Research On Stereo Matching Algorithm Based On Convolutional Neural Network

In this paper,the binocular stereo matching and convolutional neural network are studied.The principles and techniques of binocular stereo vision are introduced in detail,and the basic concepts and basic structure of the convolutional neural network are described.The convolution neural network is applied to the similarity calculation of left and right views.Then use the stereo matching algorithm to add smooth constraints,optimize the matching cost,and obtain accurate and dense disparity maps.The main work is as follows:(1)A convolutional neural network structure is designed to calculate the matching cost.The structure and parameter training effect of the convolution neural network have a direct effect on the next matching.A convolutional neural network structure is designed to calculate the matching cost.The structure and parameter training effects of the convolutional neural network have a direct impact on the next matching.Therefore,based on the analysis of the traditional matching cost calculation method and the reference to the classical convolutional neural network structure,a new convolutional neural network structure is proposed to calculate the similarity of the left and right images.This paper proposes a convolutional neural network structure using the combination of the Siamese network and the spatial pyramid pooling(SPP).It is use the corresponding regions(local sensing regions)of the left and right views as the lowest level input of the hierarchical structure.Finally,the original matching cost function of left and right views is obtained.At the same time,the influence of the convolutional neural network hyperparameter setting on the calculation result of the initial matching cost is studied.Experimental results show that this method reduces the initial matching error rate.(2)A matching cost aggregation method is proposed in this paper.Region-based local stereo matching algorithm is a simple and efficient matching algorithm,and window selection in local matching algorithm is a crucial issue.After using the convolutional neural network to calculate the initial matching cost,this paper uses semi-global matching algorithm and adaptive window matching algorithm based on diagonal search for matching cost aggregation——Anchor-diagonal window matching cost aggregation(ADCA).In order to solve the problem that the matching accuracy of weak texture region is low and the region of occlusion is prone to mismatch.n this paper,the left and right consistency detection is used to perform the mis-match point test and interpolation correction on the stereo matching first disparity.Then use sub-pixel edge extraction and bilateral filtering to optimize the parallax results to reduce the initial parallax mis-match rate and obtain dense disparity maps with higher accuracy.This method has good flexibility,high matching accuracy,and elimination of noise point effects.