基于数据驱动与语义建模的多摄像机拓扑推理与行人再识别研究

With the developing of the city video surveillance and its applications, person re-identification through the video surveillance under the natural scene become the frontier research in the field of computer vision in recent years. This project is aimed at the key problems for person re-identification system research. Time-Delayed mutual information model will be employed to address Multi-cameras Topological Inference; Matrix low-rank decomposition technique will be used to get the segmentation of pedestrian video’s foreground and multiple objective; For getting the bag of visual words features from the underlying filters based on topography independent component analysis,a shared regional conditional restricted Boltzmann machine pre-trained by common pedestrian images will be used to choose the spatial features of videos; At the same time,long short-term memory recurrent neural network is employed to learn the temporal features; In addition, some semantic attribute classifiers based will be added in the top of the deep learning network, and a transfer learning method based on comparison of attributes will be employed to satisfy the users' interactive search requirements. Finally, a person re-identification system based on MapReduce parallel computation framework will be designed with the cloud platform environment and Hd intelligent camera.

随着城市视频监控的普及及其应用领域的不断扩大。自然场景下的行人再识别已成为计算机视觉领域中近年来备受关注的前沿研究。本项目旨在针对行人再识别系统的关键问题展开研究，拟采用时间延时互信息模型进行多摄像头拓扑关系推理；采用矩阵低秩分解技术实现行人视频前后景分离与多目标分割；采用拓扑独立成分分析提取底层特征，并使用经通用行人图像预训练过的共享的区域条件受限玻尔兹曼机学习获取目标行人的空域特征，再由长短期记忆复发网络模型学习获取目标行人的时域特征，通过该深度学习网络对特征的无监督选择优化后，获取了行人视频的特征词袋；同时，在该深度学习网络顶层设计基于作者-主题模型的行人属性-类别分类器，利用高层词袋特征进行训练学习，并通过基于属性比较的行人迁移学习来实现用户的交互搜索需求。最后，拟利用MapReduce并行计算框架设计出云平台环境下高清智能相机视频行人再识别原型系统。

随着城市视频监控的普及及其应用领域的不断扩大。自然场景下的行人再识别已成为计算机视觉领域中近年来备受关注的前沿研究。本项目针对行人再识别系统的关键问题展开研究，采用时间延时互信息模型进行多摄像头拓扑关系推理；采用矩阵低秩分解技术实现了行人视频前后景分离与多目标分割；采用拓扑独立成分分析提取底层特征，并使用经通用行人图像预训练过的共享的区域条件受限玻尔兹曼机学习获取目标行人的空域特征，再由长短期记忆复发网络模型学习获取目标行人的时域特征，通过该深度学习网络对特征的无监督选择优化后，获取了行人视频的特征词袋；同时，在该深度学习网络顶层设计基于作者-主题模型的行人属性-类别分类器，利用高层词袋特征进行训练学习，并通过基于属性比较的行人迁移学习来实现用户的交互搜索需求。最后，利用MapReduce并行计算框架设计出云平台环境下高清智能相机视频行人再识别原型系统。该项目成果在交通、安检、银行、军事等领域有着广泛的应用前景。