Stereo 3-D Perception for a Robot

This thesis concludes a study of robot vision and presents an analysis of the rudimentary vision problem of modeling form in 3-Space. A stereo "snapshot" vision theory for a computer is proposed, based on an experimental implementation. Here, stereo vision is argued to be essential for modeling natural or unfamiliar domains. The "firm" perception resulting from stereopsis is second only to kinesthesia/tactility in effectiveness with the unknown. The novel mechanism introduced-as the heart of the system-is a stereopsis algorithm for growing stereo surfaces in natural scones. First, 2-D features are extracted from the stereo pair of digital images by locating patterns of change in the images' "gradient-arrow" representations. Then, by associating features in the loft imago with features in the right image, stereo regions or "matches" are made. The stereopsis process fuses the stereo images by growing contexts of matched features. Every match defines via the camera geometry a visible surface in the scene, interlocking with neighboring matches like the pieces of a 3-D jigsaw puzzle. The resultant surface molds provide a firm basis for a polyhedral model of the scene's forms.