Research On Wide Baseline Intensive Stereo Matching Method

Binocular stereo vision system is a computer vision system for the simulation of human vision. through this system, computer can obtain the 3D information of reality scene. The stereo matching is the core of the whole system. Stereo matching is to search for the correspondence between the same object points in different images pixels. However, due to the complexity of reality scene resulting in stereo matching in dealing with occlusion, weak texture and disparity discontinuous problem is still a big challenge. This paper is mainly to improve the accuracy of stereo matching algorithm as the research point.First of all, through the analysis of common matching method which based on the local window, this paper put forward there are two shortcomings. One is the window partition threshold is difficult to set, the other is match cost calculation method cannot deal with the image deformation. In order to solve the first problem, this paper presents a new method for finding support window, which is based on the regional color value fluctuation. The method using the regional color standard deviation to set window threshold, at the same time by using connectivity to find support window. Meanwhile, we designed an index to measure the rationality of the support window, and elected a reasonable threshold by the indicators. To solve the second problem, this paper proposes a new method to calculate the matching cost, which used translational correspondence method in the one-dimensional logarithmic coordinates. It can transform the proportional relationship of image line to translation. It can be convenient for calculate the correct matching cost between the large deformation image.Secondly, in order to obtain more precise disparity image, this paper proposes a multi direction scanning line optimization method, which using the consistency disparity area, rather than the color difference, as a judgment standard for disparity discontinuity. In order to obtain a larger area disparity consistency, we select the initial disparity plane to iterative merging, which based on the position relationship and the maximum plane distance.Finally, we built a test platform, and made the experiment for those algorithm. The experimental results show the effectiveness of this method. By comparison of the final disparity images and the standard disparity images also confirmed the algorithm can deal well in occlusion, weak texture and disparity discontinuities areas. And through the experiment of different baseline images also verify the stability of the algorithm in the baseline changes.