Research On Vision SLAM Algorithm For Intelligent Vehicle Considering Semantic Information Of Road Features

In recent years,the "new four" changes of electrification,intelligence,networking and sharing have become an important direction for the future development of the automotive industry,and high-grade self-driving cars have gradually entered the pilot application stage.Many countries around the world have incorporated the development of autonomous driving technology into their national top-level plans to seize the strategic high point of future automotive industry development and strengthen national scientific and technological competitiveness.The realization of high-level autonomous driving technology relies on the intelligent vehicle’s accurate cognition of the surrounding environment and its own location,so the problem of high-precision positioning of intelligent vehicles needs to be solved urgently.To address this task,this paper is oriented to road traffic scenarios and uses binocular vision sensors to realize the extraction of traffic participants and road structured information from the vehicle surroundings through a combination of deep learning and multi-view geometry,and combines the extracted semantic information with SLAM algorithms.While improving localization accuracy,a lightweight semantic point cloud map containing semantic information of road features is constructed to provide intelligent vehicles with richer environmental information to adapt to complex autonomous driving tasks.First,in order to improve the localization accuracy and robustness of SLAM algorithm in dynamic environment,a dynamic feature point rejection based method is proposed,which can effectively improve the localization accuracy and robustness of binocular vision SLAM algorithm in dynamic environment.Based on the open source ORB-SLAM2,a dynamic feature point filtering thread is added to distinguish dynamic and static regions in the image based on the multi-view geometry method and the instance segmentation neural network,and then the feature points belonging to the dynamic regions are eliminated in the tracking thread to ensure that the camera localization and map building are not affected by the moving objects in the scene.Secondly,a binocular vision lane detection algorithm on the BEV plane is studied to provide accurate lane line location information for road feature semantic map construction.The algorithm uses a virtual camera position method to ensure the consistency of the spatial relationships of the front cameras in different vehicles;the algorithm also proposes a BEV space coordinate system transformation module to realize the coordinate transformation of the front cameras to the BEV space under the BEV space to achieve fast and accurate lane line position detection in the BEV space.Finally,the high-precision intelligent vehicle motion position information is used to construct a road feature semantic map containing road feature semantic and geometric information by combining the BEV lane line detection results,which provides rich environmental information for intelligent vehicles to realize more complex autonomous driving tasks.