Research On End-to-End Binocular Stereo Matching Method Based On Deep Learning

In binocular vision technology,the stereo matching step is a crucial step,and it is also the most central step.The essence of stereo matching is to obtain the horizontal position difference of pixels in the binocular image,that is,disparity.Then the depth value of the image is calculated by the triangulation.Compared with the conventional stereo matching algorithm,the matching algorithm based on deep learning uses machine learning for matching calculation,so it has great advantages in accuracy.In this paper,the following research will be carried out on the binocular stereo matching algorithm based on deep learning.First of all,in view of the problem that the current cost construction method cannot accurately measure the similarity between feature pixels,this paper proposes a different cost construction method,which is constructed by the group-by-group SAD method.This method can calculate the similarity between pixels more accurately,and then construct the cost volume that contains more accurate similarity information.Then,in order to solve the problem that the feature information is lost in the continuous downsampling process by using stacked hourglass modules in the aggregation network,this paper proposes a densely connected multi-scale feature reuse aggregation network.The network integrates the feature information between different hourglass modules by means of skip connections,which effectively suppresses the loss of feature information caused by continuous downsampling operation.Secondly,most of the current aggregation networks use continuous encoding and decoding operations to repeatedly learn feature information,but it ignores the fact that the aggregation network needs to make full use of the multi-level context information of the cost volume for feature similarity learning.This paper proposes a 3D feature Pyramid attention aggregation network,which improves the performance of the aggregation network by learning multi-level contextual feature information on the cost volume.Finally,a disparity optimization module is designed to solve the problem of mis-matched pixels in the initial disparity map.The disparity optimization module is used to post-process the initial predicted disparity map,which reduces the mismatched pixels in the initial disparity map,and improves the accuracy of the stereo matching algorithm.This paper conducts sufficient experiments on three major datasets of stereo matching:Scene Flow,KITTI 2012 and KITTI 2015.It verifies the effectiveness of the algorithm in this paper.