Research On Semi-global Stereo Matching Algorithm Of Binocular Stereo Vision

Stereo vision is an important part of computer vision, it has become a researchhotspot. Binocular stereo matching is mainly used in the field of robot navigation,industrial measurement, the military field and so on. Binocular stereo vision mimicthe stereo perception process of human visual, obtain three dimensional space sceneinformation. In a stereo vision system, the accuracy of the stereo matching directlyaffects the effect of3d reconstruction, and since the scene illumination and noiseinterference, stereo matching becomes the most important and most difficult part ofthe stereo vision.There are a lot of stereo matching algorithms and some of them have achievedremarkable results. This paper divides stereo matching algorithm into different types,and discusse the advantages and disadvantages of various algorithms. The paperchoose semi-global matching algorithm as the research direction. This algorithm hasgood robustness, and is not sensitive to light. However the algorithm still has someproblems, such as the fuzzy of boundary, miss matched of occlusion region anddiscontinuous region.To solve these problems, this paper makes a research on cost aggregation anddisparity refinement. Combined with image segmentation algorithm to optimize theglobal energy function. In disparity refinement section, filling method for occlusionsusing the ideas of weights. The main work of this paper includes the following threeaspects:(1)Using mean-shift algorithm for Image segmentation, dividing the similarcolor of pixels into the same area. According to the priori knowledge that disparity inthe same block changes continuously and the location of the disparity mutations isusually the edge of the block, this paper adds image segmentation information to theglobal energy function of semi-global stereo matching algorithm and proposes a morerational global energy function, solves the problem of give the same penalty todisparity mutations caused by different reasons. Finally achieve the purpose of improving the matching accuracy.(2)In the part of disparity refinement, this paper analyzes the problems oftraditional method, because one can not guarantee that there is at least one non-occlusion pixel in the neighbor of the occluded pixel, so the estimated disparity is notaccurate. The disparity estimation use the idea of weights, adding image segmentationinformation into the calculation of the weights. Searching for the first non-occlusionpixel in8or16directions and using the disparity of the non-occlusion pixel whichhas the largest weight as the disparity of the occlusion pixel.(3)The control parameters of mean-shift and the punishment factor of the newglobal energy function determined by the experiment. In the end, this paper tests theproposed algorithm by using the stereo matching images of the Middlebury website.After comparing with the original algorithm and other algorithms, this paper analyzeseffectiveness of the algorithm by the matching accuracy.