The Optimized Stereo Matching Algorithm Of Occlusion Regions And Low-texture Regions Based On Belief Propagation

A disparity maps could be calculated by the stereo matching algorithm of binocular vision, which is an important branch of computer vision and has significant meaning to virtual reality and automatic navigation.Compared with laser range finding and other active range measuring technology, the advantage of binocular vision is lower cost, and binocular vision can be used to calculate the depth information of large scale scene in a short time. The performs of binocular vision depends mainly on stereo matching algorithm,and therefore there are several typical mismatch phenomena. Based on belief propagation, the research work and main innovation of this paper is as follows:1 .In this paper, the binocular vision model is analyzed, and several common stereo matching algorithms are studied. The principle of stereo rectification and the evaluation criteria of the stereo matching algorithm are introduced. The theory and calculation process of the belief propagation algorithm are studied emphatically. And the belief propagation algorithm is chosen as the basis of the improved algorithm.2.The causes of occlusion phenomenon is analyzed. The left-right consistency algorithm and the improved algorithm based on edge detection are studied deeply, and the main defects of common occlusion region optimization algorithm are pointed out. The reasons for the decrease of matching precision in the low-texture region are analyzed. The low-texture region optimization algorithm based on image segmentation and plane fitting is analyzed, and the main drawback of this algorithm is that edges and textures can not be distinguished by the image segmentation algorithm,which leads to many redundant computations.3.An algorithm based on belief propagation algorithm to reduce the false matching ratio of occlusion region is proposed: The occlusion region is marked based on the results of belief propagation iteration and the left-view texture information. The energy function and iterative calculation process are rewritten. Experimental results show that the matching accuracy of occlusion region is improved by the proposed algorithm.4.An algorithm to improve the matching effect of low-texture regions is proposed. The mean shift algorithm is used to segment -the left view image. SVM is used to judge the classification of the large area segmentation region and surface fitting is used to get the .parallax of the low-texture region. It is proved that the scheme is feasible and the matching effect of low-texture region is improved obviously.