Research On Target Recognition And Tracking System Based On Laser Radar And Camera Information Fusio

With the booming development of smart car industry in recent years,unmanned driving technology has received extensive attention and research.Target recognition and tracking technology is an important part of environment perception technology in unmanned systems,and whether it has real-time accurate target recognition and tracking capability will not only affect the result of environment perception,but also affect the path planning and obstacle judgment made by the decision module.The camera can obtain rich target pixel information,but it is difficult to obtain target depth information;lidar can obtain accurate object coordinate information,but the target point cloud classification ability is poor.Therefore,this paper selects lidar and camera information fusion to build a target identification and tracking system.The research content of this paper is divided into two parts: lidar and camera information fusion target identification and multi-target tracking.In the lidar and camera information fusion target recognition,the camera part first uses the YOLOv4 deep learning algorithm to recognize the target and ensure the speed and accuracy of target recognition within the range of the computing power of the IPC.The lidar part first divides the point cloud into regions of interest and voxel filtering processing to effectively reduce the number of point clouds;for the problem of wall point clouds being misidentified as ground point clouds,a segmentation algorithm is proposed to determine the plane distance and vertical distance threshold of point clouds to effectively identify ground point clouds;for the sparse problem of distant point clouds,a clustering algorithm is proposed with a clustering radius that adaptively changes with the target distance to effectively cluster distant target point clouds.effectively clusters distant target point clouds.Finally,the information is fused using point cloud and image target-level fusion strategy to obtain target spatial coordinates and semantic information.In the multi-target tracking part,the trajectory of the target is predicted using the traceless Kalman filter,and an improved Hungarian algorithm based on the breadth-first search strategy is proposed for the problems of long computation time and target loss when encountering occluded objects in the traditional Hungarian algorithm,which not only reduces the time required for target identification and tracking data matching,but also solves the problem of target loss due to target occlusion,and improves the target tracking speed and accuracy.In this paper,the feasibility of the system is verified on the KITTI dataset and the intelligent driving platform,respectively.The results demonstrate that the improved ground segmentation algorithm can better identify ground point clouds;the improved clustering algorithm can more effectively cluster long-range target point clouds;the maximum tracking error of the improved tracking algorithm is 0.0721 m and 0.0558 m for the target spatial position in X and Y directions,respectively,with a running time of 80 ms,which has a faster running speed and higher target tracking accuracy.