Research On Vehicle Detection And Tracking Method Based On 3D LiDAR

In recent years,vehicle detection and tracking technology based on lidar has developed rapidly and has been widely used in the field of intelligent connected vehicles.However,there are still many deficiencies in the existing vehicle detection and tracking technology.The vehicle detection algorithm is affected by contour occlusion,and the accuracy of shape and heading estimation is poor;Multi-target association problems in the environment.For this reason,this topic takes the vehicle detection and tracking problem of 3D Li DAR as the research object,further refines the shape and heading of the vehicle detection results,and then performs data association and track management for each target vehicle based on the point model,which effectively improves t he shape and direction of the vehicle.and the accuracy of heading estimation,a reasonable solution to the multi-target association problem is proposed.The main work of this paper is as follows:First,in order to have both vehicle detection accuracy and inference speed,a 3D point cloud vehicle detection is performed based on the Point Pillars network.The original point cloud is converted into a pseudo image by Pillar coding,and then a 2D convolutional backbone network is used for feature learning and upsampling,and then the category,3D detection frame and heading angle of each target are predicted based on the SSD detection head;when designing the loss function,comprehensively consider the loss of target positioning,the loss of heading angle estimation and the loss of target classification and carry out reasonable weighting to effectively avoid the huge error caused by the difference of heading angle being π and the excessive interference of the mismatched frame to the matching frame features,whic h is the best way to refine the shape and The heading estimation module provides accurate detection and classification information.Secondly,in order to solve the problem that the shape and heading estimation of Point Pillars are unstable when the vehicle point cloud cluster is L-shaped,a vehicle shape refinement and heading estimation algorithm based on the point cloud cluster feature is proposed.In the heading angle traversal stage,the characteristics of the rectangular frame are used and the search st ep size is appropriately increased to improve the solution speed;in the parameter calculation stage of the target detection frame,the distribution characteristic s of the laser point cloud at different positions are considered,based on the centroid point of the point cloud cluster and the rectangular detection frame.The positional relationship of the center is designed to optimize the strategy;in the objective function design stage,the influence of factors such as the area of the matrix detection frame,the extreme difference in the number of points in the point cloud cluster,and the sum of the distances from all points in the point cloud cluster to the rectangular detection frame on the shape estimation effect is comprehensively considered.;In the output stage of vehicle shape and heading,it is judged whether it needs to be refined by setting the objective function and threshold,and then the accurate vehicle shape and heading are output.Thirdly,in view of the problem that the existing multi-target tracking methods are difficult to take into account the real-time performance and accuracy,the point model is used to improve the computational efficiency,and the multi-target vehicle tracking is carried out based on Kalman filter and interactive multi-model;for the multi-target correlation matching problem in complex environments,an improved global nearest-neighbor data association algorithm based on track length is proposed.Considering the influence of Mahalanobis distance and track length,the minimum association distance is calculated based on the Hungarian matching algorithm.The track cancellation strategy is used to manage the track status of each target vehicle.Finally,based on the real vehicle point cloud data collected by the intelligent networked vehicle,the vehicle detection algorithm,shape refinement algorithm and multi-target tracking algorithm are experimentally verified.The experimental results show that the 3D point cloud vehicle detection algorithm based on Point Pillars has high detection and classification accuracy;the shape refinement and heading estimation algorithm proposed in this paper can effectively deal with the poor stability of shape estimation when the vehicle contour point cloud is occluded and is L-shaped.It can improve the shape and heading estimation accuracy to a certain extent;the multi-target data association algorithm proposed in this paper has better association accuracy,better tracking stability,and the track management strategy has certain feasibility.