基于改进DeepSORT和FastReID的室内多目标人员跨镜识别与跟踪  

Indoor multi-object personnel recognition and tracking across camera based on optimized DeepSORT and FastReID

作　　者：赵安新[1] 杨金桥 杨浩波 史新国 付文旭 刘帅[1] 王伟峰 ZHAO Anxin;YANG Jinqiao;YANG Haobo;SHI Xinguo;FU Wenxu;LIU Shuai;WANG Weifeng(College of Communication and Information Engineering,Xi’an University of Science and Technology,Xi’an 710054,China;Shaanxi Zhengtong Coal Industry Co.,Changwu 713600,China;Shandong Boxuan Mineral Resources Technology Development Co.,Jining 272073,China;College of Safety Science and Engineering,Xi’an University of Science and Technology,Xi’an 710054,China)

机构地区：[1]西安科技大学通信与信息工程学院,陕西西安710054 [2]陕西正通煤业有限责任公司,陕西长武713600 [3]山东博选矿物资源技术开发有限公司,山东济宁272073 [4]西安科技大学安全科学与工程学院,陕西西安710054

出　　处：《西安科技大学学报》2023年第3期622-630,共9页Journal of Xi’an University of Science and Technology

基　　金：陕西省重点研发计划(2022GY-152,2023-YBGY-025);国家重点研发计划(2021YFE0105000),国家自然科学基金(52074213)。

摘　　要：为解决利用视频分析实现目标检测、目标跟踪和行人重识别存在的诸如目标框重框、遮挡等问题,提出了一种基于改进DeepSORT和FastReID的室内多目标人员跨镜追踪的方法。该方法使用YOLOv5s进行人员检测、DeepSORT进行人员跟踪、FastReID进行人员重识别。采用EIOU-NMS算法解决了YOLOv5s人员检测过程中出现的重框问题;在FastReID的特征提取网络中引入了NEUFA注意力机制,并使用优化后的FastReID的特征提取网络替换了DeepSORT原有特征提取网络,降低了DeepSORT跟踪过程中由于遮挡导致的ID跳变的次数;结合注意力机制、人员动态图像库与运动估计的方法减少了人员识别过程中因遮挡导致的人员无法识别和错误识别的次数。结果表明:EIOU-NMS算法将人员检测准确率提升了0.8%,召回率提升了0.4%;替换了特征网络后的DeepSORT将人员跟踪中的ID跳变次数降低了38.46%;结合注意力机制、人员动态图像库和运动估计方法之后,None的次数减少了79.8%,误识别次数减少了91.2%。研究结果能够提升人员识别与跟踪的准确性,降低遮挡对人员跟踪和识别带来的影响。To solve the problems such as target frame re-framing and occlusion that exist in target detection,target tracking,and person re-identification using video analytics,a method based on optimized DeepSORT and FastReID for indoor multi-target person cross-camera tracking is proposed.YOLOv5s is adopted for person detection,DeepSORT for person tracking,and FastReID for person re-identification.The EIOU-NMS algorithm is used to solve the re-framing problem that occurs during YOLOv5s person detection;the NEUFA attention mechanism is introduced in the feature extraction network of FastReID,and the optimized feature extraction network of FastReID is chosen to replace the original feature extraction network of DeepSORT,which reduces the number of ID jumps due to occlusion in the DeepSORT tracking process.The combination of attention mechanism,dynamic image gallery,and motion estimation reduces the number of unidentified and incorrectly identified persons due to occlusion during DeepSORT tracking.The results show that the EIOU-NMS algorithm improves the person detection accuracy by 0.8%and the recall rate by 0.4%;the replacement of the feature networks in the DeepSORT reduces the number of ID jumps in person tracking by 38.46%;combining the attention mechanism,person dynamic image library and motion estimation methods reduces the number of None by 79.8%and the number of false identifications by 91.2%.The research results can improve the accuracy of person identification and tracking,thus reducing the impact of occlusion on person tracking and identification.

关 键 词：多目标跟踪 局部遮挡 目标检测 行人重识别 人员图像库 运动估计 

分 类 号：TN911.73[电子电信—通信与信息系统]