Research On Vehicle Recognition And Tracking Based On The Information Fusion Of Lidar And Camera

Autonomous driving technology has received extensive attention and research in recent years.Environmental perception technology is at the forefront of the automatic driving system,and the recognition and tracking of vehicle targets on the road is an important part of the perception technology.With real-time and accurate vehicle target recognition and tracking capabilities,it helps the decision-making module to make correct and reasonable planning and judgments.lidar can provide relatively accurate point cloud information,but its ability to classify objects is low.The camera can obtain rich pixel information of the vehicle target,but the camera is a two-dimensional sensor and cannot provide the depth information of the target.Therefore,this paper fuses the information of lidar and camera to complete the recognition and tracking of vehicle targets.The scheme of this paper is divided into two parts,one is vehicle target recognition based on information fusion technology,and the other is vehicle target tracking.In the vehicle target recognition part,the original point cloud of the lidar is first filtered to reduce the density of the point cloud.Then,the constraint of the ground plane normal vector is introduced into the ground recognition algorithm to avoid the influence of the vertical surface of tall buildings in the surrounding environment.The point cloud near the lidar is denser than the point cloud far away.Based on this case,a clustering algorithm that clustering radius changes adaptively with point cloud distance is designed.The recognition of vehicle targets in the image adopts the lightweight Yolo Fastest-XL deep learning algorithm,which can ensure recognition accuracy in low computing power scenarios.Finally,the target-level fusion strategy is used for information fusion.Aiming at the problem that the vertical scanning range of the lidar is smaller than the camera’s field of view,a new information fusion index is proposed to improve the accuracy of information fusion.After the vehicle target information is obtained,the unscented Kalman filter is used to track the vehicle target in the vehicle target tracking part.The same target is marked with a unique ID.In this paper,the joint probability data association algorithm is used for data association.Aiming at the problem that the calculation time of the algorithm increases exponentially with the increase of the number of targets,this paper adopts the distance weighting method to avoid the division and calculation of feasible events and improve the calculation efficiency.Simultaneously,the improved algorithm is combined with the nearest neighbor algorithm to enable it to deal with scenes where the number of targets changes.To avoid the interference of noise points and the temporary disappearance of vehicle targets due to occlusion,the M/N test strategy is introduced in target management to buffer the confirmation and disappearance of targets and improve the robustness of the algorithm.This paper uses the KITTI data set and the campus intelligent driving platform for algorithm verification.The algorithm running platform is Ubuntu16.04+ROS Kinetic.Experimental results prove that the improved point cloud clustering algorithm has high distance adaptability and clustering accuracy,and can meet the clustering requirements of vehicle targets at different distances.The proposed information fusion algorithm can effectively solve the phenomenon that the point cloud 3D frame size is too small and the information cannot be fused.The improved multi-target tracking algorithm has higher tracking accuracy,and the maximum tracking error is 0.5117 meters.The running time of the vehicle target recognition and tracking algorithm is 70 ms,which has certain real-time performance.