Abstract: The suddenly localized crowd gathering behavior in public places is often a precursor of abnormal events, but it is difficult to be effectively detected by existing traditional computer vision techniques due to its strong randomness and inconspicuous characteristics of precursors. This paper proposes a modified LGMD-based neural network model for detecting suddenly localized crowd gathering behaviors, based on the structural properties of locust vision systems and the hazardous perception mechanism of lobula giant movement detector（LGMD）. This model perceives visual signals triggered by the crowd activity in the field of view, and then integrates the motion cues collected by each pixel based on the specific neural mechanism in mammalian retinas. Finally, a spiking threshold mechanism is built to tune the output membrane potentials of the neural network, by means of the hazardous perception mechanism of LGMD neuron. The experimental results of crowd activity videos in different scenes show that the proposed neural network can effectively detect suddenly localized crowd gathering behaviors in the field of view and generate warning signals. This paper involves the processing for dynamic visual information of crowd activities inspired by the optic nerve mechanism of biological vision systems, which can provide novel methods and creative ideas for crowd activity detections and behavior analyses for future intelligent video surveillance.

Key words: abnormal event detection, crowd gathering, LGMD model, crowd behavior analysis, intelligent video surveillance

摘要： 公共场所中的人群突发局部聚集常是异常事件发生的先兆，由于其随机性强，前兆特征不明显，现有的传统计算机视觉技术较难对其有效检测。基于蝗虫视觉系统的神经结构特性与小叶巨型运动检测器（lobula giant movement detector，LGMD）危险感知机理，提出一种人群突发局部聚集行为检测的LGMD改进型神经网络模型。该模型感知人群活动在视野域中引发的视觉信号，基于哺乳动物视网膜视觉信号处理机制整合视觉运动线索，借助LGMD神经元危险感知机理构建尖峰阈值机制调谐神经网络输出，以感知人群活动中的突发聚集行为。不同场景下的人群活动视频实验结果表明，提出的神经网络能有效检测视野域中人群突发局部聚集行为并对其预警。该文涉及生物视神经机理启发的人群活动动态视觉信息加工处理，可为智能视频监控中的人群活动检测与行为分析提供新思想、新方法。

关键词: 异常事件检测, 人群聚集, LGMD模型, 人群行为分析, 智能视频监控