Research On Stereo Matching Algorithm And Application Based On Binocular Vision

With the development of autonomous driving,unmanned aerial vehicle(UAV)navigation and other advanced technologies,binocular vision technology has become a research hotspot in recent years.The stereo matching algorithm is the core algorithm of binocular vision,and the accuracy of stereo matching will directly affect the conversion of2 D information to 3D information.Compared with the traditional stereo matching algorithm,the stereo matching algorithm based on deep learning develops rapidly and has gradually become a hot topic in the research of stereo matching algorithm.The algorithms in this paper solves the existing problems of stereo matching from the following aspects.(1)Aiming at the fact that binocular images contain complex scenes,which is easy to lead to poor stereo matching performance,a coarse-to-fine stereo matching algorithm based on multi-scale structural information filtering is proposed.In order to improve the limitations of using ordinary convolution to obtain features,the algorithm introduces multi-scale residual module to obtain multi-scale feature information and improve the stereo matching performance of textureless areas.In addition,aiming at the false matching caused by the lack of structural information,the structure attention weight generation module is proposed to enhance the model’s ability to obtain structural information and optimize the diversity of features.Finally,in order to further optimize the matching cost,the cost aggregation strategy from coarse to fine is used to enhance the supplement of inter-stage structural information.Experiments show that the proposed stereo matching algorithm can improve the stereo matching accuracy of challenging areas in complex scenes.(2)Aiming at the problem of weak generalization ability of cross-domain stereo matching,a cross-domain adaptive stereo matching algorithm based on transfer learning is proposed.In order to enhance the model’s acquisition of general domain information,a general domain feature extraction module is proposed to acquire general domain features.In addition,to reduce the false matching caused by non-stereo matching task information,a feature adapter is designed to adapt general domain features to stereo matching tasks.To solve the problem of lack of similarity information caused by single matching cost,a multi-scale matching cost construction strategy is proposed,by constructing multi-scale matching cost,the diversity of cost information is improved.At the same time,in order to adaptively adjust the cost distribution,a structure weight prediction module is proposed,which adaptively adjusts the disparity range of matching cost according to the structure weight.Experiments on multiple datasets show that the proposed adaptive stereo matching algorithm based on transfer learning can improve the cross-domain accuracy of stereo matching.(3)The current status of 3D reconstruction and target ranging is analyzed,and the application experiments of 3D reconstruction and target ranging are designed for different scenarios,which verifies the feasibility of the proposed algorithms.For the 3D reconstruction,the dense disparity map is obtained by the proposed structural filtering stereo matching algorithm,and the point cloud is reconstructed in 3D according to the obtained camera parameters and disparity map,and the experiments show that the dense point cloud obtained is basically the same as the manual measurement result.For the ranging task,the target detection algorithm is used to detect the main targets in the scene,and the disparity map of the target is obtained by the domain adaptive stereo matching algorithm,the ranging results are basically correct within the error range compared with the manual measurement results.The application experiments show that the proposed algorithm can be applied to 3D reconstruction and target ranging tasks.