基于任务联合的三维车辆检测与跟踪集成算法

3D Vehicle Detection and Tracking Integration Algorithm Based on Task Collaboration

在常规的目标检测及跟踪算法中,检测器与跟踪器按照流水线的方式级联工作,目标检测模块的漏检会降低目标跟踪模块的性能。针对该问题,目前主要方法是通过构建检测与跟踪联合算法使2种任务互为促进,进一步提高检测与跟踪精度。基于此,提出了一种车辆目标检测与跟踪任务联合框架3D Tracktor++。利用前一帧目标检测结果生成携带有身份信息的先验候选区域,引导检测器在目标出现概率较大的区域进行目标框回归并直接输出轨迹编号。为了避免对新进入场景的目标漏检且弥补先验候选区域的偏差,增加了候选区域补充模块,集成后的先验候选区域与补充候选区域通过感兴趣区域(RoI)池化、目标回归、轨迹检查等步骤,输出车辆目标位置及其轨迹编号。在KITTI数据集上的试验结果表明,与单任务目标检测方法VoxelRCNN相比,所提出的任务联合方法平均检测精度AP3D更高,其中在中等难度样本上AP3D提高2.75%;与三维多目标追踪的基准方法(AB3DMOT)相比,多目标跟踪精确率(MOTP)与平均多目标跟踪精确率(AMOTP)分别提升3.59%与0.77%。与常规“先检测后跟踪”算法相比,论文提出的联合算法具有合理性,有效提高了检测与跟踪精度。

For the conventional object detection and tracking algorithm,the detector and the tracker work in a pipelined way.The missed detection of the target detection module will reduce the performance of the target tracking module.To solve this problem,a joint detection and tracking algorithm can be constructed to make the two tasks mutually promote each other,so as to further improve the detection and tracking accuracy.A joint task framework(3D Tracktor++)for vehicle object detection and tracking is proposed in this paper.The prior candidate region with identity information is generated by object detection results of the previous frame,the detector is guided to carry out object box regression in the region with a high probability of object occurrence,and further the track number is directly output.In order to avoid missed detection of new object entering the scene and make up for the deviation of prior candidate regions,a supplementary module of candidate regions is added.The prior candidate region and supplementary candidate region are integrated,and vehicle object positions and track numbers are output through RoI pooling,target regression,track inspection and other steps.The experimental results on KITTI data set show that compared with the single task object detection method(VoxelRCNN),the average detection precision(AP3D)of the proposed joint task method is higher,and the AP3D is improved by 2.75%for the medium difficulty samples.Compared with the basic method(AB3DMOT)for 3D multi-object tracking,MOTP and AMOTP have increased by 3.59%and 0.77%,respectively.Compared with the conventional"detect before tracking"algorithm,the proposed algorithm is reasonable,and improves the detection and tracking accuracy effectively.