基于激光雷达的三维多目标检测与跟踪

3D multi-target detection and tracking based on LiDAR

在自动驾驶应用中,为提高基于激光雷达的三维(3D)多目标跟踪精确度,降低计算时间成本与系统复杂性,提出了一种快速准确的实时3D多目标跟踪技术-PV3DMOT。该方法先结合体素和点云进行全面特征学习,以实现快速准确的3D目标检测,然后用3D卡尔曼滤波器进行目标状态预测与更新,并利用马氏距离与贪心算法完成数据关联,最终实现高效的3D多目标实时跟踪。经自动驾驶KITTI数据集实验测试,该方法的检测结果在中等与困难类别中的准确率相比Voxel R-CNN算法提升了0.2%,跟踪结果相比AB3DMOT算法,其关联准确率提升了1.16%,联合召回率提升了3.44%,能有效提高智能驾驶中3D多目标检测跟踪的精确度。

In order to improve the accuracy of 3D multi-target tracking based on LiDAR and reduce the cost of calculation time and system complexity in application of auto-driving,a fast and accurate real-time 3D multi-target tracking technology,PV3DMOT,is proposed.In this method,a comprehensive feature learning is carried out by combining the voxel and point cloud to detect the 3D target quickly and accurately.Then,the 3D Kalman filter is used to predict and update the object state,and the Mahalanobis distance and greedy algorithm are used to complete the data association,so as to realize the real-time 3D multi-target tracking.After testing on autonomous driving KITTI dataset,it is found that the accuracy of experimental detection results in moderate and difficult categories is 0.2%higher than that of Voxel R-CNN algorithm,the association accuracy of tracking results is 1.16%higher than that of AB3DMOT algorithm,and the association recall rate is 3.44%higher.Therefore,this method can effectively improve the accuracy of 3D multi-target detection and tracking in intelligent driving scenes.