Open Collections

Abstract

This work presents a new obstacle detection algorithm that uses Gabor filters. The task performed by this algorithm is the detection of moving and stationary obstacles from an autonomous vehicle undergoing predominantly rectilinear motion. Image measurements from stereo cameras are used to extract three-dimensional properties of viewed objects and of the vehicle. Properties such as depth and motion are used to predict if (and when) the object will collide with the vehicle. Three inherently difficult problems associated with the estimation of depth and motion from stereo images are solved. (1) Stereo and temporal correspondence problems are solved using predictive matching criteria. (2) Segmentation of the image measurements into groups belonging to stationary and moving objects is achieved using error estimation and the "Mahalanobis distance." (3) Compensation for transient rotations produced by a shaking camera is achieved by internally representing the inter-frame (short-term) camera rotations in a rigid-body dynamical model. These three solutions possess a circular dependency, forming a "cycle of perception." A "seeding" process is developed to correctly initialize the cycle. An additional complication is the translation-rotation ambiguity that sometimes exists when sensor motion is estimated from an image velocity field. Eigenvalue decomposition is used to detect such ambiguity. Temporal averaging using Kalman filters reduces the effect of motion ambiguities. The obstacle detection algorithm operates correctly in a variety of difficult conditions such as: stereo images with different brightness; image sequences with large image velocities; transient sensor rotations; and concurrent object and sensor motion. Under these difficult conditions, the obstacle detection algorithm presented in this thesis is able to identify moving objects, and distinguish between obstacles that will collide with the vehicle and objects that will pass safely by the vehicle.