视频异常事件检测中的群体特征感知研究

Accompanied by the progress of urbanization and the increase of population mobility, public security problem in hot spots of public place is becoming more serious. The timely and accurate detection of abnormal events is essential to avoid or weaken the crowd disasters. This project intends to consider crowd feature perception as the breakthrough point of crowd abnormal event detection, and specific research three problems - the extraction of crowd feature, the disaggregation of crowd feature, and the spatial-temporal modelling of crowd feature. First, the compound prior is used to coordinate the sociality and randomness of individuals in the crowd. And the sparse coding in deep hierarchies is used to extract progressively the crowd feature through the multilayer and local clustering method. Then, based on the information entropy analysis, the network flow graph of the spatial-temporal correlation of crowd feature is established. So the crowd feature is disaggregated by solving the problem of multi-spatial coverage and multi-temporal continuation based on the network optimization and the discriminative sparse coding. Last, for the diversity of spatial-temporal coupling of abnormal crowd feature, the resonance between the space model and time model of crowd feature is introduced into the spatial-temporal model. And this resonant spatial-temporal model is studied to avoid the detection blind area of some kind of spatial-temporal coupling of crowd feature. This research project is expected to realize the breakthrough of the core theory in crowd feature perception and provide the basic technical support for crowd abnormal detection.

伴随着城市化进程及人口流动性的增加，热点公共场所的群体安全问题日益严重。避免群体灾害的发生，消弱群体灾害的蔓延和扩大，及时准确的群体异常事件检测至关重要。本项目拟以群体特征感知作为群体异常事件检测的切入点，具体研究群体特征抽取、群体特征解集和群体特征空时建模三个问题。拟利用复合型先验协调群体中个体的社会性与随机性，基于深度结构框架下的稀疏编码，研究多层逐级局部聚类的群体特征抽取方法；基于信息熵分析建立群体特征时空关联网络流图，拟通过网络优化和鉴别稀疏编码解决群体异常事件的多空间覆盖和多时间延续问题，实现群体特征的解集处理；针对群体异常事件空时耦合关系的多样性，在群体特征空间模型和时间模型之间引入谐振，拟研究谐振式群体特征空时模型，避免对某类空时耦合关系形成检测盲区。本项目的研究成果整体有望在群体特征感知方面实现核心理论的突破，形成群体异常事件检测的关键技术支撑。

群体异常检测在许多领域都具有巨大的应用价值。群体异常检测是一种特殊的二分类问题，具有正常和异常两类极不对称的特性，且群体正常特征具有复杂多样、随时间动态变化的特性。.本项目针对群体特征多样性和复杂性，研究基于约束的群体特征抽取，群体特征的多尺度表述，以及群体特征空时建模三个问题。首先，群体特征相对于个体特征，其鲁棒性和稳定性更强。利用抽取的复杂多样的个体特征，基于场景或任务的先验或模型约束是改善群体特征鲁棒性和稳定性的有效手段。本项目研究了复合稀疏先验、协稀疏、线性稀疏、泊松分布，构造数据和DPM等约束问题。其次，群体特征存在尺度变化上的多样性。因此本项目的研究分别从图结构和深度网络两方面研究了特征多尺度抽取问题。最后，社会属性决定了个体特征区别于群体特征，却影响着群体特征。利用普通的隐变量模型描述群体特征和个体特征的相互关系，个体特征对群体特征的反向影响作用无法得到充分体现。因此，本项目研究了带环路的隐状态模型，以此描述群体特征和个体特征的相互影响：一方面群体特征决定个体特征的变化，另一方面个体特征反过来会诱导群体特征的改变。