Stereoscopic recovery and description of smooth textured surfaces

Abstract A stereo algorithm called Needles is described, specialized to deal with smooth textured surfaces. Constraints of local surface smoothness and global surface continuity are used to solve the correspondence problem. The algorithm is edge based. First the left image is divided into square patches, and a disparity histogram of the potential edge matches is constructed in each patch. Above-threshold peaks in the histogram are passed into a Hough transform, which fits a plane to a subset of the potential matches lying around the peak, forming local hypotheses for the range and orientation of the visual surface along with the edge matches. Next, hypotheses in adjacent, overlapping patches are connected if they share enough common matches. A region growing procedure locates large areas of mutually connected hypotheses, corresponding to continuous, possibly overlapping surfaces. When surfaces overlap, the largest one is chosen. Needles has been integrated with AIVRU's TINATOOl vision system1 to run on a Sun workstation, and implemented in parallel on the MARVIN Transputer network2. Results are presented for two stereopairs, and compared with physical measurements.