Research On Stereo Matching Algorithm Based On Binocular Vision

Binocular stereo vision is one of the important technologies in the field of computer vision.By processing images captured by binocular cameras,the disparity map of the image is obtained,and then the three-dimensional information of the object in the image is further calculated.It can be used for tasks such as reconstructing the three-dimensional model of the object and measuring object distance.Stereo matching technology,as a core step in binocular stereo vision,still has the problem of poor matching results in certain pathological areas,which affects the accuracy of binocular vision measurement and 3D reconstruction.So this article studies two high-precision deep learning based stereo matching algorithms and applies the improved algorithms to stereo matching and 3D reconstruction visualization of fan blade images.(1)A stereo matching algorithm based on multi attention.In response to the problem of low matching accuracy of current stereo matching algorithms for reflective pathological areas in images,a multi-spectral attention module is embedded in the residual feature extraction network of the Pyramid Stereo Matching Network(PSMNet)algorithm.Different twodimensional discrete cosine transforms are used instead of a single average pooling method to extract different frequency feature information,enhancing the network’s representation ability.At the same time,introducing a coordinated attention mechanism in the pyramid pooling module can not only maintain the long-distance dependence of direction perception,but also better capture the relevant position information lost during the pooling process,improving feature extraction ability.The experimental results on the KITTI2015 dataset show that the improved multi attention stereo matching algorithm MANet in this paper is less susceptible to interference caused by reflections compared to the benchmark PSMNet algorithm.The comparative experiment of reflective regions on the KITTI2012 dataset shows that the improved algorithm in this paper has more accurate disparity prediction results for reflective pathological regions.(2)A stereo matching algorithm based on multi feature aggregation.Aiming at the problem of high mismatch rate in edge ill conditioned areas of images,considering the important influence of cost aggregation network on disparity prediction,3D attention mechanism is introduced into the stacked hourglass network of PSMNet algorithm,and the generated weight is assigned to the advanced feature channel to focus on more important Semantic information.At the same time,in order to fully utilize the contextual information of the image,the contextual information generation module is improved to extract feature information around the region of interest.Finally,a variety of feature information aggregation modules are designed to combine the details of low-level features,Semantic information of high-level features and context information,optimize the constructed cost body,and improve the feature aggregation capability of the network.The test results on the Scene Flow dataset and KITTI dataset show that compared with the benchmark PSMNet algorithm and other advanced matching algorithms,the improved multi feature aggregation stereo matching algorithm MFANet in this paper has reduced endpoint error and mismatch rate to varying degrees,and can better estimate disparity in edge regions of the image.(3)Stereo matching and reconstruction visualization of fan blade images.This article designs a parallax estimation and reconstruction visualization software for binocular cameras,which can complete camera calibration,blade image acquisition and correction,stereo matching,and blade 3D reconstruction.Combining the feature extraction network of MANet algorithm with the cost aggregation network of MFANet algorithm,the MA-MFANet algorithm is proposed and applied to stereo matching of blade images.Then,the 3D reconstruction of blade images is performed using Open3 D tools.The results indicate that the MA-MFANet algorithm is more accurate in predicting disparity in the edge area of blade images compared to the benchmark PSMNet algorithm,and can reconstruct the reflective area of the blade.