Open Collections

Translational visual servoing control of quadrotor helicopters

University of British Columbia

Vision based control enables accurate positioning relative to a stationary or moving target of quadrotor helicopters independent of on-board sensor accuracy. The use of vision for helicopter control opens indoor and GPS-denied environments as an area of operation, which currently poses challenges to these systems. This thesis presents a full vision-based quadrotor helicopter controller for tracking a stationary or slow-moving target. The full controller used three control loops. The outer control loop converts an image feature error into UAV velocity commands. The quadrotor is a nonlinear system that has highly coupled position and attitude dynamics. The second loop performed a feedback linearization on these dynamics and translated the desired velocity into attitude commands. The inner loop was an on-board attitude and altitude controller, which was part of the ARDrone system and which was analyzed via frequency domain system identification for its dynamics. A linear quadratic Gaussian (LQG) controller, consisting of a Kalman filter for state estimation and a linear quadratic regulator was used to control the linearized system. The visual servoing control scheme was image based which convergences towards the desired configuration independent of on-board sensor accuracy and in the presence of camera calibration errors. The visual servoing features, used to control the translational degrees of freedom, sphere based on the virtual sphere approach, which uses two points in 3D space to create a virtual sphere. Control of the helicopter in the x and y direction was linked to the center location of the virtual sphere and distance to the target was controlled by the virtual sphere radius. The target features were two points, detected by colour and shape based detection methods. Adaboost.MRT, a boosting method for multivariate regression is proposed as part of this thesis. Adaboost.MRT is based on Adaboost.RT and extends the original to multivariate regression, improves boosting’s noise sensitivity, and improves the singularity in the misclassification function. A variance-scaled misclassification function is proposed and the threshold parameter is expanded to accommodate vector output. The proposed method is tested on eight datasets and displays a similar or better accuracy than the original Adaboost.RT or the base learning algorithm.

Text

2014-01-29

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/45919

Mechanical Engineering

University of British Columbia

UBCO

http://creativecommons.org/licenses/by-nc-nd/4.0/