Research On Target Detection And Tracking Algorithm Based On 3D Lidar For Autonomous Driving

Autonomous Driving is an emerging technology that acquires environmental information through sensors and realizes vehicle control by computer system.A mature autonomous driving system can accomplish completely without human intervention.In recent years,with the increase of people's demand,the development of autonomous driving technology has entered the blowout period.With the continuous exploration of autonomous driving technology,the requirements of environmental perceived reliability are also increasing.As one of the most important data acquisition sensors for autonomous driving vehicles,3D Lidar has the characteristics of high precision and strong anti-interference ability.Target detection and tracking based on 3D Lidar is an important part of the environment perception of autonomous driving vehicles.The vehicle is enabled to detect surrounding objects and estimate their trajectories,predicting the state of motion,so as to make autonomous driving vehicles follow the specific route and avoid collisions with other objects.The research results of this paper rely on the robot research group of Jilin University to develop a target detection and tracking algorithm for unmanned driving.The Velodyne HDL-32 E 3D Lidar is used as the main data source.The grid map is used to complete the target clustering and extract the feature information of the target.The result of target detection realizes the tracking function of the target through data association and state estimation operation.The main research work is as follows:(1)The target detection module maps the 3D point cloud data into a 2D grid map and according to the correlation between grid cells and the height information in the statistics,filtering out irrelevant point cloud data such as ground points and noise points.In point cloud clustering,in order to solve the problem that the unstable clustering effect caused by the increasing distance between target and lidar and the gradual sparse density of point cloud scanned to target.In this paper,a method is proposed to set the threshold of the number of point clouds required for the clustering target adaptively according to the distance between the target and the lidar,which improves the clustering effect of distant targets and reduces the probability of mis-clusters on the near-field noise.Then,the feature informations of the targets in the clustering result are extracted by the minimum circumscribed rectangle method.(2)For the data association in target tracking,in order to solve the problem of fastly and effectively correlating many-to-many data,this paper proposes a method of merging the Nearest Neighbor Filter and Joint Probability Data Association Filter to correlate target data at different moments.The algorithm can reduce the computational complexity while effectively dealing with data associations in complex situations.(3)For the state estimation in target tracking,in order to obtain accurate target state information,this paper establishes an Extended Kalman Filter model suitable for tracking targets in a 2D coordinate system,and corrects the state information of the current time target and predicts the state of motion of the target at the next moment.(4)Design and implement a trajectory manager to store the trajectory data of all tracked targets at different times and update the target state in the trajectory in real time.