Research On Local Stereo Vision Matching With High Accuracy

Stereo vision matching technology is one of the most important research directions of machine vision.The main target of stereo vision matching is to get the correspondence pixel point from two or more pictures of the same scene,and generate a disparity map of the scene.Combining the spatial geometric relation and the disparity map generated by stereo matching,a depth map for 3D reconstruction can be generated.The accuracy of stereo matching directly determines the effect of 3D reconstruction.According to the disparity optimization method,the stereo matching can be divided into global stereo matching algorithm and local stereo matching algorithm.Although the global stereo matching algorithm obtains the optimal estimate disparity value by minimizing the global energy function,but the time complexity of the global algorithm is very high,so it is not suitable for real-time applications.The local stereo matching algorithm using the Winner Take All disparity selection strategy and local window cost aggregation method,so the local algorithm has low time complexity and fast running speed.With the emergence of adaptive window and the adaptive weight local stereo matching method,the local algorithm gets better disparity map,even better than some of the global algorithm,which makes the local stereo matching algorithms get more and more widely practical applications.Firstly,this thesis studied and improved the local stereo matching algorithm based on tree structure,using the cost value generation transmission character of minimum spanning tree(MST),makes the pixel of the whole image be able to participate in the cost aggregation.In the traditional method of generating tree cost function AD-Gradient,vertical gradient is added to improve the vertical change and the matching effect of the edge region;In addition,matching errors and their propagation are effectively reduced by adding multiple weights to the join edges of the spanning tree.Experimental results show that the proposed method has obvious advantages of matching accuracy,and the time cost is not obviously increased.Secondly,according to the existing algorithms only consider the adaptive weight matching similarity measure to calculate the weight between the point and the surrounding pixels,but ignore the pixel itself credibility problem,this thesis proposed an improved weight algorithm based on pixel classification and color segmentation.The proposed algorithm classified pixels into stable points and unstable points,and then using the results pixel classification and color segmentation to improve the cost aggregation weight value,so that the pixels with high reliability can propagate to lower confidence regions in the process of cost aggregation.Experimental results show that the disparity map obtained by the proposed algorithm has higher accuracy and better edge,especially in low texture and occlusion area.Meanwhile,the complexity of the algorithm increases little.Finally,this thesis proposed a non-local multilevel disparity refinement method based on tree structure.Disparity values were re-estimated by the MST obtained during cost aggregation.Left-right consistency and edge consistency detection were used to detect the inconsistencies and edge differences in the left and right disparity maps.Sub-pixel refinement is used to obtain sub-pixel precision,and bilateral filtering was used to smooth the disparity values of flat regions and maintain stable edges.Experimental results show that the proposed multilevel disparity refinement algorithm can achieve more stable edge,and the accuracy of the algorithm is improved greatly at sub-pixel level,and the mismatching rate is significantly reduced.Due to using the generated cost aggregation MST,the time complexity increases very little.