非可控环境行人再识别综述  

Overview of person re-identification in unconstrained environments

作　　者：冯展祥 朱荣 王玉娟 赖剑煌[1] FENG Zhanxiang;ZHU Rong;WANG Yujuan;LAI Jianhuang(School of Data and Computer Science,Sun Yat-sen University,Guangzhou 510006,China;Xinhua College of Sun Yat-sen University,Guangzhou 510006,China)

机构地区：[1]中山大学数据科学与计算机学院,广东广州510006 [2]中山大学新华学院,广东广州510006

出　　处：《中山大学学报（自然科学版）》2020年第3期1-11,共11页Acta Scientiarum Naturalium Universitatis Sunyatseni

基　　金：国家自然科学基金(61902444);国家自然科学基金(U1611461);中山大学新华学院第四批博士导研计划。

摘　　要：最近几年,随着深度学习理论和行人再识别方法的发展和成熟,行人再识别技术取得了很大的突破,在理想条件下取得了很高的识别精度。但是,当前行人再识别算法在非可控环境的识别精度还比较低,距离实际可用还有很长的距离。非可控环境行人再识别面临许多挑战,包括训练样本不足、光照剧烈变化、行人遮挡和开集测试等,严重降低了行人再识别算法的性能。文章对非可控行人再识别技术,尤其是对小样本、可见光-红外、遮挡和开集行人再识别技术的近期进展、使用的数据库进行阐述,并分析了相关技术存在的问题和未来的发展趋势。In the last few years, with the development of deep learning theory and person re-identification(re-id) methods, re-id techniques has achieved great breakthrough and gained high recognition accuracy in constrained environments. However, the existing re-id approaches perform poor in unconstrained environments and are still far from practical applications. There are many significant challenges in unconstrained environments, including lack of training samples, dramatic illumination variations, person occlusion and open-set tests, which significantly decreases the performance of re-id models. In this paper, we will introduce the latest improvements, the involved datasets, the existing problems and the future trends of the unconstrained person re-identification techniques, especially for unsupervised re-id, visible-infrared re-id, occlusion re-id, and open-set re-id.

关 键 词：行人再识别 非可控环境 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]