A Robust Stereo Matching Algorithm Based on Adaptive Relaxation
基于自适应迭代松弛的立体点对匹配鲁棒算法

Establishing stereo image correspondence is a vital and the most difficult problem for binocular stereo measure system. An adaptive relaxation algorithm for dealing with feature point stereo matching is proposed. In this algorithm, a correspondence support function is constructed based on disparity gradient limit, and the matching can be achieved through a relaxation optimizing procedure. A new scheme for dynamically updating the relaxation parameter during the matching procedure is introduced, by which the FAR (false acceptance rate) & FRR (false rejection rate) can be significantly reduced compared with other algorithms with parameters being fixed. After the relaxation procedure, the disparity gradient limit is re-imposed to further filter out false correspondences. It has been validated by experiments that this strategy can efficiently reduce FAR at a modest cost of increase in FRR, which meets the requirement of restrict FAR limitation in many industrial applications. The algorithm has been used in a binocular stereo measurement prototype system, and its robustness and effectiveness is affirmed by subsequent stereo reconstruction.