Research On Stereo Matching Technology Of Long Focal Length Binocular Vision Based On Deep Learning

Compared with the typical stereo vision system,the long focal length stereo vision system can maintain the resolution of 3D imaging at a longer shooting distance and has great potential in the application of 3D reconstruction of large infrastructures at medium and long distances.However,the problems of image blur and stereo rectification failure brought by long focal length will severely affect stereo matching,the core procedure of 3D reconstruction.The research in this thesis aims to solve the key technologies of stereo matching in long focal length stereo vision.With deep neural network algorithms,the study is conducted in three aspects: the construction of datasets for stereo matching,the stereo matching based on a deblurring network,and the unconstrained stereo matching based on an optical flow network.The main contributions are as follows:1.In-depth study of the imaging characteristics of long focal length stereo vision,analysis of the causes of unbalanced defocus blur,and unrectified stereo images.By adding normalized defocus blur and unconstrained stereo perturbation to the FlyingThings-Stereo dataset,respectively,an unbalanced defocus stereo dataset and an unconstrained stereo perturbation dataset are constructed.The constructed datasets successfully train the stereo deblurring network and unconstrained stereo matching network.2.To reduce the influence of unbalanced defocus blur on stereo matching in long focal length systems,a stereo matching method based on a stereo deblurring network is proposed.We successfully train a binocular deblurring network named BL-Net(Binocular Linkage Network)by the unbalanced defocus stereo dataset,and construct a similarity-enhanced stereo deblurring loss function to improve the sharpness consistency of the same scene in the left and right images after deblurring.The BL-Net effectively improves the matching accuracy of the subsequent stereo matching network.3.Aiming at the problem that a long-focal length stereo vision system is difficult to perform stereo rectification by real-time calibration,an unconstrained stereo matching method is proposed to search in the range of the 2D full image.The network is called EGOF-Net(Epipolar Guided Optical Flow Network).The basic assumptions in unconstrained stereo matching are defined,and the architecture of the optical flow estimation network is modified accordingly.The epipolar guidance module is added,and the matching result can be obtained with only one network structure without stereo rectification.4.Based on the theoretical analysis of long focal length stereo vision,a deblurring network BL-Net and unconstrained stereo matching network EGOF-Net are applied together to establish a set of long focal length unconstrained stereo vision experimental system in which the left and right cameras can be rotated independently.A series of 3D reconstruction experiments are carried out on the concrete surface with weak texture information by the built system.Also,the system is applied to the 3D scan part of the Tianjin TV Tower and achieves a 3D point cloud with good quality.