Development of a Risk Space Prediction Model Based on CCTV Images Using Deep Learning: Crowd Collapse

Abstract

Crowd disasters are not limited to events but can occur at any time and place where large crowds are densely gathered. The most significant factor contributing to injuries in mass gatherings is pressure. The likelihood of injuries due to pressure significantly increases when the crowd density exceeds 5-6 individuals per square meter. Once a crowd disaster occurs, it becomes challenging for rescue and medical personnel to access the affected area, potentially exacerbating the situation. However, since there is no single clear solution to address crowd disasters, there is a need for a system that can detect and analyze them in advance or in real-time. This research aims to contribute to the proactive detection and analysis of various crowd disasters, focusing on improving strategies for disaster response. In this study, we utilized a dataset of 50 videos from 35 individual incidents to propose a framework that classifies the risk levels into Crowd Crush, Crowd Wave, and Crowd Collapse. Additionally, we analyzed the feasibility of real-time detection using P2PNet (Point to Point Network), Yolo (You Only Look Once) v8, and DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithms, moving away from prior still image-based detection methods. Our research outcomes provide new evidence for the feasibility of real-time detection using DBSCAN in Crowd Disaster scenarios. Moreover, the findings from this study can serve as valuable reference material for upcoming research, particularly emphasizing the algorithmic analysis of crowd collapses.