Real-Time Stampede Risk Prediction from Crowd Videos Using YOLOv8 and Spatio-Temporal Modeling

Abstract

Stampedes pose a major risk to people's lives in densely populated settings, such as public gathering places, transit hubs, and places of worship. In order to avoid stampedes and fatalities, it is crucial to promptly recognize unstable crowds. In this paper, we offer a computer vision system that uses YOLOv8-based person detection to analyze crowd behavior and estimate the probability of a stampede. To ascertain how a crowd is acting collectively, the system makes use of data on time, motion patterns, and crowd density. Furthermore, we employed a spatial zone-based analytic technique to pinpoint regions of the scene with the highest potential risk of a stampede. Finally, we have included a way to build a temporal persistence on the population in order to reduce the incidence of false alarms and stabilize risk estimations over time. Through a real-time web-based interface, our technology generates visual analytics output and multi-level risk alerts. The suggested method can distinguish between crowds that are behaving normally, at the warning stage, and in a high-risk state, according to experimental results acquired from recorded videos of crowds. This creates a useful decision support tool for crowd monitoring applications.