智能视频监控关键技术:行人再识别研究综述  

Key technology for intelligent video surveillance:a review of person re-identification

作　　者：赵才荣[1] 齐鼎 窦曙光 涂远鹏 孙添力 柏松 蒋忻洋 白翔 苗夺谦[1] Cairong ZHAO;Ding QI;Shuguang DOU;Yuanpeng TU;Tianli SUN;Song BAI;Xinyang JIANG;Xiang BAI;Duoqian MIAO(School of Electronic and Information Engineering,Tongji University,Shanghai 201804,China;School of Electronic Information and Communications,Huazhong University of Science and Technology,Wuhan 430074,China;Microsoft Research Asia(Shanghai),Shanghai 200232,China)

机构地区：[1]同济大学电子与信息工程学院,上海201804 [2]华中科技大学电子信息与通信学院,武汉430074 [3]微软亚洲研究院(上海),上海200232

出　　处：《中国科学：信息科学》2021年第12期1979-2015,共37页Scientia Sinica(Informationis)

基　　金：国家自然科学基金(批准号:62076184,61673299,61976160,62076182);上海科技创新行动计划(批准号:20511100700);上海市级科技重大专项——人工智能基础理论与关键核心技术(批准号:2021SHZDZX0100);中央高校基本科研业务费专项资助项目。

摘　　要：行人再识别(person re-identification,ReID)旨在解决跨摄像头跨场景下目标行人的关联与匹配,作为智能视频监控系统的关键环节,对维护社会公共秩序具有重大作用.为了深入了解行人再识别研究现状和加速推进国内行人再识别相关研究及技术落地,本文对该领域国家自然科学基金申报数量、资助力度以及地理分布情况进行统计,并针对近年来发表在国际顶级会议和期刊上的行人再识别研究进行全面梳理.具体地,首先阐述一个标准行人再识别算法流程,并总结其中3个关键技术:表征学习、度量学习和重排序优化.随后,列举了实际开放场景中面临的主要难点与挑战,并据此概括了7种开放行人再识别任务:遮挡、无监督、半监督、跨模态、场景行人搜索、对抗鲁棒和快速检索.此外,本文整理了标准行人再识别和开放行人再识别的代表性数据集,并且对一些代表性行人再识别算法进行比较.最后本文对行人再识别的未来发展趋势进行展望.Person re-identification(person ReID)aims to solve the association and matching of target person across cameras and scenes,which is a key link of intelligent video surveillance systems and plays a significant role in maintaining social public order.In order to understand the development status of person ReID technology and accelerate the implementation of person ReID research and applications,this paper provides statistics on the number of applications,funding intensity and geographic distribution of NSFC in this field,as well as a comprehensive review of person ReID research published in top international conferences and journals in the past decade.The paper starts with a standard person ReID algorithm process and details three key techniques:representation learning,metric learning,and re-ranking.Then,the main challenges faced in practical open scenarios are listed,and seven open person ReID tasks are outlined accordingly:occlusion,unsupervised,semisupervised,cross-modal,end-to-end search,adversarial robustness and fast retrieval.In addition,representative datasets of standard person ReID and open person ReID are collated and some representative algorithms are compared.Finally,this paper provides an outlook on the future trends of person ReID.

关 键 词：行人再识别 智能视频分析 深度学习 表征学习 度量学习 

分 类 号：TN948.6[电子电信—信号与信息处理]