Stereo Image Correspondence Methods for Efficient Hardware Implementation

摘要

In this thesis, depth information extraction via ordinal measures for logical pattern correlation is investigated, along with matching techniques that translate well into hardware implementations. First, the ordinal correspondence metric is evaluated for complexity and performance using traditional correspondence matching techniques. The results show that the ordinal measures are very robust in stereo correspondence when paired with modest error handling techniques. Second, the ordinal measures are applied to an efficient dynamic programming matching algorithm to produce highquality disparity maps. It is shown that ordinal measures are demonstrably fast in software and that they can be adapted to an extremely fast hardware implementation to produce high quality disparity maps for very high resolution images in real time.