汽车工程 ›› 2021, Vol. 43 ›› Issue (11): 1587-1593.doi: 10.19562/j.chinasae.qcgc.2021.11.003

• • 上一篇    下一篇

基于自查询的车载多目标跟踪算法研究

陈龙1,朱程铮1,蔡英凤1(),王海2,李祎承1   

  1. 1.江苏大学汽车工程研究院,镇江  212013
    2.江苏大学汽车与交通工程学院,镇江  212013
  • 收稿日期:2021-07-19 修回日期:2021-08-16 出版日期:2021-11-25 发布日期:2021-11-22
  • 通讯作者: 蔡英凤 E-mail:caicaixiao0304@126. com

Research on Vehicle Multi-Target Environment Aware Tracking Algorithm Based on Self-Query

Long Chen1,Chengzheng Zhu1,Yingfeng Cai1(),Hai Wang2,Yicheng Li1   

  1. 1.Institude of Automotive Engineering,Jiangsu University,Zhenjiang  212013
    2.Institude of Automotive and Transportation Engineering,Jiangsu University,Zhenjiang  212013
  • Received:2021-07-19 Revised:2021-08-16 Online:2021-11-25 Published:2021-11-22
  • Contact: Yingfeng Cai E-mail:caicaixiao0304@126. com

摘要:

为兼顾跟踪性能(即MOTA、MOTP、IDSW等指标)与跟踪速率,尤其是解决视频多目标跟踪的后处理较复杂的问题,提出了一种基于自回归查询机制的多目标跟踪方法,并基于MOT20数据集进行训练和验证,验证表明,每帧图片推理用时约44 ms,多目标跟踪准确度达58.9%。将该模型集成到智能车ROS平台进行测试,结果表明,所提算法能实现复杂交通场景下的多目标实时跟踪,算法具有很好的实际应用价值。

关键词: 车载多目标跟踪, 深度学习, 自查询跟踪

Abstract:

In order to balance the tracking performance (i.e. the indicators of MOTA, MOTP and IDSW etc.) and tracking speed, especially, to solve the complexity of the post processing for video multi-target tracking, a multi-target tracking method based on autoregressive query mechanism is proposed, with training and verification conducted. The results of verification show that the inference of each frame of picture takes about 44 ms, and the accuracy of multi-target tracking reaches 58.9%. The model is integrated into the ROS platform of intelligent vehicle for testing and the results of test indicate that the algorithm proposed can achieve multi-target real-time tracking in complex traffic scenes, and the algorithm has good practical application value.

Key words: onboard multi-object tracking, deep learning, self-query tracking