群体异常行为分析中面临的挑战与相关技术

Challenges and relevant techniques in crowd abnormal behavior analysis

群体异常行为的分析与识别过程在技术层面通常面临与个体行为分析不同的挑战,如群体行为类型的定义不明确、用于群体行为分析的数据不足以及对多种类型群体行为检测难度较高等问题。针对群体行为分析的3个阶段所面临的主要问题,对群体行为数据模拟、群体行为图像特征提取和群体异常行为识别相关的研究进行了相关的归纳与介绍。其中,群体行为数据模拟方面,主要介绍不同研究中对群体行为的分类定义,以及宏观、介观和微观方式群体行为模拟的相关研究。其次,对多个研究中使用的群体行为图像特征提取方法,按照低级别运动流的特征到高级别语义特征的顺序进行介绍。最后,分别归纳群体能量模型等基于全局的群体行为识别方法,以及社会力模型等基于局部的群体行为识别方法相关研究,对比其优势及缺陷,并对群体异常行为分析与识别领域可能的研究方向进行展望。

The analysis and recognition process for crowd abnormal behaviors usually encounter different technical challenges from those for individual behaviors,such as the implicitly of crowd behavior taxonomy,insufficient video data for analysis,difficulties on detection of multiple crowd behavior types.Based on the primary issues to be tackled on the three phases of crowd behavior analysis,the state-of-the-art research works of crowd behavior synthesis,video pattern extraction and crowd abnormal behavior recognition are summarised and reviewed.For the crowd behavior synthesis,the taxonomies of crowd behavior types in different criteria are introduced and the macroscopic,mesoscopic and microscopic crowd synthesis approaches are reviewed.For the crowd video pattern extraction,research works are reviewed from the low-level flow-based patterns to the high-level sematic patterns.Finally,the global crowd behavior recognition approaches such as crowd energy model,and local approaches such as social force model are reviewed respectively.The pro and cons of these approaches are compared,and the future direction is prospected as well.