Binocular Matching Based On Stereoscopic Vision

Binocular vision is an important branch of the computer vision field, is widely used in many fields. This paper focus on the key of the binocular vision --Stereo Matching. The paper introduces some basic theoretical knowledge as well as the basic model of binocular vision stereo matching, summarizes some of the existing common stereo matching algorithm, uses an improved local stereo matching algorithm and an global stereo matching algorithm based on graph cuts, and a global stereo matching algorithm based on Bayesian theory. Meanwhile, do some experiments on the Middlebury international standard platform, analysis the experimental results compared with the results of existing matching algorithms, demonstrating the effectiveness and accuracy of the algorithm.Local stereo matching algorithm in the paper use gradient difference value as a matching cost, improving the smoothness of the disparity map, meanwhile introduce Tukey’s Biweight function to matching cost, so as to achieve the effect of suppressing noise, in the next step, optimize the disparity, obtaining a good matching results.In the global stereo matching algorithm based on graph cuts used in the paper, feature points and edge points are extracted firstly, then match feature points and edge points to obtain the disparity of these points as the initial label of graph cuts algorithm; calculate DAISY descriptors of each pixel as data item of the energy function, calculate the sum of gradient difference between the horizontal direction,vertical direction and color change direction as the smooth item, finally obtain dense disparity map by seeking the minimum energy function, achieving stereo matching.In the global stereo matching algorithm based on Bayesian theory used in the paper, during the stereo matching process, Bayesian model is introduced. At first, use MSERDoG operator to extract support points, then match these support points with pixel gray value as the matching cost, fixed window as Cost Aggregation matched, and triangulate the matched support points to create two dimension network, we take disparity of the support points, two dimension network created by triangulate support points as priori probability conditions, thus ensuring efficient disparity search space and improving the matching efficiency.