协同视频监控中的行人再辨识关键技术

项目名称： 协同视频监控中的行人再辨识关键技术

项目编号： No.61471049

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 无线电电子学、电信技术

项目作者： 赵志诚

作者单位： 北京邮电大学

项目金额： 80万元

中文摘要： 行人再辨识是在跨摄像机协同监控中，在传统基于生物特征方法失效的情况下，针对大量的空间盲区域环境实现监控对象身份一致性判别的一种新兴技术，对反恐、公共安防及平安城市的建设有着极为现实的意义。当前，研究者主要从外观特征设计和相似度度量两个方面对行人再辨识进行了研究，但由于在跨摄像机条件下，摄像机视角、光照条件、摄像机的参数都不同；此外，人体的非刚性、变化的姿态、运动及遮挡等因素都会使行人的外观出现巨大差异。因此，现有算法存在特征提取及表示方法单一、相似性度量模型泛化能力不强、重排序技术缺乏研究等不足。本项目将深度学习、局部度量学习、半监督学习、重排序等技术引入跨摄像机行人再辨识，在现有算法存在的几个关键问题上取得突破，提高行人再辨识的准率，为建立大数据协同监控平台提供支撑。

中文关键词： 行人再辨识；图像理解；特征学习

英文摘要： Person re-identification is a new identity authentication technology in spatial blind regions under multi-camera, where traditional biometrics identification technologies could not work. Nowadays the importance in anti-terrorism, public security and construction of peaceful city has largely grown while in research circle, some re-identification algorithm were proposed and they mainly focused on two aspects: the extraction and representation of appearance features and similarity metric. However, due to the huge differences in shooting angle, illumination, camera parameters, non-rigid body, posture, motion and so on, among multiple cameras, some key problems always exist such as the description and fusion of visual features, metric model with low generation power, lack of re-ranking techniques, etc. In this task, we are going to introduce deep learning, local metric learning, semi-supervised learning and re-ranking to overcome above mentioned problems, and to try to increase the precision of person re-identification and to back up the construction of collaborative video monitoring platform for big data.

英文关键词： person re-identification;image understanding;feature learning