Recognition of shape and orientation using binocular vision

This thesis examines the problem of building primitive vision elements for the remote operation of heavy equipment using minimal supervision. The ultimate goal of the vision system is to provide the robot with a world view that is compatible with the operator's view such that the operator sets goals and the robot executes low level tasks needed to accomplish intermediate objectives. A system for perceiving shape is developed that allows the robot to identify the ground surface on which it moves and to recognize obstacles and simple objects within the workspace. Stereo vision is used for this purpose. Disparity between stereo images contains distance information about a scene. The information in stereo images and the recovery of dense disparity maps is studied. A procedure for matching stereo images is given. The disparity gradient is used as the fundamental evidence of shape perceived by stereo vision and the properties for shape are analyzed within disparity gradient space. These attributes are exploited for navigation and recognition of simple shapes. The techniques developed are used to recognize cylinders and to estimate their location, size and orientation within a robot workspace. The vision process includes three phases: (1) recovery of a dense disparity map that includes sub-pixel interpolation of local surfaces, (2) deciding the location of objects within the robot workspace, and (3) confirmation of a cylinder and determining its location, radius, and pose. Example of these procedures are provided for real images. The algorithm provides a means for navigation on plane surfaces and for recognition and pose estimation of cylinders within the robot workspace. Examples described here demonstrate the feasibility and reliability of this approach. Future work will consider improvements to the low-level processes and increasing the resolution of disparity surfaces.Applied Science, Faculty ofElectrical and Computer Engineering, Department ofGraduat