Parallel Implementations of the CSBP Stereo Vision Algorithm

We developed two parallel versions of the Constant Space Belief Propagation algorithm (CSBP - one of the best stereo algorithms currently known): one in OpenMP and one in C+CUDA. For images with 640x480 pixels, the sequential version has a performance of 1.16 frames per second (FPS), the OpenMP parallel version has a performance of 3.7 FPS, while the C+CUDA version has a performance of 17.3 FPS in high-performance desktop machines. These results are important because they enable the implementation of autonomous vehicles with sensors like camera, which is one of the objectives of a PRONEX project being currently developed in LCAD-DI/UFES. One of the goals of this project is to implement an autonomous vehicle from a commercial automobile.