Research On An Improved Belief Propagation Algorithm For Extracting Disparity Map

Stereo matching is an important branch of computer vision technology, its main role is to get accurate dense disparity map. In recent years, Stereo matching has become an important direction of many computer vision research enthusiasts, and has made tremendous development and is now being used in people’s daily lives, such as target tracking, robotics,3D reconstruction fileds etc.Paper based on existing belief propagation (BP) algorithm and nearest neighbor search sampling PatchMatch algorithm’s advantages and disadvantages, considering existing stereo matching algorithm’s exists disparity map accuracy is not high, sparse and poor real-time problem. In order to balance both the BP algorithm for belief propagation model which can cleverly solve global disparity optimal algorithm through global stereo matching theory and PatchMatch algorithm which can improve the reliability of the center pixel information by the neighborhood sampling, the paper made a rigorous in-depth study. Main work includes:1. The proposed belief propagation PatchMatch-BP algorithm which based PatchMatch nearest neighbor search algorithm sampling. In order to optimize the previous iteration algorithm, the weak texture and disparity discontinuous region,in whch the center pixel to obtain news of low reliability during message propagation iteration. By constructing a center pixel patch to implement an message’s iterative process, we can get the last message iteration affect from neighboring pixels, thereby increasing the reliability of the center pixel information, ultimately improving the accuracy of the disparity map in the region of weak texture and disparity region,where the matching accuracy is not high. Experimental results show that our new algorithm in the above area prone to parallax accuracy problems, compared to the previous existing algorithms, Our algorithm has gained a significant increase in reducing mismatch rate and improving match accuracy.2. In order to solve the problem with our new algorithm (PatchMatch-BP) in the occlusion area resulting message can not be delivered correctly, the new algorithm (Occ-PM-BP) which based occlusion detection is proposed in the blocked area to improve algorithm. During PatchMatch-BP algorithm’s initializetion step in first iterative process, the pixel matching score and occlusion threshold function are defined for occlusion detection, after detecting an occluded pixels, through a series of processing to elect the pixel which can maximum it’s message support weight in the neighborhood domain of occlusion pixel, and then replace the occlusion pixel value using the elected pixel. The iterative process of the message is the same as that of the PatchMatch-BP algorithm. so as to solve the problem of high mismatching rate with disparity map in the occlusion area. Experiments show that, In the global disparity region,the matching accuracy of the improved PatchMatch-BP algorithm is further improved, the results obtained are more accurate disparity map.