Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (10): 145-152.DOI: 10.3778/j.issn.1002-8331.2011-0050

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Pedestrian Multi-Object Tracking Algorithm of Anchor-Free Detection

SHAN Zhaochen, HUANG Dandan, GENG Zhenye, LIU Zhi   

  1. 1.School of Electrical and Information Engineering, Changchun University of Science and Technology, Changchun 130022, China
    2.National and Local Joint Engineering Research Center for Space Optoelectronics Technology, Changchun 130022, China
  • Online:2022-05-15 Published:2022-05-15



  1. 1.长春理工大学 电子信息工程学院,长春 130022
    2.空间光电技术国家地方联合工程研究中心,长春 130022

Abstract: A multi-object tracking algorithm based on anchor-free detection is proposed to solve the problems of incorrect data correlation and poor tracking real-time due to detector omission and frequent occlusion in complex environment. The algorithm uses the method of predicting the thermal diagram of the target center to realize the target detection and location, and improves the problem of missing detection caused by the ambiguity of anchor frame regression. At the same time, the depth apparent feature extraction branch is embedded in the detection model, and the multi-task network of joint detection and tracking is built to improve the real-time performance. In order to solve the problem of data association error and track loss caused by occlusion of pedestrians in the tracking stage, a similarity measurement algorithm based on weighted multi-feature fusion is proposed to evaluate the matching degree of detection and tracking based on a variety of key features to significantly improve the accuracy of data association. A survival based tracking status update method is proposed to effectively recover lost track and improve tracking robustness. The tracking performance is tested on the MOT dataset. The experimental results show that the algorithm can effectively deal with occlusion and realize long-term stable tracking, taking into account both real-time and accuracy.

Key words: multi-object tracking, depth appearance characteristics, multi-feature fusion, data association

摘要: 针对复杂环境下行人目标因检测器漏检和频繁遮挡而导致的数据关联不正确、跟踪实时性差的问题,提出了一种基于免锚检测的多目标跟踪算法。算法采用预测目标中心点热力图的方法实现目标检测定位,改善了因锚点框回归歧义所导致的漏检问题。同时在检测模型中嵌入深度表观特征提取分支,构建联合检测与跟踪的多任务网络用于提升实时性。为解决跟踪阶段行人因遮挡而引起的数据关联错误和轨迹丢失问题,提出加权多特征融合的相似性度量算法,综合多种关键特征评估检测与轨迹的匹配程度,显著提升数据关联正确性;提出基于存活期的跟踪状态更新方法,有效找回丢失轨迹,提升跟踪鲁棒性。在MOT数据集上对跟踪性能进行测试,实验结果表明,算法能够有效应对遮挡,并实现长时间稳定跟踪,兼顾了实时性与准确性。

关键词: 多目标跟踪, 深度表观特征, 多特征融合, 数据关联