Research On End-to-end Stereo Matching Algorithm Based On Convolutional Neural Network

Stereo matching is one of the core technologies of computer vision.The goal is to calculate the disparity of the reference image pixels in the image pair.Stereo matching has the characteristics of flexibility and low cost,so it is widely used in autonomous driving,augmented reality and robot navigation.In recent years,convolutional neural networks achieved good results in the fields of computer vision,owing to their powerful feature-representation and function-fitting capabilities.Stereo matching algorithms based on convolutional neural networks have gradually become domestic and foreign research hotspots.This thesis mainly focuses on the end-to-end stereo matching algorithm based on convolutional neural network.(1)Aiming at the problem of poor processing effect of current algorithms on weak texture,repeated texture and slender structure regions,a stereo matching algorithm based on multiple attention mechanisms is proposed.The algorithm first uses the hourglass position attention module as the feature extraction network,because it can effectively aggregate multi-scale information and global context information to enhance feature representation,and then uses the combined volume to retain feature dimensions and similarity information,next,a cost aggregation network based on a multi-scale disparity attention module is used to aggregate feature information at different scales and different disparity levels,and optimize the cost volume;finally,the disparity map is obtained by disparity regression.Comparative experiments on three datasets Scene Flow,KITTI2012 and KITTI2015 show that the algorithm in this thesis has greatly improved the matching effect of weak texture,repeated texture and slender structure regions,compared with GC-Net and PSMNet,the endpoint error is reduced by 1.43% and 0.06% respectively,and the running speed also has certain advantages.(2)Aiming at the real-time problem of stereo matching algorithm based on convolutional neural network,a fast stereo matching algorithm based on edge consistency is proposed.The algorithm is mainly improved in three aspects,reducing the redundancy of the algorithm,improving the speed of the algorithm and the accuracy of the disparity estimation of the edge area.Use pyramid convolution to construct a pyramid residual network,and extract robust features through fewer network layers;Use a lower resolution to construct the cost volume and predict a low-resolution disparity map,and then gradually improve the disparity map through residual learning and restore it to the original image size;Reduce the number of 3D convolutions and use an edge consistency network to correct the disparity map.Comparative experiments on the two data sets on Scene Flow and KITTI2015 data sets show that the algorithm in this thesis can run in real time at a speed of 30 ms,and maintains finer edges in the final disparity map.