Research Of Stereo Matching Algorithm Based On Binocular Vision

Research has shown that about 80% of humans' access to external information is obtained through vision.Binocular vision is a system that simulates the human visual system to obtain real-world information.It is an important branch of computer vision technology.After many years of development,it is more and more widely used in robotic vision,reverse engineering,industrial inspection,and medical imaging.Stereo matching is an important part of binocular vision.The result of stereo matching directly affects the effect of 3D reconstruction.However,the stereo matching process will be affected by various factors,such as lighting,object occlusion,and weak texture,which will lead to low accuracy of disparity map.How to improve the accuracy of the disparity map while improving the realtime performance has always been the focus of scholars' research.This paper mainly studies the stereo matching algorithm.For the mismatching problems of the local stereo matching algorithm in weak texture and discontinuous disparity areas,this paper proposes improved methods,and the improvement is mainly carried out in the matching cost calculation,cost aggregation and disparity refinement stages of the stereo matching step.(1)In this paper,a matching algorithm of census transform with edge gradient is proposed,which is improved in the phase of matching cost calculation.Since the classical census transform relies on the center pixel excessively,the center reference pixel is determined by the discrimination result of the average value of the neighborhood in the support window and the center pixel.Then,the absolute difference(AD)is combined with the census transform after the average discrimination,and the edge of the image is constrained by the improved Sobel edge gradient operator,and the fusion matching cost function is constructed to improve the matching efficiency of the algorithm in discontinuous area of the edge disparity.The experimental results show that compared with the traditional census transform and AD-census transform,the discontinuous area average matching error rate of the Census transform matching algorithm with edge gradient is reduced by 7.95%and 3.0% respectively.(2)The stereo matching algorithm based on guided filtering and multi-scale cost aggregation is proposed,with emphasis on the improvement of cost aggregation and disparity refinement.The guided filter method is integrated into the Gaussian pyramid structure,and the idea of minimum spanning tree is introduced to construct a more natural similarity measurement method.The cost aggregation is carried out in multi-scale space,and the regularization term is added to enhance the consistency between scales,so as to improve the mismatch problem in weak texture area.After the disparity is calculated by the winner take all(WTA)strategy,the left and right consistency detection and background filling method detection are used to eliminate the occlusion points.The sub-pixel enhancement method is used to improve the matching accuracy of the discontinuous disparity region,and the adaptive median filter method is used to further smooth the whole disparity map.Finally,the test image data set provided by Middlebury platform is selected to verify that the local stereo matching algorithm based on guided filtering and multi-scale cost aggregation proposed in the paper can effectively improve the matching accuracy of weak texture and disparity discontinuity region.Compared with the traditional census method,the average error matching rate of the algorithm is reduced by 14.52%,and the operation time of the algorithm is increased within 5%.