Research On Key Technologies Of Binocular Stereo Vision Navigation And Positioning Of Intelligent Vehicles

An important prerequisite for the intelligentization of automobiles is high-precision navigation and positioning technology.At present,feature-based Visual Odometry is one of the most commonly used technologies in the field of intelligent vehicle positioning.However,the feature-based Visual Odometry method usually has the problem that the real-time performance is not good enough,and it is prone to false feature matching.This thesis focuses on these two issues,and the main research work is as follows:The basic principles and key theoretical foundations of the feature-based Visual Odometry method are introduced,including related theories such as definition of coordinate systems and vehicle motion,projection camera model and binocular depth estimation.This chapter also introduced the advanced Oriented FAST and Rotated BRIEF(ORB)feature point extraction and matching method,and 3D-2D motion estimation method.Aiming at solving the problem of poor real-time performance of feature-based Visual Odometry method,a Visual Odometry acceleration method based on the optical flow tracking and adaptive feature extraction is proposed.In most images,Lucas-Kanade optical flow tracking is used to obtain feature correspondences to increase the running speed of the feature-based Visual Odometry,and for the resulting loss of feature points,an adaptive feature point extraction method was proposed.When the feature point loss is small,the frame interval will be increased to ensure the real-time performance of the Visual Odometry,and when the feature points loss is large,the frame interval will be reduced to maintain the normal operation of the Visual Odometry.Aiming at solving the problem of false feature matching,a false feature matching removal method based on local image similarity was proposed.When judging whether a feature correspondence is correct,the partial images around the feature points are extracted separately,and the similarity of the two partial images is calculated by the bag of words model.If the similarity is higher than the set threshold,it is a correct match,otherwise It was rejected as a mismatch.Based on the proposed Visual Odometry acceleration method and feature mismatch removal method,the corresponding algorithms were designed respectively.The algorithm experiment test on the KITTI data set proves that the Visual Odometry acceleration method proposed in this thesis can effectively improve the real-time performance of the feature-based Visual Odometry with little change in accuracy.Also based on the KITTI data set,the proposed mismatch removal algorithm was tested.The results showed that the proposed mismatch removal method can remove most mismatches while retaining a sufficient number of correct matches.Finally,through small-scale testing,we explored the feasibility of using the Visual odometry acceleration method and the mismatch removal method proposed in this thesis to simultaneously improve the real-time and accuracy of the feature point method.It is found that the real-time performance is significantly improved in all tests and the accuracy can be improved when there are fewer feature points.