Obstacle Detection And Tracking Based On LiDAR

Environmental perception technology is an important part of the unmanned driving system.The accuracy and real-time nature of the perception technology directly affect the safe driving of unmanned vehicles.LiDAR(Light Detection and Ranging)is an active sensor that can directly obtain three-dimensional information of the environment to provide rich data for unmanned driving.Compared with cameras,LiDAR is not easily affected by external light;compared with millimeter wave radar,LiDAR is more accurate and has higher resolution,so it is widely used in the field of unmanned active collision avoidance.However,LiDAR also has disadvantages such as large amount of point cloud data and uneven distribution.When processing its data,it is necessary to consider how to deal with the challenges of low real-time and poor applicability of the algorithm.This paper focuses on the problems of inaccurate detection,unstable tracking and low real-time performance in the current main gear collision avoidance technology of LiDAR.The specific work is as follows:Introduce the working principle of LiDAR and various performance indicators of the Velodyne VLP-16 model LiDAR;introduce the data form of LiDAR Point Cloud,and complete the analysis and visualization of point cloud data based on the ROS(Robot Operating System)platform;According to the installation position of the LiDAR on the "ART"(Autonomous rail Rapid Transit)train,the NDT(Normal Distribution Transform)matching algorithm is used to calibrate the coordinate system of the two LiDARs,and the coordinate system of the point cloud data is unified.Prepare for the subsequent data processing.Aiming at the disorder and hugeness of LiDAR point cloud data,polar coordinate grids are used to process the original data,and filtering rules are set based on the grids to achieve ground point cloud filtering;For DBSCAN(Density-Based Spatial Clustering of Applications)with Noise)The clustering algorithm has the disadvantages of high time complexity and poor parameter adaptation.The adaptive search parameters are used to improve the adaptability of the clustering algorithm to the environment,and the "representative point" growth method is used to improve the neighborhood search of the clustering algorithm Speed;For the clustered point cloud clusters,the minimum envelope rectangle algorithm based on convex hull is used to extract the feature information of obstacles.To solve the problem of JPDAF(Joint Probability Data Association Filter)with a large amount of calculation in a target-intensive environment,the strategy of joint event screening and confirmation matrix decomposition is adopted to simplify the confirmation matrix of the joint probability data association algorithm,reduce the number of joint events,and improve The real-time nature of data association is improved.The forgetting factor is used to evaluate the observation noise in real time and optimize the adaptability of the filter algorithm to achieve more accurate tracking of obstacles in the environment;In order to deal with obstacles in the process of tracking,such as occlusion,disappearance,and new additions.The tracker is designed in the multi-target tracking system,and the movement status of the target is managed by introducing the number of survival times to realize the maintenance and update of the target movement information.Based on the ROS platform,the vehicle-mounted LiDAR is used for experiments on urban roads,and the roadside LiDAR is used for experiments on park roads to verify the effectiveness of obstacle detection and tracking algorithms.In the detection experiment,the detection effects of the European clustering,the DBSCAN algorithm and the improved DBSCAN algorithm are compared,and the real-time performance and accuracy of the clustering algorithm are analyzed.In the tracking experiment,the tracking effects of the traditional JPDAF algorithm and its improved algorithm are compared,and the real-time performance and tracking accuracy of the tracking algorithm are analyzed.The experimental results show that the algorithm can reliably extract the point cloud characteristic information of obstacles and perform target tracking stably.