Research On Stereo Matching Algorithm Based On Deep Learning

Binocular stereo vision has the advantages of high flexibility,low cost,and easy implementation on the problem of extracting image depth information,so it has been widely used in the fields of automatic driving,robot walking route navigation,and robotic arm grasping workpiece.The core of binocular stereo vision is to find the corresponding matching points in the left and right images captured by the two cameras,obtain the parallax,and further use the triangulation method to calculate the depth information of the image.In recent years,the convolutional neural network has developed rapidly,and it has also been widely used in binocular stereo vision.Binocular stereo matching algorithms based on deep learning have gradually become a new direction in binocular stereo vision research.This paper studies the binocular stereo vision algorithm for end-to-end convolutional neural networks,mainly focusing on studying the error problems caused by occlusion areas,reflections,weak textures and repeated texture areas in the matching process,improving the matching accuracy.In order to improve the matching accuracy,we mainly researched from the following two aspects:(1)A stereo matching algorithm based on atrous spatial pyramid pooling is propsed.We designed an end-to-end stereo matching algorithm.The core module of the algorithm is the atrous spatial pyramid pooling structure.The design of the atrous spatial pyramid pooling structure effectively extracts the multi-scale semantic information of the image,effectively improves the accuracy of the stereo matching algorithm in the feature extraction module,and the existence of the atrous convolution without the linear interpolation algorithm.This structure can improve the speed of the network processing of image pairs.(2)A stereo matching algorithm based on densely connected convolutional networks is proposed.The core of the algorithm is a densely connectedconvolutional network structure.The densely connected convolutional network strengthens the transfer of features through dense connections,connects the input and output layers of each densely connected block,and fuses multiple layers of feature information.Experiments show that this algorithm can further improve the matching accuracy.We carried out an experimental analysis on the KITTI Stereo dataset,and the results show that our proposed stereo matching algorithm effectively improves the accuracy of parallax estimation and increases the speed of network training.