基于高阶上下文建构学习的多源异质视频异常检测

Robust abnormal event detection can make the surveillance system timely and efficiently process the videos containing potential abnormal behaviors, and contribute importantly to the intelligent security systems. Specifically, event modeling and anomaly measurement are the key components in anomaly detection. With the development of the imagery technique and the demand of complex surveillance environment, multiple heterogeneous videos have jointly coexisted in the current surveillance space. However, the existing anomaly detection methods are incapable to achieve the jointly modeling and association mining for multiple heterogeneous videos. This proposal re-visits the abnormal event detection problem from multiple heterogeneous videos jointly via constructively high-order context learning. The main innovations include: 1) modeling a robustly contextual representation of event by hypergraph manifold, so as to make an achievement of virtual event registration for multiple heterogeneous videos; 2) establishing a compact and intact event representation space by constructively high-order context learning; 3) developing a hashing model on high-order contextual structure for fast event indexing; 4) achieving efficient anomaly detection by a tuple selection of high-order contextual structure. The research of this proposal not only can expand the paradigms of anomaly detection, and enrich the context learning theory. In addition, it is also beneficial for the intelligent surveillance for better public security.

异常事件检测旨在提取视频中偶然发生、存在潜在危害且不期望发生的目标行为模式，以便及时且针对性地进行视频重点处理与分析，是建立智能化安防系统的重要途径之一。其中，事件特征建模、事件异常解析是异常事件检测的两个核心内容。随着视频技术和监控需求的提升，多源异质视频协同共存已成为当前监控环境的客观现状，而现有方法缺乏对多源异质视频的联合建模。本项目着眼于事件发生的上下文特性与多源协同作用，在上下文学习理论框架下进行研究，包括：1）基于超图流形的视频源内事件上下文表达，促进后续事件层多源虚拟配准的达成；2）基于紧致完备空间建构学习，建立多源异质视频源间的高阶事件上下文关联；3）基于高阶上下文结构哈希编码，实现快速事件索引；4）基于上下文结构元组选择，构造多源视觉敏感特性的异常度量。本项目研究不仅可拓展现有研究范式，丰富上下文学习理论，同时也将促进视频监控系统智能化以更好地服务于社会公共安全事业。

项目针对1）鲁棒且易于交互的视频上下文模型；2） 多源视频上下文建构学习中的协同与交互及3）符合上下文表达结构几何性质的高效异常度量三个科学问题，从1）基于超图流形的潜在视频事件上下文表达；2）基于紧致完备空间建构学习的多源上下文关联；3）基于哈希学习的事件上下文结构高效索引；4）基于事件上下文结构元组选择的视频异常度量四方面研究内容进行研究。结合项目开展期间的应用场景及可行性，将初始设定内容及问题更加具体化，主要聚焦的研究点为： 1）多源视频中的时-空-谱上下文建模；2）多源视频数据中上下文结构的紧致融合；3）哈希索引在时空上下文搜索方面的应用及4）面向驾驶及监控环境的时空元组约束下的异常事件检测。项目研究形成了诸多有影响力的成果，建立了三个大规模事件演化分析数据集，发表了高质量学术论文20篇，其中SCI期刊论文8篇，会议论文12篇。多篇论文发表在领域内顶级期刊和会议上，项目超额完成了既定目标。