Research And Application Of Graph Matching Binocular Vision Obstacle Detection Based On Feature Fusion

Obstacle detection is a crucial application of stereo matching technology,which utilizes the disparity map obtained from a binocular vision system to identify obstacles on the road,enabling mobile robots to avoid them.The effectiveness of stereo matching directly impacts the quality of the disparity map,which subsequently affects the accuracy of obstacle detection.With the research and development of stereo matching technology,more and more stereo matching algorithms have been proposed.Among them,feature-based stereo matching methods have become one of the research focuses due to their advantages of fast matching speed,strong robustness to lighting changes,and so on.However,unlike region-based stereo matching methods,feature-based stereo matching methods do not have a standard template for matching,so the initial feature point matching is particularly important.High-quality initial matching is the basis for obtaining accurate disparity maps and is also a prerequisite for accurate obstacle detection.This thesis mainly focuses on the problem of low matching accuracy caused by deformation and noise in traditional feature point matching methods in feature-based stereo matching,which affects the quality of the disparity map.It conducts basic research and algorithm improvement and packages the research results into application.The main work of this thesis is as follows:(1)Analyzing the imaging process and principle of the camera,using the calibration method proposed by Zhang to complete the calibration experiment of the binocular camera,and determining the reliability of the obtained camera’s intrinsic and extrinsic parameters,which prepares for the rectification of left and right views and the recovery of three-dimensional information.(2)Addressing the problem of low initial feature point matching accuracy in feature-based stereo matching methods.This thesis proposes a feature fusion-based graph matching method and a graph matching method based on updating the affinity using the correlation correspondence relationship.The matching effect of feature points is tested using synthesized datasets,the CMU house dataset,natural image dataset,and the PASCAL dataset.The experimental results show that the proposed feature fusion-based graph matching method improves the matching accuracy of feature points,and adding the idea of updating affinity based on feature fusion further improves the matching accuracy.(3)In obstacle detection applications,this thesis first uses the feature point matching pairs after eliminating the mismatched pairs as seed nodes and uses the region growing method to solve the sparse disparity map problem of feature-based stereo matching methods.Then,after optimizing the disparity map,it uses a defined three-dimensional information threshold filtering strategy for obstacle detection.Finally,the obstacle detection method proposed in this thesis is tested using self-collected binocular scene images.The test results show that the proposed obstacle detection method has high theoretical reference value and engineering application value.