Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (8): 96-102.DOI: 10.3778/j.issn.1002-8331.2001-0286

Previous Articles     Next Articles

Multiple Target Tracking Using Hierarchical Data Association Based on Deep Learning

NIU Tong, QING Linbo, XU Shengyu, SU Jie   

  1. College of Electronics and Information, Sichuan University, Chengdu 610065, China
  • Online:2021-04-15 Published:2021-04-23

基于深度学习的分层关联多行人跟踪

牛通,卿粼波,许盛宇,苏婕   

  1. 四川大学 电子信息学院,成都 610065

Abstract:

Due to the uncertainty of data association between continuous frames and the insufficient discrimination of extracted appearance features, multi-pedestrian tracking is susceptible to target appearance changes, motion state changes, interference from similar objects, and the disappearance and reappearance of targets, and the problem of track ID switch occurs, which limits the performance of action detection and posture recognition based on pedestrian tracking. In order to improve the reliability of data association to reduce track ID transformation, a hierarchical data association based on track confidence is proposed. Meanwhile, the features of process layer are integrated to improve the discrimination of appearance features used for data association. Verification on the open MOT16 test data set shows that the algorithm in this paper can not only guarantee the tracking accuracy and precision, but also effectively reduce the number of track ID transformation and improve the tracking performance.

Key words: multi-pedestrian tracking, data association, appearance feature, ID switch

摘要:

由于连续帧之间数据关联的不确定性和所提取外观特征的鉴别力不足,多目标跟踪容易受目标外观变化、运动状态变化、相似目标以及目标消失再出现等干扰因素的影响,出现轨迹ID变换的问题,从而限制基于轨迹分析的行为检测、姿态识别等研究的性能。为了提高数据关联的可靠性从而减少轨迹ID变换,提出了一种基于轨迹置信度的分层数据关联方式。同时,为了提高用于数据关联的外观特征的鉴别力,提出融合了过程层的特征提取网络。在公开的MOT16测试数据集上验证表明,该算法在保证跟踪准确度和精确度的同时,有效降低了轨迹ID变换的次数,提高了跟踪性能。

关键词: 多行人跟踪, 数据关联, 外观特征, ID变换