Research On Intelligent Vehicle Localization Method Based On Multi Basic-element Fusion

As an important development direction of current automotive industry,unmanned driving technology attracts more and more automotive and technology companies to participate in its research and development.In order to realize real unmanned driving,in addition to excellent control and planning algorithm,it also needs precise positioning,which is the basis for the realization of other functions of the vehicle.With the development of machine vision,vision-based positioning method has become the focus of current research,and vision sensors are relatively cheaper,which is the first to be applied to mass production of vehicles.Visual localization is a method of pose estimation by using camera sensors,which has the advantages of great accuracy and robustness.However,it is easy to fail in positioning and tracking in scenes,such as fast motion,weak texture,weak illumination,and obvious lighting changes.In the case of missing image frames,it will also lead to the failure of the whole system.Therefore,this paper proposes a visual localization method based on multi-feature primitive in complex scenes,and verifies the accuracy and robustness of the proposed algorithm.Based on the research objectives of the paper,the main work is as follows:(1)This paper proposes a localization algorithm based on points and lines fusion.Aiming at the problem of tracking failure of the localization algorithm based on point feature caused by the obvious change of light,this paper introduces the line feature which is not sensitive to light transformation on the basis of point feature,studies the extraction and description method of line feature,introduces the extraction of LSD line feature and the description of LBD line feature.Besides,aiming at the characteristics of more line features and obvious length difference,a method for extracting long line features is proposed.Moreover,based on ORB-SLAM2 framework,the point-line feature fusion algorithm is proposed,and the complete pose estimation algorithm and local optimization method are constructed.The accuracy of the algorithm is verified by public data sets.(2)This paper proposes a multi-path fusion method based on visual features.The multi-path fusion is introduced including loop detection and pose transformation theory in different coordinate systems;This paper introduces the differences between multi-path fusion and relocation methods in theory,and proposes a loop detection method based on spatial consistency,which improves the recall rate of loop detection;Under the premise of correct relocation,path fusion is realized to complete the correct pose transformation of camera in multi segment path.The effectiveness of the proposed method is verified by public data sets,and the amount of pose information is improved greatly.(3)The binocular stereo vision positioning model is designed through two monocular cameras,and the positioning test platform is built based on Bao Jun and the model.The model of the automobile body and the imaging model of pinhole camera are analyzed to improve the test platform,so that the test platform has a better vision and runs stably.Experiments are carried out on residential roads,urban roads and a variety of roads with viaducts to block the sun at noon,which verifies that the proposed algorithm has high robustness in complex scenes such as obvious light changes.