Real-time Line Detection and Line-based Motion Stereo

Recognition of shape is one of the fundamental problems in computer vision. A number of fast line detection and line-based depth recovery algorithms have been developed on various parallel architectures to meet the requirement of real-time robotic vision. This thesis describes a parallel and hierarchical (pyramidal) approach to fast Hough line detection and line-based motion stereo. The lines are represented using their Hough parameters p and 0. The processes of line merging and matching are integrated in the pyramidal framework. To reduce the complexity of the line merging process, the Dynamically Allocated Quadtree (DAQ) is introduced to represent the Hough parameter space. The line merging algorithm using the DAQ is appreciably more efficient than the original algorithm presented by Jolion and Rosenfeld. A line-based motion stereo algorithm is also developed to recover the depth information along linear features in the scene. Live images are obtained from a single static camera and a moving belt. Line segments from the multiple motion stereo images are matched while they are merged hierarchically in the pyramid. It is shown that the problem of matching lines among the multiple images can be effectively carried out as a straight-line detection problem in a well-defined three dimensional parameter space. Experimental results from the SFU hybrid pyramid are presented for both line detection and line-based motion stereo algorithms. The pyramid architecture capitalizes on advantages of both the mesh and the pipeline architectures. It is demonstrated that the new algorithms combined with the hybrid pyramid hardware environment are suitable for real-time object recognition.