An algorithm for parallel stereo matching based on dynamic programming

This paper proposes an algorithm for parallel stereo matching based on dynamic programming and presents a massively parallel implementation of this algorithm. Our implementation can use many processors per scanline, compared to a naive approach of one processor per scanline. This is important because typical images contain 256 to 1024 scanlines, while massively parallel machines can have many more processors. We also introduce a method of handing inter-scanline inconsistencies that is very well suited for parallel implementation. The method only increases the total amount of processing needed to solve the stereo matching problem by a small fraction. A parallel implementation of both the dynamic programming algorithm and the inter-scanline inconsistency correction algorithm are fully discussed.