人群行为分析及异常事件检测是视频图像分析领域的前沿课题，由于常规中等规模人群的运动较为松散且个体间相互遮挡，所以难以用传统的基于宏观统计特性或微观个体轨迹、姿态分析的方法来恰当的描述此类人群。本项目通过探究宏观与微观特性间的联系，提出在介观层面构建因果认知网络以分析人群行为的理论思想。主要内容包括：(1)介观团体分割：探索人群运动的流场形态表达，依据非稳定流场可视化的方法将人群运动表达成图像纹理，通过纹理聚类算法分割运动小团体；(2)智能体跟踪：确定体现团体间社会关系及心理影响的智能体运动准则，并结合外观特征跟踪多目标，为复杂网络的构建提供动力学参数；(3)因果认知复杂网络分析：借鉴认知学中的因果概念，构建介观团体间的因果动力学模型，根据Granger因果检验评估介观团体间的因果关联度以确定网络的结构，并剖析人群行为与网络功能特性间的相关性，最终实现人群行为的准确表达与分析及异常事件检测。

Crowd behavior analysis and abnormal event detection has been attracting increasing attention in the field of video image analysis. Tradition methods are based on macroscopic statistical characteristics or microscopic individual trajectories and postures analysis, which are not suitable for describing conventional medium scale crowd because the pattern of movement in these crowds are loose and individuals occluded each other. This proposal explores the relationship between the macroscopic properties and microscopic characteristics of crowds. Therefore, a theory is proposed in mesoscopic level to analyze crowd behavior by constructing a complex network based on the causal cognition model. There are three main methods: (1) motion segmentation of mesoscopic groups: exploring the representation of crowd motion flow, describing the crowd motion as image texture by unsteady flow field visualization method, and segmenting the moving of small groups by a texture clustering algorithm; (2) agents tracking: determining the rules of agents motion according to the social relationships and psychological impacts of different agents, tracking multiple agents by combining with appearance features of these agents to provide kinetic parameters for the construction of complex network; (3) analysis of causal cognition complex networks: constructing the causal dynamics model of mesoscopic groups according to the causal concept of cognitive science, evaluating the causal relation of mesoscopic groups based on Granger causality method to determine the structure of network, and analyzing the relationship between crowd behavior and characteristics of complex network. Therefore, an effective crowd behavior analysis and abnormal event detection can be achieved by the proposed method.