基于文本查询的行人重识别关键技术研究

Video investigation has become the new normal of the public security criminal investigation. To develop the efficiency of video investigation, researchers carry out the research of person re-identification by image searching, namely searching the correct person images from the multi-camera view via the image of suspect. However, there lacks of the image of suspect in the early stages of the public security investigation, person re-identification by image searching is hard to carry out. In another viewpoint, eyewitness testimony is not only the important basis of video investigation but also the key clues of criminal investigation. Inspired by this, from the bran-new viewpoint that searching the correct person images using the textual descriptions of suspect, this project carry out the research of the key technologies of person re-identification based on text querying. To conquer the new challenges such as text testimony possess incomplete information and noise, a set of new methods have been proposed especially in the key links such as feature extraction, similarity measurement, rank optimization. First, person is represented by salient semantic features which are further reconstructed through correlation analyzing of the semantic features. Second, joining the constraints of forward and backward distances, and then constructing the relative distance function for similarity measurement. Finally, initial ranking lists are fused based on both the viewpoints of salient and global semantic features representation. This project proposes a novel theoretical framework of person re-identification, which will be of great significance for improving the practical performance of person re-identification in video investigation.

视频侦查已成为公安侦查破案方式的新常态，为提高其工作效率，学者们开展了“以图搜图”的行人重识别研究，利用嫌疑人图像自动查找多摄像头画面中相匹配的行人图像。然而，公安侦查初期往往缺少嫌疑人图像，不能进行“以图搜图”的行人重识别。换个角度看，目击者口供也是视频侦查的重要依据和侦查破案的关键线索。受此启发，本课题从利用口供中嫌疑人的文本描述查找相匹配的行人图像这一全新视角，开展基于文本查询的行人重识别关键技术研究。针对口供文本信息不完备和存在噪声等新挑战，在特征提取、相似性度量和排序优化等关键环节提出一整套新方法。首先，对行人进行显著语义特征表示，并通过语义关联分析进行特征重构；其次，加入顺向和逆向的距离约束，构造相对距离度量函数用于相似性度量；最后，综合显著语义和全局语义的双重视角进行排序融合。本课题研究提出了新的行人重识别理论框架，对于提升行人重识别技术在视频侦查中的实际效能具有重要意义。

随着我国城市视频监控系统的全面建设，视频侦查已成为公安侦查的关键技术支撑。现有视频侦查模式主要通过人海战术研判嫌疑目标的多摄像头活动轨迹，不仅费时费力且效率低下，亟需针对嫌疑目标（特别是行人）的自动检索技术。然而，现有相关技术主要是“以图搜图”的行人重识别，难以适应侦查初期缺少嫌疑人图像的应用场景。针对这一问题，本课题从目击者口供这一早期侦查的关键线索入手，利用口供文本查找相匹配的行人图像，开展基于文本查询的行人重识别关键技术研究，并取得了以下结果：1）在特征提取上，提出基于关联语义的特征扩展与基于显著语义的特征表示方法。对口供文本进行语义补全及特征重构，提升了口供文本的信息完备性与鲁棒性；2）在相似性度量上，提出基于双向尺度学习的相对距离度量方法。通过构造顺向与逆向的双向距离约束，提高了信息不完备条件下相似性度量函数的判别力；3）在排序优化上，提出基于显著语义与全局语义的排序融合方法，进一步增强了行人重识别的性能。本课题共发表SCI/EI检索论文16篇，包括IEEE TGRS/TMM/TCSVT等中科院二区以上SCI期刊论文5篇，ICCV/ECCV/AAAI/IJCAI/MM/ICASSP等CCF推荐的A/B类会议论文7篇；获授权发明专利4项。本课题探索了基于文本查询的行人重识别理论框架，对于提升行人重识别技术在视频侦查中的实际效能具有良好意义。