人群运动行为突变识别模型的建立，对借助摄像机实现群体性突发事件的自动监测预警，具有重要意义。研究一种正确的、可以定量化和普适性应用的人群运动行为突变识别的方法，是人群运动行为突变自动识别亟待解决的难题之一。为解决这一问题，本文将人群运动行为作为研究对象，利用人群中的个体运动速度构建了人群的微观状态空间，以香农熵模型为基础，构建了评价人群宏观状态的熵模型。结合数值模拟和典型视频实验，通过对人群熵值突变的分析和建模，实现了人群运动行为突变的自动识别。 主要研究内容及成果如下：1、构建了具有普适性的人群运动行为宏观状态熵模型。采用统计物理学方法，从人群中个体的运动速度出发，构建了人群的微观状态空间，借助对香农熵、玻尔兹曼熵和克劳修斯熵的概念分析，构建了表达人群宏观状态的熵模型，从而建立了人群宏观状态与微观状态的关系。2、验证了个体运动速度作为人群微观变量的合理性。基于社会力模型，模拟了人群有序、无序及其突变行为，考察了人群宏观状态与微观变量的关系，发现人群宏观状态与速度分布存在一定的关联性，进而验证了采用个体速度对人群微观状态空间进行构建的合理性。3、建立了人群运动行为突变与熵变间的关系。通过模拟实验验证了人群宏观状态与熵值的关系，即人群越无序熵值越高；并提出了“一阶熵变”与“二阶熵变”人群状态变化的分析方法，对人群有序－无序状态突变进行量化分析，提出了基于熵变的人群运动行为突变识别准则，并初步探讨了人群宏观状态突变的机制。4、分析了人群运动行为突变识别方法的适用性。通过典型视频实验，采用熵变分析方法，识别出人群的“浪涌”现象、“走-停”突变、“不拥堵-拥堵”突变以及因设置警戒引发的人群突变，验证了本文构建的熵模型能够适用于高密度人群复杂状态突变的识别。本文提出的“人群微观状态空间”和“微观状态发生概率”，可以将人群的宏观状态与个体的速度等物理学特征直观地联系起来。利用本文的基于熵模型的人群运动行为突变识别方法，可服务于政府反恐处突中，提升其对突发事件的监测预警能力。

Identifying the terror attack, illegal public gathering or other mass events risks by utilizing cameras is an important concern both in crowd security area and in pattern recognition research area. A primary concern is to identify the crowd mutation caused by above events or risks. However, crowd mutation is a complex process process, and thus, a universal model to correctly describe the crowd macro state model is needed. This model-based, quantitative and universal method is the current research gap. This paper provides a method to construct crowd behavior microstates and the corresponded probability distribution using the individuals’ velocity information (magnitude and direction). Then an entropy model based on Shannon Entropy was built up to describe the crowd behavior macro state. Simulation experiments and video detection experiments were conducted. The following conclusions are drawn:1) A crowd macro state entropy model was constructed. From the point view ofstatistical physics, after the construction of individuals’ and crowd micros states, and the confirmation of each micro state’s probability distribution, the concept of Shannon entropy, Boltzmann entropy and Clausius entropy were extended to be used in crowd mutation recognition.2) Individuals’ velocity was verified as the crowd’s micro variable. The ordered state, disordered state and the mutation from one state to the other was simulated by the simplified social force model. The relationship between crowd macro state and the micro variable was investigated. It was found that: when the crowd was in an ordered state, the distribution of velocity direction and magnitude were more concentrated in some partitions disordered state; while in an disordered state, the velocity direction and magnitude were more even or normal. Thus, the reasonability of adopting velocity to construct the micro state was verified.3) The relationship between the crowd mutation and the crowd entropy mutation was constructed. Simulation experiments were conducted. It was concluded that in the disordered state, the crowd entropy is higher; while in ordered state, the entropy is lower. The first-order entropy mutation and the second-order mutation were introduced as the tools to quantify the crowd mutation. This paper tried to explore the physical mechanism of crowd mutation after the experiments. The possibility to detect the crowd mutation by recognizing the entropy mutation was validated. 4) The above crowd entropy model and the tools were applied to the more complex crowd mutation recognition at high density. The crowd “surge” phenomenon, “stop-and-go” mutation, “non-congestion-to-congestion” mutation and the crowd mutation caused by police cordon were recognized successfully. It was verified that the entropy was applicable at high density scenarios and was more applicable than the order parameter model. By the definition of the “crowd micro state” and the probability of each state in this paper, the crowd macro state was connected visually with the individual velocity. The entropy based recognition of crowd mutation can be served in terror-attack, and enhance the ability of governments’ emergency monitoring and early warning.