基于激光雷达的3D实时车辆跟踪

doi:10.19562/j.chinasae.qcgc.2021.07.008

摘要/Abstract

摘要：

3D多目标跟踪算法是智能车辆感知算法的重要组成部分，现有跟踪算法多与检测算法耦合以提高精度，导致算法实时性不足。针对此问题，本文中提出一种基于激光雷达的3D实时车辆跟踪算法。首先，对于激光雷达检测结果杂波较少的工况，提出结构精简的双波门GNN关联算法，有效提升其关联速度及精度；其次，优化关联向量与关联距离，既保证了算法的普适性，又提升其跟踪精度；最后，针对3D目标运动情况使用3D IMM?KF算法解决了3D机动目标的跟踪问题。基于公开数据集KITTI，本文算法在获得266.1 FPS跟踪速度的前提下可实现81.55%的MOTA精度；基于自研无人车平台进行面对遮挡工况的验证，结果表明本算法具有良好的目标跟踪及关联性能。

关键词: 无人车, 激光雷达, 数据关联, 多目标跟踪

Abstract:

The 3D multi?object tracking algorithm is an essential part of the intelligent vehicle perception algorithm. The existing tracking algorithm is mostly coupled with the detection algorithm to improve the accuracy， resulting in insufficient real?time performance. To solve this problem， a 3D real?time vehicle tracking algorithm based on lidar is proposed. Firstly， for the working conditions with less clutter in the detection results of lidar， a double?validation gate GNN algorithm with a simple structure is proposed to effectively improve its correlation speed and accuracy； secondly， the correlation vector and correlation distance are optimized， which improves the tracking accuracy while ensuring the generality of the algorithm. Finally， the 3D IMM?KF algorithm is used to solve the tracking problem of 3D object with changing dynamics. The proposed algorithm achieves a MOTA accuracy of 81.55% at a tracking speed of 266.1 FPS according to the public data set KITTI. Based on the self?developed unmanned vehicle platform， the verification of facing occlusion conditions shows that the algorithm has good object tracking and correlation performance.

Key words: unmanned vehicles, lidar, data association, multi?object tracking

王海,李洋,蔡英凤,孙恺,陈龙. 基于激光雷达的3D实时车辆跟踪[J]. 汽车工程, 2021, 43(7): 1013-1021.

Hai Wang,Yang Li,Yingfeng Cai,Kai Sun,Long Chen. 3D Real⁃Time Vehicle Tracking Based on Lidar[J]. Automotive Engineering, 2021, 43(7): 1013-1021.

图/表 11

图 1

图 2

图 3

表 1

图 4

表 2

表 3

表 4

表 5

图5

图 6

参考文献 21

1	SVENSSON L， SVENSSON D， GUERRIERO M， et al. Set JPDA filter for multitarget tracking［J］. IEEE Transactions on Signal Processing， 2011， 59（10）：4677-4691.
2	BLACKMAN S. Multiple hypothesis tracking for multiple target tracking［J］. IEEE Aerospace and Electronic Systems Magazine， 2009，19（1）： 5-18.
3	KONSTANTINOVA P， UDVAREV A， SEMERDJIEV T. A study of a target tracking algorithm using global nearest neighbor approach［C］. Proceedings of the International Conference on Computer Systems and Technologies，2003：290-295.
4	FARUQI F A， DAVIS R C. Kalman filter design for target tracking［J］. IEEE Transactions on Aerospace and Electronic Systems， 1980， 4：500-508.
5	WAN E A， MERWE RVAN DER. The unscented Kalman filter for nonlinear estimation［C］. Proceedings of the IEEE 2000 Adaptive Systems for Signal Processing， Communications， and Control Symposium （Cat. No. 00EX373），2000： 153-158.
6	SHERRAH J， RISTIC B， REDDING N J. Particle filter to track multiple people for visual surveillance［J］. IET Computer Vision， 2011， 5（4）： 192-200.
7	GENOVESE A F. The interacting multiple model algorithm for accurate state estimation of maneuvering targets［J］. Johns Hopkins APL Technical Digest， 2001， 22（4）： 614-623.
8	BASER E， BALASUBRAMANIAN V， BHATTACHARYYA P， et al. Fantrack： 3d multi⁃object tracking with feature association network［C］. 2019 IEEE Intelligent Vehicles Symposium （IV）. IEEE， 2019： 1426-1433.
9	ZHANG W， ZHOU H， SUN S， et al. Robust multi⁃modality multi⁃object tracking［C］. Proceedings of the IEEE International Conference on Computer Vision， 2019： 2365-2374.
10	FROSSARD D， URTASUN R. End⁃to⁃end learning of multi⁃sensor 3d tracking by detection［C］. 2018 IEEE International Conference on Robotics and Automation （ICRA）. IEEE， 2018： 635-642.
11	徐国艳，牛欢，郭宸阳，等. 基于三维激光点云的目标识别与跟踪研究［J］. 汽车工程， 2020， 42（1）： 38-46.
	XU Guoyan， NIU Huan， GUO Chenyang， et al. Research on target recognition and tracking based on 3D laser point cloud［J］. Automotive Engineering， 2020， 42（1）： 38-46.
12	谢德胜，徐友春，王任栋，等. 基于三维激光雷达的无人车障碍物检测与跟踪［J］. 汽车工程， 2018， 40（8）： 952-959.
	XIE Desheng， XU Youchun， WANG Rendong， et al. Obstacle detection and tracking for unmanned vehicles based on 3D laser radar［J］. Automotive Engineering， 2018， 40（8）： 952-959.
13	XIANG Y， ALAHI A， SAVARESE S. Learning to track： online multi⁃object tracking by decision making［C］.Proceedings of the IEEE International Conference On Computer Vision， 2015： 4705-4713.
14	娄新雨，王海，蔡英凤，等. 采用64线激光雷达的实时道路障碍物检测与分类算法的研究［J］. 汽车工程， 2019， 41（7）： 779-784.
	LOU Xinyu， WANG Hai， CAI Yingfeng， et al. A research on an algorithm for real⁃time detection and classification of road obstacle by using 64⁃line lidar［J］. Automotive Engineering， 2019， 41（7）： 779-754.
15	SHI S， GUO C， JIANG L， et al. Pv⁃rcnn： point⁃voxel feature set abstraction for 3d object detection［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2020： 10529-10538.
16	SHI S， WANG X， LI H. Pointrcnn： 3D object proposal generation and detection from point cloud［C］. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition， 2019： 770-779.
17	RACHMAN A S A. 3D⁃lidar multi object tracking for autonomous driving［D］. Delft： Delft University of Technology， 2017.
18	文龙. 基于车载激光雷达的无人驾驶车辆障碍物检测与跟踪技术研究［D］. 长春：吉林大学， 2020.
	WEN Long. Research on obstacle detection and tracking technology of autonomous vehicle based on lidar［D］. Changchun： Jilin University， 2020.
19	CHIU H， PRIOLETTI A， LI J， et al. Probabilistic 3D multi⁃object tracking for autonomous driving［J］. arXiv Preprint， arXiv：， 2020.
20	BERNARDIN K， STIEFELHAGEN R. Evaluating multiple object tracking performance： the CLEAR MOT metrics［J］. EURASIP Journal on Image and Video Processing， 2008， 2008： 1-10.
21	LI Y， HUANG C， NEVATIA R. Learning to associate： hybridboosted multi⁃target tracker for crowded scene［C］. 2009 IEEE Conference on Computer Vision and Pattern Recognition. IEEE， 2009： 2953-2960.

算法名	指标
算法名	MOTA	MOTP	IDS	FRAG	FPS
mmMOT^［8］	74.07%	78.16%	10	125	4.8
FANTrack^［9］	74.30%	75.24%	35	202	23.1
本文方法	81.55%	77.11%	77	132	266.1

关联算法	指标
关联算法	MOTA	MOTP	IDS	FRAG
MHT	27.46%	78.70%	2312	2451
JPDA	68.07%	75.10%	82	191
GNN	81.55%	77.11%	77	132

匹配设置	指标
匹配设置	MOTA	MOTP	IDS	FRAG
欧氏距离（不包含朝向角θ）	79.65%	77.46%	132	195
马氏距离（只包含位置信息）	81.38%	77.19%	66	123
马氏距离（包含朝向角θ）	81.12%	77.19%	66	134
马氏距离（不包含朝向角θ）	81.55%	77.11%	77	132

匹配设置	指标
匹配设置	MOTA	MOTP	IDS	FRAG
f_s=3、f_min=3 f_d=2、f_max=2	80.44%	77.38%	16	86
f_s=5、f_min=3 f_d=3、f_max=2	80.16%	77.46%	17	98
f_s=3、f_min=2 f_d=5、f_max=3	80.36%	76.96%	66	108
f_s=3、f_min=2 f_d=3、f_max=2	81.55%	77.11%	77	132

波门算法	指标
波门算法	MOTA	MOTP	IDS	FRAG	FPS
单波门	78.63%	75.43%	14	104	200.2
双波门	81.55%	77.11%	77	132	266.1