Locally Adaptive-Weighted Stereo-Matching Method Based On Binocular Vision

Binocular vision system, which can obtain space information of the scene, conducive to further-more scene analysis and path-planning of robot, compare with the monocular vision, is a method of robot simulation of the human visual system and the research focus of the three-dimensional reconstruction and field of machine vision. Stereo matching problem of the core issues of binocular vision, stereo matching due to the occlusion problem becomes difficult to solve. This article working on a scheme that gets a stereo matching algorithm efficient that achieves the disparity map of high accuracy and the method of s mismatching and occlusion area detection, and on the improved algorithm experiments were carried out. We analysis and reviews of stereo matching algorithm through reading a lot of papers and information,. The paper has made the following research:(1) The paper proposals a cost weighted function that based on geodesic distance between segmentations, uses to calculate weights of pixels belong to different segmented regions in the support window. Combines the adaptive-weighted method and image segmentation priori, the proposal proposes a new color differences function based on geodesic distance between segmentations that used to calculate weights of pixel in local support window. We segment the image when begin matching, regard support window of local method as a collection of different segmented regions and uses the geodesic distance function to calculate weights for every segmentation, calculate the matching cost and finally use the WTA method to determine disparity value. In the end, we do a few experiments to verify the feasibility and effectiveness of the algorithm.(2)The paper studies about occlusion detection and occlusion filled based on adaptive-weighted method. Local matching method in the occlusion area and low texture region often exist mistake matching, and difficult to add smoothness constraints in the match, the disparity map results possible to produce non-smooth disparity value, and the area of non-smooth disparity is bigger than the occlusion region. This paper uses Left-Right Consistency (LRC) to detect the error match, and combines with the Distinctive Similarity Measure (DSM) to detect the occlusion, end through analysis the principle of occur occlusion filled occlusion areas. The paper uses the standard stereo image to carry out the experiment, compares and analyzes the experimental results in the end.