Research On Theory And Application Of Traffic Environment Perception Based On LiDAR Point Clouds

The rise of major applications such as smart transportation and automatic driving has placed an urgent need for accurate,real-time,3D(three-dimensional)semantic percep-tion of the traffic environment.Laser scanning technology is one of the best technical methods for large-scale 3D measurement,providing a data foundation for 3D perception of large-scale traffic environments.Existing traffic-oriented laser scanning point cloud applications mainly focus on the extraction of static road objects and the modeling of high-precision maps.The traffic environment is centered on the vehicle operation and the driver.Therefore,the visual perception assessment of the driver and the 3D modeling of the traffic dynamic process are important scientific issues in the traffic environment perception,and the related research work is less.In this dissertation,base on the laser scanning 3D point cloud,from the perspective of 3D scene perception,the visually rec-ognizable computing model of traffic signs in static traffic scenes and the dynamic 3D reconstruction of vehicles in traffic environment are studied in depth.the study The main research contents and innovations are as follows(1)Proposed a recognizability field calculation model for traffic sign based on LiDAR point cloudsAiming at the problem of driver's assessment of the visual perception of traffic signs in the traffic environment,based on the principle of retinal imaging,an evaluation model of the visibility of traffic signs at a certain viewpoint is proposed.Combined with the sight distance and the installation specification of traffic signs,an evaluation model of the recognizability of traffic signs at a certain viewpoint is constructed.For the first time,the definitions of visibility,visibility field,recognizability and recognizability field are given,and their exact calculation method is given.The results of this research can significantly improve the efficiency of the traffic facility management department's visual perception maintenance of traffic signs(2)Constructed a large-scale vehicle pose estimation datasetAiming at the shortage of data samples for vehicle pose estimation in traffic surveillance video,a large-scale rendering dataset Autonomous Driving for Smart City(ADSC)was constructed to meet the perspective,2D-3D matching key points and the registration of image to point clouds.The data set can be used as training data and evaluation benchmark for vehicle pose estimation,vehicle detection,vehicle semantic segmentation and dynamic traffic 3D reconstruction.(3)Proposed a model-driven vehicle 3D pose reconstruction algorithmAiming at the problem that the existing pose estimation algorithm relies on the camera in-trinsic parameters,thus causing the poor adaptability between different cameras,we pro-pose a model-driven deep learning network architecture based on the image and 3D point clouds registration.The architecture can estimate the 6-DoF(six Degree of Freedom,6-DoF)pose of multiple vehicles in the global coordinate system from the surveillance image,and make the vehicle 3D pose only related to the position of the 2D key point,and is independent of the intrinsic and extrinsic parameters of cameras,thus ensuring that the network architecture is applied to all surveillance cameras.Experiments show that the algorithm is superior to the traditional algorithm in the scale evaluation of central 2D key points,wheel 2D key points and 3D IoU(Intersection over Union)(4)Proposed a dynamic reconstruction algorithm of traffic environment combining static point clouds of objects and surveillance videoAiming at the problem that the detection results for image key points are susceptible to texture blur,shadow,and illumination,based on the model-driven vehicle 3D pose recon-struction algorithm,this paper further studies how to make reasonable use of the elevation and roll angle of the vehicle provided by the ground point cloud,and the mutual constraints between the key points in the 3D model to optimize 2D key point prediction results.An algorithm of dynamic reconstruction of traffic environment combining static point cloud and surveillance video is proposed.Experiments show that the proposed algorithm is su-perior to competitive algorithms in the scales of center 2D key points,wheel 2D key point accuracy,3D orientation accuracy,and 3D IoU.The algorithm can be used for 3D re-construction of traffic accident,autonomous driving.It is of great significance for traffic accidents responsibility determination and intelligent traffic construction.