Surfaces from stereo: an integrated approach

Stereo vision provides the capability of determining three-dimensional distance of objects from a stereo pair of images. The usual approach is to first identify corresponding features between the two images, then interpolate to obtain a complete distance or depth map. Traditionally, the most difficult problem has been to correctly match the corresponding features. Also, occluding and ridge contours (depth and orientation discontinuities) have not been explicitly detected and this has made surface interpolation difficult. The approach described in this thesis is novel in that it integrates the processes of feature matching, contour detection, and surface interpolation. Integration is necessary to ensure that the detected surface is smooth. The surface interpolation process takes into account the detected occluding and ridge contours in the scene; interpolation is performed within regions enclosed by these contours. Planar and quadratic patches are used as local models of the surface. Occluded regions in the image are identified and are not used for matching and interpolation. The approach described is fairly domain-independent since it uses no constraint other than the assumption of piecewise smoothness. A coarse-to-fine algorithm is presented that requires no human intervention other than an initial rough estimate of depth. The surface estimate obtained at any given level of resolution is used to predict the expected locations of the matches at the next finer level. As the final result, a multiresolution hierarchy of surface maps is generated, one at each level of resolution. Experimental results are given for a variety of stereo images.