Research On 3D Multi-object Detection And Tracking Technology Based On Lidar Sensor

As self-driving car travel becomes a hot topic of social concern,self-driving environment sensing technology is becoming more and more challenging.At present,lidar,vision camera,radar and other sensors have been widely used in the self-driving environment perception,but the detection effect of vision camera is not good in strong light or weak light environment,and the detection error of radar is large.The 3D multi-object detection based on lidar tracking technology can be in a strong or weak light environment for 3D object realizes the accurate detection and tracking.It is helpful to improve the safety performance of self-driving system.This study focuses on the three-dimensional multi-object detection and tracking based on lidar as follows:In order to solve the problem of abnormal points and noise points in lidar point cloud data,the point cloud data collected in practice are preprocessed.Spherical linear interpolation method is adopted to improve the motion compensation,and using plane fitting ground point to remove,through filtering and fusion,radius of voxel filter,filter method to point cloud data filtering process,and considering point cloud data has its inherent characteristics,a grid clustering method based on polar coordinate system is used to cluster the filtered point cloud data.Finally,combined with Hough line detection method,3D bounding box is used to accurately frame the object.For the problems of low accuracy and high computational cost of 3D multi-object detection based on lidar,a 3D multi-object detection algorithm based on PV(Point and Voxel)neural network model was proposed.The voxel-based method is used to select candidate regions,and then the 3D information of voxel and point cloud is integrated in an efficient way through layered feature extraction and set abstraction based on point cloud key points,which improves the accuracy of 3D multi-object detection in complex scenes.Finally,KITTI data set was used to train and test the model.Compared with Voxel R-CNN algorithm,the detection accuracy of equal difficulty improved by 0.26%,and that of difficult difficulty improved by 0.09%.For the problems of low association accuracy rate and association recall rate of3 D multi-object tracking based on lidar,and high system complexity,a fast and accurate real-time 3D multi-object tracking technology,PV3 DMOT,was proposed.The tracking method combining Mahalanobis distance and greedy algorithm is studied,and the best data association method is proposed.Based on this method,the3 D multi-object matching and tracking is realized.The 3D Kalman filter is used to predict and update the object state,and the best correlation method is proposed to realize the real-time tracking of objects.Finally,KITTI data set was used for testing.Compared with AB3 DMOT algorithm,its association accuracy rate increased by 1.16%and association recall rate increased by 3.44%,which effectively improved the accuracy of multi-object tracking in self-driving scenes.The virtual driving simulation platform based on CARLA and ROS is used to manage and control vehicles and pedestrians in the scene through the Traffic Manager module,and a custom Python API interface is used to connect the server and the client.Besides,the method in this study is used to conduct real-time detection and tracking of three-dimensional multi-object in the scene for special test scenes.This scheme is helpful to improve test efficiency and reduce test cost.A self-assembled experimental platform was built,and external parameter calibration of lidar,inertial measurement unit(IMU),radar and vision camera was realized.Finally,a real vehicle experiment was carried out with the proposed method in the campus,and the average accuracy of multi-object detection was 83.2%,and the tracking accuracy was 67.1%.