Stereo Matching Algorithm Based On Binocular Vision

Binocular stereo vision is to observe the same scene from two viewpoints to acquire the perceptual images at different viewing angles,and calculate the parallax between the pixels of the image by triangulation principle to obtain the threedimensional information of the scene.Stereo vision technology mainly involves steps such as camera calibration,image preprocessing,stereo matching,and 3D reconstruction.Stereo matching is the most important and most difficult step.The main purpose is to obtain the corresponding disparity map by the corresponding algorithm to obtain the corresponding disparity map between the reference image and the target image.According to the disparity map information and the triangulation principle.Get the depth information of the scene.This thesis introduces the binocular stereo vision theory knowledge and the research status at home and abroad,from two aspects of the traditional method and deep learning of stereo matching technology is studied,the main work is as follows:1、In order to solve the problem that the traditional cross-scale stereo matching algorithm lacks reasonable constraint on parallax relation and cannot model the corresponding relation between low texture and repeated texture region,this thesis proposes a cross-scale random walk stereo matching algorithm to realize the effective aggregation of matching cost in multi-scale space and whole area.The experimental simulation results of the Middlebury data set show that,compared with the traditional cross-scale stereo matching algorithm,the algorithm in this thesis can effectively reduce the weighted average mismatching rate of scene images in all regions and non-occluded regions by 1 percentage point and 3 percentage point respectively to obtain a highprecision parallax map.2、In order to solve the problem of too many PSMNet network parameters and too long calculation time,this thesis proposes a densenet-based stereo matching algorithm,which realizes effective compression of network structure through feature reuse and bypass setting,so as to reduce the number of network parameters and improve the timeliness of the algorithm.At the same time,aiming at the problem that the current cnn-based stereo matching algorithm cannot make good use of the context information,this thesis adopts ASPP module to extract the feature information of different scales of the image,increase the sensor field,and make the predicted parallax map have more detailed information.Experimental simulation results of Scene Flow and KITTI 2015 data set show that the algorithm in this thesis is 14.92% faster than PSMNet in running speed.