视频监控场景下的行人再识别

Person re-identification is an important research area in the field of video surveillance. Person detection, feature extraction and re-identification algorithms are three key technologies for person re-identification. Existing person detection algorithms fail to take full advantage of temporal information in video datasets. Features of persons are not accurate enough for real applications, and lack interpretability. Meanwhile, re-identification algorithms are limited by their representation abilities. As a result, existing methods cannot meet actual demands in aspects of precision and robustness. Different from previous efforts, this project will model the movement between adjacent video frames, and extract features from both spatial and temporal dimensions, in order to improve the speed and accuracy of person detection. In addition, for the sake of enhancing the interpretability of features learned by deep neural networks, this project will study the problem of feature learning through incorporating person attributes. In detail, we will study the construction of person attributes and the designing of the network based on these person attributes. Furthermore, we will study bidirectional projection methods for metric learning, so as to improve the representation and discrimination abilities. This project will not only bring new ideas and new methods for person re-identification, but also provide technical supports for the development of relative research areas in video surveillance.

行人再识别旨在从非重叠的多个摄像机中找到相同的感兴趣目标，是视频监控领域一个重要的研究方向，行人检测、行人特征描述及再识别算法是其中的三个关键核心问题。现有方法在检测行人时未能充分利用视频数据的时间维信息，对行人的描述不够精确，缺乏可解释性，再识别算法也受到模型表达能力的限制，因此最终的再识别结果在准确性和可靠性方面尚不能满足实际需求。为此，本项目将建立视频数据相邻帧间的运动模型，构建行人在空间和时间维的特征描述，提高行人目标的检测速度和精度。其次，为了增强行人深度网络所学习特征的可解释性，本项目将研究联合人物属性的行人特征学习。其中包括视觉相关人物属性的构造问题，和基于人物属性的行人深度特征网络设计问题。在此基础上，研究双映射度量学习算法，提升行人再识别模型的表达和判别能力。本项目的研究不仅为行人再识别领域带来新思路和新方法，所获成果还为视频监控相关领域的发展提供理论支撑。

行人再识别旨在从非重叠的多个摄像机中找到相同的感兴趣目标，是视频监控领域一个重要的研究方向，行人检测、行人特征描述及行人再识别算法是其中的三个关键核心问题。现有方法受到模型表达能力的限制，在检测行人时未能充分利用视频数据的时间维信息，对行人的描述不够精确，缺乏可解释性，导致最终的再识别结果在准确性和可靠性方面不能满足实际需求。因此，本项目围绕监控场景下的行人再识别应用需求展开，研究视频监控场景下的行人再识别技术。主要研究了包括针对监控视频的行人检测技术、人物属性及再识别相关的行人特征提取技术、行人相似性相关的双映射度量学习技术等。经过三年的研究，本项目取得的研究进展和成果已经发表在国内外学术刊物，包括8篇国际期刊和2篇会议学术论文。此外本项目已申请专利4篇，共计培养4名硕士研究生，5名博士研究生。本项目的研究所获成果不仅为行人再识别领域带来新思路和新方法，还能为视频监控相关领域的发展提供一定的理论支撑。