2 Camera Placement 90 5.3.3 Host Computer 1 5.3.4 Target 95.4 Software Implementation 3 5.4.1 Interfacing with Qualisys hardware 94 5.4.2 Program Initialization 95 5.4.3 The Main Program Loop5.4.4 Simulation and subroutine Softcam 96 5.4.5 Diagnostic subroutines5.5 Summary 97 Chapter 6: Experimental and Simulation Results 99 6.1 Introduction6.2 Evaluation by Simulation 96.3 Experimental Testing 102 6.3.1 Experimental Procedure 103 6.3.2 Experimental Results 4 6.4 Summary 107 Chapter 7: Conclusions and Recommendations 118 7.1 Conclusions 117.2 Future Work and Recommendations 120 Bibliography 122 Appendix A: Motion Capture Systems 13A.l Abbreviations 13A. 2 Video Based Motion Capture Systems Comparision Tables 132 Appendix B: Homogeneous Transformations 136 B. l Basic affine transformationsB.2 Roll Pitch Yaw Translation Transformation 137 B.3 Yaw Pitch Roll Translate Transformation 8 B. 4 Pose extraction from triangulation results 139 Appendix C: Direct Linear Transformation (DLT) 141 C. l Direct Linear Transformation Camera Model 14C.2 DLT Calibration Equations 145 C. 3 Triangulation using DLT 8 Appendix D: Initialization of the State Error Covariance for the EKF 150 D. l Nomenclature 15D.l.l Initial State Error Covariance Model 151 iv Appendix E: Fully expanded observation model equations 157 Appendix F: Detection and Rejection of False Data 164 F.l Nomenclature 16F.2 Validation Test (Detection) 16F.3 Removing false data (Rejection) 6 V List of Tables Sensor System Comparison 10 Units of Figures 6.1 - Figures 6.13 108 Research Video Based Motion Capture Systems 132 Commercial Video Based Motion Capture Systems 3 Commercial Video Based Motion Capture Systems: Specifications 134 Commercial Video Based Motion Capture Systems: Tracking and Data Reduction 135 vi List of Figures 1.1 Ship's coordinate frame and terminology 2 3.1 Feature point position in the Camera's and the World coordinate frames 31 3.2 Pin Hole Camera 32 3.3 Pose of Mobile Rigid Body and Target in World frame 39 3.4 Ambiguous Correspondence of Stereo Images [modified form Nalwa 93] 42 3.5 Epipolar Constraint [Nalwa 93] 44 4.1 Discrete Kalman Filter Flowchart 56 4.2 Full Discrete Extended Kalman Filter Flowchart 63 4.3 Partial Discrete Extended Kalman Filter Flowchart 8 5.1 Argo Project Tracking Algorithm 74 5.2 Image coordinates acceptance and rejection 83 5.3 Argo Project Hardware Configuration 7 5.4 Qualisys Calibration Target Geometry 92 5.5 NRC Tree Target Geometry 96.1 Physical Setup Represented in Simulation 109 6.2 Simulation Results: Pose 110 6.3 Simulation Results: Pose Error6.4 Simulation Results: Pose Velocity Ill 6.5 Simulation Results: Pose Velocity Error Il6.6 Simulation Results: Pose Acceleration 112 6.7 Simulation Results: Pose Acceleration Error 116.8 Physical Setup of Experimental Testing 3 6.9 Experimental Pendulum Geometry 116.10 Experimental Results: Pose 114 6.11 Experimental Results: Pose Velocity6.12 Experimental Results: Acceleration 115 6.13 Experimental Results: Roll Angle 6 6.14 Experimental Results: Frequency Comparison of Roll Angle 117 6.15 Experimental Results: Frequency of Argo Tracked HeaveC. 1 Pin Hole Camera 141 C.2 Feature point position in the Camera's and the World coordinates frames 143 vii List of Abbreviations and Nomenclature Supporting Organizations UBC-ME University of British Columbia, Department of Mechanical Engineering C-CORE Centre for Cold Ocean Resources Engineering MUN Memorial University of Newfoundland MUN-OERC Memorial University of Newfoundland, Ocean Engineering Research Centre NRC National Research Centre of Canada NRC-IMD National Research Centre of Canada, Institute for Marine Dynamics NSERC National Science and Engineering Research Council of Canada General terms CG Centre of Gravity DLT Direct Linear Transformation EKF Extended Kalman Filter IR Infra Red CCD Charge Coupled Device LED Light Emitting Diodes PSD Position Sensitive Detector PSP Position Sensitive Photodiode ARGO Project terms MRB Mobile Rigid Body RB Rigid Body NDOF number of degrees of freedom, default = 6 NORD numerical order of kinematics model, default = 3 M total number of cameras used in system viii m camera index number N NM nm n u\ - 11 total number of markers on target frame marker index number number of observed markers in a image observed markers index number horizontal image coordinate in camera m for marker n vertical image coordinate in camera m for marker n DLT parameters for camera m Coordinate frame terms Note: All reference frame coordinate systems refered to in this document follow the right handed convention unless other specifically noted. Positions .AAA. \XB> VB' ZB> .WW w. (Xn » V/i ' J .WWW. \xc>yc>zc> . c c t\ \xn' Vn' Zn ' , m m (Xn ^n' Zn' . T T T. (Xn> 3V Zn> . MRB MRB MRB. yxn ' vn ' zn ' . W W W . VXM/?B' yMRB> ZMRB) . MRB MRB MRB. \XT 'yt 'ZT ) position of reference frame B in the coordinate system of reference frame A position of marker n in the World coordinate system position of a Camera reference frame in the World coordinate system position of marker n in a generic Camera coordiante system position of marker n in the coordiante system for camera m position of marker n in the Target coordinate system position of marker n in the Mobile /?igid Body coordinate system position of the Mobile 7?igid Body reference frame in the World coordi nate system position of Target reference frame in the Mobile /?igid Body coordinate system ix Orientations roll angle of reference frame B about the x axis of reference frame A QB pitch angle of reference frame B about the about y axis of the reference frame A yaw angle of reference frame B about z axis of the reference frame A orientation of a Camera reference frame with respect to the World coordinate frame orientation of camera reference frame m with respect to the World coordinate frame www VMRB' ®MRB> §MRB orientation of the Mobile i?igid Body reference frame with respect to the World coordinate frame $B w nw .w Vc>Qc>*bc w nw ,w Coordinate system transformations BRPYTA BYPRTA BHA Roll-Pitch-Yaw-Translation homogeneous transformation from reference frame A to reference frame B Yaw-Pitch-Roll-Translation homogeneous transformation from reference frame A to reference frame B homogenous transformation from reference frame A to reference frame B Vectors 4 VB\C vector describing pose of reference frame B with respect to reference frame A vector from reference frame or point B to reference frame or point C with re spect to reference frame A Kalman Filter Terms q number of state variables = NORDNDOF = 18 r number of estimated observations = 2-M-N k discrete time index (k e integer set) / Identity matrix a linear system dynamics model for single DOF [NORD x NORD] Ak linear system dynamics model [q x q] p variance matrix of system noise for single DOF [NORD x NORD] Ck linearized observation model [r x q] Ck(sic' ^) nonlinear observation function [rxi] ek observation-estimatation error [rxi] ek estimated observation-estimatation error [rxi] ek* state estimation error [pi] gk observation vector based on actual state [rxi] gk estimated observation vector based on state estimate [rxi] 4-gk estimated observation vector based on preliminary state estimate [rxi] gk observed data vector (incoming image coordinates vector) [NM x 1] Kk Kalman gain matrix [q x r] r error covariance matrix for a single DOF [NORD x NORD] Pk estimate (error) covariance matrix [q x q] Pk^ intermediate estimation filter covariance [q x q] Qk variance matrix of system noise vector [q x q] Rk variance matrix of observation noise vector [r x r] sk actual state vector [q x 1] xi estimated state vector [q x 1] ».* preliminary state estimate [q x 1] it observation noise vector [rxi] <rtn vm ^nw' ^>nv horizontal and vertical observation noise for camera n marker m system noise vector [q x 1] 0>1..6 system noise for a given degree of freedom T sample period £<y> expected value of y xii Acknowledgements I thank Dr. Sander Calisal, my supervisor, in the Department of Mechanical Engineering, Univer sity of British Columbia (UBC-ME), for his support during the course of this research. I thank Dr. Ray Gosine of the Centre for Cold Ocean Resources Engineering (C-CORE) and Me morial University of Newfoundland (MUN) Faculty of Engineering and Applied Scence, for su pervision and sponsoring my work. I greatly appreciate the opportunity to work with a world class organization such as C-CORE. I thank Dr. Elizabeth Croft of the Department of Mechanical Engineering, University of British Columbia (UBC-ME) for her patience, her valuable advice regarding the Kalman filter, and her ed itorial input. I thank Mr. Mike Sullivan of the National Research Council of Canada Institute for Marine Dy namics (NRC-IMD) for his assistance with the experimental phase of the project and his willing ness to discuss the problems associated with motion tracking of floating objects. I thank the Ocean Engineering Research Centre (OERC) at Memorial University of Newfoundland (MUN-OERC) for use of their target tracking video system hardware for the trials at the NRC-IMD. I thank Mr. Ronan Oger for his support and the many discussions we had. I would also like to include my family and friends for their encouragement that made this possible. I thank Mr. Luca Filipozzi, Mr. Tim Lam, and Mr. Allan Liggins, my father, for thier support and editorial input. This work was done in cooperation with the Centre for Cold Ocean Resources Engineering, Intel ligent Systems Group (C-CORE) and with the National Research Council of Canada, Institute for Marine Dynamics (NRC-IMD). Partial financial support for this work was provided through a of NSERC research grant held by Dr. Gosine. The equipment used in the experimental work belongs to NRC-IMD, C-CORE, MUN-OERC and UBC-ME. xiii Preface The Argo Project is named after the constellation Argo visible in the southern hemisphere. This collection of points of light represent Jason's ship Argo, from Greek mythology. The Argo Project tracks the rigid body motion of a ship model in manoeuvring tests by observing reflective markers, namely points of light, strategically placed on the model. xiv In memory of Gordon Liggins. XV Chapter 1: Introduction Chapter 1: Introduction The Argo Project explores the problem of quantitatively assessing the low frequency, large am plitude rigid body motions of an untethered model ship in manoeuvring and seakeeping tests. The Argo Project makes use of video based motion capture technology and the extended Kalman filter (EKF) to track the trajectory of the pose, namely the position and orientation, of a model ship in the reference frame of a test tank. 1.1 Motivation Life aboard small commercial and passenger vessels in heavy seas is uncomfortable and dan gerous to passengers, crew and vessel alike. Passenger vessels, ferries, pleasure craft, and fishing vessels must have good ride characteristics (normally referred to as seakeeping characteristics) in order to adequately provide comfort for passengers and safe working platforms for crew. Even more important than seakeeping properties is the safety of vessels in rough seas. The September 28th, 1994 [BNS 94] sinking of the passenger ferry Estonia off the Swedish coast illustrates the po tential dangers of sea transportation. Every year in British Columbia, Canada, several vessels are lost due to rough sea conditions. Despite these hazards, the dynamics of vessels in rough seas are not fully understood. Proper assessment of the seakeeping properties through model tests allows designers to evaluate potential hullforms and if necessary make any changes to the vessel prior to construction. Traditional seakeeping test protocols have the model ship pushed forward along the length of a towing tank in a head sea. During these tests the model ships are instrumented to record heave, pitch, drag force and vertical accelerations at the bow and centre of gravity (CG). Figure 1.1 illus trates the definitions of ship motion terms. To allow for instrumentation of the model ship, its mo tions are typically constrained to move only in heave, pitch and occasionally surge. The 1 Chapter 1: Introduction assumption is made that roll, yaw and sway are minor motions and are held fixed. For quartering, beam and following seas this assumption fails. Figure 1.1: Ship's coordinate frame and terminology Using free running, unconstrained model ships and floating bodies, investigators can achieve a closer representation of a ship's response to wind, wave and current excitation in model scale. The requirements of the system to measure the motion of an unconstrained floating body in a ma noeuvring tank are: •must be non-contact so as not to constrain the floating body. •must be capable of being used near water. •must cover a large test volume (30 m X 30 m X 2 m). •must not cause any electromagnetic interference with other sensors or data acquisition systems being used, •must measure large amplitude, low frequency (< 10 Hz) motions, •must have sufficient resolution and accuracy to yield useful data, •must capture the six degree of freedom (6 DOF) motions of the floating body, •must have a high signal to noise ratio, •must have a resolution of 1:5000 or better. The intended use of the data being collected is to gain insight into the motions crew members may experience on a full scale ship. The data may be used to evaluate different designs in com parative studies. The data could also be used in a feed back loop if the investigators were testing heading control in different sea states. 2 Chapter 1: Introduction Of the many non-contact motion measurement systems currently available, the video/cine pho-togrammetry method is one of the very few to meet the above requirements. Video/cine photo-grammetry, an established technique in fields such as biomechanics, is a method for extracting motion from a sequence of images by extracting dimensional information from individual images from single and multiple points of view. 1.2 Motion Tracking The problem of using a sequence of images, either film or digital, to analyse motion is a well studied problem that has been approached using a variety of methods. The problem is divided into two separate areas: (i) photogrammetry, the science of extracting dimensional information from the images, and (ii) tracking, i.e, quantifying the motion. The system presented in this thesis tracks the motion of a target composed of a set of markers that is attached to a model ship. In generic terms this is a mobile rigid body (MRB). The pose (position and orientation) of the MRB can be moni tored because pose of the target is known with respect to a coordinate frame of the MRB, which, in the case of a model ship, has an origin at the centre of gravity (CG). Systems similar in function to the one being presented are already being used in ocean engi neering test facilities worldwide to study the motion of ship models and models of other floating structures [MARTNTEK 97, HMD 5 & 6, & Sirehna 97]. At least two such systems are in use at NRC-IMD, where the most current system has been developed in house [Sullivan 93 & 97]. This system has been used to monitor wave excited motions for a variety of platforms, including, but not limited to, life rafts, aqua-culture pens, a tension leg platform and ship models. The work pre sented in this thesis is not intended to duplicate the system designed at NRC-IMD, but rather to explore and evaluate an alternate method that can be used in conjunction with it. The proposed system differs from the existing NRC-IMD system by being able to track an MRB with a minimum of one video camera to a maximum limited only by the hardware required to interface the cameras to the host computer. Also, the MRB does not have to remain in the field 3 Chapter 1: Introduction of view of all cameras in the system, as long as it remains in the view of at least one. In effect the proposed system requries a minimum amount of data, but can make use of all the data made avail able to it. On the other hand, the NRC-IMD system is a typical two camera stereo system where the MRB must remain in the field of view of both cameras for it to maintain tracking the target object. The advantages of going to a redundant multi-video-camera system is that it increases the test volume while maintaining resolution and can increase the accuracy and reliability of the meas urements made during the test. 1.3 Thesis Structure This thesis is broken down into 7 chapters, including this introduction, and a set of appendices that contain additional details of the algorithms developed and used for this project in both simu lation tests and the analysis of experimental data collected in a controlled motion study. Chapter 2 discusses the selection of a machine vision approach to tracking the motions of a model ship, as well as existing vision based tracking systems that are either under development in research facilities or are commercially available. Special attention will be given to those systems that are relevant to this project. Chapter 3 explains photogrammetry, the method of obtaining spatial information from images. The colinear equations, the fundamental set of equations that describe a pin hole camera positioned in space, along with the Direct Linear Transformation (DLT), a method of solving the colinear equations for use with non-metric, or uncalibrated off the shelf, cameras, are presented. Chapter 4 presents the Extended Kalman Filter (EKF) and how it is implemented to with the photogrammetry equations to generate the state vector that describes the six degrees of freedom of the mobile object being tracked. Chapter 5, discusses both the hardware and the software components of the implementation of the Argo system. Chapter 6, presents the experimental procedure used to test the Argo system and the results 4 Chapter 1: Introduction from those tests. Experiments include both computer simulation and physical testing. Chapter 7, concludes the body of this thesis and discusses potential continuations of this work. This is followed by six appendices of supporting material. 5 Chapter 2: Current Technology Chapter 2: Current Motion Capture and Analysis Technology 2.1 Introduction The noninvasive nature of photography and machine-vision sees it regularly employed to qual itatively and quantitatively evaluate the motion of a variety of systems. Motion capture is the col lection of quantitative data related to the motion of an object or point of interest with respect to a particular frame of reference. Motion analysis involves four components: (i) the processing of mo tion capture data for tracking trajectories of points of interest; (ii) the assimilation of motion data with additional data from other sensors that monitor other states of the system; (iii) the compar ison of data with previously acquired data for either repeatability or change; and, (iv) the determi nation of changes in position, orientation and the associated rates (velocity, acceleration, and jerk). The system to track the motion of a moving target is the combination of both motion capture and motion analysis. The motion capture-analysis system must be tailored to the specific motion being observed and the planned uses of the collected data. One of the goals of this research is devise a system that can track the motion of a model ship in a simulated seaway. The selection of an ap propriate motion capture system is important to the success of the project. However, the main thrust of this project is to track and analyse the motion. 2.2 Selection of a Motion Capture System Motion capture can be achieved with a variety of sensor systems that may be used separately or in concert, each with strengths and weaknesses. The sensor systems available for motion cap ture cover a wide range of working volumes from the very small to the very large, they also have a range of resolutions, accuracies, repeatabilities, response times, sample rates and unique environ mental constraints and accommodations. The applicable sensor systems include: electro-mechan ical sensors, inertial sensors and inclinometers, sonar, radar, electromagnetic, laser range scanning, differential global positioning system (DGPS), and optical imaging devices (machine-vision). 6 Chapter 2: Current Technology Selection of the proper sensor system must be dictated by the system under observation and the type of observations to be made. 2.2.1 Requirements for monitoring model ship motions For this project the system being observed is a model ship experiencing pseudo-random exci tation from waves in a simulated seaway. The ship is modelled as a rigid body; therefore, the sys tem will be observing rigid body motion. The 6 degrees of freedom (DOF) of a model ship as illustrated in Figure 1.1 are: roll, pitch, yaw, surge, sway, and heave. The model should be free to move and be unconstrained by the data collection system. All investigators would like to have per fect accuracy1, infinite resolution2, and perfect repeatability3 but, due to sensor limitations, this is not possible. For observing a model ship in a simulated seaway, a resolution of 2 mm over a range of 10 m and 0.1 degrees over a range of 30 degrees would be desirable with as much accuracy and repeatability as is possible. The motions of a model ship are expected to have an amplitude and frequency corresponding to the exciting wave. For a one metre wave height the heave would be expected to be about one metre. Angular motions are expected to have an amplitude of 0-30 degrees. Typically, the re sponse frequency of a model ship is on the order of 1 to 3 Hz. Under power, the frequency may increase if the rudder is being oscillated. The motion of the model ship should be sufficiently over sampled so that its motion and vibrations can be observed. For example ten times the expected frequency of 3 Hz would be 30 Hz. This far exceeds the Nyquist frequency constraint for observ ing dynamic systems and provide sufficient data to make the tracking algorithm converge and hold the track. The model ship will be observed in either a towing tank or a manoeuvring basin. These facil ities tend to be large and range in size from 50-200 m in length and 3-100 m in width. The test 1. The accuracy of a sensor is indicates the closeness of a measured value to the true value. 2. The resolution of a sensor indicates the smallest measurable incremental value. 3. Repeatability is the ability to reproduce the same measured value when duplicating a measurement. 7 Chapter 2: Current Technology volume should be considered to be in this range even if certain tests observe nearly stationary mod el ships, as in station keeping tests. Towing tanks and manoeuvring basins are typically indoors and have a room temperature of 15°C, except for ice tanks which can be as low as -20°C. The tanks are water filled with wave mak ers located along the walls of manoeuvring basins and at the end of towing tanks. As a result of wave generation it should be expected that the areas in the vicinity of the tests will be splashed or wet. Every effort should be made to ensure that special data acquisition equipment is not splashed as water drops can interfere with the data collection process, especially for optical based systems. Also, computers and other electrical equipment could be permanently damaged if flooded. Testing that requires specific lighting conditions may be difficult to do because the facilites that house test tanks, typically have overhead mercury vapour or fluorescent lighting, that would normally remain on during testing for safety and observation reasons. Safety considerations for the model are few except that the investigators would like to mini mize damage to the model so that it can be used in future tests. The instrumentation and data ac quisition equipment should be well protected because shock can effect calibration, accuracy, resolution, and may necessitate expensive repair or replacement. Safety of the investigators is par amount in any test procedure and therefore electrical connections should be made safe for a wet environment and all equipment should be securely fastened so that it can not fall into the water or injure anybody in the test area. Also, appropriate safety equipment should be made available dur ing the test for a wet and possibly cold environment. Creating a budget in a testing environment is always a difficult problem. A system such as the one being proposed would not be acquired for a single project; instead it would be amortized over several tests to make it economically viable. Any motion capture system capable of observing a model ship will have a large capital cost ranging from $20K to $250K depending on the facility and the type of tests currently under way and those planned for the future. 8 Chapter 2: Current Technology 2.2.2 Evaluation of sensor systems Sensor system selection must be dictated by the system under observation, what is the desired data, and in what form should the data be presented. To illustrate this point, in the case of moni toring the motions of a full scale ship a combination of DGPS, inertial systems, and possibly radar would typically be used. However, these same sensor systems are inappropriate for monitoring the motions of a model ship. These two problems, although similar, are very different because they take place in different environments and have different scales. A complete description of the motion capture problem is required before selection of the ap propriate sensor system. Using a two step elimination process, all but the valid sensor systems are rejected. First, obviously incompatible sensor systems are rejected. Second, the published per formance specifications of the remaining sensors are compared with the required specifications of the motion capture system as shown in Table 2.1. Performance specifications and general descriptions of the functionality of different sensor sys tems are published in reference texts [Borenstien 96, deSilva 89 & Karara 89], papers [Mackay 96 & Mulder 94], trade journals such as Vision Systems Design by Pennwell Publication and Photon ics by Laurin Publication, and product literature available from any manufacturer or distributor [CTDTEC 97, Dalsa 96, Qualisys 96 & Trimble 97]. As well there is vast amount of information available on the internet [Ariel 97 & Trimble 97]. Comparing the model scale ship testing requirements, summarized in the second column of Ta ble 2.1, with the available sensor systems. The electro-mechanical sensors and the elector-mag netic sensors can be rejected because they physically constrain the model ship's motion. The DGPS system is rejected because it can not be used in an indoor environment. The time of flight systems: sonar and radar are rejected because the number of degrees of freedom of the sensor is too low and the signal to noise ratio is also low. The scanning laser range finder is rejected because 9 Chapter 2: Current Technology Machine Vision Yes very small to large Yes none Yes 1:5000 1:30000 30 - 1000 Hz medium to high sensitive to lighting Inertial Systems Yes unlimited Yes none i Yes 1:1000 1:10000 10 - 1000 Hz medium prone to drift DGPS No unlimited Yes minimal m Yes 2-5m 10 cm N X cs medium to high 3 units required for 6 DOF Scanning Laser Range Finder Yes medium to very large Yes LASER radiation c-> Yes 1:2000 1:6000 10 - 40 Hz medium significant image processing TOF Sonar Radar Yes but environment sensitive medium very large Yes, but can cause false signals EMF levels CN Yes 1:20 1:100 30 - 100 Hz low prone to false or ghost signals Electro-Magnetic Sensors Yes but environment sensitive small to medium Yes, but can cause false signals EMF levels cn No 1:1000 1:1500 100 Hz medium sensitive to metal in test volume Electro-Mechanical Sensors Yes small & medium Yes with precautions minimal No 1:2500 infinite up to 10 kHz high requires heavy infra structure Full Scale Ship Testing Requirements No unlimited Yes minimal vo Yes 1:20 1:100 10 Hz high Model Scale Ship Testing Requirements Yes small to large Yes minimal vo Yes 1:1000 1:5000 30 Hz high Categories Suitable for indoor use Workspace size Can be used near water Safety Concerns # of DOF of sensor test object unconstrained accuracy resolution bandwidth Signal to noise level Other 10 Chapter 2: Current Technology the number of degrees of freedom of the sensor are too few. Additionally the laser-based system's requirement of significant image processing eliminated it from contention. A machine-vision based systems was selected because its characteristics most closely match the requirements for testing a model ship. Having gone through this process, the most appropriate sensor system for observing the motion of a model ship in a simulated seaway is a video-based tracking system possibly in combination with inertial and inclinometer sensors on board the model. The main reasons for this are these sys tems are non-contact, can cover a large test volume, can be used indoors, are safe to use near water with proper precautions, and they have sufficient accuracy, resolution, response time, and sampling rates. The on board sensors would be coupled with a master data acquisition system via a spread spectrum radio link, which could be shared with a remote control system. Due to the complexity of the overall problem, the added cost to the project and the fact that at this stage the project is a proof of concept only; the video-based motion capture was pursued and the inertial sensor system was left for future work. Typical video tracking systems capture a set of synchronized image sequences of a 3D dynamic scene, from multiple points of view. From analysing both stationary and moving feature points, also known as landmarks, tokens or object points, in the images the motion of the scene relative to some fixed reference frame can be measured. There are many commercial, video based, close-range photogrammetry, motion capture systems that have been developed and are currently avail able on the market. These commercial systems are mainly used for recording human motions for medical, animation and sports applications. Few of these systems are optimized or even capable of tracking rigid body motion. Video tracking is also being actively pursued by the research com munity for non-contact measurement in experimental testing [Johnson 90, Rediers & Wysner 83], to improve human motion analysis [Romilly 95 & Safee-Rad 87] and to aid in robot navigation, servoing and obstacle avoidance [Borenstien 96 & Wilson 93]. 11 Chapter 2: Current Technology 2.2.3 Imaging Technology Early implementations of motion capture based on image sequences was achieved through ei ther cinematography [Shapiro 77], sequenced still photography, as Muybridge did in his study of a galloping horse in 1877 [Cook 81], or multiple exposure strobe photography [Edgerton 87]. In all of these cases the images of the scene were recorded on film with a camera or group of cameras synchronized together and with the action in the scene with the appropriate lighting. Imaging technology used in film and newer electronic cameras is a very large topic and covers a wide range of techniques and equipment depending on such things as lighting conditions, the de sired frame rate, the desired image resolution, scene environment such as temperature and humid ity. Film has the highest resolution images and the highest frame rate for medium to high resolution applications. However, for the majority of applications it has been replaced by electron ic imaging because it is easier to use and does not incur film and processing costs. Film has also lost favour because of the difficulty in digitizing a film image for computer based image process ing. However, film is still the cheapest and most effective method for high resolution and high frame rate applications because of high costs for electronic imaging technologies that can approach the same image quality and higher frame rates of film. The most common electronic imaging technology is the charged coupled device (CCD). It is low cost, and can be easily interfaced to a computer through a dedicated "frame grabber" input/ output (I/O) digital signal processor (DSP) board for image processing. The CCD array sensor is most sensitive to the red and near infrared (NIR) end of the light spectrum and is available over a broad range of resolutions from 64 X 64 to 5000 X 5000 [Dalsa 96]. Scientific imaging typically has a resolution of 512 X 512. However, much of the commercial ly available motion capture systems are based on consumer and industrial CCD cameras, the most commonly available, which are interlaced composite video encoded with either PAL format, 625 video lines at 25 frames per second (fps), for most of Europe, or NTSC format, 525 video lines at 12 Chapter 2: Current Technology 30 fps, for North America and Japan. Both formats have an image aspect ratio of 4(H):3(V) [Stremler 82]. To allow for sync-signals and field and line retraces, the working image size for North American camera systems is usually 640 X 480. The majority of motion capture applications in both commercial and research efforts, use these commonly available consumer and industrial CCD based video cameras. For those systems that use infra-red (IR) markers, both passive and active ones, the camera lenses are usually fitted with an IR band pass filter. For the reflector type markers the cameras will have an IR flash mounted around the lens which is triggered by the frame rate of the camera [Motion 96, Qualisys 96 & Vicon 96]. To meet the demands of capturing higher speeds, such as a baseball pitcher's arm, newer sys tems are now using higher performance imaging and computer components, resulting in frame rates up to 1000 fps [Qualisys 96]. Some other electronic imaging devices, used for special applications where a CCD camera may not function properly, are the intensified charge coupled device (ICCD) for low light applications, silicon intensifier technology (SIT) for low light underwater imaging because of its sensitivity to blue light [Mackay 96] and the charge injection device (CID) for ultraviolet (UV) imaging and high radiation environments [CIDTEC 97]. The imaging tube is an older analogue method of capturing an image that has been replaced by CCD arrays for the majority of uses. However, there still re mains some special applications for this older technology. Another photo-optical sensor that can image a marker or landmark in the scene is a position sensing photodiode. This device, unlike the CCD array, can only view one marker at a time. A position sensing photodiode (PSP), also referred to as a position sensitive detector (PSD), is an an alogue device that returns the position of a bright spot projected onto the sensor. If more then one light source is incident to the sensor surface the sensor will return an average position of the mul tiple bright spots. The spectral sensitivity of a PSP makes it ideally suited for red and near infra red light (NIR) sources. This device has a response rate 10-15 kHz, depending on the specific man-13 Chapter 2: Current Technology ufacturer [On-Trak 97]. If the lens in front of the PSP sensor is fitted with a band pass filter that is matched with the marker light source wave length then the majority of stray light should be elim inated, minimizing noise in the signal. With two or more synchronized PSPs the position of the point of light can be triangulated similar to the method used in a conventional camera. To accom modate multiple feature points the IR light emitting diode (LED) markers are strobed in sequence, with only one lit at any given time. 2.2.4 Markers Feature points are usually represented by markers with high contrast to the test subject and the scene background [Johnson 90]. Markers are used to speed the feature point extraction process but they are not required as detailed image processing can be used to identify natural points or an op erator can manually select a point in the image. One common approach is to use passive retroreflective markers coupled with an IR strobe. An other approach uses active IR LED markers that, when viewed through a filter matched to the wavelength of the marker, produces a high contrast black and white image where the markers are easily identified. The passive markers are typically spheres, half spheres or circular dots coated in retroreflective paint that is highly reflective in the IR regime. Spherical markers are preferred be cause the apparent centroid of a sphere is independent of viewing angle. With other shapes the cen-troid can appear to shift with viewing angle. For example, the centroid of a hemisphere can appear to move by 42% of the radius of the marker when viewed at the side as compared with the view from normal to the hemisphere [Gieck 90]. A difficulty of spherical IR reflectors is that all markers appear the same in the image and cannot be uniquely identified from the raw images. Thus, the tracking software must identify them. Stray reflections from equipment or the surface of a liquid, sometimes known as ghost markers, can appear in the data set and may confuse the tracking soft ware and requiring manual intervention to complete the tracking task. Active markers such as small light bulbs and LEDs provide strong signals and typically have a 14 Chapter 2: Current Technology smaller size than the reflective passive markers. The active markers can be sensitive to observation angle, particularly LEDs, therefore; for monitoring small displacements they are acceptable, for large displacements a cluster of LEDs that can be viewed from all possible viewing angles is re quired. As with passive spherical reflective markers, small lights all appear the same in the images and a means of identifying them is required so that their positions can be triangulated. One approach used to identify individual LEDs is to strobe them in a sequence having only one lit at any one time. The advantage of this approach is that it is relatively fast, and since only one marker is illuminated at any given time correspondence between images and the marker is not an issue. A disadvantage of this method is that if the target moves appreciably over the duration of one sample iteration and each marker is illuminated once, then some measurement error will be introduced because the im age is not a "snap shot" of the marker locations at one point in time, but rather a rolling sample. This method has been implemented in commercially-available motion capture-analysis systems [Charnwood 97, NDI97, Selspot 96 & Yaman 97]. An advantage of using IR light, either with passive or active markers, is that overhead lighting need not be turned off during testing, making it safer for researchers to conduct the test. However, strong ambient levels of IR light can cause noise or clutter in the image making feature point ex traction less reliable. Many systems make use of the visible light spectrum and use markers that contrast from the scene. One popular method is to use high contrast black and white patterns that are relatively easy to extract from an image. Some systems have employed coloured and/or differing shaped markers to uniquely identify individual markers. Other systems use no markers and rely on the operator to define a feature point from the scene in the image. With no markers available for tracking, this approach requires either the operator to manually track the point of interest frame by frame or com plex image processing and tracking algorithms that can follow these points of interest. As the level 15 Chapter 2: Current Technology of complexity of the point of interest increases, the requirements for more complex imaging hard ware and image processing for detecting colour detail and/or shape of the feature point in the scene also increases. This added complexity would likely be reflected in the cost of the image capture system. As well, such systems are more sensitive to ambient light that may change the apparent colour of a point of interest or cast a shadow that may change its shape. Obtaining the image location and size of what is perceived to be a marker, namely, "feature extraction" is done through image processing. This topic is not discussed in this thesis as it is out side of the scope of the project. The image processing done in the collection experimental data for this project was accomplished in firmware built into the Qualisys video processor hardware. 2.2.5 Commercial motion capture systems There are many machine-vision based motion capture systems commercially available today. The majority of these systems have been designed for human motion analysis for kinesiology and medical studies, with a few notable exceptions. Recently the systems originally designed for the sports [Blackburn 96] and medical communities [Harris 96] are being used by the entertainment industry for animation in movie productions and action video games [Delaney 97]. These systems track individual markers attached to the patient, client or actor at strategically located points such as the hip, knee, elbow and shoulder. From this the operator can infer body movements from the trajectories of the markers attached to the test subject, captured with synchronized video cameras while the test subject performs the target physical activities [Harris 96]. Often these systems do not operate in real-time and rely on post-processing of recorded data. Unfortunately, very few of these systems are designed specifically to follow rigid body motion. Rigid body motions have ad ditional constraints that allow for speedier and more accurate tracking, with greater robustness and effective resolution. This is not to say that these systems could not be used for tracking rigid body motion but rather that they are less then optimal. Systems designed for human motion analysis track the 3 DOF of many feature points. On the 16 Chapter 2: Current Technology other hand, a system designed to track rigid body motion is intended to track the position of the origin and the orientation of a coordinate frame attached to the test object. The techniques used in these two applications are not entirely separate but different enough that software developed for kinesiology work is inappropriate to study the rigid body motions of ship models. Much of the software supplied with these systems is setup for clinical trials of patients or atheletes, and the only output available is of time histories of the motions of individual markers attached to the test sub ject. As well due to the proprietary nature of this type of software, it was not possible to obtain the necessary building blocks to adapt the existing program to track large-amplitude, low-frequency rigid body motion. Another problem with the proprietary nature of the software is that the motion analysis techniques used are not openly published forcing a researcher wanting to expand the sys tem beyond its original intended use to guess at what the company has done or to build a secondary data post processing routine. As a result, system development for this application had to start from the ground up. Fortunately, many of the mathematical building blocks for developing a rigid body tracking system are available in the public domain. Many of the manufacturers of this equipment make use of proprietary video processors with built in firmware that can not be modified by the user. The tracking routines or portions of them are usually included in this hardware forcing developers to rely on partial execution of the tracking process. This is a benefit because it will save time, but it is also a detriment as it limits what can be done and how much access one has to the raw, unmanipulated data. Before using such a system the developer will require a detailed knowledge of how the data is manipulated and of the output format so that they can link it into their work. Some commercial systems, although not optimized to track large amplitude, low frequency rig id body motion, do so by tracking the individual markers that make up the target and then reconcile their motion to that of the mobile rigid body (MRB). Other systems have been designed specifi cally for tracking rigid body motion. These systems are typically configured for industrial inspec-17 Chapter 2: Current Technology tion tasks or for integration with robotics for calibration, evaluation, and control input. There are a number of the commercial systems designed for human motion studies that have been adapted to the problem of observing the motion of model ships. The U. S. Army Corps of Engineers is using a Motion Analysis Expert Vision system at their Waterways Experiment Station in Vicksburg, Mississippi. They use this system to follow model ships in their mock-up of the low er Mississippi [Motion 97]. The Oxford Metrics Vicon 370 system is being used at the Shipping and Marine Department at Strathclyde University in the United Kingdom for capture roll, pitch and yaw of a model in their towing tank [Vicon 96]. Both Marintek of Norway and NRC-IMD have employed both the Selspot and OPTOPOS (Optical Positioning) systems to track the rigid body motion of model ships in manoeuvring tests [MARINTEK 97 & Selspot 96]. At NRC-IMD the OPTOPOS system, believed to have been made by Saab, is no longer operational. Presently NRC-IMD, MUN-OERC and NRC-Hydraulics of Ottawa, Ontario are using a system developed at TM NRC-IMD based on the Qualisys Mac Reflex system [Sullivan 93 & 97]. This system has been proven in comparative tests and is routinely used on commercial and internal research projects. For further information on specific systems the reader is directed to tables A.2, A.3, and A.4 in Appendix A: Motion Capture Systems for a comparative review of some of the commercially-available machine vision based motion capture systems. 2.2.6 Selected motion capture system The Qualisys MacReflex™ camera and video processor hardware with its built in firmware was selected for the Argo Project because it met the requirements for observing a model ship in a tank and was compatible with other on-going research projects at NRC-IMD. This system has a suitable sample rate, sufficient resolution and accuracy, does not constrain the model, can be used with overhead lighting, and has built in frame grabbers and image processing for extracting the image feature points. Only the hardware component of the Qualisys MacReflex™ system is being used because the 18 Chapter 2: Current Technology software normally supplied with the system can not track the rigid body pose of one or more MRBs in real-time. The original software is designed for doing human motion analysis and due to its pro prietary nature could not be modified to meet the needs of the project. NRC-IMD are using soft ware developed in house to acquire data from the Qualisys hardware and track the pose of a model ship [Sullivan 93 & 97]. The Argo Project is currently concentrated on the tracking algorithm and how it can be improved; therefore, it relies on the NRC-IMD software to capture and record data gathered with the Qualisys hardware. 2.3 Motion Analysis Techniques Motion analysis from a sequence of images is the extraction and interpretation of motion from those images. This data reduction process includes photogrammetry, the extraction of position data from an image, coupled with tracking and data fusion of image coordinates from multiple im age sequences. Early motion-capture-analysis projects that went beyond a purely qualitative assessment used metric cameras and an analogue electro-mechanical-optical device, known as a stereoplotter, to manually digitize film sequences frame by frame. The manually digitized image feature point cen-troids were then entered into a computer for triangulation of each point. From the triangulation results the pose of the object would be resolved using a least squares fit for each frame. With the improvement of imaging, electronics and computing ability the above described process has evolved to become more efficient, faster and less costly but essentially remains the same. The ster eoplotter is replaced with a frame grabber and automated feature point location software. Entire sequences of images can now be processed without any manual intervention by the operator. A benefit to increasing the speed and the ability to automate the process allows for the real-time track ing of motion for control applications. The majority of the information, published on the topic of motion analysis based on a sequence of images, comes from the reference texts and the university and government research community 19 Chapter 2: Current Technology as private companies tend to hold this information as a trade secrets to prevent their competition from profiting from their efforts. As a result little is openly published about the techniques and methods used in the commercial systems discussed in sections 2.2.5, 2.2.6. 2.3.1 Photogrammetry Photogrammetry is a technique for extracting three dimensional (3D) information scene from two dimensional (2D) images of the scene. Photogrammetry is regularly used for mapping, sur veying and measurement of static and dynamic systems. It is used when physical measurements can not be performed because they may damage or interfere with the system under observations, or when the range and scale of the scene make measurement impractical or impossible. Photo grammetry makes use of the geometric relationship between 3D space to the 2D representation in the image plane. This transformation is essentially a plane perspective projection with a few addi tional terms for correcting lens and image plane distortions. The methods and techniques of pho togrammetry are well developed but some advances have been made coupled with advancing technology [Karara 89]. The concept of motion analysis from a sequence of images is an extension of traditional photogrammetry to include a time vector [Faugeras 93]. The geometric relationship between 3D space and the 2D image plane is typically referred to as the camera model. The simplest camera model, typically the basis for more detailed ones, is the pin hole camera model. This model is the plane perspective projection model where the feature point in the scene, the focal point and the image point all lie on the same vector in space. A pa rameterized version of this basic camera model is called the Direct Linear Transformation (DLT). This model was designed for non-metric commercial off the shelf (COTS) cameras for still photo grammetry [Aziz 74 & Marzan 76]. Triangulation of feature points in the scene, as well as cali bration are simple, straight-forward procedures. Although the DLT was developed for still photogrammetry it has been successfully applied to motion analysis at both the research [Anglin 93, Miller 80, Safee-Rad 87 & Shapiro 77] and the commercial [Areil 97& Peak 96] level. The 20 Chapter 2: Current Technology DLT has been selected as the camera model for the Argo Project. As a single 2D image of a 3D scene does not contain enough information to determine a 3D position of a point in the scene, additional information is required. Typically this extra information comes in the form of one or more additional images taken from different points of view, covering the viewing volume. This is the stereo vision approach. An added benefit is improved resolution of the observation system as the individual cameras that make up a multiple camera system may suffer from low resolution problems along the focal axis of the camera. Using the principles of triangulation and knowing the pose of the different cameras in the world coordinate frame, the 3D position of any point can be determined from images of the scene. Systems that rely on triangulation have difficulties when a marker is occluded from view and is not visible in at least two fields of view. In this case, the position of that marker is not computed at all. Once the 3D position of feature points fixed relative to the coordinate frame of the MRB has been determined, the pose of the MRB can be found. The pose of the MRB can be found by match ing the known positions of the feature points in the coordinate frame of the MRB with the positions of those same feature points measured in the world coordinate frame. This is a nonlinear problem which can not be solved for directly; however, it can usually be solved for with an iterative least squares approach. To determine pose of a target from triangulated results, the positions of at least four non-coplanar feature points are required to be able to compute both position and orientation without ambiguity. The added constraint of knowing the location of feature points in the MRB coordinate frame make it possible for a monocular system to determine the pose of the MRB, though with lower ac curacy than a stereo-vision system. Since triangulation is not possible with a monocular system it can not determine the 3D position of an individual point in the scene. There is one exception, when the feature points are constrained to be in a plane parallel to the image plane and with known sep-21 Chapter 2: Current Technology aration between the planes. Advance knowledge of the scene can provide enough information to make a second image unnecessary for determining pose of an MRB. For pose recovery of the MRB, the positions of known feature points on the MRB are combined with the camera model re sulting in a set of nonlinear constraint equations that, when solved, yield the pose of the MRB. There are several monocular motion analysis systems, both as commercial products and as re search projects. These single camera systems have the advantage of having smaller computational requirements then stereo systems because they incorporate less hardware, less image processing, and require no correspondence matching between image sets. However, they also suffer from low resolution and potential significant measurement error in the focal axis of the camera. By capturing a sequence of images at known time intervals with synchronized cameras the con cept of motion analysis can be introduced. The motion of the MRB can be analyzed by recovering the pose of the MRB for each set of images yielding a step wise trajectory of the pose. The differ ence in pose between sequential sets of images and the frame rate can be used to determine the ve locity and accelerations terms: essentially the differentiations of the trajectory of pose with respect to time. This motion analysis method does give a history of the pose, velocity and acceleration of the MRB; however, it is prone to noise and its efficiency could be improved upon. A more effec tive approach to the problem of motion analysis utilizes tracking algorithms. A more detailed explanation of photogrammetry, the colinear theory and the Direct Linear Transformation (DLT) can be found in Chapter 3 and the accompanying Appendices B and C. 2.3.2 Tracking Tracking has been described as the recursive process of predicting, comparing, and updating [Bar-Shalom 88]. The method and implementation of the tracking, as with sensor selection, must to be dictated by the problem being addressed. Some of the concerns to be addressed are whether the system is linear or not, the number of variables to be tracked, and will the system follow an expected trajectory or is it considered to be random. 22 Chapter 2: Current Technology Tracking the full 6 DOF of a model ship along with the velocity and acceleration terms requires that the tracking algorithm must be: • able to handle the nonlinearities associated with collecting data with a camera, a nonlinear device, • able to follow a continuous unplanned trajectory, • scalable to match the amount of available data, • able to integrate data from multiple sensors, • able to maintain track of a moving target in a noisy environment, • stable in the presence of expected noise from both the sensors collecting data and the system under observation, • able to track multiple states simultaneously, namely the pose of the MRB along with the veloc ities and accelerations of pose, 6x3 = 18 state variables. The extended Kalman filter (EKF) was selected as the tracking algorithm for the Argo Project because it met all of the above requirements as well as having been demonstrated to function for tasks similar to that of tracking a model ship with video cameras. The EKF, a recursive suboptimal state estimator for nonlinear systems, is an extension of the popular tracking algorithm, the Kalman filter (KF). The KF and the EKF are both popular and well studied methods, are considered relia ble, and have been implemented successfully in a variety of applications ranging from radar track ing of aeroplanes and missiles, machine-vision tracking, and analysing and tracking financial trends in the market [Bar-Shalom 88 & 93, Bozic 94, Faugeras 93, Grewal 93, Wilson 93, Wu 88 & Zhang 92]. To track the pose of an MRB with cameras, it is not necessary to triangulate the position of markers. Feature points in the scene, the image coordinates of the markers and the known position of the markers in the coordinate frame of the MRB provide sufficient data. The triangulation proc ess is simply a data reduction technique that renders image coordinates into identifiable locations in 3-space. Since the goal is not to know the position of each marker attached to the target but rath-23 Chapter 2: Current Technology er the pose of the target, it is not necessary to go through the intermediate step of triangulation which may only increase the computation time. A set of nonlinear constraint equations that de scribe the pose of the target can be formed from: (i) the colinear equations that describe the numer ical camera model, (ii) the known target geometry, (iii) the general form of the coordinate transformation between the world reference frame, and (iv) the model ship's reference frame and the image coordinates. The six pose variables in the nonlinear constraint equations can be solved in a variety of ways. They can be solved for with an iterative least squares solution method or they can be built into a tracking algorithm such as the Extended Kalman Filter (EKF). The EKF can also be implemented to track the pose of the target based on the triangulated positions of the markers attached to the mo bile rigid body. A detailed discussion of using the EKF for tracking the trajectory of a moving tar get based on image coordinates of known feature points or triangulated feature point location for both monocular and stereo imaging systems can be found in [Faugeras 93 & Zhang 92]. This meth od has been successfully demonstrated in two separate research projects that both used the EKF and a known target geometry for a monocular system [Wang 91 & 92, Wilson 93, & Wu 88]. Both of these systems are limited to a small working volume so that small changes in position along the focal axis have a noticeable effect. This same methodology is utilized for tracking the pose of a mobile rigid body (MRB) in this project, but it is extended to include multiple cameras viewing the scene to increase resolution and working volume. There are tracking algorithms and methods other than the EKF that can track the pose of an MRB based image data captured with video cameras. An iterative least-squares approach can ac curately determine the pose of objects in a scene by either fitting triangulated feature points with the known geometry of the target or using image coordinates and target geometry as constraints in a set of colinear equations that include pose terms. This method may be more accurate than the EKF approach and does not require time for the filter to establish a track, but it does not actually 24 Chapter 2: Current Technology track the target but rather builds a sequence of pose data for each set of image frames. Also, it re turns only the target pose from which velocity and acceleration terms must be obtained through differentiation of the pose sequence. This is the approach taken by [Sullivan 93 & 97] in his work to develop a system to track the motion of a model ship. Curve fitting techniques could be applied to the pose sequence to fill in the gaps between frames and act as a low pass filter by removing some high frequency noise in the trajectory. Some work has been done to provide a fast nonitera-tive least squares solution to this problem [Weng 92]. Some other methods of following the trajectory of an MRB through video camera captured data are neural networks [Zha 95], a modified Kalman filter approach that assumes the motion follows an affine transformation [Manku], a frequency domain approach using the Fourier transformation [Lin 86], various statistical approaches such as modelling the motion as a Markov process [Aggar-wal 97], and model based tracking that is an iterative least squares minimization approach [Aggar-wal 97 & Lowe 92]. Some of these methods, although highly accurate, require large computer resources and are very difficult to implement in real-time applications at a reasonable cost. At the other extreme are very simplistic tracking methods that are easy to implement but are not very re liable. In such cases, if the signal is noisy and/or if the target moves somewhat erratically, it may loose track of the target too easily. The Kalman filter theory and that of the EKF, along with its implementation in the Argo Project, is presented in Chapter 4 of this thesis. 2.4 Related Research Projects Industry, university and government researchers working on motion capture-analysis problems typically concentrate their efforts in the areas of motion analysis and image processing. They tend to rely on commercial off the shelf (COTS) cameras and computer interfaces as they are well de veloped and are widely available. On going research and development by industry is done to im prove existing commercial systems in order to stay abreast of advancements in technology and 25 Chapter 2: Current Technology remain competitive. Researchers are working on improving human and animal motion tracking for medical, sports and entertainment-related activities. Others are working on improving rigid body motion tracking for automated inspection and for controlling and evaluating industrial, military, and exploration robots. Of the researchers working on rigid body motion, a few are working on the problem of tracking the motion of model ships in towing tank and manoeuvring basin environments [Alexander 97, Gospodnetic 92, Sullivan 93 & 97, & Veillon 96]. Another area of common research is visual ser-voing of an end effector of a robot manipulator arm. This is being explored with both a monocular camera system and the Kalman filter [Wang 92 & Wilson 93] and a stereo camera system coupled with a learning algorithm as the tracker [Zha 95]. Others have been more general in their approach, and are working on the generic problem of tracking rigid body motion from images [Lowe 92, Miller 80, Shapiro 77, Wu 88], camera calibration [Yuan 89], photogrammetry [Krishnan 92] and the combination of all three [El-Hakim 92]. The reverse problem of resolving observer's motion by object tracking is of interest for pho tography-based mapping from a moving platform, controlling autonomous robots, and the movie industry. An autonomous robot requires knowledge of where it is so that measurements made by the robot can be identified and to aid in navigation for following a preplanned path or to know if the destination has been achieved. The position of the observer can be determined from images of landmarks in the scene with both known and unknown locations [Borenstein 96, Shan 95, Yuan 89]. This is in effect an on-the-fly calibration of the cameras exterior orientation. For an autono mous underwater vehicle (AUV) [Hallan 83] explores the problem of determining the observer's position from sonar images. The source of the images is not relevant as the basic approach for tracking based on image data is common to a variety of image types. In the movie industry, it is necessary to know the observer, namely the camera, pose for integrating computer-generated ani mation with live characters and scenes. 26 Chapter 2: Current Technology Additional work is being done by researchers considering the extraction of motion from a se quence of full images using the concept of optical flow [Faugeras 93, Horn 86, Nesi 96 & Sezan 93]. Optical flow is based on the relationship between the temporal variation of image intensity and motion. This approach has some advantages over a token tracking system: it requires no mod ification of the object to be studied since no target need be attached, it can be used to study non rigid structures, and it can analyse multiple sources of motion in different parts of the scene. How ever, this method has an extremely high computational cost. Thus, this method is difficult to im plement in real-time without expensive specialized computational hardware. Furthermore, solving the correspondence between stereo images is a considerable challenge using this method. As tech nology advances and our collective understanding of artificial vision increase this will likely be the direction pursued by many as it will give a better evaluation of the motion in the scene. The reader is referred to table A. 1 in Appendix A, for a summary of some of the motion anal ysis systems under development in a research environment, reviewed as part of this research project. These projects have, for the most part, been demonstrated to function at varying levels of completeness and are typically oriented to a specific task. 2.5 Summary The task of tracking a manoeuvring model ship is a multifaceted problem. It requires the ap propriate sensors, data reduction techniques, and data assimilation techniques. For the Argo Project a multi-video-camera systems was selected as the sensor system for the primary reason that it does not interfere or constrain the model ship in any way. Secondary reasons are that this system will provide sufficient accuracy, resolution, range and sample rate for the desired experiments and data analysis. The Qualisys, Mac Reflex™ hardware was selected for the Argo Project. This hard ware consists of one to seven custom video cameras, each with a synchronized IR strobe and a vid eo processor that controls the camera and performs preliminary data reduction that extracts potential feature points and transfers them to the host computer for analysis. The data assimilation 27 Chapter 2: Current Technology and analysis algorithm combines the Direct Linear Transformation (DLT) camera model for fur ther data reduction and the Extended Kalman Filter (EKF) to assimilate and combine data from dif ferent data streams as well as estimating the trajectory of the pose of the mobile rigid body (MRB) along with the velocity and acceleration profiles for that trajectory. 28 Chapter 3: Photogrammetry Chapter 3: Photogrammetry 3.1 Introduction Photogrammetry is the technique of making reliable measurements through the use of photog raphy. With the invention of the camera in the early part of the 19th century, photogrammetry soon followed. Aime Laussedat, the father of photogrammetry, was the first to develop and use photo-grammetric techniques to produce perspective drawings of the facade of 1'Hotel des Invalides in Paris, 1849. He carried on to develop new camera equipment and to develop topographical map ping based on photography [Blachut 89]. Traditional photogrammetry is done with specialized film cameras, and stereo plotters, an opto-electro-mechanical device used to locate targets in stereo images so that its position in space could be computed. Presently there is a transition from older analogue equipment and methods to newer digital technology, resulting in hybrids of the two systems. As the cost of high resolution digital-images comes down and digital storage techniques become more efficient, the eventual transition to fully digital photogrammetry will ensue. With the establishment of photogrammetry to extract quantitative data about the scene from images, the next logical step was to analyse motion with images. Moving film and video cameras capture a sequence of still images separated by a fixed known time interval. By replaying this se quence the motion in the scene is in effect captured. By applying photogrammetric techniques to the individual images and considering the time interval separating the images; time based quanti tative data such as changing positions, velocities, accelerations can be extracted. Provided that the frame rate is sufficiently fast to fully observe the motion, this technique can be used to study the motion of almost anything. A great deal of work has been done in this area in the study of human and animal motion to improve sports skills, evaluate medical treatments, and generate human motion for animation used 29 Chapter 3: Photogrammetry in video games, movie special effects, animated movies and cartoons. The majority of motion cap ture systems commercially available are designed to develop trajectories of individual points, needed for human motion studies, but not the trajectory of the pose of a MRB with a few exceptions as noted in the previous chapter. An image is a two dimensional (2D) representation of a three dimensional (3D) scene. As a result, depth information about the scene is lost, thus 3D information can not be extracted from a single image without additional information. This information can be obtained from another image with a different point of view of the same scene, or from a priori knowledge of the 3D scene that can be used to apply additional constraints to the image. Human beings have an apparent depth perception when examining a photograph. They use an existing knowledge base of the environment that the image represents and apply a set of rules. For example, trees on the edge of a lake: in the image they appear to grow out of the water but we know that this is wrong. We know that they are growing on the shore adjacent to the lake so we infer from the picture that they are growing on the shore, although the image does not show this. 3.2 Coordinate Frame Definitions A simple interpretation of photogrammetry is that it is the combination of two geometric trans formations; a coordinate frame transformation of a point from the World coordinate system to the camera coordinate frame followed by a plane perspective projection of the points in the viewer cen tred coordinate frame onto the image plane of the camera. Figure 3.1 depicts the fundamental co ordinate frames necessary for the following descriptions and explanations of how an image is generated in the camera and how to obtain useful information from that image. In the following sections a generic camera coordinate frame will be referred to by C, later when it becomes neces sary to differentiate between camera coordinate frames the variable m will be used. 30 Chapter 3: Photogrammetry where: —c c c c Gn — (xn > yn» zn) = coordinates of feature point n in camera C — W w w w. Gn - (xn , yn , zn ) = coordinates of feature point n in the World coordinate frame. -w www. Gc = \xc, yc,zc)= position of focal point (camera frame C) in the World coordinate frame. Figure 3.1: Feature point position in the Camera's and the World coordinate frames 3.3 The Pin Hole Camera and Colinear Theory The simplest camera model used in photogrammetry is the pin hole camera, illustrated in Fig ure 3.2. The pin hole model is a simple perspective projection onto an image plane that is usually orthogonal to the focal axis. A ray of light emanating from an object point, in the scene, passes through the focus point and then intersects the inverted image plane, yielding an image point. These three points coincide on the same line, hence they are colinear. This is the basis of the co-linear theory. It is important to note that the focal point of the camera and the origin of the camera's coordinate frame are coincident. A feature point in the scene (FPn) or world space has a unique mapping to a point (IPn) in the image plane. However, the converse is not true, a point in the image plane does not have a unique mapping to a point in the world space. Due to this fact, a single camera can be calibrated from a set of known points in the scene, but the image coordinates from a single camera cannot give a unique solution for the location of a point in the scene. This unique mapping is described using the following plane perspective projection equation pair that provide image coordinates from feature points in a camera centred coordinate system, Equation (3.1) [Gosine 96 & Weng 92a]: 31 Chapter 3: Photogrammetry C , xn •/ un-up = —^- , (3.La) C , where: • (un , vn) = image coordinates of feature point n, • (up , vp) = image coordinates of principal point, • /= focal length, c c c • (xn' y«' zn) = position of feature point n in the camera coordinate system C. To be able to view a scene that is in a coordinate system other than that of the camera a transfor-Figure 3.2: Pin Hole Camera 32 Chapter 3: Photogrammetry mation is required. A feature point in the scene in the world coordinate system is transformed into the camera coordinate system so that it can be projected onto the image plane. This transformation is based on the position and orientation, the pose, of the camera in the world coordinate system. Combining this transformation and Equation (3.1) yields a new camera model based in the world coordinate system, Equation (3.2) [Aziz 74]. u = u —f n a Ju *l,-l(*n -Xc)+Rl,2(yn -yc)+R\,3JZn ~zc) R3,\(xn -Xc)+R3,2(yn -yc)+R3,3(Zn ~ ZC> (3.2.a) V = V — f n p Jv R2,\(xn -Xc) + R2,liyn -yc) + R2,3(zn ~ ZC ) KR3,\(Xn -Xc)+R3,2(yn -yc) + R3,3(zn ~ ZC ) (3.2.b) where: • (fu, fv) = focal lengths in horizontal and vertical components of image plane (typically fu =/v), IV w w • R = a [3 x 3] orthogonal rotation matrix based on (\|/c, 0C, §c), www • (\|/c, 8C, (j)c) = orientation of the Camera coordinate system in the World coordinate system, www • (xc ,yc>zc) = position of focal point (camera frame C) in the World coordinate frame, www • (xn ' yn ' zn ) = coordinates of feature point n in the World coordinate frame. To evaluate the image coordinates (w„ ,vn) in Equation (3.2) for the corresponding feature point www (*„ , yn » zn ) an assumption is made that the remaining parameters are known. These parameters are broken into two groups: interior orientation or intrinsic parameters and exterior orientation or extrinsic parameters. The interior orientations are: principal point location (the point where the fo cal axis intersects the image plane), focal length, pixel size, and systematic error correction terms for image plane distortions and lens distortions. The exterior orientations are the position of the focal point and the orientation of the focal axis and the image plane, which define the reference frame of the camera, with respect to the world coordinate frame. This is the definition of the pose 33 Chapter 3: Photogrammetry of a camera. These yet unknown terms are found through calibration. Calibration of a camera is no different than that of any other sensor system. It is the search for a transfer function relating known inputs to measured outputs. For a camera, the known inputs are the locations of feature points in the scene in the world coordinate system that the camera is to be calibrated against. The measured outputs are the image coordinates corresponding to the feature points in the scene. Solving for the unknown values in the general form of the non-linear camera model transfer function, Equation (3.2), is typically done with an iterative least squares method specific to the camera model being used. Certain experimental and industrial applications require the camera(s) to move to cover a large area. This is important to note for the application of studying the motion of model ships, where the cameras may be mounted on a moving carriage, to follow the model ship, along the length of a tow ing tank. In the case of a mobile camera system, the reference frame in which the camera(s) were calibrated, can move but the interior and exterior orientations must be fixed relative to this refer ence frame. Additional data regarding the movement of the mobile camera base reference frame would have to be included into the tracking module to recover pose, velocity and acceleration terms with respect to a fixed frame of reference. This is not considered in this work. 3.4 Systematic Errors Colinear theory represents an ideal pin hole camera; however, in reality a camera deviates from this model. This deviation from theory is a combination of systematic errors: image deformation and lens distortion. Image distortion is related to physical deformation and limitations of the media used to capture the image, film or photosensitive electronics. For film it relates to unflatness of the film during exposure and elastic or plastic deformation of the film resulting from winding, changes in temperature, or handling during processing and developing. For electronic image capture sys tems, such as a CCD array, image distortion can be related to variations in sampling rate and syn chronization, the physical alignment of the imaging surface due to manufacturing defect and 34 Chapter 3: Photogrammetry vibration, thermal stresses, and the physical limitations of the sensor saturation (over exposure) and quantization limitations (under exposure). Image distortion can be modelled with a combination of two dimensional affine transformations. Lens distortion can be divided into two components: radial and decentering. Radial distortion is the bending of the ray that connects the feature point and the focus point towards the focal axis to intersect the image point. The term "decentering" refers to the lens focal axis not intersecting the image plane at the principal point. The decentering distortion initially results from a manufac turing defect but can vary with time with handling and the environment that the lens is used and stored in. Lens distortions can be modelled with polynomials. These distortions were not considered in the current implementation of this project because it is a proof of concept prototype that will be refined in future versions. Detailed explanations of image and lens distortions can be found in [Go sine 96, Karara 89, Tsai 85 & Weng 92a]. Systematic error correction terms (dun, dvn) can be added to the camera model to improve its representation of a real camera and in turn its accuracy, Equation (3.3) [Aziz 74]. These terms are typically computed in the calibration process, through the iterative least squares methods. The im plementation of this project in its current phase did not include systematic error; however it could be computed as: u + du = u —f. n n p Ju *I,I(*H -xc)+Ri,2(yn -yc)+Ri,3(z„ ~ZC) R3,\(*n -Xc)+R3,2(yn ~yc)+R3,3(Zn ~ *C ) (3.3.a) v + dv = v — f n n p Jv R2,\(xn -Xc) + R2,2(yn -yc)+R2,3^n ~ ZC ) *3,l(*« ~xc) + R3,2(yn 'Vc) +R3,3(Zn " *C > (3.3.b) where: (un , v„) = image coordinates of feature point n, 35 Chapter 3: Photogrammetry • (up , vp) = image coordinates of principal point, • dun, dvn = systematic error correction terms for image coordinates (un , v„), • (fu, fv) = focal lengths in horizontal and vertical components of image plane (typically fu =/v), www • R = a [3 x 3] orthogonal rotation matrix based on (\|/c, 8C, §c), www • (\|/c, Qc, §c) = orientation of the Camera coordinate system in the World coordinate system, www • (xc, yc,zc)= position of focal point (camera frame C) in the World coordinate frame, www • (xn , yn , zn ) = coordinates of feature point n in the World coordinate frame. 3.5 Direct Linear Transformation (DLT) The Direct Linear Transformation (DLT) is one implementation of the colinear theory, the pin hole camera model, Equation (3.3) [Aziz 74, Karara 89, Marzan 76, Miller 80 & Shapiro 77]. The DLT was originally designed to compute feature point locations in a scene from a pair of stereo pictures, taken with non-metric cameras. The image coordinates of the feature points are measured and then the location of each feature point is computed using the DLT triangulation equations. The triangulation procedure requires image coordinates from at least two images taken from different points of view. Image coordinates from additional points of view can be included into the triangu lation, making the problem over-determined and the error can be rmnimized, using a least squares solution. This data reduction system is based on feature points in the image and not the image as a whole, on a camera by camera basis. Appendix C provides a derivation of the DLT equations, as well as the calibration and triangulation equations. Through algebraic manipulation of Equation (3.3) the basic DLT equations are [Aziz 74]: www Lxxn +L2yn +L3zn +L4 un + dun = w w w , (3.4.a) L9xn +LWyn +Ll\Zn + 1 www , L5Xn +V« +L7^ +L8 vn + dvn = —________ , (3.4.b) L9xn +Lwyn +Lnzn + 1 36 Chapter 3: Photogrammetry where: • («„, vn) = image coordinates of feature point n, www • (xn ' vn ' zn ) = feature point n in scene in World coordinate frame, • L]_jj = DLT parameters, • (dun, dvn) = systematic error correction terms for feature point n. The DLT data reduction method has several advantages: • it is based on feature points and not the whole image, • it is independent of the means used to obtain the images, • it can be used with off the shelf non-metric cameras, • it can be used with any number of cameras (1 to N), • it has a straight forward implementation. These advantages, combined with consultation with other researchers working with similar technology in an unrelated field, resulted in its selection for this project. The eleven DLT parameters, LJ.J j in Equation (3.4) [Aziz 74], can be calibrated for with a di rect method if the systematic errors are ignored, because they are assumed to be small and negligi ble, and if the locations of the calibration feature points are accurately known. An iterative least squares method must be employed for a full calibration that includes systematic error correction terms (dun, dvn) that are included in the camera model to improve accuracy. In the current imple mentation of the Argo Project, the direct method is used; however, systematic error correction will have to be included in future versions as error can be significant. 3.6 Photogrammetry and Pose Extraction By using two images taken from different points of view, the 3-D location of a point in the scene can be inferred by triangulation. Triangulation is the projection of two rays from the image points through their respective focal points and out until they intersect. The point at which the two rays intersect is the location of the feature point in the scene. One systematic error associated with 37 Chapter 3: Photogrammetry triangulation arises when the projected rays do not intersect. To deal with this problem, a best guess approach can be used. If additional data is available from another camera, the triangulation becomes an over-determined problem, and can be resolved by least squares error minimization. Triangulation is a well studied problem and the methods of stereo photogrammetry can be found in a variety of image processing and photogrammetry texts [Duda 73, Faugeras 93, Gosine 96, Horn 86 & Nalwa 93]. A significant problem in triangulation is locating corresponding feature points between two images to the feature point in the scene, namely the correspondence problem. This problem is dealt with in the next section. Pose extraction from a set of images of a mobile rigid body (MRB) requires prior knowledge of its geometry relating to identifiable and observable features on the body. For example an ob server can determine both the position and orientation of a coffee cup from the location of the han dle. However, one has difficulty determining a full description of the orientation of a drinking glass as it is symmetric and has no identifiable feature. In this project, the feature points in the scene are retroreflective spherical markers. Through a combination of special illumination, filters, high contrast black and white images and preliminary image processing the features are reduced to points that represent the centroids of the markers. The markers appear as light spots on a dark back ground, and firm-ware in the image processing units identifies them and computes their centroids along with their height and width dimensions. A spot usually corresponds to a marker but this is not guaranteed. As a result only simple pose extraction techniques can be employed, thus eliminating pose determination from shape and edge detection and other more elaborate methods [Faugeras 93 & Horn 86]. To work within the constraints of having only image centroids of markers as input and to achieve the requirements that the MRB have identifiable and observable features, a target is at tached to it. The pose of a MRB with respect to the World frame, described by vector A" in Figure 3.3, can be inferred from the pose of the target in the world frame, vector F in Figure 3.3, provided 38 Chapter 3: Photogrammetry \ Target Frame A" = pose vector of MRB in World frame, unknown. World Frame Marker n B = pose of Target in MRB frame, known from target placement on rigid body. C = position of marker n in target, known from target geometry. D = position of marker n in MRB frame, known from B + C\ E = position of marker n in World frame, measured value from triangulation. F = pose of Target in World frame. Figure 3.3: Pose of Mobile Rigid Body and Target in World frame. that the target's pose relative to the MRB's coordinate system, vector B in Figure 3.3, must be known. The target is comprised of markers with known locations in the target coordinate system. To avoid ambiguities the target must have a minimum of four non-coplanar markers to be able to determine its orientation and position, that is, its pose. The pose of an object with respect to a given frame of reference is mathematically represented, as a transformation between two coordinate sys tems. The determination of pose of a target comprised of markers with known fixed locations from a set of images of a target can be done in a variety of ways. With a multi-camera systems the position of each marker on the target can be triangulated. Once the positions of the markers making up the target have been triangulated, two different least squares methods can be employed to obtain the pose of the MRB that the target is attached to. The first method solves for the transformation from the MRB coordinate system to the world coordinate system directly. The pose is then obtained by decomposing the resulting transformation to its six basic variables. The second method uses least squares to solve directly for the six variables that describe pose from 3-N nonlinear equations, 39 Chapter 3: Photogrammetry where N refers to the number of markers in the target. See Appendix B for details on both methods. A single camera system also has the ability to measure the pose of a MRB, with reduced accu racy due to low resolution in the focal axis of the camera. By using the dispersion pattern of rays projected through the image points, and the focal point out into the test volume of the scene, along with the geometry of the target, the pose can be determined. By forcing each marker position to lie on the appropriate ray a set of constraints can be developed and the pose of the target is deter mined and in turn the pose of the MRB. A variation of this method is to combine the pose transformation of the MRB, to the world frame, and a camera model into a pair of constraint equations, for each marker visible in the image. [Wang 91] Combining the roll-pitch-yaw-translation transformation and the DLT camera model results in Equation (3.5). The derivation of this pair is located in Appendix E. T m W ,m W T m W Tm L, • x„ +L9 • y„ +L, • z„ +L_ Un ~ 7^ W 7m W 7m W ~ ' (i.O.a.) L9 'xn +LlO-yn +L\\-Zn+l Tm W T m W Y m W , m m _ L5 -xn +L6 ' vn + Ll ' zn + L8 n , L9 -xn +LWyn +L\\-Zn + 1 where: W ,,W . /nW . MRB xn = co&($MRB)co&(QMRB)xn WW WWW MRB (3 5 c) + (-sin(^B)cos(\irM/?B) + cos((|)M/jB)sin(e^B)sin(\|/MOB))yn ' ^ • • 1 WW WWW MRB + (sin((|)M/jB)sin(^/?fi) + cosWMRB)sm(QMRB)cos(yMOB))zn w + XMRB W . ,,W . /r.W . MRB yn = sm(<\>MRB)cos(QMRB)xn WW WWW MRB + (cos((|>MflB)cos(\|/MrtB) + sin(4)M/?B)sin(eM^fl)sin(i|/M^))y/J WW WWW MRB ' + (-cos((()M^B)sin(\j/M/?B) + sin(^B)sin(eM^B)cos(\i/M/?B))z„ w + yMRB (3.5.d) W . ,nW N MRB ,aW . . . W . MRB ,~ c . zn = -sm(eMRB)xn + cos(QMRB)sm(\vMRB)yn (3.5.e) W W MRB W + COS(QMRB)COS(VMRBK + Z-MRB This method has the advantage that it requires only four non-coplanar markers to be visible in 40 Chapter 3: Photogrammetry the image, and any additional visible markers are used to minimize the error in pose estimate. If a marker is not visible for any reason, the system of constraints can still be used to estimate the pose. An additional benefit is that data collected with any additional cameras can easily be added to the set of constraint equations. With the previously mentioned multi-camera methods, a marker had to be visible in at least two images so that its position in the scene could be triangulated. With this method if a marker is only visible in one image it will still be used to provide two constraint equa tions. This method yields up to 2-M-N nonlinear constraint equations with six unknowns, M is the number of cameras used and N is the number of markers used. The six unknown pose terms may be solved for using a least squares method. A multi-camera system has a greater computational cost because it has to build a correspond ence map for the feature points for each image, and produces more constraint equations which must be solved. The advantage of this method is that it increases resolution and measurement volume, while lowering the measurement error. This combination of a pose transformation and a camera model is the basis of the observation model implemented in the extend Kalman filter pose trajectory tracker, discussed in chapter 5. The extended Kalman filter (EKF) is presented in chapter 4. 3.7 The Correspondence Problem The correspondence problem is matching the image of a feature with the corresponding feature in the scene. This is a significant problem and must be addressed before information about the scene can be recovered. Stereo triangulation relies on an image point in image A that has at most one unique matching image point in image B. The correspondence problem is an ambiguous one as depicted in Figure 3.4. As shown there are three possible triangulation solutions for the loca tions of three feature points, feature sets (1,2, 3), (4, 6, 8) and (5, 7, 9). Once paired, the set of image coordinates need to be identified in order to establish correspondence of target points. Sin gle camera and some multi-camera systems require a correspondence map between points in the 41 Chapter 3: Photogrammetry Field of View Bounds \ Left Focus Point Right Focus Point Figure 3.4: Ambiguous Correspondence of Stereo Images [modified form Nalwa 93] image and the feature point in the scene associated with that image point. The identity of an image point is crucial to obtaining correct and useful information about the scene from a set of images. Researchers have applied varied solutions to the correspondence problem. This section re views some of the methods relevant to the implementation of this project. The reader is directed to the literature for further details of the methods presented and descriptions of those not mentioned here, [Faugeras 93, Nalwa 93 & Zhang 92]. Some additional problems that can make the correspondence problem more difficult to solve are: [Zhang 92] • Occlusion: A feature point may be hidden from view from the target interacting with its envi ronment, the fixed scene, another target, or itself: markers can line up along a line of sight making only one marker visible for two. In Figure 3.4, if the feature point set (4,8,9) is consid ered although all of these features are inside the bounds of the FOV of both cameras feature 4 is not visible in the right image because it is occulded by feature 9, as they lie on a common 42 Chapter 3: Photogrammetry line of sight. Disappearance: Disappearance is when a feature leaves the bounds of the field of view (FOV) of a camera. This prevents the matching of that feature with other images, from another point of view that still have the feature in view. This is definitely a concern if an individual camera can not cover a large test volume with its FOV and maintain sufficient resolution to preform the test. It may be necessary to have multiple cameras, with crossing fields of view to fully cover a large test volume, so as to maintain sufficiently high resolution. If a marker is visible in only one image, there is no correspondence between images and its position can not be trian gulated. For example in Figure 3.4, feature 3 is visible in the FOV of both cameras; however, if feature 3 moves to the position of feature 10 then it will leave the FOV of the right camera and will have disappeared from view. Ghosts: Phantom or ghost markers are points visible in an image but have no corresponding valid feature point. They may be visible from multiple points of view and as a result can have correspondence between them but still they are noise and need to be filtered out before passing data onto the tracking or triangulation procedures. They can be caused by stray reflections off a shiny piece of metal or the surface of a liquid. This a particular problem when doing tests of a model ship in waves. If in Figure 3.4, feature 10 is considered a ghost marker and the feature set (1, 2, 3) are the real features then this will pose no problem when examining the right image but in the left image two problems will arise. The first problem is there are too many features in the left image for the number of expected features; therefore, which are the valid images. The second problem is that if the images of features 3 and 10 appear to overlap in the left image because of close proximity of the lines of sight for these two features then the centroid extraction of the feature could be shifted away from the real centroid of features 3 towards that of feature 10, thus corrupting the measured value taken from the image. Another scenerio that could cause a problem is if feature 10 is still considered a ghost marker and the feature set (1, 43 Chapter 3: Photogrammetry Left Focus Point Right Focus Point Figure 3.5: Epipolar Constraint [Nalwa 93] 3, 8) are the real features. In this case the position of feature 10 could be taken as the position of feature 3, as feature 3 is occluded by feature 8 resulting in no extra features in the left image. By accepting feature 10 as feature 3 in the left image an error is introduced that not only will cause incorrect results, it may also cause an instability in the tracking algorithm when feature 3 emerges from behind feature 8. • Absence: A feature may not be visible in the FOV that should be. The lack of this feature is not due to occlusion, disappearance, or error in feature extraction. For example, if active markers such as LEDs are being used and one of the LEDs should fail to light for some reason during a test, this would be absence. • Appearance: A feature may move into the FOV that was not previously visible. The new fea ture may cause confusion with existing features already visible. Appearance can result from an occluded feature becoming visible, an intermittent active marker returns from absence, or a disappeared feature returns to the field of view. To aid in solving the correspondence problem between two images, a geometric constraint called the epipolar constraint can be used as shown in Figure 3.5. A ray is project out into the scene 44 Chapter 3: Photogrammetry through the image point, the focal point and the feature point. This ray is then projected onto the image plane of the second camera, appearing as a line. This line is the epipolar line and the image point in the second camera corresponding to the one in the first image must lie on that line. The epipolar line may be thought of as the intersection of the epipolar plane and the second image plane. The epipolar plane is defined by three points: the focus points for the first and second im ages and the image point in the first image [Faugeras 93, Nalwa 93 & Zhang 92]. If the range of depth of the feature point in the scene is known and bounded then by projecting these boundaries onto the epipolar line in the second image, this limits the section of the epipolar line in which the image point may be located. An additional constraint that may sometimes be used depending on the scene and camera place ment is monotonic-ordering. Monotonic-ordering is when the left to right order of feature points in one image have the same left to right ordering in the second. This is shown in Figure 3.4 for the correspondence solution of feature set (1, 2, 3). However, this constraint is violated by the other possible solutions, for feature sets (4, 6, 8) and (5, 7, 9). Some systems take advantage of colours and unique geometric shapes to differentiate markers and aid in the correspondence process. Constraints from lighting and image intensity as well as edge continuity are also used in determining correspondence between images of the features and those in the scene. Hard corners, the intersection of two lines in an image can provide readily iden tifiable and observable features in a scene. Their geometry allow them to be easily identified in different images, providing correspondence between images and the scene. These cues and others are unfortunately lost and can not be exploited in the Argo project because the imaging system used returns only image coordinate data of suspected marker images. For a single camera system the correspondence problem involves identifying those feature points in the scene that correspond to the image points. The geometric constraints used to match points in a stereo system can not be applied to a single camera system. The method used in the 45 Chapter 3: Photogrammetry Argo project is to make an estimate of the image coordinates with known correspondence and com pare the estimates with the incoming image, [Allen 96 & Wang 91]. The incoming marker images are identified by matching them up with their corresponding estimate using a nearest neighbour search. The estimate of the image could be as simple as the previous image in the sequence or one gen erated from a dynamics model coupled with and observation model, as implemented in the EKF tracking module. This method has some definite advantages; it removes ghost markers in the im age because they have no corresponding estimate. As well, it identifies occluded or disappeared markers when the estimate has no corresponding image, and the estimate-measurement error is scaled down to meet available data. Also, if a valid marker appears in the image by entering the FOV, or emerging from an occluded position, the estimate of this feature will identify it and it will then be included into the list of valid measurements. 3.8 Summary This chapter contains an overview of the theory of how an image of scene is generated and how to obtain information about the scene that it is observing. The first section of this chapter presented a brief history of photogrammetry and an introduction to the remaining sections. The pin hole cam era and colinear theory section describes the function of a pinhole camera and how it can be used to model a real world camera. The systematic errors section explained the deviation from theory observed in real world cameras including lens and image distortions. The Direct Linear Transfor mation (DLT) is a linearized camera model based on the pin hole camera model. The DLT camera model was selected for the implementation of this project because it was designed for off the shelf non-metric cameras and it is based on feature points rather then the whole images captured by the cameras. This model is well suited to the camera and video processor hardware this project is in tended to be used with. Photogrammetry and pose extraction from the images captured is a funda mental component of tracking the pose of a mobile object undergoing MRB motion. Incorporating 46 Chapter 3: Photogrammetry the DLT camera model into a tracking algorithm, such as the Kalman filter discussed in the next chapter, allows for the tracking of a target's trajectory based on camera collected data. The final section dealt with the correspondence problem. This significant problem must be dealt with before any information about the scene can be extracted from the image. If the correspondence problem is dealt with incorrectly the interpretation of the image may yield appearantly valid data, yet be in correct. 47 Chapter 4: State Estimation, Tracking and the EKF Chapter 4: State Estimation, Tracking and the Extended Kalman Filter 4.1 Introduction As described in Section 1.2, the purpose of the Argo Project is to track of the pose of a scale model ship using multiple video cameras. In general terms, the problem is trajectory tracking of a randomly moving target using data collected with multiple nonlinear sensors in a noisy environ ment. This type of tracking problem is commonly tackled with an extended Kalman filter (EKF) [Faugeras 93]. To make the text generic to tracking the trajectory of an object based on video im ages, the scale model ship will be referred to as the mobile rigid body (MRB). The extended Kalman filter (EKF) was selected as the tracking algorithm for the Argo Project because it can be used to model nonlinear systems, and has been demonstrated to function in sim ilar vision-based token tracking systems [Faugeras 93, Hosie 95 & Wilson 93]. The term token can refer to feature points or markers, lines that define edges of rigid objects, and even 3 dimensional objects such as a cube [Faugeras 93]. In 1960, R. E. Kalman presented a new recursive approach to data filtering based on the Weiner filter [Kalman 60]. The Kalman filter (KF) is a linear recursive filter that generates an optimal es timate of the state of a dynamic system from a model of the system's dynamics and a set of direct and/or indirect measurements of the state variables. The state of the system refers to a set of vari ables that describe the inherent properties of the system at a particular instance. The discrete Ka lman filter is ideally suited to computer implementation because it is linear and it samples data at discrete intervals and models system dynamics as discrete events. The KF has been employed for a variety of tasks: data smoothing, trajectory tracking, forecasting and prediction of future trends, multi-sensor data fusion, and system parameter identification [Bar-Shalom 93]. The flexibility of this mathematical construct has seen it successfully applied in a wide variety of unrelated sectors: 48 Chapter 4: State Estimation, Tracking and the EKF military, finance, medical, industrial, and meteorology [Bar-Shalom 93 & Grewal 93]. As a result of this wide spread use, the KF is well studied and well documented in the literature [Anderson 79, Bar-Shalom 88, Bozic 94, Chui 91, Grewal 93 & Welch 97]. In the early 1960's following Kalman's revolutionary paper, S. F. Schmidt and his team at the National Aeronautics and Space Administration (NASA) [McGee 85] began to explore the possi bility of extending the Kalman filter (KF) to the application of nonlinear state estimation problems. As a result of this work, Schmidt introduced the extended Kalman filter (EKF). The EKF linearizes the non-linear state estimation problem by evaluating partial derivatives of the nonlinear state and observation constraint equations at the values of the estimated state variables for each iteration [McGee 85 & Schmidt 70]. Linearization of the problem permits the application of the linear Ka lman filter equations to the task of state estimation. In recognition of Schmidt's work, the EKF is also known as the Kalman-Schmidt filter [Grewal 93]. Similar to the KF, the EKF is a well studied technique and is well documented in literature [Bar-Shalom 88 & 93, Chui 91, Grewal 93 & Welch 97]. The EKF is required for the Argo project because of the nonlinear observation model relating pose of the mobile rigid body (MRB) to the image coordinates captured from video images. The basis for the observation model was presented in Chapter 3 along with a more detailed derivation in Appendix E. The EKF has been successfully applied to the problem of tracking the trajectory of a mobile rigid body (MRB) from a sequence of images [Faugeras 93, Hosie 95, Wang 91, Wil son 93, Wu 88]. The large number of authors of literature regarding Kalman filters, and the wide range of topics covered have resulted in variations in terminology and variable names. Some of the notable vari ations are: • the terms plant and dynamics model are used interchangeably. • the terms observation and measurement are used interchangeably, although, by definition, 49 Chapter 4: State Estimation, Tracking and the EKF measurements are observations related to the state of the system. • a(sk , k) and f(sk , k) are the non-linear functions that represent the plant/dynamics model. • Ak, Fk, and <&k are the variable names used for the linear/linearized dynamics matrix models. • c(sk , k) and h(sk, k) are the two variable names used for the non-linear observation function. • Ck and Hk are the variable names used for the linear/linearized observation matrix models. • Typically xk represents the state vector and zk is used for the measurement vector. To avoid confusion with the position variables Xpand Zp, the state and measurement vectors are repre sented hereinby sk and gk respectively. This substitution of variable names is also found in lit erature. This chapter will introduce the portions of the linear Kalman filter and the extended Kalman filter relevant to the Argo Project. The theory of the linear KF is presented in section 4.2 because it is the basis for the EKF, presented in section 4.3. Section 4.4 discusses design and practical con siderations relating to the implementation of the KF and EKF. 4.1.1 State Variables The state of a system describes the condition or inherent properties of the system being mod elled at a particular instance, represented by a set of j state variables. These state variables are the parameters in the state-space equations that numerically model the system. In general, a jth order ordinary differential equation can be transformed into a system of j, 1st order ordinary differential equations by variable substitution and order reduction. The resulting system of equations are the state equations, with j state variables, and the system is then termed to be in state space. These linear equations can then be manipulated with linear algebra. The number of state variables repre sents the degrees of freedom of the dynamic system being modelled [Grewal 93]. In general, the state variables are collected together into a single ./-element state vector, sk, where the subscript k is the base variable that describes the current discrete interval. For the application of tracking the pose of an MRB, the position and orientation, along with the corresponding velocity and acceler ation terms, are the state variables describing the MRB. 50 Chapter 4: State Estimation, Tracking and the EKF 4.2 Linear State Estimation and The Discrete Linear Kalman Filter The Kalman filter (KF) is a linear recursive optimal state estimator that can be configured for filtering, smoothing and predicting the state of a system. By taking into account the previous state, system dynamics, observations of the system, and statistical models of the expected observation noise and perturbations to the trajectory of the state the KF makes an optimal estimate of the current or future state of the system. The discrete Kalman filter is indexed at regular length intervals making it suitable for imple mentation on a digital computer. The incoming data is sampled at each interval; therefore, the data sampling rate is built into the filter. The system's dynamics are modelled as discrete events at each interval. The indexed base variable, represented by the subscript k, can be any independent varia ble provided that it only advances in one direction with constant interval length. In the majority of applications, the base variable is time, but it could be any advancing independent measurable quan tity such as distance travelled measured by a car's odometer. 4.2.1 Linear stochastic system model The state of the system being modelled evolves in time according to the discrete first order re cursive difference equation, Equation 4.1 [Bar-Shalom 88]. sk = Ak_lsk_l+Bk_luk_l+(dk_l , (4.1) Here, sk is the state vector, uk is the deterministic input vector and to^ is the dynamics noise vector. The dynamics model, Ak, is a set of state space equations that describes the rate of change of the state of the system. This model is used to generate the current state of the system, sk, based on the previous state, sk_\. The input matrix, Bk, defines the relationship between the deterministic input variables and the state variables. For the application herein, there is no control input1, uk; there fore, the Bk_\uk_\ term is dropped from Equation 4.1. Background noise that may corrupt the gen-1. Any further reference to the variable u in this document, unless otherwise specifically stated, is not related to a control input but rather to the horizontal image coordinate, where the subscript would be assigned a numerical value or a variable other than k. 51 Chapter 4: State Estimation, Tracking and the EKF eration of the state and uncertainty in the dynamics model is modelled as an additive, zero-mean, Gaussian white noise vector, (£>k. The output of a system will be either directly or indirectly related to the state of the system. The relationship that governs the output of the system defined in Equation 4.1 is modelled by Equa tion 4.2 [Bar-Shalom 88]. Sk = CkSk + Dkuk + ^k , (4.2) Here, gk is the output vector of the system, h,k is the measurement noise vector, sk is the state vector and uk is the deterministic input vector. The relationship between the current state, sk, and the ob-servable output, gk, is the observation model , Ck. The inverse of the observation model relates measurements taken of the system to the state of system. The direct transmission matrix, Dk, app-plies deterministic inputs directly to the output of the system. Since, there are no deterministic in puts, uk, for this application, the Dkuk term is dropped from Equation 4.2 and no longer referenced herein. Noise in the data collection process and the level of uncertainty in the sensor is modelled as an additive, zero-mean, Gaussian white noise vector, Dropping the deterministic input terms from Equations 4.1 and 4.2, yields the two models that are used in the following discussion of the linear Kalman filter: sk = Ak_xsk_x+tok_x , (4.3) 8 k = cksk + St • (4-4> 4.2.2 System Noise and Uncertainty Uncertainty in the two system models, dynamics and observation, is modelled as noise. The observation noise vector, c^, and the dynamics noise vector, (Ok, are assumed to be additive, uncor rected (defined by Equations 4.5 through 4.7) and zero-mean Gaussian white noise, (defined in Equation 4.8). 2. In some instances in the literature the observation model is referred to as the measurement model. 3. In most literature related to the Kalman filter the observation noise is referred to by the greek variable vk, but to avoid confusion with the vertical image coordinate variable vk, t,k is used in place of v^. 52 Chapter 4: State Estimation, Tracking and the EKF 2 = J = k) for all; and k , (4.5) yo j*kj 2 £<co,a)[> = fa"> >J = k) for all; and it , (4.6) E(^(dkT) = 0 for all; and it , (4.7E(^k) = E(ak) = 0 for all k , (4.8) It is unlikely that a linear state space model exists that perfectly represents a real world system. A real system has imperfections, potential cross coupling effects, and possible nonlinearities that might not be included in the state space model. Physical systems are also susceptible to environ mental uncertainties. Examples of environmental noise are the effect turbulence has on a flying aircraft, and the effect that chop has on the motion of a ship. In both cases the environmental ef fects have a noticeable local effect, but on the scale of the overall trajectory of the craft these effects can be modelled as small scale, high frequency noise. Such factors listed above will cause a deviation between the linear dynamics model of how the states interrelate and how they actually interrelate in the real world. The expected effect of these factors is modelled as noise, <% and added to the model to better reflect the real world situation. The level of noise must be estimated by the modeller and may have to be adjusted if its magnitute is too large or too small, to reflect the real world. Unfortunately, the magnitude of the noise at any given interval is not a measurable quantity, and the expected value of the noise, E((ak), is zero, as defined in Equation 4.8. Althought the expected value of the noise is zero, the variance of this noise is an assignable value that describes the probable range of the magnitude of the noise. The definition of the covariance of the dynamics noise vector, co^, is: Qk = E{((*k-E(ak))(<»k-E(ak))T) , (4.9) Substituting Equation 4.8 into Equation 4.9 reduces the dynamics covariance to: <2* = £<cW> , (4.1053 Chapter 4: State Estimation, Tracking and the EKF Observations made of the system output also contain some uncertainty which is modelled as noise, ^, and included into the observable output model, Equation 4.4. The uncertainty in this process is the combination of short comings in the observation model, Ck, and the physical limi tations and characteristics of the sensors and data acquisition equipment used to measure the sys tem output. Potential sources of inaccuracy in the observation model, Q, are vibrations resulting from not perfecdy stiff members, possibly producing unforeseen cross-coupling terms. Another potential source of inaccuracy is imperfect manufacturing of the system that may yield a differing geometry from that expected in the model. Data collection in the observation of the system outputs is subject to its own imperfections. Electronic instrumentation is prone to electronic interference that corrupts the signal, and small nonlinearities in the sensor that can change with time and/or tem perature. Care must be taken when estimating the signal to noise ratio as this will effect the con fidence in the observed outputs of the system. Since it is not possible to evaluate the actual magnitude of the noise, ^k, at each interval, the expected value of this noise, E(^k), is equal to zero, as defined in Equation 4.8. As with the dyanmics noise, the vairance of the observation noise is quantifiable. The covariance of the observation noise vector is defined as: Rk = E(£k-E(^k-EmT) , (4-11) Substituting Equation 4.8 into Equation 4.11 reduces the observation covariance to: Rk = E(UkT) , (4-124.2.3 State Estimation To evaluate the state, sk, and the corresponding output, gk, of a system using Equations 4.3 and 4.4 it is necessary to know the uncertainty terms, co^ and <^k, at each interval k. However, as indi cated above, this is not possible. Therefore, the best that can be achieved is to estimate4 the state, Sk, and output, gk, using the expected value of the noise terms. Also, since the state propagation equation is recursive, the initial state, s0, for k = 1, is required to start the process. With no way to 4. The hat symbol, A, above a variable name indicates that it is an estimate of the real value. 54 Chapter 4: State Estimation, Tracking and the EKF evaluate accurately the initial state the best that can be done is to use an estimated value, So. Ap plying this argument to Equations 4.3 and 4.4 yields Equations 4.13 and 4.14. h = Aksk.l+E((Ok) , (4.13) gk=Ckh + E(^k) , (4.14Substituting equation 4.8 into equations 4.13 and 4.14 reduces them to: h = Aksk_i , (4.15) h = Cksk , (4.16An evaluation of the effectiveness of the state estimation process described in Equation 4.15 may be done by comparing the estimated output, gk, generated by Equation 4.16, with the actual output, gk, of the physical system. ek = gk-h , (4.17) The Kalman filter (KF) uses knowledge of the expected uncertainty in the estimation process as well as the error between the estimated and measured outputs of the system, which indicate the overall signal to noise ratio at the current interval, to generate an optimal state estimate. The KF operates in a predict-correct loop sequence. The estimate of the current or future state of the system is determined by first making an initial guess as to what the current or future state of the system will be. A correction factor, based on observation of the system and the statistics associated with the initial guess and the observation processes, is added to the initial state estimate to produce a final estimate of the current or future state. As depicted in Figure 4.1, the discrete Kalman filter loop starts with the preliminary estimate of the current state5, sk , based on the previous state es timate, sk _ j, with the linear Equation 4.15, where Ak represents the set of linear state space equa tions that make up the dynamics model. 3h = Akh-\ , (4-18) t t The preliminary state estimate, sk , is then used to estimate of the output values, gk , using the 5. The superscript dagger, f, indicates that this variable is a preliminary estimate only. 55 Chapter 4: State Estimation, Tracking and the EKF Previous State: sk-i ~< Current State Estimate: sk 1 Preliminary State Estimation - t A -sh = Aksk_i $k = $k +Kkek Sk Observation Estimation gk = Cksk gk ^k = gk-gk tl Kkek Kalman Gain Kk = Pl^CkT[CkPk^CkT + Rk\' gk Ri PS Get Data Vector gk State Covariance Update I Delay I k-l State Covariance Prediction P^ = Ak_\Pk_\Ak_l +Qk_x Qk-i i i Figure 4.1: Discrete Kalman Filter Flowchart. 56 Chapter 4: State Estimation, Tracking and the EKF observation model, Ck, that relates the state of the system to observations made of the system, equa tion 4.19. _*f = , (4.19) The estimated error, , is the difference between values measured from observations made of the system6, gk, and those predicted from the preliminary state estimate, gk , Equation 4.20: h = gk-gk = ~gk-Cksk , (4.20) The optimal estimate of the current state, sk, is computed by adding to the preliminary state estimate, Sk , a correction factor, namely, the estimated error, e~k, scaled by the Kalman gain ma trix, Kk, as shown in Equation 4.21 [Bozic 94]. . sk = s^ + KkCgk-Ckh^ , (4-21) This recursive process repeats itself for each discrete interval. 4.2.4 The Kalman Gain and the State Error Covariance The Kalman gain matrix, Kk, scales the error between observed and estimated outputs, ek from Equation 4.20, to minimize the mean square of the state-estimation error. The state-estimation er ror, ek*, is the difference between the actual state, sk, and the state estimate, Sk, generated from the KF, as shown in Equation 4.22 [Welch 97]. ek* = sk-h , (4.22) The definition of the covariance of the state-estimation error is shown in Equation 4.23. Pk = E{(e*-E(ek*))(ek*-E{ek*))T) , (4.23) Since the KF minimizes the state-estimation error, the desired and expected value of the error is zero. E(ek*) = 0 , (4.24) Using Equations 4.22,4.23 and 4.24, the covariance of the state-estimation error, otherwise known 6. The tilde symbol, -, over the output variable g indicates that it is the vector of measured values that may be corrupted by the data collection process. 57 Chapter 4: State Estimation, Tracking and the EKF as either the updated, or a posteriori, state error covariance, reduces to Equation 4.25 [Welch 97]. Pk = E({sk-sk){sk-sk)T) , (4.25) Another important and similarly defined covariance term that is indicative of the accuracy of the preliminary state estimate, sk , is the predicted, or a priori, state error covariance, shown in Equa tion 4.26 [Welch 97]. /y = E((sk-3k)(sk-$k)T) , (4-26) These two covariance terms indicate how accurate the filter currently is and how it should be shifted if necessary to minimize the state estimation error. Their influence on the overall estima tion process of the KF is achieved by controlling the Kalman gain. The Kalman gain, Kk, represents the relative confidence between the state estimation and state observation processes and how this relationship effects the overall state error. The confidence of the state estimation and the state observation processes are defined as the dynamics noise covari ance, Qk, and the observaiton noise covarinace, Rk, respectively. Based on these two covariance models the Kalman gain, Kk, is adjusted as the state error covarinace, Pk, varies. A detailed expla nation and derivation of the Kalman gain, Kk, can be found in many texts related to the KF [Bar-Shalom 88 & 93, Bozic 94 & Grewal 93]. Inside the dashed box in Figure 4.1, a small recursive loop that self tunes the Kalman gain exists within the greater KF loop, which is itself recursive. To compute the Kalman gain, Kk, the predicted, or a priori, state error covariance, Pk\ is used in Equa tion 4.27 [Bozic 94]. Kk = P^JiC^cJ + R^ , (4.27) The calculation of the a priori state error covariance, Pk\ is expressed in terms of the updated, or a posteriori, state error covariance, Pk, as in Equation 4.28 [Bozic 94]. =Ak_lPk_]Ak_lT+Qk_l , (4.28) Closing the loop is the calculation for the a posteriori state error covariance, Pk, by referring to the Kalman gain, Kk, Equation 4.29 [Bozic 94]. 58 Chapter 4: State Estimation, Tracking and the EKF Pk = V-K^PJ , (4.29) This recursion process allows the filter to retain information about past state estimates and regu lates the KF by adjusting the confidence level placed on observations and initial state estimates from the dynamics model. In effect, the state error covariance matrix, Pk, acts as the memory of the filter, recording how past events influenced the filter's operation and the confidence levels of estimation and observation. The error covariance tends to increase as noise or instability in the dy namics of the system increase. Changes in the relative confidence levels between the estimation and observation processes is reflected in the Kalman gain, Kk. An increasing Kalman gain, Kk, indicates a reduction in the confidence level of the preliminary state estimate, sk , by placing higher confidence in the measured values that describe the system. A large gain will tend to make the KF have a rapid response to changes in measured values. Cor respondingly, a small gain indicates a high level of confidence in the state estimation process and little correction is required and/or there is low confidence in the measured values recorded. A small gain will make the KF sluggish and slow to respond to changes that are observed, as it will tend to keep with the a priori state estimate [Bar-Shalom 93]. Because of this, the Kalman gain should not be too small as to not respond to changes in trajectory and cause the filter to diverge, conversely it should not be too large as it will follow spurious noise in the trajectory that can lead to instability. 4.2.5 Kalman filter regulation If the error, ek, between the initial estimate of the outputs, g^, and the observed values, gk, + increases, the confidence in the initial state estimate, V > decreases, provided that the confidence in the measured values remains constant. This results in an increase in the state covariance, Pk, resulting in an increase in the Kalman gain, Kk. This has the effect of increasing the correction t factor added to the initial state estimate, , to compute the final state estimate, . If the expected variance of the dynamics noise is lower than the actual variance of the distur-59 Chapter 4: State Estimation, Tracking and the EKF bances, then the magnitude of the dynamics covariance matrix, Qk, will be too small. If this is the case, then confidence in the preliminary state estimate, sk , will be too high, forcing the state esti mation correction, based on observations, to be smaller than necessary for an optimal state esti mate. This has the effect of making the filter less responsive to changes in the trajectory of the target that were not anticipated by the design of the dynamics model. If the anticipated signal to noise ratio of the measurement of system outputs is actually too low, then the observation covari ance model, Rk, will be too small. In this case, the confidence in the measured outputs, g#, is too high. This has the effect of making the preliminary state estimate, , not significant enough in the overall estimate of the current state, sk. When the confidence in observations of the system is too high then the filter will tend to follow spurious trajectories resulting from noise artifacts from the data collection process. If Rk and Qk are both too big then all disturbances are accepted as being within parameter specifications resulting in a sluggish system that lags changes in trajectory. If Rk and Qk are both too small then the confidence in the models are too high and the filter will follow spurious noise and potentially be unstable. Since Kk is proportional to the ratio of Qj/Rf., if system noise levels increase, then the confi dence in the dynamics model, Ak, is lessened, implying an increase in the confidence of observa tions made. This change in confidence is manifest by increasing Kk which has the effect of increasing the correction factor to the preliminary state estimate, sk . If the measurement noise increases then Kk decreases and the confidence in the observations, gk, decreases and more weight is given to the state estimate. If it is possible to monitor the noise levels, the observation and dy namics covariance models can be adjusted up or down, tuning the filter to match the expected noise levels and produce the optimal trajectory estimate. Because the ability of the KF, used as a tracker, to follow a changing trajectory is sensitive to the expected noise levels, it is necessary select the values of these noise with great care. Another important factor governing the success of the KF to generate an optimal state estimate are the initial 60 Chapter 4: State Estimation, Tracking and the EKF conditions used to start the tracking process. 4.2.6 Kalman Filter Initialization Special care must be taken when initializing the KF, otherwise it may fail or at the very least take several more iterations than expected to establish track of the target. For the first iteration of the KF the index base variable is set to one, k = 1. To initialize the KF a starting state, s0, and an initial state error covariance, P0, are required. Also required are the dynamics model, A0, and the initial dynamics and observation noise covariance models, Q0 and R}, respectively. The starting state, s0, should be as close as possible to the actual state of the system being tracked. It can be determined by using a known starting point, from a best guess of the operator, or by taking meas urements of the system and running them backwards through the observation model. The initial state error covariance is more difficult to estimate as it depends on the dynamics of the system, the relationship between measured values and the state of the system and the expected sources of noise. A derivation of the initial state error covariance, P0, used in the implementation of the Argo Project can be found in Appendix D. To generate the three initial covariance models (PQ, QQ and Rj) an estimate of the sources of noise and the expected noise levels are necessary. Special care must be taken when estimating noise levels as they control the sensitivity and confidence of the filter to track the trajectory. If the expected noise levels and sources of noise do not change with time, they are considered constant. 4.3 Nonlinear State Estimation and the Extended Kalman Filter (EKF) In the event that the system dynamics model and/or the observation model are nonlinear, the direct use of the conventional Kalman filter for state estimation is not possible. However if the problem is first linearized, the linear Kalman filter theory and equations can be applied to state es timation of a nonlinear system. Nonlinear state propagation takes the general form of Equation 4.30. The nonlinear relation-61 Chapter 4: State Estimation, Tracking and the EKF ship between system output and the current state is represented in its general form by Equation 4.31. Equations 4.30 and 4.31 are analogous to their linear counterpart Equation 4.1 and 4.2. sk = a(sk_vuk_l,(Qk_l,k) , (4.30) gk = c{sk,uk,^k,k) , (4.31As before with the linear system, k is the index base variable, sk is the state vector, uk is the deter ministic input vector which again is unused and will be ignored in the following discussion, (dk is the dynamics noise vector, gk is the system output vector, and t,k is the observation noise vector. The function a, in Equation 4.30, represents a set of nonlinear state dynamics equations that relate the previous state of the system, sk.x, to current state, sk. There is one state dynamics equation for each state variable. The function c, in Equation 4.31, represents a set of nonlinear equations that describe the current output of the system, gk, based on the current state of the system, sk. There is one equation for each observable output of the system. Removing the unused deterministic inputs, uk, and assuming that the dynamics and observation noise terms are linear, additive, white and conform to the rules previously stated in Equations 4.5 through 4.8, Equations 4.30 and 4.31 become. sk = a(sk_vk) + (Hk , (4.32) gk = c(sk,k) + ^k , (4.33Making the assumption that the noise is linear and additive allows the use of the noise covari ance matrices defined earlier in Equations 4.10 and 4.12 to be used in the same manner in the EKF as in the KF. Comparing Figures 4.1 and 4.2, it is obvious that there are four differences between the struc tures of the KF and EKF loops. This comparison also shows the remaining structure of the EKF is taken directly from the KF. The first difference is that the preliminary state estimation process is now a nonlinear process, as shown in Equation 4.34. & = a(sk-uk) , (4.34) 62 Chapter 4: State Estimation, Tracking and the EKF Previous State: $k- i Current State Estimate: sk A Preliminary State Estimation Sk = a(sk_x,k) Sk = Sk +1 - t Sk 1 A r Observation Estimation 8k = c(sk\k) Linearization of Observation Model ck~Ysc{s'k) s = sk - t gk ek = h-gk + i gk + Kkek Kalman Gain KK = PK^CKT{CKPK^CKT + RK] -l RL Get Data Vector State Covariance Update IS Linearization of Dynamics Model lk-l Delay k-l PS State Covariance Prediction = AK_LPK_LAK_L +Qk_i Qk-Figure 4.2: Full Discrete Extended Kalman Filter Flowchart. 63 Chapter 4: State Estimation, Tracking and the EKF Secondly, the estimate of the output, gk , based on the preliminary state estimate, sk , is now a nonlinear relationship described by Equation 4.35. The two most notable differences between the KF and the EKF are the addition of the two linear ization processes to approximate the dynamics model, Ak, and the observation model, Ck, for com puting the state error covariance and the Kalman gain. The justification for these approximations is presented in the following subsection. Other than these four differences, the EKF functions and behaves in a very similar manner as the conventional KF, described in the previous section. The EKF tends to be less robust than the KF because of errors introduced in the linearization process. This requires that the equations used to model the real system be as accurate as is reasonably possible. Tthe anticipated signal to noise ratios for both the dynamics and observation models must also be carefully selected. 4.3.1 Linearization Since it not possible to know the actual state of a system, Equations 4.32 and 4.33 can not be evaluated. However, if an estimate of the last state of the system exists, a Taylor series expansion can be used to approximate the function a by evaluating it about the last estimate of the state, sk _ j. Similarly an approximation of the function c can be done by evaluating the Taylor series about the preliminary estimate of the current state, sk . HOT represents the higher order terms which are ignored. gk = c(sk,k) (4.35) a (sk_],k) = a(sk_hk)+-=^a(s,k)\s = ^ (sk_1-sk_l) + HOT (4.36) (4.37) (4.38) 64 Chapter 4: State Estimation, Tracking and the EKF gk = c(h\k)+Tc(s,k)\ ^(sk-sj) + $k , (4.39) OA \s - St The derivatives of Equations 4.38 and 4.39, with respect to the state estimates that their Taylor se ries expansions were evaluated at, are: ds k = ^-a(s,k)\ „ , (4.40) as is = sic-i Sk-l S = Sk- 1 <>8k ds = ^-c(s,k)\__4t , (4.41) t UA is — Sk s = sk The linear equivalents to Equations 4.34 and 4.35 are Equations 4.15 and 4.16, and are repeated for the reader's convenience as Equations 4.42 and 4.43. sk = Ak$k-i , (4-42) gk = Cksk , (4.43These two equations have the general form of a line that passes through the origin. In the case of Equation 4.42, the dynamics model, Ak, can be considered as the slope of the line and therefore the derivative of the current state estimate, sk, with respect to the previous state estimate, sk_ \. Therefore in the linearization process, it is reasonable to expect that the Jacobian of the vector of nonlinear state equations, a(sk_ \,k), with respect to the previous estimated state vector, sjc_ \, can be assigned as the dynamics model, Ak, as shown in Equations 4.44. The approximated dy namics model, Ak, is then used in the calculation of the state error covarinace prediction, Pk\ Ak~^-a(s,k)\ , , (4.44) K OS >s = Sk-i For the output estimation, Equation 4.43, the observation model, Ck, is the slope for the linear re lationship between the independent variable, the current state estimate, s^, and the estimated out put, gk. As with the dynamics model it is reasonable to assign that the Jacobian of the vector of nonlinear output equations, c(S^, k), with respect to the current estimated state vector, sk, as the slope in the observation model, Ck, shown in Equation 4.45. This approximation of the observation 65 Chapter 4: State Estimation, Tracking and the EKF model is substituted into the linear KF equations in place of the linear version when computing the Kalman gain and the state error covariance update. Ck~^-c(s,k)\ . , (4.45) * as \s = sk 4.3.2 About the Extended Kalman Filter The extended Kalman filter (EKF) is a nonlinear state estimator that takes its general structure from the conventional Kalman filter. The EKF linearizes the nonlinear function(s) about the cur rent estimated state at each discrete interval. The EKF is a sub-optimal filter, unlike the KF which is an optimal state estimator. An optimal nonlinear filter is difficult to realize and is typically com putationally intensive [Bar-Shalom 93]. For the sake of practical implementation this accommo dation is made and as a result the filter is detuned. As has been shown for the EKF, the dynamics and observation models are linearized about an estimated state; however, it is possible for a nominal state to be used in place of the estimated state. This process, normally termed as the linearized Kalman filter (LKF), has the advantage that the filter gains and the state error covariance can be computed in advance, off-line of the real-time op eration. The disadvantage of this method is that it requires a priori knowledge of the state trajec tory and it is not very robust against nonlinear approximation errors. If the actual state deviates from this trajectory; then the filter will either fail because of divergence or it will yield unreliable results. For this reason the LKF was rejected for the Argo Project, as it is intended to track the pose of any MRB through any trajectory, without a priori knowledge of the trajectory and it should be considered random. The EKF linearizes about an estimated trajectory which leads to more accu rate trajectory estimates as it adapts and reacts to perturbations to the trajectory. The main disad vantage of the EKF is that the Kalman gains and the state error covariances can not be precomputed requiring on-line/real-time computations. These are the reasons that the EKF tends to be more ro bust than the LKF but more difficult to realize in real-time applications 66 Chapter 4: State Estimation, Tracking and the EKF 4.3.3 Argo Project EKF requirements For the Argo Project, the dynamics of the ship motions are modelled with linear kinematics equations, while the observation model that relates camera data and the pose of the MRB is non linear. This means that the requirements for the state estimation process are for a partial discrete extended Kalman filter where only the observation model is linearized. The EKF loop required for the Argo Project is depicted in Figure 4.3. 4.4 Design and implementation considerations The KF and the EKF can both fail to track the trajectory of a target for a variety of reasons. The two main modes of failure of the KF and EKF are failure to establish track of the trajectory and failure to maintain track of the trajectory. When the filter fails to maintain track of the trajectory it is said to diverge; this is an all encompassing term that has many possible causes. Failure can be the result of a single problem but is often the reslut of a combination of problems. The EKF tends to be more sensitive than the KF to many of these problems because of the linearization process, making it less stable. Some of the causes for failure of the KF and EKF are listed below [Grewal 93]: • The (E)KF can diverge if the filter gain, K, becomes small and the observation data are still significant and needed to correct the state estimate. • The (E)KF uses the dynamics model, A, to approximate how an actual physical system changes state. If this model does not sufficiently represent the physical system, the filter will fail or yield suspect results. An inappropriate dynamics model, Ak, will yield a false preliminary state estimate, sk , that in turn will cause a false estimate of the observable system outputs, gk • It will also cause an error in the state error covariance prediction, Pk^. • The second model used by an (E)KF is the observation model, Ck, that relates measurements made of the dynamic system to the state of the system. A mismatched observation model, Ck, 67 Chapter 4: State Estimation, Tracking and the EKF Previous State: sk- Current State Estimate: sk 1 Preliminary State Estimation $k = Aksk-\ sk = Sk + Kkek - t + Observation Estimation gk = c(sk,k) Linearization of Observation Model ckssds'c('s'k) ~ t s = sk gk ek = gk~ gk Kkek Kalman Gain Kk = PfcfiCffc' + R^ gk Pj Get Data Vector gk State Covariance Update Pk = V-KkW HE a Delay | F k-l State Covariance Prediction Pk' = Ak_\Pk-\^-k-\ +Qk-\ Qk-Figure 4.3: Partial Discrete Extended Kalman Filter Flowchart. 68 Chapter 4: State Estimation, Tracking and the EKF 4-will yield a false observation estimate, gk , and will also cause errors in the state error covari ance update, Pk, and the Kalman gain, Kk. An error in the observation estimate, gk , will result in an error in the estimated observation error term, ek, that will affect the accuracy of the cur rent state estimate, sk. If the linearized dynamics and observation models do not sufficiently represent the original nonlinear relationships, the EKF will fail. This problem may be overcome by adding higher order terms into the linearization process. However, this should be avoided as this would fur ther deviate the filter from Kalman's original work. Improper estimation of the expected noise levels in either the observation and dynamics noise covariance models (Q and R) can lead to either: • an unstable filter that will react to noise and spurious trends if the estimated noise levels are too small. Any noise outside the expected bounds will be perceived incorrectly as a change in trajectory. • a sluggish filter that lags the actual trajectory if the expected noise levels are too high as small changes in trajectory or measured values can be perceived as allowable noise. Not until the change in trajectory becomes significant will the filter start to adapt to the change and then slowly. In effect the (E)KF acts as a low pass filter: the more sluggish the filter becomes the lower is the cut-off frequency. Some ill-conditioned problems relating to implementation of the (E)KF are [Grewal 93]: • Errors or large uncertainties in the values of the dynamics, observation and statistical mod els have not been allowed for in the design of the Kalman filter and may cause it to diverge. • Large numerical ranges of values in the state variables, measured values, or the matrix models indicate inappropriate scaling or dimensional units used in the design of the filter. This can lead to round-off errors. • The intermediate matrix [CkPk^Ck + Rk] in the Kalman gain calculation could be non-69 Chapter 4: State Estimation, Tracking and the EKF invertible. • Large dynamics and observation models have large matrix dimensions that are computa tionally and memory intensive to manipulate. The number of arithmetic operations increases by the square or the cube of the matrix dimension, depending on the operation. Each additional arithmetic operation introduces a round-off error and increases the compu tation time of each iteration through the loop. This can be a limiting factor in real-time sys tems. • Poor machine precision can also result in large round-off errors and lead to numerical insta bility of the filter. • The filter can fail to maintain track if the incremental changes between samples are too large. This can be the result of too low a sampling rate or if disturbances are too large. • False data or falsely interpreted data can lead to confusion and the filter may track an uni-tended trajectory. • A good initial estimate of the state of the system is required otherwise the filter may never acquire the track. The KF and the EKF can be implemented in a parallel structure for parallel processing. This will improve the computational time performance to meet the scheduling requirements of real-time filtering [Chui 91]. 4.5 Summary The Kalman filter is a linear optimal state estimator, ideally suited for tracking trajectories and trends of a dynamic system under either direct or indirect observation. The KF is a very flexible tool and can be implemented to track, to smooth data, to extract a signal from a noisy background and data fusion, and to combine data gathered from multiple sources to observe a system. The Ka lman filter can not be directly applied to the problem of tracking a MRB using multiple video cam-70 Chapter 4: State Estimation, Tracking and the EKF eras because of the nonlinearity of the observation outputs of the system. However, the KF can be applied to the study of a nonlinear system by incorporating a linearization step. The resulting filter is sub-optimal because of the linearization process implemented, and is referred to as the extended Kalman filter (EKF). 71 Chapter 5: Argo Project Implementation Chapter 5: Argo Project Implementation 5.1 Introduction As stated before, the goal of the Argo Project is to track, in real-time, the rigid body motion of a model ship operating in the simulated sea of a towing tank or manoeuvring basin. The current implementation of this goal is at the proof of concept stage. The computer program can track sim ulated and experimental data, off-line. The system has been designed around the Qualisys MacReflex™ hardware so as to be compat ible with on-going research programs at NRC-IMD, NRC-Hydraulics, and MUN Faculty of Engi neering. With a common hardware base, different projects can concentrate on various functions and resources can be shared. The research component of this project is in the software that performs image data reduction and tracks the motion of the mobile rigid body (MRB) in the test volume. This research effort in volves the combination of video technology, photogrammetry, state estimation and trajectory tracking. The necessary background on these different areas has been presented in the previous three chapters and accompanying appendices. The video technology allows for capturing snap shots in time of a dynamic system under observation. Photogrammetry allows for the interpreta tion of the pose of the MRB at a particular instance in time. Also, photogrammetry shows the re lationship between observable outputs of a system and the corresponding image coordinates recorded, through what is known as a camera model. State estimation and trajectory tracking fuses the data collected from the different video cameras together, introduces the time element, and gen erates the best estimate of motion of the MRB. 72 Chapter 5: Argo Project Implementation 5.2 Methodology From image coordinates of known feature points of an MRB captured with multiple synchro nized video cameras, the Argo Project tracks the six degrees of freedom (6 DOF) of motion of an MRB inside a known test volume. The basis for the tracking algorithm used in the Argo Project is the partially extended Kalman filter (EKF), presented in the previous chapter. This algorithm gen erates an estimate of the state variables, the pose of the MRB, from observable outputs of the sys tem, the image coordinates corresponding to known feature points on the MRB. A flowchart of the tracking algorithm of the Argo Project is shown in Figure 5.1. It is necessary to use the EKF rather than the regular Kalman filter (KF) because the observable outputs do not have a linear re lationship to the state variables that describe the system. This algorithm, along with the associated model, and state variables are described in the following subsections. 5.2.1 State Variables and the System Dynamics Model For the Argo Project, the state of the system is the pose of the MRB with respect to an inertial world frame of reference, along with the first and second order derivatives of pose with respect to time. For tracking the motion of a model ship, the inertial frame of reference is defined by the ge ometry of the towing tank or manoeuvring basin. The reference frame of the model ship has its origin at the centre of gravity (CG) of the model ship as defined in Figure 1.1. The six pose terms combined with the associated six velocity terms and six acceleration terms yield 18 state variables, shown in vector form in Equation (5.1). W -W -W W AW -W W ;W ~W VMRB VMRB VfMRB ^MRB "MRB "MRB YMRB ^MRB §MRB (5.1) w -W ..W W .W ..W W -W -W XMRB XMRB *MRB y\lRB yMRB yMRB ZMRB ZMRB ZMRB where: \)/ = roll angle, 6 = pitch angle, § = yaw angle, x = surge displacement, y = sway displacement, z = heave displacement. 73 PL S S 3 3 8 » H) (D n 11 CD co O CD II — 3 CD 3 O -r- O 8 9r a. CD Q. CD CO I—»• 3' 2. CD Q TJ 3- _ o -a <_ 3 Si _> CD' o —•* o' 3 D CD 03 *< -Q CO I*- . £ CD O o < 03 I*-I I 03 3 o CD I + I*-I CD Q. o' •—»--i O' 3 to I O Tl Tl 03 d 03 _§ 5 9 * 5L o CD < CD 0Q> 0Q> I T 03 CD 5. 3 CD 5 - Q) -a cS 8 co s; o o — CO ° 7$ CD 9J O CQ 3 a CD ~ &0> <~1 m o z a CT o ~ CO 3 3 CD = Si 2 3 => o -o QL CD On o c CT 3 CO CD CD 03 3 N 03 _. 03 ig S o o -* Q. CD CV O CD 03 CD Q. CD 03 CD _ a ^ a CD <D O ^ Si I TJ — CD 3 3 03 "2 CO «—t-a CD m CO I—^ 3' a 6' 3 —1-&5> —t-+ O 03 03 o CD 3 CD a 5' 3 CO 3' c_ 03 o' 3 TJ — CD < O c CO CO 03 CD I o CD 03 *< X o c — — CD 3 CO a CD Chapter 5: Argo Project Implementation To make the tracking module capable of following any random motion of an MRB, the dynam ics model assumes uniformly accelerated linear motion for each degree of freedom with no cross-coupling terms. This simple kinematics model was selected for four reasons: (i) it is a linear rep resentation of virtually any motion provided the step-size is small enough, (ii) it requires no a pri ori knowledge of the system dynamics and can accommodate a changing system dynamic, (HI) it simplifies and minimizes computations as no linearization is required, and (iv) it has been demon strated to function, provided the target motion is sufficiently over-sampled [Wilson 93]. The uni formly accelerated linear motion dynamics model for a given degree of freedom is shown in Equation (5.2). a = where: 1 X 0 1 0 0 (5.2) • x = sample period of the observations made of the system, • a = the [3 x 3] matrix that describes the 3rd order linear motion of a single degree of freedom. With no cross coupling, the single degree of freedom model is repeated for each degree of freedom in the system dynamics model. a 0 0 0 0 0 0 a 0 0 0 0 0 0 a 0 0 0 0 0 0 a 0 0 0 0 0 0 a 0 0 0 0 0 0 a (5.3) Replacing this simplified dynamics model with a more detailed one, specific to the system un der observation, would improve the state estimation process. However, this limits the universality of the tracker for use with systems with dynamics that differ from those modelled, and for systems 75 Chapter 5: Argo Project Implementation where the dynamics change during the test period. The situation of a system with changing dy namics is relevant to seakeeping tests of a model ship because of the free surface effects of water on deck or in a compartment will change the system's mass and mass distribution. Furthermore, building a specific dynamics model also requires extensive a priori knowledge of the system and how it will perform under all possible conditions that may be experienced during a test. 5.2.2 Observable Outputs and the Observation Model The observable outputs of the system are measurable physical responses , that either directly or indirectly describe the state of the system. Since the Argo Project is a non-contact measurement system, it is not possible to directly measure the pose of the MRB. Hence, an indirect observable output is necessary. The indirect method chosen for the Argo Project is to use the image coordi nates of feature points of known locations on the MRB. The observation model relates the state of the system, the pose of the MRB, to the observable outputs, the image coordinates of feature points on the MRB. The feature points on the MRB are high contrast markers that make up a target that is rigidly mounted to the MRB. The observable outputs are the image coordinates corresponding to the cen-troids of these markers. A minimum of four non-coplanar markers are required to define a volume that has a unique pose solution. Therefore, it is necessary that the target attached to the MRB con tain a minimum of four non-coplanar markers. Having additional markers in the target is desirable to add a level of redundancy that will potentially lower the overall measurement error and will pro vide enough data if one or more markers are occluded from view. Ping-pong-ball-sized spheres were selected as the markers because their centroid has the same apparent position from all angles of view. These spheres are coated with retroreflective paint, giv ing them a high contrast appearance compared with the MRB and the background of the scene. Markers are used in place of feature points native to the MRB because they are readily identi fiable in an image and they are independent of the MRB. In general, the use of standardized feature 76 Chapter 5: Argo Project Implementation points, independent of the MRB, is advantageous because it is not necessary to reconfigure the fea ture point extraction process for each new MRB tested. Rather than attaching individual markers directly to the MRB, it is recommended that the markers be attached to the MRB as a collective group, known as a target. The markers have known positions in the 3D space of the target's coordinate frame. Knowing the pose of the target's coordinate frame relative to that of the MRB's allows for the determination of the marker locations in the MRB coordinate system. This transformation is shown in Equation (5.4). Hence the investigator, studying the motion of an MRB, need only know the pose of the tar get, rather than having to survey the location of each marker on the MRB for each test schedule. This makes the experimental setup simpler, speedier and more accurate. MRB MRB MRB MRBHTarget (5.4) where: * MRBHrarget lS a roll-pitch-yaw-translation homogeneous transformation from the target coor dinate frame to the MRB coordinate frame, defined in Section 2 of Appendix B. The relationship between the state of the system and the observable outputs is known as the observation model. In this project, the observation model is the combination of two separate trans formations. The first transformation converts the locations of the target markers from the MRB's coordinate frame to the world coordinate frame, as shown in Equation (5.5). The world coordinate frame is the base reference frame that the pose of the MRB is measured relative to. 77 Chapter 5: Argo Project Implementation W w yn w n 1 where: = WHMRB MRB MRB MRB (5.5) WHMRB is a roll-pitch-yaw-translation homogeneous transformation from the MRB coordinate frame to the world coordinate frame, defined in Section 2 of Appendix B. The second transformation is a camera model that relates three dimensional (3D) points in the world coordinate system to the two dimensional (2D) coordinates in the image plane. For the Argo Project the Direct Linear Transformation (DLT) was selected as the camera model. Details of the DLT and the reasons for its selection are presented in Chapter 3 and Appendix C of this document. The combination of the DLT and transformation of marker coordinates from the MRB reference frame to the world reference frame results in the nonlinear observation model, represented by ck(sk>k) m Figure 5.1. The definition of the observation model is shown in Equations (5.6.a) through (5.6.e). ,in W Tm W rm W T m L\ Xn +L2 +L3 -Zn +L4 L9 - xn +L\Q-yn +Lll-Zn +1 vn = L5 - xn +L6 + L1 'Zn + L8 L9 ' xn +L\0-yn +L\\-Zn + 1 where: w Xn = W W MRB *« - COS(^MRB)COS(QMRBK WW W + (S™(<$>MRB)COS(VMRB) + cos((|)^B)sin( QML + (sin(<t)^/fS)sin(\|/^r^ • — w YV + XMRB ^ MRB' n W WWW MRB WMRB) + C0^MRB)SIN(QMRB)SIN(^MRB))yn W WWW MRB FMRB) + cos((t>M/?B)sin(eM/?fi)COS(VM/fB))^ (5.6.a) (5.6.b) (5.6.c) 78 Chapter 5: Argo Project Implementation w • (J* N raw ^ MRB /c c ^ yn = sm(^MRB)COS(QMRB)Xn > (5-6-d) WW WWW MRB + (cos((t>Mfl5)cos(\|/MOB) + sin(^B)sin(e^B)sin(\|/^s))y„ WW WWW MRB + (-cos($MRB)sm(\\fMRB) + sin(^,MRB)sm(QMRB)cos(^MRB))zn w + yMRB W . ,aW . MRB /CtW . • . W . MRB , . zn = sm(QMRB)xn +cos(QMRB)sm(\TiMRB)yn , (5.6.e) W W MOB W ^V0S(QMRB)C0S(VMRBK + ZMRB w w w w w w (VMRB> ®MRB' §MRB> XMRB> VM/?B' ZMRB) XT tne state variables that represent the pose of the MRB's coordinate frame relative to the world's coordinate frame. ("«' vn ) are tne image coordinates of marker n, in the image plane of camera m, a pair of observable outputs. h"l_ J , are the DLT parameters for camera m. The eleven DLT parameters, L, _ n , for each camera are determined through calibration. A complete description of the calibration procedure is given in Section C.2. The nonlinear observation model function, ck, generates the vector of observable outputs, gk, using the following loop structure: for m = 1 to M for n = 1 to N gt[(2xiVx(m-l) + 2xn-l),l] = u™ (5.7.a) gk[(2xNx(m-l) + 2xn), 1] = v™ (5.7.b) end end As part of the EKF execution, the nonlinear observation model is linearized by taking the gra dient of the observation outputs, gk, with respect to the state vector, s^., as shown in Equation (5.8), C* s I,-,4 • <5'8> The resulting linearized observation model, C^, is a matrix that is comprised of alternating rows corresponding to the alternating pattern of the (w, v, w, v) of the observation output vector, gk. The 79 Chapter 5: Argo Project Implementation general form of the rows that make up the linearized observation model, Ck, are as follows. Row 2xAfx(m-l) + 2Xrt-l of the linearized observation matrix is shown in Equation (5.9.a), I^VMRB d0M/?B d$'MRB UAMRB uy MRB U<-MRB Row 2xNx(m-l) + 2xn of the linearized observation matrix is shown in Equation (5.9.b), 0 0 w 0 0 du „ w 0 0 du n „ „ dun _ dun 2 W MRB 0 0 3 w WMRB 0 0 dz w 0 0 (5.9.a) °VMRB 0 0 3v 89 w MRB 0 0 0 0 3v 0 0 ^VWRB C>XMRB ^MRB 0 0 3v! 3z MRB 0 0 (5.9.b) Since the observation model has no direct relationship with the state variables corresponding to the higher order terms, velocities and accelerations, the partial derivatives of the observations with re spect to these terms are set to zero. A detailed derivation of the linearization of the observation model is found in Appendix E of this document. 5.2.3 Expected Errors and Covariance Models 5.2.3.1 System Dynamics Noise The system dynamics model described in Equations (5.2) and (5.3), is a third order 6 DOF sys tem. It was decided to account for inaccuracies in the dynamics model and perturbations to the system trajectory by assuming it could be modelled by white noise introduced into the dynamics model through the acceleration term. This is shown for a single degree of freedom in Equation (5.10), 1 X ~2 0 1 x 0 0 1 xk- 1 0 Xk-l + 0 _ G)v Xk-\ X (5.10) It follows from the dynamics model that this assumtion is the same for each degree of freedom with the expected noise levels being adjusted independently for each degree of freedom. The noise vec tor for a third order 6 DOF system is shown in Equation (5.11), 80 Chapter 5: Argo Project Implementation (x)k = [o 0 co, 0 0 co2 0 0 co3 0 0 co4 0 0 co5 0 0 co6] (5.11) The definition of the dynamics covariance from Chapter 4, is repeated here as Equation (5.12), Qk = E((*kakT) . (5.12) From this definition the dynamics covariance, Qk, for a third order, 6 DOF system is an [18 x 18] square matrix. Just as there is no cross-coupling in the dynamics model and by the definition that the noise is uncorrelated, there are no cross-coupling terms in the dynamics covariance. The gen eral form of the dynamics covariance matrix for a single degree of freedom is described in Equation (5.13.a), 0 0 0 0 0 0 0 0 a; CO (5.13.a) The single degree of freedom model is repeated for each additional degree of freedom as shown in Equation (5.13.b), Qk = (5.13.b) P 0 0 0 0 0 0 p 0 0 0 0 0 0 p 0 0 0 0 0 o p 0 0 0 0 o o p 0 0 0 0 0 o p 5.2.3.2 Observation Noise Observation noise represents inaccuracies in the observation model and the noise in the signal coming from the sensor and data acquisition system. In the Argo Project, the assumption is made that two types of observable output exist: the horizontal image coordinate and the vertical image coordinate. A further assumption may be made that these two types of observable outputs are col lected with different sensors, even thought they originate from the same image plane. This as sumption is made to account for different horizontal and vertical resolutions in the camera. The 81 Chapter 5: Argo Project Implementation observation noise vector alternates with the same (w, v, u, v) pattern of the observation model as shown in Equation (5.14), <-1 p\ y\ pi ptn ptn pM pM Slw Slv $2H S2y — hnH ••• hNH (5.14) 2 2 2 Op = Op = Op The definition of the observation covariance from Chapter 4, is repeated here as Equation (5.15), Rk = E(vkvkT) . (5.15) For the Argo Project, it is assumed that the variance in observations made in the horizontal and vertical axes are constant and equal, (5.16) This assumption is made because the technical manual for the Qualisys camera system indicates one variance for both horizontal and vertical [Qualisys 92]. It is also felt that variance due to the observation model itself is likely to be similar for both axes. Each observed image coordinate has two components, horizontal and vertical. Therefore, the observation covariance matrix is square with the dimensions [2 • M• Nx2 • M• N], where M = number of cameras and N = number of markers on the target. The diagonal elements of this matrix are the variance of the observations, as shown in Equation (5.17), 0 0 0 0 0 0 2 0 0 0 0 0 0 2 0 0 0 0 0 0 2 0 0 0 0 0 0 2 °* 0 0 0 0 0 0 (5.17) 5.2.4 Data Acceptance and Rejection, and Pair Matching The vector of estimated image coordinates, gk , assumes that all markers are visible and that there are no ghost markers present. To be able to compute the error, ek, between the estimated, 82 Chapter 5: Argo Project Implementation gk. , and observed, gk, image coordinates, the two vectors must have the same length (same number of observation) and have the same order. To ensure that the two vectors correspond, a data acceptance and rejection process coupled with pair matching and sorting routines are required. Figure 5.2 illustrates some of the difficulties associated with the acceptance and rejection process. / ^ ^ 6 \ IT *1 ] / 1 ) . l?o / I 2* ^ / 1 ( )8 i ' V ^ " \ ( )9 y r )10X / , \K ' \ 5« » 1 4 / Let: • The filled circles (1, 2, 3,4, 5) represent markers at the estimated image coordi nates. • The hollow circles (6, 7, 8, 9, 10) repre sent markers at the observed image coordinates. • The large dashed circles are the acceptance regions of the estimated image coordinates. Figure 5.2: Image coordinates acceptance and rejection Examining Figure 5.2, there are some obvious pairing assignments between the estimated im age coordinates and the observed image coordinates. At the same time, it is also appearant that there are both estimated and observed image coordinates without pairs that must be rejected. Also, it illustrates an area of ambiguity, where the pair matching is not conclusive. These markers will be rejected. Initially all estimated and measured observations are assumed to be rejected. As observations are accepted they are so tagged. At the end of the process all observations that have not been ac cepted are filtered out. The acceptance process begins by computing the radial distance between an estimate image co ordinate and all of the observed image coordinates. This list of radial distance is then tested to de termine how many distances fall inside a preset tolerance. Considering estimated marker #1, only one observed marker #7 falls into the circle of acceptance; therefore, they are temporarily accepted as a valid pair. For estimated marker #2 and observed marker #6, this is also the case. For esti-83 Chapter 5: Argo Project Implementation mated marker #3 two observed markers, #8 and #10, fall inside the acceptance region; however, both can not be valid. In this situation the estimated marker #3 and both the observed markers #8 and #10 are rejected because the pair matching is inconclusive. Estimated marker #4, only has ob served marker #10 inside the acceptance circle, but since marker #10 has already been rejected in a previous validation test, estimated marker #4 is also rejected. If marker #4 had been tested prior to testing marker #3, marker #4 would have been paired with marker #10. But when marker #10 failed because of the double acceptance for marker #3, both marker #10 and marker #4 would then be rejected. Estimated marker #5 has no corresponding observed marker, possibly due to occlu sion; therefore, it is rejected. Any remaining observed markers, for example marker #9, that were not paired with a unique estimated marker or were not rejected because of overlap, are considered as ghosts and are summarily rejected. This leaves estimated marker #1 paired with observed mark er #7 and estimated marker #2 paired with observed marker #6. All others in the image are rejected and removed from the observation vector along with the corresponding rows in the linearized ob servation model, Q, and the observation covariance model, Rk. The last task is to ensure that the order of the incoming observation vector, gk, matches that of the estimated observation vector, gk . Image coordinates in a video image are referenced left to right, top to bottom, with the origin of the image plane in the upper left corner. Referring back to Figure 5.2, the order for the estimated markers is #1 follows #2; however, their observed pairs are in the reverse order, #6 precedes #7 because #6 is one row above #7. As a result the incoming observation vector, gk, is sorted to match the order of the estimated observation vector, gk , so that the error between measured observations and estimated observations can be computed. This process is repeated for each separate field of view in a given frame. For example, for a three camera system there are three images per frame. A more detailed discussion of the data acceptance and rejection, pair matching, and sorting may be found in Appendix F. 84 Chapter 5: Argo Project Implementation 5.2.5 Initialization of Tracker and State Error Covariance To initialize the tracking process, it is necessary to provide an initial state estimate, so, and state error covariance, PQ. Every effort should be made to provide the best possible initial estimates as this will help the EKF establish track sooner. The initial dynamics model, A0, the dynamics co-variance, QQ, and the observation covariance, /?], are considered constant and are defined based on known and estimated parameters. All other terms in the tracking algorithm are computed from these basic terms. During experimental and simulation tests conducted with the Argo Project, the target is held still for the first ten seconds to establish the initial pose. The initial pose is computed by taking the observation model, Equations (5.6.a) through (5.6.e), and solving for the six pose terms, w w w w w w (VMRB> QMRB> §MRB> XMRB' yiARB' ZMRB) ' using a least squares method. When determining the initial pose, the target is assumed to be stationary; therefore, the initial velocity and acceleration terms of pose are zero. As a result, the initial state estimate, SQ , is only a function of the initial pose of the MRB. The initial state error covariance, P0, is based on the dynamics model and the expected dynam ics and observation noise. Since the state error covariance is a self adjusting entity within the track er, any inaccuracies in this initial guess will be self corrected with the execution of the tracker. For the models used in this project, it was determined that an appropriate initial estimate of the state error covariance, P0, is as shown in Equations (5.18.a) and (5.18.b). A detailed derivation of this initial estimate of the state error covariance is given in Appendix D. The initial state error covari ance for a single degree of freedom is: 85 Chapter 5: Argo Project Implementation r = 2 2 °£ X a% 2— 3— ' x — 3— aco + 6^ (5.18.a) XT X The single degree of freedom is repeated for each additional degree of freedom. The 6 DOF initial state error covariance model is: rooooo oroooo oorooo oooroo ooooro o o o o o r (5.18.D) 5.2.6 Incoming Data and Simulated Data Generation Figure 5.1 shows that the Argo Project may be configured for two possible sources for the data stream of image coordinates. The intended use of the program is to process recorded experimental data gathered from a set of synchronized video cameras. However, for testing purposes, simulated data is fed to the tracker in place of recorded experi mental data. The simulated data stream is similar in form to a data stream captured from the camera system and is generated independently of the main tracking algorithm. It uses only the target ge ometry and the current frame number as inputs to this subroutine. The process for generating sim ulated data is first to generate the pose of the MRB in the world coordinate system for the current image frame. Using this pose, the positions of markers in the MRB coordinate system are trans formed into the world coordinate system. With the marker locations in the world coordinate sys tem they are then projected onto the image plane. The current version of the simulator uses a 86 Chapter 5: Argo Project Implementation simple plane perspective projection camera model to generate the image coordinates. It is planned to upgrade this model to be more representative of the data gathered with real cameras. This cam era model is completely separate from the one used in the observation model, although it is very similar in operation. 5.3 Hardware Implementation The hardware components necessary for the Argo Project are: • a host computer. • video processors that act as frame grabbers and also do preliminary image processing. • synchronized video cameras equipped with strobes. • a target of known geometry attached to the MRB. A schematic of the configuration of these components is shown in Figure 5.3. RS-422 Daisy Chain Network Video r Processor 1 Video Camera 2 Processor 2 Video Processor m Host Computer Marker Mobile Rigid Body (Model Ship) Figure 5.3: Argo Project Hardware Configuration 87 Chapter 5: Argo Project Implementation 5.3.1 Video Cameras and Processors The video cameras and video processor hardware used for this research is the Qualisys MacRe-flex™ system. The reasons for this selection are presented in Chapter 2. The Qualisys MacRe-flex™ system supports one to seven video processor and camera pairs linked to the host computer through a single local area network. The video camera used in the Qualisys Mac Reflex™ system is a modified industrial CCD vid eo camera with PAL video encoding. The system has sample rates that can be set up to a maximum of 50 Hz, twice the frame rate of the PAL standard of 25 frames/sec. Qualisys employs a trick that effectively doubles the frame rate of the camera by treating each field of a frame as a separate im age. A full interlaced video frame is comprised of two fields. A field is every other line of a full interlaced video frame, in effect the first field would be all the odd numbered lines in a frame and the second field is all the even number lines of the same frame. When these two fields are meshed together the full frame is formed [Greaves]. This results in a 604(H) X 294(V) image, which has the same horizontal resolution as a full frame but only half the vertical resolution of a full frame [Qualisys 92]. The camera passes to the video processor a nearly all black image except for some high intensity bright spots that most likely represent markers visible in the scene. To achieve this the camera has four features not normally found on industrial video cameras: • A flash, made of directional IR LEDs clustered around the lens, is used to flood the field of view (FOV) for that camera with IR light. The retroreflective nature of the markers reflect this light back in the same direction it originated. Since the flash surrounds the lens, most of the returning light is captured by the lens lowering the power required for the flash. The net effect of this is that high intensity spots in the image likely correspond to markers in the scene. • The lens is fitted with an IR band pass filter. Therefore, only IR light sources will be seen in the image. These light sources can be reflections from the IR flash or from IR sources such as LEDs or LASERS. • A high speed electronic shutter, l/1250th of a second [Qualisys 92] synchronized with the flash is used. This short exposure time allows only high intensity light sources to register in the 88 Chapter 5: Argo Project Implementation image. • Variable gain is used to control light amplification. High intensity light sources in the image are reduced to normal levels and all others elements of the picture become dark. [Qualisys 92] Together, these features eliminate everything from the image that is not a strong IR source or reflector. This eliminates most concerns about ambient lighting and allows researchers to work safely in a well lit environment. High intensity circular patches in the image are perceived to be images of markers in the scene. However, this is not guaranteed as wave crests and shiny surfaces such as a plated bolt or a piece of aluminium in the scene may also return a signal, termed as a ghost marker. The determination of whether the return is a valid marker is handled by the data acceptance-rejection tests, mentioned previously in Section 5.2.4. The proprietary Qualisys video processor acts as camera controller and power supply, frame grabber, and performs feature extraction using built in firmware. The video processor first grabs the high contrast black and white image and identifies light blobs in the image as potential features. Then, with subpixel interpolation, it determines the centroid, height and width of each potential feature. These data are captured to the computer with the acquisition and camera control software where they are stored or processed. TM An Apple Macintosh computer is the intended host computer for the Qualisys Mac Reflex system. As a result, communication between the video processors and with the host computer is done via a daisy chained RS-422 serial connection, native to the Apple Macintosh. This daisy chained network runs at the maximum 500 kbaud set in the video processor hardware. The rela tively low band width of this serial RS-422 network limits the number of markers that can be tracked at any one time. The host computer uses the serial network to send a reset command to all the video processors and an acquire command to the master video processor. The master video processor receives the 89 Chapter 5: Argo Project Implementation acquire command and starts an internal clock, with an adjustable frequency (default is 50 Hz), that in turn triggers the slave units to simultaneously take a picture of the scene. Each video processor simultaneously captures an image, and through dedicated hardware and internal software, returns the image coordinates of the centroid of each image blob above a preset intensity threshold along with the corresponding height and width. After completing its image processing, it will transmit to the host computer the camera identification code, the number of bright spots found in the image, the frame number and the data for each bright spot in the image. This data is used for further data reduction and tracking by software on the host computer. Additional details on the communication protocols, the data stream structure and the physical characteristics of the Qualisys Mac Reflex™ system hardware can be found in their technical man uals [Qualisys 92]. Recentiy, Qualisys has made available 60 Hz systems that are based on the NTSC standard and they now support the PC environment. They are also considering replacing the RS-422 serial con nection with a higher speed network to increase the data bandwidth, allowing for more markers to be tracked. 5.3.2 Camera Placement The placement of cameras to observe the motion of an MRB is important for a successful test. There are two components to camera placement, the position and the orientation of the camera. There are several points to consider when choosing camera placement. The test volume that the MRB travels in is defined by the field of view (FOV) of the multiple video cameras monitoring the scene. Selecting the FOV, through camera placement and the lens focal length, is a compromise between the desire to have high resolution and to have a large test volume. The larger the field of view of the camera, by placing it farther away from the scene or using a wide angle lens, increases the test volume and lowers the resolution with which the scene is observed. Conversely, a narrow, close field of view forces the testing volume to be small but 90 Chapter 5: Argo Project Implementation increases the resolution with which the scene is observed. To avoid distortion of the image the camera should be placed so that during a test the target appears in the centre of the image. Because a component of lens distortion is a function of radius, the farther away the target is from the focal axis, the more pronounced the distortion becomes. The proper camera orientation can help eliminate or alleviate some of these problems. Human beings like to see images right way up and assume automatically that an image is unless there are other indicators to the contrary. However, the computer has no preconceived notions of how an image should appear. Because of this a camera may be placed on its side or at an odd angle if it helps to view the scene better. For example, consider a regular video image, that is wider than it is tall, used to observe a target that has most of its travel in the vertical axis. It would be more suit able to place the camera on its side in order to align the image axis with the greatest resolution to the axis of greatest interest or of largest displacement in the image. Also, efforts should be made to minimize or eliminate obstructing from the view of a camera any part of the scene that the MRB is likely to travel. 5.3.3 Host Computer The host computer collects the data from the video processor units and runs the tracking soft ware. The host computer selected for the Argo Project is an Intel x86 processor family based per sonal computer (PC). The reasons for this selection were to use existing computer equipment and to maintain compatibility with the system under development at NRC-IMD. A Quatech MPA 200 RS-422 card was used to interface the Qualisys video processors to the host computer. 5.3.4 Target The target is comprised of a minimum of four non-coplanar, spherical, retroreflective markers at known locations in the 3D space of the target coordinate system. The markers used in the target are either spherical, hemispherical or disks that are roughly the size of ping pong balls and are coat ed in a retroreflective paint which is highly reflective of light in the near-IR wavelengths. The 91 Chapter 5: Argo Project Implementation sphere is the preferred marker shape as the position of its observed centroid is fixed for all angles of view. In the case of a hemisphere, viewing it from the side will shift the centroid away from the flat edge by a factor of 0.4244 x the radius of the hemisphere [Gieck 90]. Disk markers, on the other hand, can not be viewed from the side at all. The markers are mounted to a frame to form a target. The target frame structural members should be painted flat black with no shiny fastners or brackets left exposed to try and eliminate the possibility of stray reflections that could result in ghost markers. Similar care should be taken when preparing the MRB for testing. Although, there is no need to paint the MRB flat black. The target is attached to the MRB with a known pose relative to the MRB's frame of reference. This allows for the mapping of the marker locations into the MRB reference frame. The target ge ometry, the marker locations in the target coordinate system, should be carefully selected to avoid symmetry in the layout. This is necessary to avoid confusion when attempting to identify markers TM in a image. A poor example of a target is the Qualisys calibration frame that has six markers defining the comers of a vertical equilateral triangular prism, shown in Figure 5.4. In an image, this geometry appears the same upside down or when rotated ±120° about its vertical axis. In ad-Blue Corner Stay Wires Qualisys Calibration Target Marker locations (units are mm) X Y Z 1 ( 123, 0, 662) 2 ( 446, 773, 662) 3 ( 954, 107, 662) 4 ( 123, o, 88) 5 ( 446, 773, 88) 6 ( 954, 107, 88) [Qualisys ??] Figure 5.4: Qualisys Calibration Target Geometry 92 Chapter 5: Argo Project Implementation dition the markers on this target frame are equally spaced making it even more difficult. This is the target frame that was used in the experimental testing of this project. A more suitable target is one where the markers are randomly placed and the structure of the target does not occlude the markers. An example of this is the NRC tree target, shown in Figure 5.5. 5.4 Software Implementation The motion tracking software component of this system is a set of subroutines collectively known as the program Argo. The program tracks the 6 DOF rigid body motion of an arbitrary MRB with a known target attached to it. This program has two main modes of operation: simula tion and processing experimental data. The tracking algorithm implemented is a scalable modified EKF loop with selectable data sources: computer generated image coordinates or image coordi nates captured from the Qualisys hardware. The Argo program is not currently stand alone and operates in the MATLAB™ environment. MATLAB™ is an interpreter-based high level programming environment that is relatively user NRC-IMD Target Geometry Marker locations (units are mm) • 2 P 3 1 X Y Z 1 ( 533, 4, 510) 4 (0, 0, 0) 2 ( -19, 14, 432) Q /I 3 ( 1, 421, 404) 4 (-584, -2, 2) f 5 \. 7 5 ( -3, -736, -3) 6 ( 2, 207, -190) z 7 ( 215, 0, -194) 8 ( 2, -322, -302) \ 9 ( -420 0, -407) Mounting Base / [Sullivan PC] Figure 5.5: NRC Tree Target Geometry 93 Chapter 5: Argo Project Implementation friendly and permits efficient coding of algorithms. The disadvantages of this environment are that it can not produce a stand alone product and that the program's execution may be slower then an optimized and compiled version of the same program. To turn this research effort into a useable product, the software will have to be ported to a compiler based programming language such as C, C++ or Pascal. There are four reasons for this: (/) the program must be stand alone, (ii) the program must be able to interface with the system hardware to read the data streams from the video proces sor units, (iii) the program must be able to operate under a real-time operating system, and (iv) se curity against theft of intellectual property and unauthorized modification of the program is greatly improved with a compiled version because the source code is not directly available to the user. The program's different operation modes are controlled by a configuration vector that contains option flags. These option flags control everything from where the data stream originates, which graphs to generate, which diagnostic routines to run, and other internal house keeping chores. The program queries the user with prompts such as what data file to load and where to store the proc essed data. This program can also operate in a batch mode where the program queries a batch file rather than the user, freeing them from having to be present while the computer processes data. 5.4.1 Interfacing with Qualisys hardware Since, Argo has been written in the MATLAB™ environment, it is not possible to communi cate with the Qualisys hardware in real time; therefore, existing external programs were used to capture and store the image coordinates generated by the video processors. The program used to capture the image coordinates and then store them to binary data files was written and developed at NRC-IMD [Sullivan PC]. This program is part of their in house research effort in motion track ing of model ships. An additional software utility furnished by NRC-IMD translates these binary data files to ASCII text files which can then read by the Argo program. 94 Chapter 5: Argo Project Implementation 5.4.2 Program Initialization The Argo program starts with a multi-step internal initialization process that: • clears system memory; • loads the configuration vector containing option flag values that control the operation of the program; • if in batch mode, it loads the batch file and deactivates all further user prompts (this option allows the program to run for extended periods of time without direct supervision by the user); • starts the screen activity log to record all user inputs and program text outputs • logs the start time; • defines global and system variables specific to the mode of operation; • calibrates the cameras for the session; • if processing recorded data, it loads the image coordinate data stream history into memory. 5.4.3 The Main Program Loop To be able to process a series of image sequences,the tracking loop is nested inside a main loop that indexes through a set of simulations or data files to processed. If the source of the image se quence is a recorded data file, then it is loaded into memory. If triangulation of a stationary target is being done, the marker correspondence lists for each camera are loaded. These correspondence lists have to be determined by hand in advance of running this option. The target geometry, target pose with respect to MRB space and marker locations in target space, are loaded and the marker locations are transformed into MRB space. With all the appropriate data loaded, there are three options: generate a scatter plot of the entire image sequence, triangulate a stationary object, or track a moving object. At the beginning of tracking any sequence, the initial conditions for the EKF are set based on a triangulation of the starting pose of the target. The EKF tracking loop is indexed with the current frame number. At the beginning of each cycle through the loop, a request is made for the next frame, and all the image coordinates associated with that frame. If the program is running in sim-95 Chapter 5: Argo Project Implementation ulation mode the frame number is sent to the Softcam program and a set of image coordinates is returned; if running in data processing mode, the next frame is pulled from the frame history. Once the data stream has been preprocessed by the data acceptance and rejection routines, it is then passed onto the EKF loop for current state estimation. Each state vector is recorded and stored in a file for later analysis. 5.4.4 Simulation and subroutine Softcam Softcam is a subprogram inside the Argo program that generates a data stream of image coor dinates with the same form as those captured from the hardware. Softcam, although contained in the main program, is independent in operation. The interaction between Softcam and Argo is lim ited to Argo passing the current frame number as an input to Softcam and receiving the image co ordinate vector as the output. Using the frame number and a known frame rate, a subroutine internal to Softcam uses a set of hard coded equations of motion for each degree of freedom to gen erate the next pose that the MRB will experience in the simulation. This generated pose is then used to compute the 3D position of the markers in world space. The camera model used to generate the image coordinates of these markers is plane perspective projection, a simple pin hole camera model with no lens distortion. This camera model is completely independent of the one that is part of the observation model in the EKF tracker. 5.4.5 Diagnostic subroutines The Argo program contains several diagnostic service routines to help in the assessment of new algorithms. These routines would have to be removed from a final product because they greatly impede computational efficiency. A scatter plot of the full image sequence allows the operator to check for ghost images in the image sequence. This plot is also useful for generating the image correspondence lists needed for triangulation and calibration. Triangulation is done as a check of calibration and target placement. Triangulation can be done 96 Chapter 5: Argo Project Implementation on a single or averaged image if the correspondence lists are provided. This does not necessarily preclude triangulating the locations of markers on a moving target but a correspondence list would have to be generated for each image. As a result, this is usually reserved for stationary targets. As a double check, the inter-marker distances for the known geometry and the triangulated markers are compared and non-dimensionalized to give a single measure of error in the triangulation. 5.5 Summary The Argo Project, using an EKF based tracking algorithm, tracks the six degrees of freedom motion of an MRB following a continuous and random trajectory with data gathered from a set of synchronized video cameras. The tracking software takes the stream of image coordinates corre sponding to the centroids of potential markers as input to the tracking algorithm. If a potential marker in the image can not be conclusively identified through a series of tests and paired with its corresponding estimate, it is rejected by the data acceptance and rejection routine. This software is flexible in its design because it can track the trajectory of an MRB from a se quence of image coordinates, independent of the source. Any image capture system, film or video, that will return the centroids of potential markers in the image is sufficient. Certain dependencies will have to be built into the software for specific image capture systems; however, these modifi cations are external to the core of the program that will remain unchanged. TM The software component of this system has been developed in the MATLAB environment for off line execution. There are plans to port this program to the C language so that it can be com piled for real-time operation. The advantages to this system are: • Any number of cameras, from one to as many as the hardware can support, can be used to study a scene. The more cameras that are used, the greater the reduction in the error in the state estimate. 97 Chapter 5: Argo Project Implementation Ghost images that may appear in the image are handled. Occluded markers are handled provided that not too many are hidden from view. Since a sin gle camera can be supported, if a marker is only visible in one view, the image coordinates are still valid data points and they are included into the measurement vector. Combining the above listed advantages also allows for the possibility of an increased test vol ume. With a multiple camera system the target can go out of the view of one camera and come into the view of another while constantly remaining in the view of a third camera and still be tracked. This means that test resolution need not be sacrificed because of a large test volume. Many cameras could be combined together to cover the large area and not compromise on the resolution of the data. 98 Chapter 6: Experimental and Simulations Results Chapter 6: Experimental and Simulation Results 6.1 Introduction The Argo Project software was tested by both simulation and processing of data collected from physical experiments. Simulations have been ongoing through out the project to evaluate how new features and refinements affect the system's performance. During the summer of 1996, physical experimental tests were conducted at the NRC-IMD facility in St. John's, Newfoundland, in con junction with C-CORE's Intelligent System Division. The data collected in these tests were sub sequently post-processed using the Argo software to track the motions of the moving target. The results of experimental and simulation testing of the Argo Project indicate that the concepts behind this research effort do work, but need further refinement to make this work into a practical tool for towing tanks. It appears that the tracked oscillations tends to lag physical oscillations in excess of 1/2 hertz, while frequencies higher than 20 hertz are filtered out of the signal. 6.2 Evaluation by Simulation Simulation tests with the Argo software were done, by generating a stream of image coordi nates in a form similar to the data stream, that comes from the video processor units. These simu lated image coordinates are generated by: (i) calculating the pose of the simulated MRB based on a set of known equations of motion where the frame number is used as the time variable; (ii) mapping the locations of the target markers from the MRB coordinate system to the world coordinate system; (iii) projecting the marker positions in the world coordinate system onto the image plane, through a camera model separate from the one used in the tracking algorithm. This procedure functions entirely separate from the main tracking program. The only common 99 Chapter 6: Experimental and Simulations Results information between the two programs are the current frame number, the target geometry, and the pose of the target with respect to the MRB's coordinate system. The physical setup represented by the simulation discussed herein is depicted in Figure 6.11. The test volume is defined by three cameras observing the scene. The cameras are modelled as 1/2 inch CCD video cameras, each with an image plane measuring 6.4 mm wide by 4.8 mm tall. The camera located at (0 m, 0 m, 4 m) is fitted with a 10.0 mm lens with a horizontal field of view of 35.5°. The remaining two cameras are fitted with 8.5 mm lenses that have a horizontal field of view 31.5°. All cameras have a downward looking angle of roughly 20° below the horizon. The target used in this test is the NRC Tree Target, shown in Figure 5.5. This target is attached to the MRB with a position of (0 m, 0 m, 1 m) and an orientation of (0°, 0°, 0°), relative to the MRB's coordinate system. With a combination of three cameras (Af = 3) and nine markers on the target (N = 9), the system has a maximum of (2 • M • N) 54 constraint equations defining the pose of the MRB in the world coordinate system. The results of this simulation are presented in Figures 6.2 through 6.7. To avoid cluttering the figures, the units of the data are listed in Table 3.1. In these figures, the solid lines represent the known inputs and the dashed lines denote the estimated tracked output of the EKF. The particulars of this simulation are: • Simulation duration is 60 seconds; • Sampling rate is 100 frames/second; therefore, the simulation is represented by 6000 frames; • The generated roll motion is corrupted with white noise, with a signal to noise ratio of 75; • Roll motion (rotation about the x axis) is generated by using a sinusoidal motion, with an amplitude of 5 degrees, at a frequency of 0.10 hertz, about 0 degrees mean angle; • Pitch motion (rotation about the y axis) is generated by using a sinusoidal motion, with an amplitude of 10 degrees, at a frequency of 0.13 hertz, about 0 degrees mean angle; • Yaw (rotation about the z axis) is held fixed at 10 degrees; • Surge motion (motion along the x axis of the world coordinate system) is generated using a 1. Note: figures in this chapter are included at the end of the chapter due to their size and number 100 Chapter 6: Experimental and Simulations Results sinusoidal motion with an amplitude of 2.5 metres, at a frequency 0.05 hertz, and with an offset of 8.5 metres; • Sway motion (motion along the y axis of the world coordinate system) is generated using a sinusoidal motion with an amplitude of 2.0 metres, at a frequency of 0.05 Hz, and with an off set of 9 metres; • Heave motion (motion along the z axis of the world coordinate system) is generated using a sinusoidal motion with an amplitude of 0.4 metres, at a frequency 0.25 hertz, and with an offset of 0.4 metres. The results show that the tracked pose trajectory closely matches that of the known trajectory, as shown in Figure 6.2. There is some error in each channel of pose, but it should be noted that the error is at least one order of magnitude less than the signal being observed, as shown in Figure 6.3. The roll error shows the high frequency noise contained in the input signal, indicating that the EKF filtered it out of the signal. The yaw error shows how error in one channel can effect another. Prior to frame 500, the target is essentially motionless, with the exception of a slow drift in the roll and pitch terms. During this time period the error in yaw is minimal, after frame 500 the target starts to move, and immediately the level of yaw error increase. This increase in error is an artifact of the EKF tracking process. The low frequency motions in the x-y plane (surge and sway) show ex cellent position tracking. The higher frequency motion of heave shows a slight over-shoot com pared with the known heave motion. It is felt that this error could be reduced by adjusting the gain of the EKF. This simulation also tested the robustness of the Detection and Rejection function of the Argo program as marker four (n = 4) moved in and out of view of camera one (ra = 1). The system suc cessfully managed to deal with a marker that disappeared from view and then reappeared later on. The Detection and Rejection function also rejected data if it could not be conclusively identified for purposes of correspondence mapping. 101 Chapter 6: Experimental and Simulations Results 6.3 Experimental Testing The physical experimental testing of the Argo Project consisted of using combinations of two to four Qualisys MacReflex™ Motion Sensor systems observing a target attached to the bottom of a swinging pendulum, as shown in Figure 6.8. The data streams from the Qualisys video processors, are captured to the memory in the host computer and then stored in binary coded files. These files were later translated to ASCII format so that they could be read into the MATLAB™ programming environment for processing by the Argo software. The software used for controlling the Qualisys hardware and capturing the incom ing data streams, as well as the translations software was provided by NRC-IMD. The Qualisys calibration frame was used as the target, and is shown in Figure 5.4. Provided that all six markers on the target (N = 6) are visible, in all four video cameras (M = 4), then the maximum number of constraint equations defining the pose of the target is (2 • M • N) 48. The pendulum used in these experiments was an existing aluminium swing-frame, with the Qualisys calibration frame attached to the base of the frame, as described in Figures 6.9. This swing-frame, originally designed to determine the radius of gyration of ship models, was made available by NRC-IMD. Lead bars were added to the base of the swing-frame, to increase the swing period and to decrease the damping effect. By adding the lead bars the effective length of the pendulum was approximately 2.23 metres. The pivot axis of the pendulum was 4.0 metres above the floor. An accelerometer based inclinometer were used to log the angle of the swing-frame. The in clinometer data were sampled at 100 Hz using a data acquisition system that was completely sep arate from, and not synchronized with, that of the camera system. The data gathered from the inclinometer can be used to corroborate the amplitude, frequency, and rate of decay of the tracked results. Unfortunately because the video stream and accelerometer data captures were not synchro-102 Chapter 6: Experimental and Simulations Results nized it is not possible to evaluate the phase of the tracked output compared with a secondary in dependent source output. 6.3.1 Experimental Procedure At the beginning of each test session, the cameras recorded images of the target with the pen dulum held stationary, in the plumb condition, for calibration purposes. Also when ever any of the cameras were moved either intentionally or accidentally, or if the cameras were left unsupervised, images of the stationary target were recorded for recalibration of the cameras. With the calibration images logged, the swing-frame was then setup for either free and forced swing tests. For the static tests, the swing-frame was held fixed at a known angle. This test was done pri marily for calibration purposes but was also used for triangulation and pose extraction tests. For the majority of the static tests all four cameras were used; however, some of the tests were done with two and three camera configurations. For free swing tests the swing-frame was set to an arbitrary start angle, which was recorded with a portable digital inclinometer. The swing-frame was held in place by a string tied to a fixed anchor point. The camera and the accelerometer data acquisition systems were started and logged data for 5 to 15 seconds prior to the release of the swing-frame. The string was then cut to release the pendulum. This method was simple, did not interfere with the pendulum, and did not introduce any artificial motions. For the forced swing tests, the pendulum was pushed by hand in an irregular fashion producing a psuedo-random trajectory. In both forced and free dynamic tests, the camera and the accelerometer data acquisition sys tems were started and logged data for 5 to 15 seconds while the pendulum was stationary to estab lish a starting state for the EKF. To test different modes of operation and the error handling capabilities of the software, differ ent test configurations were tried. Some of the tests were conducted using combinations of two 103 Chapter 6: Experimental and Simulations Results cameras and three cameras, even thought all four cameras were available. This was done, to try and observe if the additional data from more than two cameras, improved the ability to track the target. The tests were inconclusive and should be repeated both in simulation and by additional physical tests. Some of the reasons for not being able to draw any conclusions from these tests are that, not enough runs could be analysed, because the software was not able to track the trajectories of many tests runs. Occlusions were also tested by intentionally placing panels between a camera and the target. As the target swung some of the markers would move in and out view of the camera. Error handling for ghost markers was also tested by intentionally placing an extra marker in the scene for some of the tests. To test a random trajectory, a few tests were conducted where the swing-frame motion was forced by pushing the swing-frame with an erratic motion. 6.3.2 Experimental Results During the three days of experimental testing, there were a total of 78 tests recorded. The 78 tests can be broken down into the follow three classes: 50 were static tests, 23 were dynamic free swing tests (regular motion), and 5 were dynamic forced swing tests (irregular motion). Unfortunately, of the 28 dynamic tests recorded, the Argo software successfully tracked only two dynamic free swing tests. It is believed that the failure to track the pendlum's motion, for the majority of the tests, can be attibuted to three reasons: (i) The camera model used in the observation model, do not include a lens distortion model. This is significant because, as the target moves away from the focal axis of the camera, the radial lens distortion comes into effect. (ii) The camera model in the observation model was likely not sufficiently calibrated for the test ing being done. This oversight in calibration was not realized until after the testing was com pleted and further testing was not possible. At the time of the testing the level of required calibration was still unknown and as a result, only a basic calibration was done. The calibra tion used was based on image sequences of the target, attached to the swing-frame hanging at a zero degree incline. This method yielded a calibrated test volume that was too small for the tests being conducted. As the target moved away from the region of the image that was cali-104 Chapter 6: Experimental and Simulations Results brated, the estimated observations became less accurate. In the future the entire test volume should be calibrated. (iii) The EKF may not have been tuned adequately for the system under observation. As a result, only runs that did not see the target move far away from the calibrated region of the image, near the centre of the image, were tracked with the current version of the software. Be fore any further testing is attempted all three of the above deficiencies must be dealt with. As an example, of the output from the Argo tracking software, for one of the successful natural free swing tests is shown in Figure 6.10, 6.11 and 6.12. The test results presented in Figures 6.10 through 6.12 are for a swing-frame swung in the y-z plane, observed by four cameras with no oc clusions and no additional ghost markers. Again, to avoid cluttering the figures the units of the data are listed in Table 3.1. Figure 6.10, shows the time history of the pose of the target. With the swing-frame constrained to swing in the y-z plane, the only expected motions are a rotation about the x axis (roll) and linear displacements along the y and z axes. The remaining pose terms are considered to be fixed at their initial values. The tracked results are as expected for a free swinging pendulum. The tracked roll (rotation about the x axis) and sway (movement along the y axis) motions are sinusoidal with an exponential decay, as expected. The additional data gathered from the inclinometer, an independent secondary source, can par tially corroborate the tracked output. Figure 6.13, shows that amplitude, decay rate, and frequency of the tracked roll angle of the target compared with the measured roll angle are very closely matched. The decay function superimposed on both plots in Figure 6.13, is represented in Equation (6.1): _LO ii -0.0029T //r \\fD = ±8.13e , (6.1) The decay rate of -0.0029 was determined by inspection only and not computed using energy meth-105 Chapter 6: Experimental and Simulations Results ods. Figure 6.14 shows that the swing frequency of tracked roll data and the measured data align at approximately 0.34 Hz. The tracked heave (movement along the z axis) motions are more significant as they indicate that the motion being tracked is that of a pendulum. The troughs of the heave signal are flattened at the same level for each cycle, while the peaks have a naturally damped exponential decay. This is the expected motion for heave, because for a pendulum of fixed length the lowest point of the swing is constant and does not change as it slows. The effect of forcing the troughs of the heave signal to have the same value has the results that the heave signal contains a small secondary higher frequency component. Figure 6.15, a plot of the heave power spectrum vs frequency, shows the heave signal and its first harmonic. The main frequency of approximately 0.34 Hz corresponds to the swing of the pendulum, and the frequency of the first harmonic at approximately 0.68 Hz, with a smaller energy, resulting from the truncated troughs of the heave signal. The small ripples in the yaw motion of the target, can be attributed to small vibrations that may be caused from misalignment in the two pivot points of the pendulum. This motion falls within experimental error and is considered as noise. Examining the rates of pose, shown in Figure 6.11, the roll and sway terms are obviously the derivatives of the corresponding terms in Figure 6.10. The orientation and position terms of roll and sway respectively, can be described by a negative cosine function with exponential decay. As expected, the derivative of these terms can be represented by a positive sine function with an ex ponential decay. The rate of pitch and yaw have increased noise levels, but this is expected when differentiating a signal. The heave velocity signal is of the greatest interest because the small step at zero speed corresponds to the flattened trough in the heave displacement signal. Figure 6.12 shows the accelerations of the target. The heave acceleration signal is the most sig nificant to note because it shows the changes in acceleration corresponding to the flattened trough of the displacement signal. 106 Chapter 6: Experimental and Simulations Results 6.4 Summary While the Argo software has its limitations it has been demonstrated to function both in simu lation and processing experimental data gathered from physical testing. The tracking algorithm can track low frequency (<l/2 Hz), large amplitude (<10m or <30°) motion for all six degrees of freedom. The dynamic response of the tracker to changes in trajectory is slow, causing it to lag the system that is under observation. Tuning the EKF tracker by refining the expected noise levels in the dynamics and observation models, could increase the dynamic response of the Argo Project from 1/2 Hertz to 5 Hertz. Another possible modification that could improve the dynamic response of the tracker is to replace the linear dynamics model in the EKF with one that is based on the equa tions of motion for a pendulum. This effort has the draw back that the tracker will loose the ability to follow a non-oscillatory trajectory. This requires a compromise between an all purpose tracker that will have a lower dynamic response and one with a high dynamic response that is designed for a specific task. 107 Chapter 6: Experimental and Simulations Results Table 3.1: Units of Figures 6.1 - Figures 6.13 variable description variables units angular displacement roll (\|/), pitch (0), yaw (<()) degrees [°] linear displacement surge (x), sway (y), heave (z) metres [m] angular velocity (V. M) degrees per second [7s] linear velocity (x, y, z) meters per second [m/s] angular acceleration (\|/,9,(|)) degrees per square second [7s2] linear acceleration (x, y, z) metres per square second [m/s ] 108 Chapter 6: Experimental and Simulations Results Camera #3 15 15 Figure 6.1: Physical Setup Represented in Simulation. 109 Chapter 6: Experimental and Simulations Results 1000 2000 3000 4000 O5l0t=" CO ° 0 1000 2000 3000 4000 |*10--co 5 -oj-<D 1 > OB O 1000 2000 3000 4000 1000 2000 3000 Frames 4000 5000 5000 5000 5000 6000 6000 6000 6000 6000 6000 Figure 6.2: Simulation Results: Pose. Figure 6.3: Simulation Results: Pose Error. 110 Chapter 6: Experimental and Simulations Results Figure 6.4: Simulation Results: Pose Velocity. Figure 6.5: Simulation Results: Pose Velocity Error. ill Chapter 6: Experimental and Simulations Results 20 r Frames Figure 6.6: Simulation Results: Pose Acceleration. 20 r _21 i i i i i i 0 1000 2000 3000 4000 5000 6000 Frames Figure 6.7: Simulation Results: Pose Acceleration Error. 112 Chapter 6: Experimental and Simulations Results Figure 6.8: Physical Setup of Experimental Testing. Figure 6.9: Experimental Pendulum Geometry. 113 Chapter 6: Experimental and Simulations Results o rr 10 0 -10 £ 0 -10 I 0 -10 I CD25 S 2 CO CO CO 1.5 2.51 2 1.5 CD > CO -1 0) 1 X0.5 1 0 -- 1 -1000 I 2000 3000 4000 1 5000 1 1 1 1 1 0 r~ 1000 2000 3000 4000 5000 . _ — _ i I I I I 0 r~ 1000 2000 3000 4000 5000 i i i i t 0 r* 1000 MAAAA 2000 /www 3000 i/WWW 4000 wwwi 5000 wwwww 0 1000 2000 3000 4000 5000 I i i I i 1000 2000 3000 Frames 4000 5000 Figure 6.10: Experimental Results: Pose. _ 20 0 -20 ° o DC u Pose Velocity AAA/WVWW^ 1000 2000 3000 4000 5000 sz 20 £ 0 °- -20 0 1000 2000 3000 4000 5000 3 CO 20 -20 1000 2000 3000 4000 5000 CD CO 1 S 0 CO 1000 2000 3000 4000 5000 1°r 1000 2000 3000 4000 5000 CD > co o 1000 2000 3000 Frames 4000 5000 Figure 6.11: Experimental Results: Pose Velocity. 114 Chapter 6: Experimental and Simulations Results 500 r o rr -500 500 _ 0 a. -500 0 500 r -500 f o CO CO 2° CD > co o 1000 2000 3000 4000 5000 1000 2000 3000 4000 5000 1000 2000 3000 4000 5000 1000 2000 3000 4000 5000 1000 2000 3000 4000 5000 1000 2000 3000 Frames 4000 5000 Figure 6.12: Experimental Results: Acceleration. 115 Chapter 6: Experimental and Simulations Results 20 40 60 80 Time [Seconds] 100 120 10 DC 55 0 •*—> CD E 1 "5 o c -10 I I 1 i ; ; ; ; _ 1 1 1 - MM -- M1111M IJjjjH illiiiiiiiii^ IT in _ > * > * * * * *——l 1 1 _J I I I I I 20 40 60 80 100 120 Time [Seconds] Figure 6.13: Experimental Results: Roll Angle. 116 Chapter 6: Experimental and Simulations Results x 10 3.5 o CD CD | 2 o CD Q-co o Q. 0.5 I I I I I 1 - Inclinometer Roll Angle Argo Tracked Roll Angle '. . . J V i i i i i 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Frequency [Hz] Figure 6.14: Experimental Results: Frequency Comparison of Roll Angle. Figure 6.15: Experimental Results: Frequency of Argo Tracked Heave. 117 Chapter 7: Conclusions and Recomendations Chapter 7: Conclusions and Recommendations 7.1 Conclusions The Argo Project has been successfully demonstrated to track the large amplitude, low frequen cy, rigid body motion of a target with known geometry, without constraining its motion. This tool, although developed specifically to evaluate the response motions of model ships in towing and ma noeuvring tanks, is shown to be a very effective non-contact 6 DOF target trajectory measurement tool that could be applied to numerous situations. From a review of non-contact position and motion measurement devices, machine vision was selected as the sensor for this project over other measurement systems for two reasons: (/) the char acteristics of machine vision most closely matched the requirements of the project, and (ii) ma chine vision had been successfully implemented in similar systems already in use in towing and manoeuvring tanks facilities around the world. This work approaches the problem of tracking the rigid body motion of a model ship, directly from image data, rather than triangulating the positions of individual markers and then solving for the pose as many other existing systems do. The pose of a target with known geometry can be solved for, from a set of nonlinear constraint equations, that use data collected from an individual or multiple cameras. This set of nonlinear constraint equations are collectively known as the ob servation model and are based on a 6 DOF coordinate transformation and the direct linear transfor mation (DLT) camera model. An extended Kalman filter (EKF) is used to track the rigid body motion of the target. The EKF was selected because it is capable of state estimation based on indirect non-linear observation of state variables, multi-sensor data fusion, and signal noise reduction. The multi-camera-observa tion model is incorporated into the EKF to relate the incoming image coordinate-data to the state 118 Chapter 7: Conclusions and Recomendations variables that describe the pose of the model ship. The new tracking algorithm resulting from this research effort was tested with computer sim ulation and with data gathered from physical experiments. In the numerical simulation tests the system successfully tracked a dynamic system but was limited to cyclic motions of less than 0.25 Hz because of a following phase lag between the actual and tracked trajectories. During the phys ical experiments the system observed the motion of a target of known geometry attached to the free end of a pendulum. The results of these tests were satisfactory as a preliminary validation, but also indicated possible sources of errors, as the tracked data did not perfectly match with the expected results. Some of the potential sources of error in the experimental test results, from both simulation and physical testing are: (i) calibration error, (ii) lack of stiffness in the test rig, (iii) numerical lin earization errors, and (iv) the EKF was not optimally tunned for the system under observation. This research effort has direct application to the marine testing industry and could be the basis of a task specific tracking system. It is believed that such a system can be assembled with off the shelf components that would be flexible in its design. With small changes external to the core tracking algorithm, the system could be reconfigured and optimized to match a variety of motion measurement tasks. The deficiencies identified in this prototype system will be dealt with in future versions of the tracking algorithm. As well, incorporating additional functionallity into the system will lead to a fully working tool that can be used in towing and manoeuvring tanks to evaluate the response motions of a model, with confidence in the results. 119 Chapter 7: Conclusions and Recomendations 7.2 Future Work and Recommendations It is the opinion of the principle researcher of this project that, although it has been demonstrat ed to function within the confines of controlled experiments, there is additional work required to make it a fully functioning product. Much validation testing of this program will be required be fore this system can be employed with confidence by researchers as a measurement tool. Some of the improvements and additions that should be considered are: • Adding lens distortions to the observation model to improve accuracy of the estimates of the observable outputs when the target in the image moves away from the focal axis of the camera. • Refinement of the calibration method, to make it more user friendly and more accurate, is nec essary to increases the effective working volume. The current observation model and calibra tion method proved problematic when processing experimental data that travelled away from the centre of the working volume. • Trajectory track recovery should be included in a future version. If the tracking algorithm loses track of the target the program should go into a search mode to re-establish the track if possible. In a self propelled model seakeeping test it is conceivable that the target could be momentarily occluded from view such that insufficient data is gathered and the EKF will loose track. Presently, once the program has lost track of the trajectory it terminates in a semi-con trolled manner. • Improving the data correlation and validation algorithm will make the tracking process more robust in handling noise in the images. • Upgrading the camera model in SoftCam to be more representative of how a real camera gen erates an image of a scene is necessary for the development and testing of a more accurate observation model. 120 Chapter 7: Conclusions and Recomendations The ability to track multiple mobile rigid bodies simultaneously is desirable when looking at the interaction of two floating bodies, such as a tug and barge experiment in a towing tank. This could be achieved by adding a second parallel EKF and modifying the pair matching process to split the input vector in two, to feed into the EKF pair. Tracking the pose of a moving camera with respect to a fixed set of feature points in the scene. This problem is very similar to the calibration of the external orientation of a camera. Modify the observation model so that the camera group can be moved relative to a fixed refer ence frame. This is necessary if the cameras are to be mounted to a carriage and are tracking a model ship with respect to a fixed point. This should be attacked from two fronts: odometry from the carriage control and from fixed feature points in the scene that the vision system could track its own position. A complicated problem that may be more academic than necessary is to track the motion of a mobile rigid body with independently moving cameras. This may have no immediate use but with increasing demands for tele-presence by organizations like NASA, the entertainment industry and the medical community, it is likely to be needed in the near future. The program should be ported to a compiler based programming language, most likely C, so that it can be made stand alone and can be made to operate under a real-time operating system. 121 Bibliography Bibliography [Aggarwal 97] J. K. Aggarwal and Q. Cai, "Human Motion Analysis: A Review"," Proceed ings of IEEE Computer Society Workshop on Motion of Non-Rigid and Articu lated Objects, 90-102, 1997. [Alexander 97] Gregory H. Alexander, "An Automated Ship Model Tracking System (Manoeu-vering Experiments without a Purpose-Built Facility)," 4th Canadian Marine Hydromechanics and Structures Conference, Ottawa, ON, Canada, June 25-27, 1997. [Allen 96] Ronald L. Allen, "Strategies for Real-Time Motion Analysis", in Real-Time Im aging: Theory, Techniques, and Applications, P. A. Laplante and A. D. Stoy-enko (editors), IEEE Press, New York, NY, USA, 1996. [Anderson 79] Brain D. O. Anderson and John B. Moore, "Optimal filtering," Prentice Hall, New Jersey, USA, 1979. [Anglin 93] Carolyn Anglin, A Functional Task Analysis and Motion Simulation for the De velopment of A Powered Upeer-Limb Orthosis, Master's Thesis, Deartment of Mechanical Engineering, University of British Columbia, Canada, 1993. [Ariel 97] Product literature and manual for Ariel Performance Analysis System, Ariel Dynamics Inc., California, USA, 1996. [Aziz 74] Y. I. Abdel-Aziz and H. M. Karara, Photogrammtric Potentials of Non-Metric Cameras, Univeristy of Illinois at Urbana-Champaign, Department of Civil En gineering, March 1974. [Bar-Shalom 88] Yaakov Bar-Shalom and Thomas E. Fortmann, Tracking and Data Association, Academic Press Inc., Florida, USA, 1988. [Bar-Shalom 93] Yaakov Bar-Shalom and Xiao-Rong Li, Estimation and Tracking: Principles, Techniques, and Software, Artech House Inc., Boston, MA, 1993. [Blachut 89] Teodor. J. Blachut and Rudolf Burkhardt, Historical Development ofPhotogra-metric Methods and Instruments, Ameriacan Society for Photogrammetry and Remote Sensing, 1989. [Blackburn 96] Dave Blackburn, "Wired Athletes," IRIS Universe, No. 36, Silicon Graphics Inc., 1996. [Beer 84] Ferdinand P. Beer and E. Russel Johnston Jr., Vector Mechanics for Engineers Statics and Dynamics 4th Ed., McGraw-Hill Book Company, New York, 1984. [BNS 94] Baltic News Service, "Estonia Passenger Ferry Sinks, More Than 750 Feared Dead", 12:00, September 28,1994. 122 Bibliography [Borenstien 96] [Bozic 94] [Charnwood 97] [Chui 91] [CIDTEC 97] [Columbus 97] [Cook 81] [Coombes 96] [Dainis] [Dalsa 96] [deSilva 89] [Delaney 97] [Duda 73] [Dynascope 96] J. Borenstein, H. R. Everett and L. Feng, "Where am I?" Sensors and Methods for Mobile Robot Positioning, containted on Mobile Robotics Lab 10-year An niversary Research Report CD, University of Michigan, 1996. S. M. Bozic, Digital and Kalman Filtering 2nd Ed., Wiley, New York, 1994. Product literature regarding CODA mpx 30, Charnwood Dynamics Ltd., Leicestershire, UK, 1997 C. K. Chui and G. Chen, Kalman Filtering With Real-Time Applications 2nd Ed., Springer-Verlag, New York, 1991. Product literature for CIDTEC cameras and sensors, CIDTEC, New York, USA, 1997. Product literature from Columbus Instruments for thier VIDEOMEX-ONE, VIDEOMEX-V and VIDEOMEX-X products, Columbus Instruments Interna tional Corporation, Columbus, Ohio, USA, 1997. David A. Cook, A History of Narrative Film, W.W. Norton & Company, Inc., New York, USA, 1981. Jennifer Coombes, "Seeing in Space Shuttle Astronauts Get a Better View with ASVS," article in QNX News, Vol 10, No. 3 & 4, 1996. Andrew Dainis and Maris Juberts, "Accurate Remote Measuremnet of Robot Trajectory Motion," authors from: Industiral Systems Division Center for Man ufacturing Engineeirng, National Bureau of Standards, Gaithersburg, MD 20899, source and date of publication unknown. Dalsa 1996-1997 Databook for CCD Image Capture Technology, Dalsa Inc., Ontario, Canada, 1996. Clarence deSilva, Control Sensors and Actuators, Prentice-Hall, New Jersey, 1989. Ben Delaney, "The Science of Motion Capture," Innovation3, No. 40, 52-56, Silicon Graphics Inc., summer 1997. Richard O. Duda and Peter E. Hart, Pattern Classification and Scene Analysis, John Wiley & Sons, USA, 1973. Product literature for Dyanscope: Optical System for Motion Analysis, Sirehna, Nantes, France, 1996. [Edgerton 87] Harold E. Edgerton, Electonic Flash, Stobe, 3rd Ed., MIT Press, Mass., USA, April, 1987. 123 Bibliography [El-Hakim 92] [Fang 83] [Faugeras 93] [Fu 87] [Gieck 90] [Golub 89] [Gosine 96] Sabry F. El-Hakim, Nicolino J. Pizzi and David Westmore, "The VCM Auto mated 3-D Measurement System - Theory, Application and Performance Eval uation," in Applications of Artificial Intelligence X: Machine Vision and Robotics, SPIE Vol. 1708, 460-482, 1992. J. Q. Fang and T. S. Huang, "Estimating 3-D Movement of a Rigid Object: Ex perimental Results," Proceedings of the Sth International Joint Conferences on Artificial Intelligence, Karlsruhe, West Germany, Vol. 2,1035-1037, August 8-12,1983. Oliver Faugeras, Three-Dimensional Computer Vision A Geometric Viewpoint, MIT Press, Massachusetts, USA, 1993. K. S. Fu, R. C. Gonzalez & C. S. G. Lee, Robotics: Control, Sensing, Vision, and Intelligence, McGraw-Hill, New York, NY, USA, 1987. Kurt Gieck and Reiner Gieck, Engineering Formulas 6th Ed., McGraw-Hill, West Germany, 1990. Gene H. Golub and Charles F. Van Loan, Matrix Computations 2nd Ed., Johns Hopkins University Press, Baltimore, MD, USA, 1989. Ray Gosine, Lecture Notes 1996, Memorial University, Faculty of Engineering, 1996. [Gospodnetic 92] Slobodan Gospodnetic, John Phillips and Michal D. Miles, Computer Vision System for Motion Measurement of Small Floating Bodies, Dominis Engineer ing Ltd. Ontario, Canada, 1992. [Greaves] John O. B. Greaves, "Interlacing Error in Video Systems: Hardware and Soft ware Issues", Motion Analysis Corporation, publication data unknown. [Grewal 93] Mohinder S. Grewal and Angus P. Andrews, Kalman Filtering Theory and Practice, Prentice Hall, New Jersey, USA, 1993. [Hallan 83] John Hallan, "Resolving Observer Motion by Object Tracking," Proceedings of the 8th International Joint Conferences on Artificial Intelligence, Karlsruhe, West Germany, Vol. 2, pp. 792-798, August 8-12,1983. [Harris 96] Gerald F. Harris and Peter A. Smith eds., Human Motion Analysis: Current Ap plications and Future Directions, IEEE Press, New York, USA, 1996. [Horn 86] Berthold Klaus Paul Horn, Robot Vision, MIT Press and McGraw-Hill, USA, 1986. [Hosie 95] Robin Hosie, Predicting Ball Trajectories in Uncontrolled Environments, hon ours Undergraduate Thesis, Curtin University of Technology, Perth, Western Australia, January 1995. 124 Bibliography [IMD 5] IMD News #5, National Research Council of Canada, Institute for Marine Dy namics, publication date unknown. [IMD 6] IMD News #6", National Research Council of Canada, Institute for Marine Dy namics, publication date unknown. [Johnson 90] Howard Johnson, "The Mechanical Engineer and the Transsition to Image Analysis," Advanced Imaging, NY, USA, November, 1990. [Kalman 60] R. E. Kalman, "A New Approach to Linear Filtering and Prediction Problems," ASME Journal of Basic Engineering, Vol. 82, Series D Number 1, pp 35-45, 1960. [Karara 89] [Kinetic 97] [Krishnan 92] [Lavic 93] H. M. Karara ed., Non-Topographic Photogrammetry 2nd Ed., American Soci ety of Photogrammetry and Remote Sensing, 1989. Product literature and personal communications regarding Eagle Eye, Kinetic Sciences Inc., Vancouver, British Columbia, Canada, 1997. Radha Krishnan, H. J. Sommer Ul and Peter D. Spidaliere, "Monocular Pose of a Rigid Body USing Point Landmarks," CVGIP Image Understanding, Vol. 55, No. 3, May, pp 307-316, Academic Press, 1992. Product literature for Lavic: Three dimensional motion real-time analysis sys tem. (Biomechanics video motion capture), Emtec Co. Ltd., Tokyo, Japan, 1993. [Lin 86] Zse-Cherng Lin, Thomas S. Huang, Steven D. Blostein, Hua Lee, and E. A. Margerum, "Motion Estimation from 3-D Point Sets With and Without Corres pondences," Proceedings IEEE Computer Society Conference on Computer Vi sion and Pattern Recognition, Miami Beach, FL, USA, 194-201, June 22-26, 1986. [Lowe 92] David G. Lowe, "Robust Model-based Motion Tracking Through the Integra tion of Search and Estimation," International Journal of Computer Vision, 8:2, 113-122, 1992. [Mackay 96] David Mackay, "Underwater Cameras - SIT or CCD?", Sea Technology, Vol ume 37, No. 9, 35-38, Compass Publications, September, 1996. [Manku] Gurmeet Singh Manku, Pankaj Jain, Amit Aggarwal, Lalit Kumar, and Sub-hashis Banerjee, "Object Tracking using Affine Structure for Point Correspond ences," paper published on internet, Dept. of Computer Sciences and Engineering, Indian Institute of Technology, New Delhi, India, date of publica tion unknown. [MATLAB 92] MATLAB Reference Guide Ver 4.0, The MathWorks, Inc., USA, 1992. 125 Bibliography [MARINTEK 97] Information from MARINTEK world wide web home page. http://www.marintek.sintef.no/ document relating to OPTOPOS http://brage.marintek.sintef.no/marintek_eks/innovati.html document relating to Selspot http://brage.marintek.sintef.no/marintek_eks/hydrolab.html [Marzan 76] G. T. Marzan and H. M. Karara, Rational Design for Close-Range Photogram-mtry, Univeristy of Illinois at Urbana-Champaign, Department of Civil Engi neering, January 1976. [McGee 85] Leaonard A. Mcgee and Stanley F. Schmidt, "Discovery of the Kalman Filter as a Practical Tool For Aerospace and Industry," NASA Technical Memoran dum 86847, NASA Ames Research Center, November 1985. [Miller 80] Norman R. Miller, Robert Shapiro, and Thomas McLaughlin, "A Technique for Obtaining Spatial Kinematic Parameters of Segments of Biomechanical Sys tems from Cinematographic Data," Journal of Biomechanics, Vol. 13, 535-547, 1980. [Motion 96] Product literature for Motion Analysis (Biomechanics video motion capture), Motion Analysis Corp., California, USA, 1996. [Motion 97] Personal coirespondance with Motion Analysis Corporation regarding use of their product by U. S. Army Corps of Engineers, 1997. [Mulder 94] Alex Mulder, "Human Movement Tracking Technology," NSERC Hand Cen tered Studies of Human Movement Project, Technical Report 94-1, Simon Fra-ser University, B.C., Canada, 1994. [Nalwa 93] Vishvjit S. Nalwa, A Guided Tour of Computer Vision, Addison-Wesley Pub lishing Company, USA, 1993. [NDI97] Product literature for Optotrak: 3D Motion Measurement System, Northern DigitalTric, Ontariio, Canada, 1997. [Neptec 96] Presentation material for Advanced Space Vision System (ASVS), Neptec De sign Group Ltd., Ontario, Canada, 1996. [Nesi 96] P. Nesi, "Real-Time Motion Estimation", in Real-Time Imaging: Theory, Tech niques, and Applications, P. A. Laplante and A. D. Stoyenko (editors), IEEE Press, New York, NY, USA, 1996. [On-Trak 97] Product literature for SiTek Electro Optics Position Sensing Detectors, On-Trak Photonics Inc., California, USA, 1997. [Paul 81] Richard P. Paul, Robot Manipulators: Mathematics, Programming and Con trol, MIT Press, Cambridge, Mass., USA, 1981. 126 Bibliography [Peak 96] [Qualisys 92] [Qualisys 96] [Rediers] [Romilly 94] [Safaee-Rad 87] [Sasazawa 91] [Schmidt 70] [Selspot 96] [Sezan 93] [Shan 95] [Shapiro 77] [Simon 90] Product literature for Peak Motous System (Biomechanics video motion cap ture), Peak Performance Technologies Inc., Colorado, USA, 1996. Qualisys AB, Qualisys Position Sensor Technical Reference Manual, Partille Sweden: Qualisys, 1992. Product literature for Qualisys Position Sensor System & Multi Traks, Qualisys AB, Sweden, and Adaptive Optics of Ameraica (AOA), Hamilton Standard, Mass., USA, 1996. Benny Rediers and Randall J. Allemang, "Video Sensing in Data Acquisition Systems," Structural Dynamics Research Laboratory, Department of Mechani cal and Industrial Engineering, University of Cincinnati, supplied by Motion Analisys, publication data unknown. D. P. Romilly, C. Anglin, R. G. Gosine, C. Hersler and S. U. Raschke, "A Func-tinoal Task Analysis and Motion Simulation for Development of a Powered Up per-Limb Orthosis", IEEE Transactions on Rehabilitation Engineering, Vol. 2, No. 3, September 1994. Reza Safaee-Rad, Functional Human Arm Motion Study with a New 3-D Meas urement System, Master's Thesis, Department of Electrical Engineering, Uni versity of Manitoba, Canada, 1987. Shigeo Sasazawa, "Non-Contact, Real-Time, 3-Dimensional-Motion Analyz ing System by Color Vision," Journal of Robotics and Mechatronics Vol. 3, No. 3, 207-210, 1991. S. F. Schmidt, "Comptational techniques in Kalman filtering," in Theory and Applications of Kalman Filtering, AGARDograph 139, NATO Advisory Group for Aerosspace Research and Development, London, February 1970. Product literature for Selspot: 3D Motion Analysis, Selspot AB, Sweden, 1996. M. I. Sezan and R. L. Lagendijk (editors), Motion Analysis and Image Sequence Processing, Kluwer Academic Publishers, Boston, USA, 1993. S. Shan and J. K. Aggarwal, "Autonomous Robot Navigation Using Fish-Eye Lenses," in Proceedings of 1995 Third International Conference in Computer Science, Hong Kong, December, 1995. Robert Shapiro, "Direct Linear Transformation Method for Three-Dimensional Cinematography," The Research Quarterly, Vol 49, No. 2, 197-205, 1977. Christophe Simon, Application of the Kalman Filter to Iceberg Motion Fore casting, University of British Columbia, Dept. of Mechanical Engineering, Master's Thesis, July 1990. 127 Bibliography [Sirehna 97] Product and services literature for Dynascope, Sirehna, Nantes, France, 1997. [Stremler 82] Ferrel Stremler, Introduction to Communications Systems 2nd Ed., Addision-Weley Publishing Company, Massachusetts, USA, 1982. [Sullivan 93] Micheal Sullivan, "Design of a Real-Time Motion Tracking System," 1993 Ca nadian Conference on Electrical and Computer Engineering, Vancouver, BC, Canada, September 14-17, 1993. [Sullivan 97] Micheal Sullivan, "Optical Tracking at the Institute for Marine Dynamics," 4th Canadian Marine Hydromechanics and Structures Conference, Ottawa, ON, Canada, June 25-27, 1997. [Sullivan PC] Peersonal correspondence between author and Micheal Sullivan of the National Research Council of Canada, Institute for Marine Dynamics, 1994-1997. [Tau 96] Product literature from Tau Corporation (visual tracking systems), California, USA, 1996. [Trimble 97] Product literature regarding GPS and DGPS technology, Trimble Navigation Ltd., Sunnyvale, California, USA, 1997. [Tsai 85] Roger Y. Tsai, "A Versatile Camera Calibration Technique for High Accuracy 3D MAchine Vision Metrology using Off-the-Shelf TV Cameras and Lenses," Research Report RC 11413, IBM Research Division, San Jose, California, USA, October 1, 1985. [Veillon 96] A. Veillon, J. M. Aillaud and P. Brunet, "Submarine Depth Control Under Waves and Experimental Approach," International Symposium Warship '96, Naval Submarines 5, "The Total Weapon System", London, UK, June 12-14, 1996. [Vicon 96] Product literature for Vicon Motion Systems (Biomechanics video motion cap ture), Oxford Metrics, Oxford, England, 1996. [Walton 86] James S. Walton, "The Accuracy and Precsion of a Video-Based Motion Anal ysis System," Proceedings of the 30th International Technical Symposium on Optical and Optpelectronic Applied Sciences and Engineering, Volume 693: "High-Speed Photography, Videography and Photonics IV," San Diego, CA., August 17-22, 1986. [Walton] James S. Walton, "Expertvision: A Video-Based Non-Contact System For Mo tion Measurement," Motion Analysis Corporation, publication date unknown. [Wang 91] Jiang Wang, Optimal Estimation of 3D Relative Position and Orientation for Robot Control, master's thesis, University of Waterloo, Canada, 1991. 128 Bibliography [Wang 92] Jiang Wang and William J. Wilson, "3D Relative Position and Orientation Es timation Using Kalman Filter For Robot Control," Proceedings of the 1992 IEEE International Conference on Robotics and Automation, Nice, France, pp. 2638-2645, 1992. [Welch 97] Greg Welch and Gary Bishop, "An Introduction to the Kalman Filter," paper published on the internet, Department of Computer Science, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599-3175, September, 1997. [Weng 92] Juyang Weng, Paul Cohen and Nicolas Rebibo, "Motion And Structure Estima tion from Stereo Image Sequences," IEEE Transactions on Robotics and Auto mation, Vol. 8, No. 3, 362 - 282, June, 1992 [Weng 92a] Juyang Weng, Paul Cohen and Marc Herniou, "Camera Calibration with Dis tortion Models and Accuracy Evaluation," IEEE Transactions on Pattern Anal ysis and Machine Intelligence, Vol. 14, No. 10, 965-980, October, 1992. [Wilson 93] Willian J. Wilson, "Visual Servo Control of Robots Using Klaman Fitler Esti mates of Robot Pose Relative to Work-Pieces," In K. Hashimoto, editor, Visual Servoing, World Scinetific Publishing Co. Pte. Ltd., Singapore, 71-104, 1993. [Wu 88] J. J. Wu, R. E. Rink, T. M. Caelli, and V. G. Gourishankar, "Recovery of the 3-D Location and Motion of a Rigid Object Through Camera Image (An Extended Kalman Filter Approach)," International Journal of Computer Vision, 3, 373-394, 1988. [Wysner 83] Shirley Worth Wysner, "Tracking high-speed movement in three dimensions: As eas as (X,Y, Z)," Search, Vol. 18, No. 3, General Motors Research Labora tories, July-August, 1983. [Yaman 97] Product literature from Yaman Ltd., Tokyo, Japan, 1997. [Yuan 89] Joseph S.-C. Yuan, "A General Photogrammetric Method for Determining Ob ject Position and Orientation," IEEE Transactions on Robotics and Automation, Vol. 5, No. 2, April 1989. [Zha 95] Hongbin Zha, Toyoshi Onitsuka and Tadashi Nagata, "Self-Organization Based Visuo-Motor Coordination for a Real Camera and Manipulator System," IEEE International Conference on Systems, Man and Cybernetics, Vol. 4, Vancouver, B.C., Oct. 22-25, 1995. [Zhang 92] Zhengyou Zhang and Oliver Faugeras, 3D Dynamic Scene Analysis: A Stereo Based Approach, Springer-Verlag, Berlin, Germany, 1992. 129 Appendix A: Motion Capture Systems Appendix A: Motion Capture Systems This appendix contains a summary of some of the motion capture-analysis research projects, as well as many of the commercially available systems. The varied nomenclature indicates the wide variety of methods, equipment and applications of this technology. A.l Abbreviations General N/A Not Applicable NP Not Published Marker Ty 3es CAIR Cyclic active IR LEDs HILP High Intensity Light Points ED Edge detection EC&CT Edges, Corners and Circular targets HCBW High Contrast Black and White Markers HCCM High Contrast Colour Markers RRS Retroreflective Spheres and Hemispheres SFSI Substraction of fixed scene from image Tracking Method EKF Extended Kalman Filter MBT Model Based Tracking MKF Modified Kalman Filter NN Neural Network NM No motion tracking, still scene only PO Track position only not vel. or accel. PPE Previous position as estimate PD Differentiation of pose trajectory Camera Model DLT Direct Linear Transformation PPP Plane Perspective Projection CL Colinear Equations NLE Nonlinear Equations Camera Format CCD Charge Coupled Device Camera (B&W) CCD-clr Colour CCD Camera CCD-IR Black and white CCD camera fitted with an near IR filter 3LCCD 3 Linear Array Charge Coupled Devices in a single camera DAC Diode Array Camera F Film Cine-Camera PSD Position Sensing Device CID Charge Induction Device 130 Appendix A: Motion Capture Systems Lighting NOL Normal overhead lighting IRS Infra-red strobe Primary Uses of System HMA Human motion analysis Ob Observation of a dynamics system PM Tracking particle motion RC Robot calibration RVS Robot visual servoing TMS Tracking model ships or floating bodies DP Dynamic positioning of submarine Host Platform PC IBM 100% compatible personal computer Mac Apple Macintosh personal computer SGI Silicon Graphics workstation Sun Sun workstation FG Frame grabber PP Proprietary processor 131 Appendix A: Motion Capture Systems A.2 Video Based Motion Capture Systems Comparision Tables Table A.1: Research Video Based Motion Capture Systems Publication Reference Number of Cameraos Stereo Triangulation Feature Points/ Marker Type Tracking Method DOF Rigid Body Motion Prinary Uses Image Format Camera Model Commercial Systems Alexander 1. HILP PP 3 TMS CCD NP Anglin 2 X HCBW PP N/A HMA CCD DLT *Argo Project 1-m ** RR EKF 6 TMS CCD DLT Qualisys Dainis X CAIR PO 6 RC PSD DLT El-Hakim X EC&CT PD N/A Ob CCD CL Fang 1 ED PD 6 Ob CID NLE Gospodnetic 2 X HCBW NP 6 TMS CCD NP Hosie 1 SFSI EKF 3 PM CCD PPP Krishnan 1 NP NM 6 DP NP PPP Lowe 1 ED MBT 6 Ob CCD PPP Manku 1 ED MKF 2 Ob CCD NP Miller 2 X HCBW PD 6 Ob F DLT Shapiro 2 X HCBW PD 3 PM F DLT Sullivan 2 X RR PD 6 TMS CCD PPP Qualisys Veillon 2 X RR NP 6 TMS CCD NP *Wilson,Wang 1 Hole EKF 6 RVS DAC PPP *Wu 1 HCBW EKF 6 Ob Video PPP Zha 2 ? NP NN 6 RVS CCD NLE * These projects use similar tracking methods as the Argo Project. ** depending on mode of operation; may or may not use triangulation 132 Appendix A: Motion Capture Systems o U B c O c0~ CN - ON Z S X cj co c X cj ® CO c X O o CO D. To c < c o £ 6 o o o q c/5 </> 73 CO 3 3 CT CT 3 d CO CO O M E2 S cs d <? U ck 2 S W O CO U of £ <=> c 2 <»s - ON c <o CO W> 2 So « 2 5H S o '3) '•5 a s t» <u u s a « U c I -a cu PQ o .2 "3 u <u E & o V 3 H CO 6 o X O a. CO .SP '> co O O 'crt To c CO :ion ics o c E ha uman D3UI0I X PQ 9" Is co •a o E CO X o U E o x IE CJ O E o CJ 3 T3 C cO X o U E o x s o CO Ul To o o x o o CJ o •a 2 o X o I— o cj cfl c CJ C 'E a. .g x g 7j E T3 o o x E cO X o u E o X c _o o E a. x o E =8 c CO X o CJ E o X CO Ui X To o o X o o CJ a. 3 T3 3 E '5 a. £ x 73 "> I "8 x E CJ T3 00 o 73 o 73 08 X o X o U CJ E x o U CO -O cO c CO U CO •O co c CO U CO CO < co °8 c cu T3 CJ CO CO a. co < on CO a. CO co T3 CO C CO u 3 O & Q. 3 C/2 Group c 60 Desi ec Nepl Ltd. CJ 'o CJ o O E co C >> Q 73 o '^•1 x C/l CJ o o CO 1 Scrt <2 Oc c ^ X CO O CJ rCJ CU H o U To o PQ CO 3 CO !i $< PQ O a o S g C 3 T3 CJ t-l a X O 2 J 6 U pq 3 a HH CO 3.1 U 5 J c CO • CO cO '5b Q CJ x C o CJ E CO z o 3 T3 O CJ CJ _ cO p •s ^ CJ c/2 C S= CO O <: > CJ pq 60 CO W CJ 1 s CO CJ E to £ CO « .22 _! >^ c c < < CJ o o Q CJ CM CJ X pq X CJ c CJ o CO -1— o CM E 2 O 3 > CO x CJ E o CJ •a I o a. O •a o & O 133 Appendix A: Motion Capture Systems CU Oi —I c cs >> o c3 3 CJ O < C "•3 CS CJ _ 5 cu s a S3 u e o T5 cu O (U .s CU o U — 3 a H UUOJJBId JSOH SJ9>[JBI^ jo jaquin^ JO 9d*i 8uijtj8n ajduiBS BJ3U1B3 JO 3d/(_ SBJ3UIB3 jo jaquin^ Z T3 O cu Z u OH PH z £ u pa ^ u u o z OH Z Q CJ U aj o g. E 8 1/5 c e C3 O < > N # CN o o o X o o o SO 0) w <o "bO CS W O + U OH CL, E X O z o SO Q U O ON • CS u CJ 5 E g B I-!— _ 73 < < •n CN a u u to a. o o CA re c >-. Q CM Z u cu o o cs =3 o so 6 in a u u SO i CS o _! C3 O CU O O o CS CU cu + c 3 00 O 00 00 Oi C< 00 oi i—i + O z a. x a oo Oi l-H + o z o so *S o pi + Q U U x a> C oi o cs 4— o o o ® cn co co cs o cs ® o co OH OH + u cu oo Oi oi 00 oi + O Z o o o <*! o cs pi + Q U U cs o CU E 2 O 3 £ 2 cu Z 0H OH + CJ cu + o in oo Oi oi 00 oi l-H + •J O Z o SO o SD o I Q U U u u o z o SO o • Q U U o —I > o o o > O PH «* in o as o\ Cu OH < CJ bo CJ 60 Q .5 X H Q I _ O o o o m cs cs 8 m cs cs CJ CU SD in cs Q CJ U CO 2 * 2 n, O > o UH in cs o d > O OH # m os o\ U cu o cs o z X Q oo CU O a. ^ Cfl CCJ 00 < o u CL, O 00 43 3 Hi c/1 C/1 O UH CD O O X3 134 Appendix A: Motion Capture Systems uon -BinSuBux o Z CU CD CU CU CD CU cu cu >- o Z cu cu cu SuppBJ} Xpog pigitf cu CU o z o Z o Z o Z djj * O CU Z >i o Z o Z o Z o Z o Z cu O Z SuppBJX JUIOJ o Z o Z cu cu cu cu cu cu CU CU cu cu cu cu S ,0 "•3 u S TS CU Q e e u b H on e <U CA CM S a U s .© I T3 cu 03 o CU s CU E £ o u — 3 cs H SuppBJJ, cu o Z-o Z cu cu cu cu cu cu cu cu o O cu D. 00 C3 '3 3 •o o E c CS CM O <*-o H *S c cd CM i2 3 3 S" & 3 3 O O CU cu 3 3 60 60 O o 13 13 n c C3 CO CN CN CU 60 60 CU 60 60 CU CU C 3b Ii O JS 73 U o Z <N cn + 00 CN X ® S" NO .tt —c bp x -3 NO NO N 2 NO -tt —i 60 x -3 NO NO X M >o CN ® -t-» IS CN NO OO o Z o Z o 60 c 3 C3 E cu CM *S >-. W -| g •I s E 9 e u CM 04 z z o X cu c T3 P ca Q Q CI CM Z CM z CM Z cu & w, cu E cu _> o ca cu Ji — o. E oo 8 cu B. u. cu M E cu > cu & CU M E cu > 60 cu £ t E oo 3 cu 60 §• e c bo S o 5 a CM z cu 5H cu cu cu cu CM z cu CM z CM z C3 _ E 3 c3 S CM z H Q Q CM z CM z CM z CM z CM Z CM z CM Z cu E ca z T3 O cu o _ ca g cu 00 a 8 co c c ca o < > cu 03 60 ca W cu 1 6 E to <£ to 5 .2 >^ cu 73 < < Q o 2 ca cu CM tt cu CL. X 03 x cu 53 0i o ca o t-< CM & 3 cu H E 3 O 3 o r-cn e o o o cu T3 o I o -4—* Q. O M 2 o o. O c o s Q « ca 135 Appendix B: Homogeneous Transformation Appendix B: Homogeneous Transformations In the study of ship motion, the transformation of coordinate frames from the world's to the w ship's uses a roll, pitch, yaw rotation sequence; where roll is the angle ) about the x axis of w w the world frame, pitch is the angle (9S ) about the y axis of the world frame and yaw ((j)5 ) is the rotation about the z axis of the world frame. The world frame is a right handed coordinate system for the z axis up is positive. The transformations presented in this appendix were derived using the symbolic math proces sor, Maple V R4. B.l Basic affine transformations Rotation about x axis (roll) 10 0 0 .0 cosy -sin\i/ 0 rotx = T T 0 sin\|/ cos\|/ 0 0 0 0 1 (B.l) Rotation about y axis (pitch) cos6 0 sinO 0 .0100 rot = y '-sine 0 cose 0 0 0 0 1 (B.2) Rotation about z axis (yaw) Translation cos<|) —sine}) 0 0 . sih<b cosd) 0 0 rot. = z 1 1 0 10 0 0 0 1 (B.3) trans = 1 0 0 x 0 1 0 y 0 0 1 z 0 0 0 1 (B.4) 136 Appendix B: Homogeneous Transformation B.2 Roll Pitch Yaw Translation Transformation This transformation is typically used to describe the pose of a boat or an airplane in the world frame. Transformation for a point in frame B into frame A. RPYT = trans • rotz • roty • rotx , ARPYTB = c<() • cQ -sty • c\\f + c<() • sQ • s\y sty • s\\f + c<|) • sQ • c\\f xB sty • c8 cty • c\\f + sty • sQ • s\\f -cty • s\\f + sty • sQ • c\(/ yB -sQ c6 • s\\f cQ • c\|/ zB 0 0 0 1 (B.5) (B.6) where: A A c\|/ = cos(\|/B) , cQ = cos(9fi) , As\\f = sin(\|/B) , sQ = sin(9fl) , This transformation can be decomposed to its basic terms: [Fu 87] Orientation (rotation) terms1: RPYT, cty = cos(tyB) , sty = s'm(tyB) . tyAg = atan l, l RPYT 2, 1 A RPYT3 2 = atan£py^' 3, 3 TEMP = [rot7] [RPYT] , A -TEMP3 , 6* = atan TEMP, \ Position (translation) terms: xB = RPYT, 4 , yAB = RPYT24 , 4 = RPYT3,4 . (B.7.a) (B.7.b) (B.7.C) (B.7.d) (B.7.e) (B.7.f) (B.7.g) 1. For implementation use the MATLAB™ atan2 function for more reliable results. 137 Appendix B: Homogeneous Transformation B.3 Yaw Pitch Roll Translate Transformation This transformation is typically used to describe the pose of a robot end effector in the world frame. Transformation for a point in frame B into frame A. YPRT = trans • rotx • roty • rotz , YPRT^ = ctycQ -stycQ sQ ctysQs\\t + styc\\i - stysBsy + ctycy -ctys\\f yB A - ctysQc\\f + stys\\f stysQc\\f + ctys\\f ctyc\\f zB 0 0 0 1 (B.8) (B.9) where: A A c\|/ = cos(\j/g) , cG = cos(9g) , As\\f = sin(\|/B) , sB = sin(9B) , This transformation can be decomposed to its basic terms: [Fu 87] Orientation (rotation) terms: A -YPRT, 2 ** = ^ YPRT,', ' A YPRT2 3 = ^YPRT^ TEMP = [YPRT][rotz] , A TEMP, 3 e* = atanr£MPT; Position (translation) terms: A XB = YPRT, 4 , yAB = YPRT24 , 4 = YPRTXA , cty = cos(tyB) , sty = sm(tyB) . (B.lO.a) (B.lO.b) (B.lO.c) (B.lO.d) (B.lO.e) (B.lO.f) (B.lO.g) 138 Appendix B: Homogeneous Transformation B.4 Pose extraction from triangulation results Pose determination of one coordinate system with respect to another can be determined if po sitions of several points are knonw in both coordinate systems. A direct method can be used where the positions of feature points in the first frame are placed in a matrix, Equation (B.ll), and the positions of the same feature points in the second frame are also place in a matrix, Equation (B. 12). [AMP] = [BMP] = A A A A xx x2 • • xn • • XN A A A A y\ yi • • yn • • yN A A A A Z\ z2 • • Zn • • ZN 1 1 . . 1 . . 1 B B B B X\ x2 • • xn • • XN b B B B y\ y2 • •• yn • • yN B B B B z\ z2 . .. zn . • ZN 1 1 . .. 1 . . 1 (B.ll) (B.12) It is important to ensure that the order that the feature points are placed is the same in both matrices. The general model for transforming several point in one frame of refernce to another is given in Equation (B.13). [AMP] = AHB • [BMP] , (B.13) In this case, the transformation, AHB, is the unknown. To solve for the unknown the postion matrix for the original frame is inverted and multiplied with the postion matrix of the destination frame. To determine orientation a minimum of four non-coplanar points are used. If more then four points are used, the postion matricies will not be rectangular. In this case the psuedo-inverse is used [Gol ub 89]. [AHB] = [AMP] • [[[BMP]7[BMP]]1 [BMP]7] , (B.14) The resultant [4 x 4] transformation matrix, AHB in Equation (B. 14), can then be decomposed into 139 Appendix B: Homogeneous Transformation its basic terms using the desired transformation model. The least squares method uses the nonlinear equations for each feature point transformation as a constraint equation and solves for the six basic variables. Each feature point has three constraint equations, of the form of Equations (B.15), (B.16) and (B.17), specific to the transformation model being solved for. The roll, pitch, yaw, translation transformation produces the following constraint equations for a given point n. A B B B A xn = cty • cQ • xn + (sty • c\\f + cty • 50 • s\\f) • yn + (sty • s\\f + cty • sQ • cx\f) • zn + xB , (B.15) A B B B A yn = sty • c0 • xn + (cty • c\)/ + sty • sQ • s\p) • yn + (-cty • s\\f + sty • sQ • c\j/) • zn + yB , (B.16) zAn = -sQ • xBn + c6 • sy • yBn + c9 • c\|/ • + zB , (B.17where: A A A c\\f = cos(\|/g) , c9 = cos(0s) , cty = cos(tyB) , As\\f = sin(\|/s) , 59 = sin(9B) , sty = sin(tyB) . 140 Appendix C: Direct Linear Transformation (DLT) Appendix C: Direct Linear Transformation (DLT) The Direct Linear Transformation (DLT) was developed as a linearized camera model for use in photogrammetry for non-metric off the shelf cameras, in the late 1960's and early 1970's. The DLT was developed to triangulate the positions of objects from image coordinates extracted from stereo images with a stereo-photo-comparator or analytical plotter [Karara 89 & Miller 80]. C.l Direct Linear Transformation Camera Model The derivation of the Direct Linear Transformation (DLT) is based on the pin hole camera model and the colinear theory which is shown in Figure C.l and given Equation (C.l), repeated here from chapter 3 for convenience of the reader. A detailed presentation of the DLT can be found in [Aziz 74, Karara 89, Marzan 76 & Shapiro 77]. Figure C. 1: Pin Hole Camera 141 Appendix C: Direct Linear Transformation (DLT) Equation (C. 1) [Gosine 96 & Weng 92a] describes a plane perspective projection in a viewer cen tred coordinate system. u + du — u = (C.l.a) c (C.l.b) where: (un, vn) = image coordinates of feature point n in camera C, (up, vp) = image principal point in camera C, • (fu, fv) = focal lengths in horizontal and vertical planes of camera C, • (dun, dvn) = systematic error terms corresponding to image coordinates n in camera C, c c c • (xn, yn, zn) = coordinates of feature point n in camera coordinate frame C. (Note: All the above terms are in the camera frame.) The model, presented in Figure C.l and Equation (C.l), uses a viewer centred coordinate sys tem. If the world coordinate system and the camera system do not coincide, a transformation is re quired to move feature points (FPn) from the world frame to the camera frame, whose origin is the focal point, so that they can be projected onto the image plane This transformation is vector addi tion, followed by a rotation to re-orient the new vector from the world frame to the orientation for the camera frame. This transformation is represented in Figure C.2 and Equation (C.2). The rota tion transformation in Equation (C.2), is a rotation about the z axis of the world frame, followed by a rotation about the y axis of the world frame, and then a rotation about the x axis of the world frame. Details of its derivation can be found in Appendix B. From this point this rotation trans formation will be represented as the matrix R. 142 Appendix C: Direct Linear Transformation (DLT) C c where: c(j)c6 -sfycQ sQ cfysQsy + sfycy - sfysQsy + c<pc\j/ -c<|).s\|/ w W Xn ~ xc W w yn - yc w w }n ~ (C.2) c\|/ = cos(\|/c) c9 = cos(9c) s\[f = sin(\|/c) sQ = sin(0c) w cos((()c) w sm(<t>c) www ( (v)/c, Qc, §c)= orientation of camera frame C in the World coordinate frame.) Camera Coordinate Frame World Coordinate Frame -w -w -c Gn = Gc + Gn -c -w -w •'• Gn = Gn - Gc where: Feature Point n in 3-D scene —c c c c Gn = (xn ,yn,zn) = coordinates of feature point n in camera C -W w w w. Gn - (*„ , yn , zn ) = coordinates of feature point n in the World coordinate frame. -w www. Gc - \*c> yczc) = position of focal point (camera frame C) in the World coordinate frame. Figure C.2: Feature point position in the Camera's and the World coordinates frames Combining Equations (C.l) and (C.2) results in a camera model that includes both the intrinsic parameters from Equation (C.l) and the extrinsic parameters, camera pose, from Equation (C.2). The results of this combination can be found in Equation (C.3) [Aziz 74]. u + du = u —f n n **p J u f r> / W W. n , W W. n , W wA Rl,l(*n -Xc)+Rl,2(yn ~yc)+Rl,3(zn ~zc) r> , w w. n , w w. n . w w. R3,\(xn ~xc)+R3,2(yn ~yc)+R3,3(zn ~zc) (C.3.a) vn + dvn vp~fv f n < w n , W W. n . W HO R2,\(Xn -Xc) + R2,2(yn -yc)+R2,3Jzn ~ zc) , w w. _ . w w. n . w w. R3,l(xn ~xc)+R3,2(yn ~yc)+R3,3(Zn ~ zc) (C.3.b) 143 Appendix C: Direct Linear Transformation (DLT) Through algebraic manipulation Equation (C.3) can be represented as: www _ Lrxn +L2-yn +L3-zn +L4 un + dun - W , W , W , ' L9-xn +L10-y„ +Ln-zn +1 (C.4.a) www L5-xn+L6-yn+Lrzn+L8 vn + " vn ~ T W T W T W , ' (C.4.b) where: • L]_JJ = DLT parameters. The expanded form of the individual DLT parameters are as follows in the Equation (C.5): www Let a = -(R3 xxc + R3 2yc + R3 3zc) , L\ = (uPR3,i-fuR\,\)/a . (C.5.a) L2 = (upR3,2-fuR\,2ya . (C.5.b) L3 = (Mp^3. 3 -/u^l, 3>/0t • (C.5.c) www LA = up+fu(R\,lxC + Rl,2yc + Rl,3ZCya • (C.5.d) ^5 = (V*i,i-/v-*2.i)/a ' (C.5.e) L6 = (VpR2,l-fVR2,2)/a • (C.5.f) (C.5.g) ' w w w L8 = vp +MR2, \XC + *2, 2>C + *2, 3^C )/a > (C.5.h) L9 = R3, ' (C.5.i) ^10 = R3,2/'a ' (C.5J) Lll = %3/a ' (C.5.k) 144 Appendix C: Direct Linear Transformation (DLT) The DLT parameters in Equation (C.5) [Aziz 74] can be decomposed to obtain the camera ex trinsic and intrinsic calibration parameters as expressed in Equation (C.6) [Aziz 74]. Let P = -1/J(L29 + L2l0 + L2u) , up = (L1L9 + L2L10 + L3L11)p , VP = (L5L9 + L6Lio + L7Ln)P2 . fu = J(L2+L22 + L23)$2-u2p , fv = J(L25+L26 + L2)V2-v2p , 6^ = asin(L9P) , w \\fc = atanC-Ljo/Ljj) , = HupL9-Lx)/fu , w w tyc = acos(/?j 1)/cos6c , w xc w yc w L2 L3 -l ~L~4 = h ^10 1 (C.6.a) (C.6.b) (C.6.c) (C.6.d) (C.6.e) (C.6.f) (C6.g) (C.6.h) (C.6.i) C.2 DLT Calibration Equations The DLT camera model, described by Equations (C.4), projects the positions of a feature point in the world coordinate system onto the image plane of the camera. The eleven DLT parameters for a given camera m, j j, can be determined with either a direct least squares method or with an iterative least squares method. The direct method assumes that the distortion correction terms are small and are therefore ignored. The iterative least squares method solves for the linearized DLT parameters as well as the non-linear distortion correction terms if they are included in the camera model. The iterative least squares method is more rigorous and will yield a better calibra tion at the cost of increased computation. Using the direct method to supply the initial guess for 145 Appendix C: Direct Linear Transformation (DLT) the least squares problem should reduce the number of iterations required to obtain a solution. The calibration procedure is to place a minimum of six markers at known positions in the world coordinate system in the field of view of the camera being calibrated. For calibrating cameras cov ering a large test scene, where a given camera can not view the entire scene, different markers can be used for each camera as long as the same world coordinate system is used to describe the posi tion of the markers. The image coordinates corresponding to the centroids of the projections of the markers onto the image plane are then recorded. Knowing both the marker positions and the cor responding image coordiantes leaves only the eleven unknown DLT parameters in the constraint Equation (C.4). The direct method of solving for the DLT parameters, L™_ x,, puts the set of constraint equa tions into matrix form and then isolates the unknown paramters. The matrix form of the constraint equations is shown in Equation (C.7). w xx w y\ w z\ 1 0 0 0 0 m W U j X j m W -uxyx m W m ux 0 0 0 0 w xx w y\ w z\ 1 m W ~vlxl m W ~v\y\ m W -vxzx ^2 m vl w x2 w yi IV Z2 1 0 0 0 0 m W ~u2 x2 m W -u2y2 m W ~u2 z2 ^3 m u2 0 0 0 0 w x2 IV y2 w z2 1 m W ~V2X2 m W ~^2 m W ~V2Z2 , m L4 m v2 ^5 w Xn w yn w zn 1 0 0 0 0 m W —U X n An m W ~unyn m W ~unzn L6 Lm Lm m U n 0 0 0 0 w Xn w yn w zn 1 m W ~vnxn m W ~vnyn m W —V Z n^n m vn ^9 w XN w w yN zN 1 0 0 0 0 m W ~UNXN m W -^N m W ~UNZN Jm Mo m uN 0 0 0 0 w xN w w yN zN 1 m W ~VNXN m W -vNyN m W ~VNZN_ Lm m (C7) The general form of the above calibration equation is: [MP][L] = [/C] 146 Appendix C: Direct Linear Transformation (DLT) where: • [MP] = marker parameters matrix (2N by 11) • [L] = DLT parameters vector (11 by 1) • [IC] = image coordinates vector (2N by 1) Premultiplying both sides of this equations by the inverse of the marker paramter matrix, [MP] iso lates the unknown parameters on one side of the equation. Due to the rectangular nature of the marker parameter matrix, [MP], a psuedo-inverse is necessary to solve Equation (C.8). The psuedo-inverse is an overdetermined least-squares solution [Golub 89]. Including the psuedo-inverse into Equation (C.8) the form of the solution becomes Equation [MP] l[MP][L] = [MP] l[IC] = [L] (C8) (C9). [L] = ([MP]T[MP])~l[MP]T[IC] , (C.9) 147 Appendix C: Direct Linear Transformation (DLT) C.3 Triangulation using DLT The advantage of the DLT is that triangulation of the position in the world coordinate system of any marker is reduced to a simple matrix operation. To triangulate a marker's location a mini mum, of two points of view are needed. Additional data from other images can be included to min imize the error in the triangulation. <4 r1 1^ (L2 j1 N T 1 1^ -Lll"n) , 1 K ~L9vn) -L10Vn) (L7 "Lllvn) (vi -L9un) (L22 T2 2. (^3 ,2 2. -Lll"J / 2 (u -Ll) ~L9vn) <A T2 2. r2 2\ w xn ( 2 -Lg Un) (Lm2 j m tn, t -j in ~L\0un) ^L3 -Lnun) W yn w - / m K -o (i? -L9 v„) (Le -L10v„) (L7 "Lllvn) i m (vn -L9 un) -Ll0un ) (L3 ~LUUn) ( M -Lf) (if ~L9 vn ) TM M. ,TM -Ll0vn ) (L7 ~Lnvn ) ( M (CIO) The general form of the above calibration equation is: [A5][Pf] = [BC] where: • [AB] = triangulation matrix (2M by 3) • [Pt] = point position vector (3 by 1) • [BC] = triangulation vector (2M by 1) It is assumed that the marker can be seen in all fields of view (FOV). In the event that the mark er does not appear in an image but is still visible in at least two images, the rows in the triangulation matrix and vector corresponding to the image where it is not visible are removed. The position of marker n in the World coordinate frame is found by inverting the triangulation matrix and then post-multiplying with the triangulation vector, Equation (C. 11). 148 Appendix C: Direct Linear Transformation (DLT) [Pt] = [AB]~\BC] , (CM) The rectangular nature of the triangulation matrix, [AB], requires a pseudo-inverse, Equation (C.12), to invert it. [Pt] = [[[AB]T[AB]]~\AB]T][BC] , (C.12) 149 Appendix D: Initialization of the State Error Covariance for the EKF Appendix D: Initialization of the State Error Covariance for the EKF This appendix presents the initialization of the state error covariance matrix, PQ, necessary to initiate the extended Kalman filter (EKF). The initial state error covariance matrix, PQ, depends on the dynamics of the system and the observations made of that system. In the case of the Argo Project the dynamic model assumes a constant acceleration linear relationship between states. With the Argo Project it is not possible to directiy observe the state of the system. The observations made, measures the pose of the mobile rigid body (MRB) and can not measure any of the higher order terms of the pose. Since the dynamics model is the same for each degree of freedom, the derivation of the intial state error covariance matrix, PQ, is based on a single degree of freedom and repeated for each one. An effort is made to make the initial state error covariance matrix, P0, as accurate as possible but as the EKF adjusts the state error covariance matrix, PK, it is not necessary that the initial value perfectly match the system, as it is only a starting point. D.l Nomenclature NDOF number of degrees of freedom (default = 6) NORD numerical order of model (default = 3) M number of cameras N number of markers on target frame q number of state variables = NORD • NDOF = 18 r number of estimated observations = 2- M • N k discrete time index (k e integer set) r error covariance matrix for a single DOF [NORD x NORD] Pk estimate (error) covariance matrix [q x q] sk state vector [q x 1] estimated state vector [q x 1 ] 5* observation noise vector [r x 1 ] co3 system noise at interval 3 for a given DOF 150 Appendix D: Initialization of the State Error Covariance for the EKF T sample period E{ ) expected value operator D.l.l Initial State Error Covariance Model Observations of the system only provides position information and all higher order state vari ables, such as velocity and acceleration, are inferred from the position data. The observed data is assumed to contain some white noise, represented by £,k. x3 = x3 + c;3 , (D.l) The inferred velocity term is: (x3 + ?,3)-(x2 + 1~2) x3 = , (D.2) The corresponding acceleration term is: (*3 + £3)-(*2 + 52) (*2+52)-(*l+5l) *3 = 1 ^ . (D-3) By collecting like terms Equation (D.3) can be reduced to: (x3 + ^3)-2(x2 + ^2)-(Xl + ^) x3 = 2 ' (D-4) 1 The EKF generates an estimate of the position term that corresponds to the above observation terms: X3 = x3 , (D.5) The estimate of the velocity term is: X<j h = , (D.6) The dynamics model used in the Argo Project assumes that deficiencies in the estimation process are modelled as noise introduced in the estimate of the acceleration term is: X-^ ^2 ^ 1 ~" 1 1 x3~ 2*2 ~ x\ /r. X3 = +C03 = 2 + <°3 ' (D-7) 151 Appendix D: Initialization of the State Error Covariance for the EKF The general form of a third order state vector for a single degree of freedom system is: s = (D.8) Therefore, the observed state of the system from Equations (D.l) through (D.4), and Equation (D.8) is: s3 = x3 + ^3 (x3 + ^3)-(x2 + ^2) X (x3 + £3) - 2(x2 + %2) + (*! + ^) (D.9) The estimated state vector for a single degree of freedom is the collection of Equations (D.5), (D.6) and (D.7), is: *3 = "^3 "^2 + CO, (D.10) 152 Appendix D: Initialization of the State Error Covariance for the EKF Taking the difference between the observed and estimated states of the system, yeilds a vector of state error terms: s3-s3 a b c (D.ll) where: a = (x3 + ?,3)-x3 = £3 , (D.12) b = (x3 + ^3) — (x2 + ^2) X3—X2 £>3~^>2 (D.13) c = (x3 + £3) - 2(x2 + $2) + (*,_ + ) c3-2x2 + xl ] ^3-2^2 + ^ "2 + (03 C03 ,(D.14) J As previously stated in Section 4.2.4, the definition of the state error covariance is: T Pk = E((sk-sk)(sk-sk) > The result of combining Equations (D.l 1) through (D.14), and Equation (D.15) is: (D.15) P = E(a > E(ab) E(ac) E(ab) E{b2) E(bc) E(ac) E(bc) E(c2) (D.16) where: 2 K2 a = S3 » (D.17) . £3 ^2^3 ab = — x x (D.18) ac = ^-2^3+^3 -CD3I; 3^3 ' (D.19) 153 Appendix D: Initialization of the State Error Covariance for the EKF bc = -- 3^ + - — + 2-^ - + — , (D.20) TXT T T T T * = 1-2— + -2 . (°-21) T T X 2 $3 .$2$3 J1S3 »3^3 A »3^2 $1 G>3$1 2 C2 = 4-r+2^-2-2- + 4-1-4-^ + 2-2- + ^--2- + a)3 , (D.22) Assuming that the above noise terms are uncorrelated, then the following constraints are ap plied for determining the expected values of the above terms. 2 . _ , E(^fiTk) = p for all; and , (D.23) = (°m for ally andifc , (D.24) V 0 ,j^kj £<£/D/> = 0 for all; and Jk , (D.25) E{^k) = E((Ak) = 0 for all *> , (D.26) Applying the constraints defined in Equations (D.23) through (D.26), the expected values of the products in Equations (D.17) through (D.22) are: £<«2> = a\3 , (D.27) E{b2) = -1 + -T ' (°-28) X T ^2 E{ab) = , (D.29) 2 £<ac> = -y , (D.30T 154 Appendix D: Initialization of the State Error Covariance for the EKF 0> Of E(bc) = 242 + 43 • (D31) 2 2 2 . Oe Of Op „ E(c2) = 4 + 442 + 4 + a203 , (D.32) X XX Assuming noise levels remain constant the following constraints can be added, Equations (D.33) and (D.34): ali= aL= ai= cl - (D34) Applying Equations (D.33) and (D.34) to Equations (D.27) through (D.32), the expected values of the "error products" can be further reduced to Equations (D.35) through (D.40). E(a2) = a\ , (D.35) 2 E(b2) = 2-\ , (D.36x 2 E(ab) = ^ , (D.37) ^2 E{ac) = -f , (D.38x 2 o> £<fcc> = 3^ , (D.39) x 2 £<c2> = 6^ + 0* , (D.40x 155 Appendix D: Initialization of the State Error Covariance for the EKF Resulting in the initial estimate of the state error covariance for a third order, single degree of free dom, system: r = 2 2 °* X 2 . X 2 2 o> 2— z 2 o> 3— J 3 X X 2 2 o> 3 — J 3 ,°cj 2 6-? + °o> X T (D.41) For a third order, six degrees of freedom system Pk is an [18 x 18] square matrix. With the current system and observation covariance models, the general form of the intial state error covar iance, repeats the single DOF model for each subsequent DOF. It should be noted that this model does not include cross-coupling terms. rooooo oroooo oorooo oooroo ooooro ooooor (D.42) 156 Appendix E: Fully Expanded Observation Model Equations Appendix E: Fully expanded observation model equations This appendix contains all the equations necessary to generate the estimated image coordinates vector and the cooresponding linearized observation model, implemented in the EKF tracking module. All of the equations presented herein were derived using a symbolic math processor soft ware pacakage, Maple V R4. Some of the basic equations used herein are presented in Appendices B andC. The estimate of the image coordinates of corresponding markers for each camera and the line arized observation matrix are generated using the estimate of the state vector from the dynamics model in the EKF and the known geometry of the target attached to the MRB. Since the camera model is calibrated to the world coordinate frame, it is necessary to transform the known marker positions in the MRB coordinate frame to the World coordinate frame as expressed in Equation (E.l) This transformation, a roll-pitch-yaw-translation, uses the estimated pose of the MRB in the World coordinate frame as its parameters. w Xn W yn w Zn 1 where: w cty • c6 -sty • c\\f + cty • sQ • sx\f sty • s\\f + cty • sQ • c\j/ xMRB w sty • cQ cty • c\\f + sty • sQ • s\\f -cty • s\\f + sty • sQ • c\|/ yMRB -sQ cQ • s\\f c8 • c\|/ 0 w cty = cos(tyMRB) w sty = sin(tyMRB) w ZMRB 0 0 1 MRB MRB MRB W c\|f = cos(yMRB) w cQ = cos(8woB) w sy = sin(vM„„) w sQ = sin(0M„„) (E.l) {x^RB, y„RB, z^RB) = position of marker n in th MRB coordinate frame. www (y„ ,yn,zn) = positon of marker n in the World coordinate frame. w w w w w w (VMRB> QMRB> §MRB> XMRB> yMRB' ZMRB) = Pose of MRB coordinate frame in the World coordinate frame. 157 Appendix E: Fully Expanded Observation Model Equations The marker locations in the world coordinate frame are transformed into image coordinates for each point of view using the DLT camera model expressed in Equation (E.2) [Aziz 74]. Tm W T m W T m W Tm m L\ -xn+L2-yn + L3 ' zn + L4 , L9 -xn +Lw-yn +Lu-zn +1 rm W Tm W T m W Tm Lc • x„ + Ls • y_ + Ln • z„ + Ls m _ 5 n o J n I ~n o n , ,. Vn ~ ~n\ W Tm' W Zm' W 7 ' (Z.Z.D) L9 -xn +LlO-yn +LU-Zn + 1 where: • {u™, v™) = image coordinates of marker n in camera m. • L"l_ J j = DLT parameters for camera m. www • (xn ' yn > zn ) = position of marker, n in the world coordinate frame, defined in Equation (E. 1). Explicitly the terms that describe the position of the marker in the World frame from knowing the pose of the MRB with respect to the World frame are shown in the equation group E.3. This is merely an expansion of the terms from equation E.l. W /xW s ,aw . MRB _ , xn = cos(<|)M/?B)cos(eM/?B)x„ - , (E.3.a) W W W W ' W MRB + (-sin((|>M/fB)cos(\|/Mrtfi) + cos(tyMRB)sm(QMRB)sm(\\fMRB))yn WW WWW MRB + (sin((()M/jS)sin(i|/M/jB) + cos(<\,MRB)sm(QMRB)cos(\\fMRB))zn w + XMRB yn = sm(<bMRB)cos(QMRB)xn , (E.3.b) WW WWW MRB + (COS((|>M/JB)COS(\|/M/JB) + sin(<^>MRB)sm(QMRB)sm(\\iMRB))yn WW WWW MRB + (-cos($MRB)sm(\\fMRB) + sm(<^MRB)sm(QMRB)cos(\\fMRB))zn w + yMRB W . fPtW . MRB /aW x . , W . MRB „ » . z = -sin(9w„B)x„ +cos(eMoB)sin(\|/M„B)y„ , (E.3.c) ^ IVI I\D /r.W /r.W v • * W M zn = -sm(QMRB)xn +cos(QMRB)sm(yMRB)yn W W MRB W + cos(QMRB)cos(\\fMRB)zn + zMRB Equations (E.l) and (E.2) are combined to estimate the image coordinates of a marker n in camera m using the estimated MRB pose. 158 The obser-Appendix E: Fully Expanded Observation Model Equations The image coordinate terms \u'n', v"n° J form the observation estimation vector, vation estimation vector is assembled using loop structure: for m = 1 to M for n = 1 to N gkf(2xNx(m-l) + 2xn- 1) = u™ , g/(2x/Yx(m-l) + 2xn) = v™ , end end The linearized observation model is obtained by taking the first partial derivative of the observation estimation vector with respect to the state vector, Equation (E.5) [Grewal 93]. (E.4.a) (E.4.b) . t s = sk = Tsc(s,k) (E.5) s = sk Since the observation equations are based solely on pose terms and do not include any higher order terms the derivatives of the image coordinates with respect to these higher order terms are zero. The resulting matrix is made up of alternating rows, corresponding to the alternating un and vn terms in , gk . The rows are of the following form, Equations (E.6) and (E.7): Row 2xAfx(m-l) + 2xn-l of the linearized observation matrix, Ck. du" dy w MRB 0 0 ae w MRB 0 0 dui dty w MRB 0 0 du" dx w MRB 0 0 dumn dy w MRB 0 0 du" dz w MRB 0 0 (E.6) Row 2 x N x (m - 1) + 2 x n of the linearized observation matrix. dv" °VMRB 0 0 dv" ae w MRB 0 0 ^ m D<VMRB 0 0 dx w MRB 0 0 3v' dy w MRB 0 0 dz w MRB 0 0 (E.7) The explicit forms of the partial derivative terms in these two vectors are expressed in the Equa tions (E.8) through (E.13). 159 Appendix E: Fully Expanded Observation Model Equations du™ {L^T% + L2T9 + L3TXQ) Txj(L9T% + L™QT9 + Tx0) —TT:— = 5 , (E.8.a) ^MOB (I?2'7 + I™r6 + L717'5 + 1) (L^ + L^^ + L^ + l) dv'n _ (L^r8 + L™T9 + I%Tl0) Tl2(L9Ts + L'"QT9 + L'[\ Tl0) dvZoB ~ ^Ti + ^0T6 + 4Ti 75 + 1) (L?r7 + LmxJe + 7/5 + 1) (E.8.b) where: www Tx = sin((|)MOB)sin(eMOB)sin(\)/A/OB) WW www T2 = -cos((t)AfOB)sin(\|/MOB) + sin(<j)MOB)sin(eMOB)cos(\)/A/OB) www T3 = cos(0MOB)sin(0MOB)sin(\|/MOB) WW www TA = sin(())MOB)sin(\|/MOB) + cos((l)MOB)sin(eMOB)cos(\)/MOB) . ,NW . MOB /nW . . , W N MOB TAW S , W . MOB W 7/5 = -sin(9MOB)x„ +cos(9MOB)sin(\|/MOfi)y„ + cos(9MOB)cos(\|/MOB)z„ +zMOB • v /AW . MOB , , , W N , W N _ N MOB ™ MOB W T6 = sin((|)MOB)cos(eA/OB)x;j +(cos((|)MOB)cos(vi/MOB) + r1)yn + T2zn +yMOB W W MOB W W MOB MOB T7 = cos(tyMOB)cos(QMOB)xn +(-sm(tyMOB)cos(yMOB) + T3)yn +TAzn +x{MOB)W _ _ MOB , . , , W . J W rri \ MOB Tg = TAyn + (sm(tyMOB)cos(yMOB)-T3)zn MOB W W T9 = Tiyn + (-C0S WMOB) COS (VMOB) " ^1 )hn)MOB W W MOB W W MOB Tm = cos(9MOB)cos(\)/MOB)y„ - cos(9MOB)sin(\|/MOB)zn Tn=LmxT1 + Lm2T6 + Lm3T5 + L: 160 Appendix E: Fully Expanded Observation Model Equations <*un (L™T\6 + L2T\l+Lr3T\Z> ri9(L9:ri6 + LH):ri7+Liniri8) **ZRB (l9 T15 + L"0T14 + LMN r13 +1) (ijr15 + L™ r14 + L" r13 +1 )2 dtC _ (L5 r!6 + L6ri7 + ^lT\%) 720(L97'l6 + L10ri7 + Ll"lTis) doJJiw ~ (^15 + ^14 + Ln 713 + 1) (I?r15 + I* 714 + J13 + 1) (E.9.a) (E.9.b) where: _ . /nW N M/fB ,AW . . , W . MRB ,aW . , W . M/fB , W r13 = -sm(0MWB)x„ + cos(eM/fB)sin(\|/M/?B)y„ + cos(eM/fB)cos(\|/MJ?B)zw + zM^B 7,4 = sin((|)MOB)cos(eMOB)x„ WW WWW MOB + (cos((|>MOg)cos(\|/MOB) + sin(<|)MOB)sin(eMOB)sin(\|/MOB))y„ WW WWW MOB + (-cos((|)MOB)sin(\|/MOB) + sin((t)MOB)sin(eMOB)cos(\(/MOB))z„ w yMOB T • /Aw \ /ow N. MOB T\5 = sm((t>MOB)COS(eMOB)^ WW WWW MOB + (-sin((l)MOB)cos(v|/MOB) + cos(())MOB)sin(eMOB)sin(\)/MOB))y„ WW WWW MOB + (sm((!)MOB)sin(VMOB) + cos(4>WOfi)sin(eMOB)COS(VMOB))^ W + xMOB W W MOB WWW MOB Tl6 = cos(<bMOB)sm(QMOB)xn + cos((()MOB)cos(eMOB)sin(\)/MOB)y„ WWW MOB + cos ((|>MOB) cos (6MOB) cos (\fMOB )zn W W MOB WWW MOB T]7 = -sin(<t>WOB)sin(8MOB)x„ + sin WM0B)co& (QMOB) sin (yMOB)yn WWW MOB + &in(<>MOB)COS(eMOB)COS(X\fMOBK . ,nW N MOB . ,AW . W N MOB . /nW N , W . MOB T18 = sm(QM0B)xn -sin(eM0B)sin(\j/M0B)y„ -sin(9MOB)cos(\|/MOB)z„ r19 = Lj r3 + L2 r2 + L3 7/j + L4 ^20 = ^5 ^3 + ^6 ^2 + ^7 3"i + ^8 161 Appendix E: Fully Expanded Observation Model Equations L7(_728) + L2 T30 Tn(L9 (~T2s) + LWT3<)) WMRB L9T23 + V22 + LnT2l + 1 (L^T23 + LmlQT22 + L* r21 + 1) dv™ Lm(-T2S) + LmT30 Tl3(Lg(-T28) + L™0r30) where: 7/21 = (sin(<|iB)sin(^ T22 = (-sin((|iB)cos(\^ T23 = cos(tyMRB)sm{QMRB)xn WWW T24 = Sm(tyMRB)^N(QMRB)C0^MRB) WWW T25 = sm(tyMRB)sm(QMRB)sin(\[fMRB) (E.lO.a) (E.lO.b) W W MRB i26 = sm(<\>MRB)sin(QMRB)xn „ . ,NW , MRB ,~W . . , W . MRB , .^W . N. M/fB W "*27 — _sln( "A/^B^n + COS{"MRB' SM^MRB)yn +CO^MRB)COS^MRB)Zn + ZMRB T2% = T26 + (cos(((>^B)cos(\|/^fi) + T25)y^RB + (-cos((|)^5)sin(\|/^B) + T2A)z^RB w ?29 = ^28 + yMRB ^30 = ^23 + ^22 + ^21 W ^31 = ^30 + XMRB — L\ T>x\ + L^ T*)Q + Li T<y-j ~¥ LA ^32 ~ ^1 ^31 + ^2 ^29 + ^3 ^27 + ^ T - JMT +TMT +JmT +J"1 1 33 ~ -^5 7 31 + ^6 •* 29 + ^7 27 + ^8 162 Appendix E: Fully Expanded Observation Model Equations L1 (L"l T36 + L2 T35 + L3 T36 3 w MRB raT jmT ,jmT . M> 1 36 + MoM5 + MlM4 + 1 ra ra ra 2 (M> M6 + M0M5 + M1M4 + 1) , ra L5 ("Mi M6 + ^6 M5 + ^7 M4 + Mi )M> MRB LmT +LmT +LmT +1 M> 1 36 + MO-1 35 + M11 34 + 1 2 (M> M6 + M0M5 + M1 M4 + *) M (LmT + LmT +LmT +LM)LM ^M 1 36 + M 1 35 + M -136 + M /Mo WMRB M> 1 36 + MO-1 35 + M1 -1 34 + 1 ra ra ra 2 (M> M6 + M0M5 + Ml M4 + * ) ra M ra .T^T +/mT + Jm\Jm vM Y 36 + M> J 35 + M 1 34 + Mi /Mo WMRB rmT +JmT +TmT +1 M> M6 + MoM5 +Ml-*34 + l ra ra ra 2 (M> M6 + MoM5 + M1 M4 + 1) T ra L3 (M M6 + M: M5 + Mi M6 + M- )Ml AZMRB jmT +JmT +JmT +1 Mi 7 36 + MO-1 35 + Ml7 34 + l ra ra ra 2 (M> M6 + MoM5 + MlM4 + 1) 3v,» M (^5 M6 + ^6 M5 + ^7 M4 + ^8 )M 1 DZZRB L9T36 + ^ToMs + LU T34 + 1 (i^36 + L7oMs + L7l T34 + 1 ^ (E.ll.a) (E.ll.b) (E.12.a) (E.12.b) (E.13.a) (E.13.b) where: - -*i<e:„»»r"+coS(esfiB)si„(rfRj)yrs+^^-(vi^r8+ Ms = sm(<l>i»/^)cos(e^)j:B + (cos((|)^B)cos(\|/^B) + sin(^B)sin(e^B)sin(\|/^B))y™ + (- cos((j)^B)sin(\)/^B) + sin((t)^B)sin(e^B)cos(\j/^B))z™ w yMRB r36 = cos(<|)M^B)cos(eM/jB)^ + (- sin(())^B)cos(i|/^B) + cos((j)^B)sin(e^B)sin(i|/^B))yfRB + (sin(<j)^B)sin(\|/^B) + cos((|)^/?B)sin(e^B)cos(i|/^/?B))zf/fB , w XMRB 163 Appendix F: Detection and Rejection of False Data Appendix F: Detection and Rejection of False Data This appendix presents an algorithm implemented in the Argo Program that deals with the com bined problems of correspondence, data validation and noise rejection. This algorithm is based on a series of validation tests and if an observed and estimated marker pair pass all of the tests then they are accepted as being valid. This method is slow but function reliably. Future versions of the program will have to include improvements to this section of code to improve performance to al low it to be used in real-time. F.l Nomenclature NDOF number of degrees of freedom (default = 6) NORD numerical order of model (default = 3) M number of cameras N number of markers on target frame NM number of observed markers in a image q number of state variables = NORD • NDOF r number of estimated observations = 2-M-N k discrete time index (k e integer set) ck linearized observation model [r x q] variance matrix of observation noise vector [r x r] h incoming image coordinates vector [NM x 1] A t gk estimated incoming image coordinates vector [rxi] F.2 Validation Test (Detection) The validation test tries to pair up the estimated marker with the corresponding observed mark er. This process should detect and reject phantom observed markers and estimated markers that do not appear in the observed image. Correspondence lists for both the estimated image coordinates and the observed image coordi nates are initially set to zero, 0, indicated no pairing for the corresponding image coordinates. For 164 Appendix F: Detection and Rejection of False Data an observed marker paired with an estimated marker the observed data list will have the identifi cation number of the estimated marker and the estimated data list will have the position of the ob served marker in the list of observations made for that image and the camera identification number that the image was captured with. In the event that either and observed or estimated marker is re jected the entry in the list corresponding to that marker is set to negative one, -1. An estimated marker can be rejected if it is paired with two or more observed markers, because of an overlap, or it is not paired at all with an observed marker, as a result of occlusion. An observed marker can be rejected if it has no corresponding estimate, because it is a phantom marker, or if it and another marker share space in the tolerance circle of an estimated marker. This can occur if two markers appear to be colliding an overlap in the image but not necessarily in test volume space. Rather than keeping the ambiguous data it was decided to reject all of the data associated with an overlap. This is one reason why it is necessary to have more then four markers on the target making the system overdetermined. Each estimated marker is compared with each observed marker for a given image. This neces sary to check for overlap, phantom markers, and different ordering of image coordinates in the ob served data vector with that of the estimated data vector. Different ordering can result from the left to right top to bottom stream that the observed image data is captured in. If for example a mark er appeared on the left side of the image one line below another marker to the right of it would second in the sequence. If the estimation places it on the same line as the one on the right then it will precede the one to the right in the estimation stream. The difference of one line in a 420 line image is relatively small and falls within the estimation tolerance. For each estimated marker the radial distance is computer for each observed marker in the im age. This radial distance is compared with a setable radial tolerance. If the radial distance is less than the radial tolerance then it is considered a matching pair. Additional tests are used to deter mine whether this is a valid pair. 165 Appendix F: Detection and Rejection of False Data IF radial distance < radial tolerance THEN • IF no existing match has been made for either the current observed or estimated markers THEN accept both as a pair and mark the observed list with the identification number of estimated marker and the estimated list with the identification number of the observed marker. • ELSE IF the current estimated marker is already rejected THEN reject the current observed marker. • ELSE IF the current observed marker is already rejected THEN reject the current esti mated marker. • ELSE IF both the current observed and estimate markers are already paired with differ ent partners THEN reject both the current observed and estimated markers along with the their already existing partners. • ELSE IF the current estimated marker is already paired with another observed marker THEN reject the current estimate and both the current observed and existing observed markers. • ELSE IF the current observed marker is already paired with another estimated marker THEN reject the current observed marker and both the current and existing estimated markers. END F.3 Removing false data (Rejection) Once the incoming, gk, and estimated, gk, data vectors have been tested any image coordinate pair from either data set that failed the battery of validity tests is removed from the data vectors and the filter is scaled down appropriately. If an estimated marker is not validated by pairing it with an observed marker then: • the corresponding image coordinate pair is removed from gk • • the corresponding rows and columns in the observation covariance model, Rk, are removed. • the corresponding rows in the linearized observation matrix, Q, are removed. • its null identity is removed from the estimation identification list. • number of estimated markers, r, is decremented by one. If an observed marker is not validated by pairing it with an estimated marker then: • the corresponding image coordinate pair is removed from gk . • its null identity is removed from the observation identification list. • the number of observed markers, NM, is decremented by one. With the rejected data removed the observed and estimated data vectors should now be the 166 Appendix F: Detection and Rejection of False Data same length and the Kalman filter observation model is scaled to match. If the dimensions do not match then an error has occurred and the program is terminated. If the dimensions of the vectors match then the observed data vectors is sorted into ascending order using two keys view number and marker number, from the observed data list. The sort is a bubble sort modified for a twin key sort. 167
- Library Home /
- Search Collections /
- Open Collections /
- Browse Collections /
- UBC Theses and Dissertations /
- Argo Project: Machine vision based motion capture for...
Open Collections
UBC Theses and Dissertations
Featured Collection
UBC Theses and Dissertations
Argo Project: Machine vision based motion capture for tracking the trajectory of the pose of a mobile.. Liggins, Geoffrey Allan 1998-05-19
pdf
Page Metadata
Item Metadata
Title | Argo Project: Machine vision based motion capture for tracking the trajectory of the pose of a mobile rigid body |
Creator |
Liggins, Geoffrey Allan |
Date Issued | 1998 |
Description | The objective of the Argo Project is to develop a tool that will track in real-time the motion of unconstrained, self-propelled, model ships in seakeeping tests done in towing tanks and manoeu-vring basins. To meet the unconstrained requirement, the tracking system must be non- contact and can not interfere with the operation or motion of the model ship. An additional operating requirement is that the sensor must cover an area in excess of thirty square metres. An optical based sensor was selected as it satisfied these constraints. Tracking the motion of the model ship is achieved with a predictive, extended Kalman filter (EKF), using feature point extraction from multiple synchronized images. The EKF is used because it can readily integrate and filter multiple noisy data sets. As well, it can generate an estimate of the pose, namely the position and orientation, of the model ship relative to the reference frame of the test tank. While this project is focused on ship tracking there are many other applications for a system of this kind. TM The system under development makes use of the Qualisys camera and video processor hardware that extract image feature points and return them to a host computer. The incoming image TM feature points are then fed into tracking software developed in MATLAB . The tracking software uses estimates of the image to do feature point correspondence and sorts the incoming data vector into the expected order. The sorted data vector is then used as the input vector for the EKF which computes the photogrammetric equations and computes the state vector for the pose of the mobile object being tracked. This work is being undertaken at the University of British Columbia (UBC), Deparment of Mechanical Engineering, Maritime Engineering and Naval Architecture Research Laboratory. The organizaitons that assisted in this research effort are the Centre for Cold Ocean Resources Engineering, Intelligent Systems Group |
Extent | 8336710 bytes |
Genre |
Thesis/Dissertation |
Type |
Text |
File Format | application/pdf |
Language | eng |
Date Available | 2009-05-19 |
Provider | Vancouver : University of British Columbia Library |
Rights | For non-commercial purposes only, such as research, private study and education. Additional conditions apply, see Terms of Use https://open.library.ubc.ca/terms_of_use. |
DOI | 10.14288/1.0080905 |
URI | http://hdl.handle.net/2429/7936 |
Degree |
Master of Applied Science - MASc |
Program |
Mechanical Engineering |
Affiliation |
Applied Science, Faculty of Mechanical Engineering, Department of |
Degree Grantor | University of British Columbia |
Graduation Date | 1998-05 |
Campus |
UBCV |
Scholarly Level | Graduate |
Aggregated Source Repository | DSpace |
Download
- Media
- 831-ubc_1998-0266.pdf [ 7.95MB ]
- Metadata
- JSON: 831-1.0080905.json
- JSON-LD: 831-1.0080905-ld.json
- RDF/XML (Pretty): 831-1.0080905-rdf.xml
- RDF/JSON: 831-1.0080905-rdf.json
- Turtle: 831-1.0080905-turtle.txt
- N-Triples: 831-1.0080905-rdf-ntriples.txt
- Original Record: 831-1.0080905-source.json
- Full Text
- 831-1.0080905-fulltext.txt
- Citation
- 831-1.0080905.ris
Full Text
Cite
Citation Scheme:
Usage Statistics
Share
Embed
Customize your widget with the following options, then copy and paste the code below into the HTML
of your page to embed this item in your website.
<div id="ubcOpenCollectionsWidgetDisplay">
<script id="ubcOpenCollectionsWidget"
src="{[{embed.src}]}"
data-item="{[{embed.item}]}"
data-collection="{[{embed.collection}]}"
data-metadata="{[{embed.showMetadata}]}"
data-width="{[{embed.width}]}"
async >
</script>
</div>
Our image viewer uses the IIIF 2.0 standard.
To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0080905/manifest