Stereo-matching of Line-segments Based on a 3-dimensional Heuristic with Potential for Parallel Implementation

This work provides a segment-based alternative to the edge-based stereo algorithm already existing in the TINA system. Our starting point is the algorithm of Ayache & Faverjon (1985). It attempts to recover groups of linematches having low local disparity variations. In our implementation of Ayache & Faverjon's algorithm matches were built quickly, but glaring mistakes occurred and the general behaviour was difficult to control. Therefore the constraints on the continuity of the world implicitly used in the original approach were reformulated to achieve more robust matching, more particularly by requiring mutual support between reconstructed 3-D segments. A new algorithm has been designed that is suitable for parallel implementation, where 3-D matches and their neighbourhood relationships are explicitly computed, cliques found (objects) and uniqueness enforced.