Research On Stereo Matching Algorithm Based On Binocular Vision

80% of the source information is the eyes,and binocular stereo vision technology is a system that simulates the visionl system to obtain real-world information,and is a research hotspot in computer vision.Stereo matching has a wide range of applications,such as target detection,unmanned aerial vehicles,face matching at ticket gates in scenic spots,etc.Its research focuses on two aspects: matching accuracy and matching speed.At present,researchers have done a lot of research in improving the matching accuracy and running speed of disparity maps,but some stereo matching algorithms still have some shortcomings.This article improves the matching algorithm for these problems.The main contents are as follows:improves the matching algorithm for these problems.1)A new semi-global stereo matching algorithm is proposed.The algorithm adds two kinds of tree structure into the path cost aggregation.The classical semi-global algorithm has some shortcomings in the path selection of cost aggregation.Neither in eight directions or sixteen directions can fully take care of the pixels of the entire image,so the mismatching rate is high in the sparse texture region and the pixel discontinuity region.In this paper,a change was made at the cost aggregation path to change the original aggregation path into four,making full use of the pixels around the matching points to increase the matching accuracy and save time.2)By analyzing the advantages and disadvantages of using the traditional Convolutional Neural Network(CNN)to calculate the initial matching cost,the network structure of the CNN is improved.This paper proposes a network structure(SPP?CNN)that combines CNN with Spatial Pyramid Pooling(SPP).Aiming at the problem of inconsistent size of the image input from the network,a pyramid pooling layer is added to the final convolution layer.The input data of SPP?CNN is two left and right images.The network finally obtains the original matching cost function through a series of convolution processes.At the same time,the influencing factors of the calculation result of the initial matching cost were studied.Experimental results show that this method improves the initial matching error rate.3)A matching cost aggregation method combining a semi-global stereo matching algorithm and an adaptive window matching algorithm is proposed.The choice of the size of the matching window will determine the matching accuracy of the localmatching algorithm.In this paper,after calculating the initial matching cost using the SPP?CNN network,a semi-global stereo matching algorithm and an adaptive window matching algorithm are used for the cost aggregation of the matching.In the stereo matching process,problems such as unsatisfactory matching of sparsely textured regions,occlusions,and other regions are prone to parallax and have a high mismatch rate.In the disparity post-processing stage,left-right consistency detection,sub-pixel enhancement,and filtering are used to correct the disparity mismatching points of the disparity map generated after stereo matching in order to alleviate the mismatch rate of the initial disparity map and further obtain a higher Dense disparity map of accuracy.Experimental results show that this method has better flexibility,can obtain disparity maps with higher matching accuracy,and eliminates the influence of noise points.The improved semi-global stereo matching algorithm combined with tree structure in this paper reduces the initial cost aggregation path from sixteen to four,and makes full use of the pixels around the point to be matched.Combined with the convolutional neural network structure of the space gold tower,the problem of inconsistent size of the input image is solved.Compared with the traditional algorithm,the matching accuracy has been significantly improved,and the improved algorithm has also improved the matching efficiency to a certain extent compared with the traditional algorithm.