东北大学学报(自然科学版) ›› 2020, Vol. 41 ›› Issue (12): 1673-1680.DOI: 10.12068/j.issn.1005-3026.2020.12.001

• 信息与控制 •    下一篇

基于生成对抗网络的多目标行人跟踪算法

魏颖1, 徐楚翘1, 刁兆富1, 李伯群2   

  1. (1. 东北大学 信息科学与工程学院, 辽宁 沈阳110819; 2. 辽宁科技大学 电子与信息工程学院, 辽宁 鞍山114051)
  • 收稿日期:2020-03-30 修回日期:2020-03-30 出版日期:2020-12-15 发布日期:2020-12-22
  • 通讯作者: 魏颖
  • 作者简介:魏颖(1968-),女,辽宁本溪人,东北大学教授,博士生导师; 李伯群(1970-),男,辽宁鞍山人,辽宁科技大学教授.
  • 基金资助:
    国家自然科学基金资助项目(61871106); 辽宁省重点研发项目(2020JH2/10100029).

A Multi-target Pedestrian Tracking Algorithm Based on Generated Adversarial Network

WEI Ying1, XU Chu-qiao1, DIAO Zhao-fu1, LI Bo-qun2   

  1. 1. School of Information Science & Engineering, Northeastern University, Shenyang 110819, China; 2. School of Electronic and Information Engineering, University of Science and Technology Liaoning, Anshan 114051, China.
  • Received:2020-03-30 Revised:2020-03-30 Online:2020-12-15 Published:2020-12-22
  • Contact: LI Bo-qun
  • About author:-
  • Supported by:
    -

摘要: 多目标跟踪领域中,在背景复杂、目标遮挡、目标尺度和姿态变换等情况下,容易出现目标丢失、身份交换和跳变等问题.针对这些问题,提出了一种基于检测的多目标跟踪算法,使用改进的YOLO人体人脸关联算法,对当前帧待检目标进行分类和位置检测,使用生成对抗网络构建特征提取模型,学习目标的主要特征以及细微特征,再运用生成对抗网络生成多目标的运动轨迹,最终融和目标的运动信息和外观信息,得到跟踪目标的最优匹配.在MOT16数据集下的实验结果表明,提出的多目标跟踪算法具有较高的精确度和鲁棒性,对比目前身份交换和跳变最少的算法,跳变的次数少了65%,准确度提高了0.25%.

关键词: 多目标跟踪, 生成对抗网络, 目标检测, 路径预测, 特征融合

Abstract: In the field of multi-target tracking, the problems of target loss, identity exchange and switch are easy to occur under the conditions of complex background, target occlusion, target scale and attitude change. To solve these problems, a multi-target tracking algorithm was proposed based on detection. A human body and face association algorithm based on YOLO was used to classify and detect the position of the current frame objects, and the feature extraction model based on generative adversarial network was proposed to learn the main features and subtle features of the objects. Then the generative adversarial network was used to generate the motion trajectories of multiple targets, and finally the target’s motion and appearance information were merged to obtain the optimal matching of the target tracking results. The experimental results show that the multi-target tracking algorithm proposed is both accurate and robust. Compared with the current algorithms with the least ID switch, the number of ID switch is 65% less and the accuracy is improved by 0.25%.

Key words: multi-target tracking, generative adversarial network, object detection, path prediction, feature fusion

中图分类号: