The Study Of Stereo Matching Using Belief Propagation

Because of realization of new technology, with the development of computer technology, more and more applications with higher computational complexity are be realized, especially in a number of computer vision issues. Due to the inspiration of the human binocular vision, how accurate the 3D in-depth information from pairs of camera will be concerned by a large number of researchers. The main purpose of this paper is to discuss and study how to speed up Belief Propagation (BP) based on the use of graphics cards.After reviewing the technological development of binocular vision in recent decades, instead of applying many popular algorithmic, using Belief Propagation to solve the problem is a best choice to solve stereo matching. Before this step, the binocular camera calibration and gray correlation algorithm will be applied to build an initial environment of stereo matching. Belief Propagation theory involves several major probability and graph theory in several areas. After instruction of the difference of Bayesian Network, Belief Propagation based on Markov Random Field (MRF) has been described, and Gibbs Energy Field is the theory cornerstone of Belief Propagation. As the Belief Propagation algorithm is highly parallel, it is nature to use graphics cards to accelerate the whole algorithm. We promote the Belief Propagation from space to time, and propose a Belief Propagation algorithm based on temporal image sequences which can greatly reduce the number of iterations of convergence.Because of commonality and high efficiency of CUDA computing, it's decided to use this NVIDIA tools to solve the parallel computation problem.At last, through comparing the experimental result of CPU and GPU method of stereo vision, we believe that CUDA parallel computing to accelerate this algorithm is feasible and highly effective. Then the result of pictures and data which comes from the temporal image sequences algorithm demonstrates the feasibility of the algorithm. Finally, we use multi-threading technology to optimize the algorithm which can accelerate the speeds to about 20ms per frame.