【摘要】 运动人群在交叉路口或通道形成独特的运动行为模式——人群汇流，易引发诸如拥挤推攘、跌倒踩踏等潜在公共安全风险，然而目前尚未有针对该人群汇流检测的计算模型研究工作报道。针对该问题，提出了一种生物启发的人群汇流行为检测神经网络(CCBDNN)。基于蝗虫视觉神经结构特性，该神经网络感知视野域中人群活动引发的视觉运动线索；借助哺乳动物视网膜方向感知机制提取人群局部运动方向线索；通过蝗虫小叶巨型运动检测器(LGMD)危险感知机理，构建尖峰调谐机制并输出表征人群汇流行为感知的神经尖峰响应。系统性实验研究表明，CCBDNN能有效检测视觉场景中的人群汇流行为，并产生具有独特偏好特性的输出响应。该工作涉及生物视感神经机制启发的动态视觉信息加工处理，可为人工智能中的人群活动检测与行为识别研究提供新方法、新思想。更多还原

【Abstract】 Moving crowd in public place can generate a unique movement pattern known as crowd convergence behavior at intersections or passageways, which can lead to potential public risks, such as crowding, pushing, and trampling.However, no computational models have been reported against to detecting crowd convergence.To fill this gap, this paper investigated a bio-inspired artificial visual neural network which named crowd convergence behavior detection neural network(CCBDNN).Based on the visual neural structure characteristics of locusts, CCBDNN perceives visual motion cues produced by crowd activities within the field of view and extracted local motion direction cues of crowd using the direction perception mechanism of mammalian retinas, and then constructed a spike mechanism and output neural spiking responding to represent the perception of crowd convergence behavior by the danger perception mechanism of LGMD neurons in locusts’ vision systems.Systematic experiments show that CCBDNN can effectively detect crowd convergence behavior in visual scenes and produce output tuning with unique preference characteristics.This paper is involved the dynamic visual information processing inspired by biological neural me-chanisms, which can provide new methods and ideas for crowd activity detection and behavior recognition in artificial intelligence.更多还原