Research Of Stereo Matching Using Belief Propagation Based On Image Segmentation

Stereo matching obtains a dense disparity map by searching one-to-one correspondence between projection points in the projected images from different viewpoints of the same actual scene. It is the core part of stereo vision . However, considering the influence of noise, occlusion, depth discontinuities, textureless regions, and other factors, stereo matching is still a challenging work to obtain a dense disparity map with high accuracy.From the perspective of improving the matching accuracy and reducing the time complexity, this paper focuses on the stereo matching algorithm. The traditional stereo matching methods using belief propagation algorithm based on pixels have two deficiencies: high computation and matching errors easily induced by single pixel. To make up for the two deficiencies, this paper proposes a novel belief propagation algorithm based on image segmentation by combining image segmentation and belief propagation. The algorithm has three main steps: image segmentation, disparity plane pattern fitting and disparity plane pattern assignment. For the problem of different parameters producing different accuracies of the segmentation, an evaluation index is proposed to measure the accuracy of the segmentation. Then more suitable segmentation parameters are selected by using the proposed index and higher accuracy of segmentation is obtained. For the another problem of reducing the accuracy of disparity plane pattern due to the initial disparity with low accuracy, a multi-window local stereo matching based on segmentation constraint is proposed to calculate the initial disparity and a high-accuracy disparity map is obtained. Meanwhile an evaluation index is used to measure the accuracy of the disparity detection, then reliable pixels and accurate disparity plane patterns are obtained. In global assignment, the way of transmitting message and belief is modified based on segmentation and the optimal assignment of disparity plane pattern is obtained. The time complexity is reduced by the proposed belief propagation algorithm based on image segmentation which uses segmented regions instead of pixels and disparity plane set instead of disparity search space, and assigns a disparity plane pattern to each segment instead of assigning a disparity value to each pixel. And the capability to deal with textureless regions and occlusions is improved with the segmentation constraint.Experimental results show that the method of measuring the accuracy of the segmentation provides an effective guidance for the exact location of image edges. The relatively accurate disparity plane patterns are obtained with the combination of the multi-window matching algorithm and the evaluation index of measuring the accuracy of the disparity detection. The belief propagation algorithm based on image segmentation has lower computation. The relatively accurate distribution of disparity plane pattern is obtained by it. Finally, an accurate dense disparity map is obtained.