汽车工程 ›› 2024, Vol. 46 ›› Issue (6): 956-964.doi: 10.19562/j.chinasae.qcgc.2024.06.002

• • 上一篇    下一篇

驾驶场景下结合运动速度以及外观特征的多类多目标跟踪方法

王海1(),丁玉轩1,罗彤3,邱梦2,蔡英凤2,陈龙2   

  1. 1.江苏大学汽车与交通工程学院,镇江 212013
    2.江苏大学汽车工程研究院,镇江 212013
    3.江苏理工学院,常州 213001
  • 收稿日期:2023-11-23 修回日期:2024-01-04 出版日期:2024-06-25 发布日期:2024-06-19
  • 通讯作者: 王海 E-mail:wanghai1019@163.com
  • 基金资助:
    国家自然科学基金(52225212);江苏省重点研发项目(BE2020083-2)

A Multi-class Multi-target Tracking Algorithm Combining Motion Speed and Appearance Features in Driving Scenarios

Hai Wang1(),Yuxuan Ding1,Tong Luo3,Meng Qiu2,Yingfeng Cai2,Long Chen2   

  1. 1.School of Automotive and Traffic Engineering,Jiangsu University,Zhenjiang  212013
    2.Institute of Automotive Engineering,Jiangsu University,Zhenjiang  212013
    3.Jiangsu Institute of Technology,Changzhou  213001
  • Received:2023-11-23 Revised:2024-01-04 Online:2024-06-25 Published:2024-06-19
  • Contact: Hai Wang E-mail:wanghai1019@163.com

摘要:

基于相机传感器的多目标跟踪算法是自动驾驶的重要组成部分。驾驶场景下,基于交并比进行前后帧关联的方案一直存在大量的身份切换,此现象在对向来车以及自车转弯时更加明显。本文将目标在二维平面上的运动速度作为扩展匹配空间的变量,设计了基于目标速度变化的交并比算法:Velocity IoU,从而优化前后帧目标关联。同时,使用自监督的外观模型提取不同目标的外观特征编码。基于上述的运动模型以及外观模型,提出了一种互补的关联策略,最终实现驾驶场景下多类别多目标跟踪。在BDD100K上验证了该方法,对应mMOTA为45.2,mIDF1为55.2,IDs为8 793,优于大部分跟踪算法。

关键词: 自动驾驶, 多目标跟踪, 多类别

Abstract:

Multi-target tracking algorithms based on camera sensors are crucial to autonomous driving. In driving scenarios, traditional association schemes based on Intersection over Union(IoU) of front and back frames are subject to a great deal of ID switches, which is more pronounced in the case of opposing traffic and self-turning vehicles. In this paper, the target's motion speed in the 2D plane is taken as a variable to extend the matching space to design IoU based on the target's speed change: the Velocity IoU, so as to optimize the front and back frame target association method. Meanwhile, a self-supervised appearance model is used to extract the appearance features of different targets. Based on the above motion model as well as the appearance model, a complementary association strategy is proposed, which ultimately achieves multi-category multi-target tracking in driving scenarios. The method is validated on BDD100K, with corresponding mMOTA of 45.2, mIDF1 of 55.2, and IDs of 8 793, which outperforms most tracking algorithms.

Key words: autonomous driving, multi-target tracking, multi-category