摘要
行人重识别是近年来计算机视觉领域的热点问题,经过多年的发展,基于可见光图像的一般行人重识别技术已经趋近成熟.然而,目前的研究多基于一个相对理想的假设,即行人图像都是在光照充足的条件下拍摄的高分辨率图像.因此虽然大多数的研究都能取得较为满意的效果,但在实际环境中并不适用.多源数据行人重识别即利用多种行人信息进行行人匹配的问题.除了需要解决一般行人重识别所面临的问题外,多源数据行人重识别技术还需要解决不同类型行人信息与一般行人图片相互匹配时的差异问题,如低分辨率图像、红外图像、深度图像、文本信息和素描图像等.因此,与一般行人重识别方法相比,多源数据行人重识别研究更具实用性,同时也更具有挑战性.本文首先介绍了一般行人重识别的发展现状和所面临的问题,然后比较了多源数据行人重识别与一般行人重识别的区别,并根据不同数据类型总结了5类多源数据行人重识别问题,分别从方法、数据集两个方面对现有工作做了归纳和分析.与一般行人重识别技术相比,多源数据行人重识别的优点是可以充分利用各类数据学习跨模态和类型的特征转换.最后,本文讨论了多源数据行人重识别未来的发展.
Person re-identification(Re-ID)has been a popular and well-investigated topic in computer vision community.However,current researches have a relatively ideal assumption that person images are captured under a sufficient light condition and with high-resolution.Although most researches can achieve very exciting performances,they are not suitable for practical applications.Since practical conditions are a little complicated,and there are multiple sources to represent persons'appearance.In this paper,we focus on the multi-source person Re-ID,which refers to the problem of using multiple sources of data for person re-identification.Compared with general person Re-ID methods,multi-source person Re-ID researches are more practical,yet more challenging in reality.We need to face challenges caused by domain gap among different data sources,such as low-resolution images,infrared images,depth images,text information and sketch images.In this paper,we start with a brief introduction of general person Re-ID.The differences between general and multi-source person Re-ID are then compared.Five types of multi-source person Re-ID are further analyzed and summarized.From these discussions,it will become evident that several advantages exist in multi-source person Re-ID over general person Re-ID methods,as the former can make full use of data sources to learn cross-modality feature transformation.Finally,the future trends of multi-source person Re-ID are discussed.
作者
叶钰
王正
梁超
韩镇
陈军
胡瑞敏
YE Yu;WANG Zheng;LIANG Chao;HAN Zhen;CHEN Jun;HU Rui-Min(National Engineering Research Center for Multimedia Software,School of Computer Science,Wuhan University,Wuhan 430072,China;Hubei Key Laboratory of Multimedia and Network Communication Engineering,Wuhan University,Wuhan 430072,China;Collaborative Innovation Center of Geospatial Technology,Wuhan 430072,China;National Institute of Informatics,Tokyo 1638001,Japan)
出处
《自动化学报》
EI
CSCD
北大核心
2020年第9期1869-1884,共16页
Acta Automatica Sinica
基金
国家重点研发计划项目(2017YFC0803700)
国家自然科学基金青年项目(61801335,61876135)
湖北省自然科学基金群体项目(2018CFA024,2019CFB472,2018AAA062)资助。
关键词
多源数据行人重识别
跨模态
度量学习
特征模型
统一模态
Multi-source person re-identification(Re-ID)
cross-modality
metric learning
feature model
modality unifying
