融合二维姿态信息的相似多目标跟踪  

Similar multi-target tracking algorithm combining two-dimensional pose information

作　　者：雷景生 李誉坤 杨忠光 LEI Jing-sheng;LI Yu-kun;YANG Zhong-guang(College of Computer Science and Technology,Shanghai University of Electric Power,Shanghai 200090,China)

机构地区：[1]上海电力大学计算机科学与技术学院,上海200090

出　　处：《计算机工程与设计》2020年第10期2969-2976,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(61672337)。

摘　　要：针对传统机器视觉技术在包含相似移动目标的监控视频中的跟踪缺陷,提出一种多目标跟踪框架,解决因目标轨迹丢失及身份变换导致的无法对目标进行准确实时跟踪的问题。采用YOLO v3算法作为目标检测器,用OpenPose算法提取每一帧画面中移动目标的姿态信息。利用Deep SORT跟踪算法结合二维姿态信息完成相邻帧之间的目标匹配,实现存在相似目标工作场景下的相似多目标跟踪。实验结果表明,所提方法能够有效提高相关场景下多目标的识别率与跟踪的准确性,在对目标发生遮挡时算法鲁棒性方面有明显改进。Aiming at the tracking defects of traditional machine vision technology for surveillance videos containing similar moving targets,a multi-target tracking framework was proposed to solve the problem of inaccurate real-time tracking of targets due to target trajectory loss and identity change.The YOLO v3 algorithm was used as the target detector,and OpenPose algorithm was used to compute the human posture information in each frame.The Deep SORT tracking algorithm was combined with the 2D pose information to complete the target matching between adjacent frames,so as to achieve similar multi-target tracking in the similar target working scene.Experimental results show that the proposed method can effectively improve the recognition rate and tracking accuracy of multi-targets in related scenes,and significantly improve the robustness of the algorithm when targets are occluded.

关 键 词：视频摘要 相似多目标跟踪 运动目标检测 人体姿态 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]