摘要
行人重识别是实现跨摄像机场景大范围追踪行人的关键技术,利用该技术可以把行人的碎片化多场景轨迹连接起来.本文首先回顾了行人重识别的发展,列举了目前行人重识别研究的主要难点和挑战.然后进一步介绍了作者所在研究团队针对行人重识别发展的非对称度量学习理论,及基于非对称度量理论和思想所开展的面向开放性行人重识别的非对称行人重识别建模.与现有用于行人重识别的度量学习算法相比,现有算法通常忽略了摄像机特征变化的特性,而非对称度量的优点是可以学习具备建模不同视域特点非一致性能力的特征变换.非对称建模除了应用在一般的行人重识别问题上,还可以应用在跨模态行人重识别、低分辨率行人重识别、基于属性与图像匹配的行人重识别、无监督行人重识别和不完整行人重识别等问题上.最后,本文讨论了行人重识别未来的发展.
Person re-identification（RE-ID） is critical and crucial for tracking people across multiple camera views, and hence, the piece-wise tracklets of each person from different locations can be connected. In this paper,we first review the development of person RE-ID and present its challenges. Subsequently, we introduce our recent development on asymmetric distance metric learning and the asymmetric person RE-ID modeling of the largely unsolved open-topic problems. Existing metric learning methods for person RE-ID usually ignore the characteristics of feature transformations between different camera views. The advantage of asymmetric metric is that it can model inconsistent feature transformations between different camera views. Except for being applied to general person RE-ID problem, asymmetric model can also be applied to cross-modality RE-ID, low-resolution RE-ID, attribute-image RE-ID, unsupervised RE-ID and partial RE-ID. Finally, we discuss the future development of person RE-ID.
作者
郑伟诗
吴岸聪
Weishi ZHENG;Ancong WU(School of Data and Computer Science, Sun Yat-sen University, Guangzhou 510006, China;School of Electronics and Information Technology, Sun Yat-sen University, Guangzhou 510006, China)
出处
《中国科学：信息科学》
CSCD
北大核心
2018年第5期545-563,共19页
Scientia Sinica(Informationis)
基金
国家自然科学基金(批准号:61522115)资助项目
关键词
视频监控
行人重识别
行人跨视域追踪
度量学习
非对称
visual surveillance
person re-identification
cross-view person tracking
distance metric learning
asymmetric modelling
