A SMART PREDICTIVE SYSTEM FOR HUMAN STAMPEDE DETECTION

Abstract

Human stampedes in public events pose significant safety risks, and early prediction is crucial for preventing accidents and ensuring crowd safety. This project presents a Smart Predictive System for Human Stampede Detection and Crowd Safety Management that leverages computer vision techniques to analyze crowd density and perform real-time human counting. The system utilizes a large-scale annotated dataset of 19,241 images sourced via Roboflow to fine-tune a YOLOv9 (You Only Look Once version 9) model for crowd monitoring across two object classes: head and person. The primary objective is to predict stampede risks when crowd density exceeds a threshold of 20 individuals in a single camera frame, signaling a potential hazard. A web-based front-end interface is developed using HTML, CSS, and JavaScript for user interaction, while the back-end is implemented using Python and Flask for model deployment and real-time prediction. By integrating advanced computer vision with a rule-based risk classifier, the proposed system achieves a mean average precision ([email protected]) of 85.6% on the test set, outperforming YOLOv5s (78.2%) and YOLOv8s (81.3%) baselines. The findings aim to improve crowd control measures, reduce the risk of stampedes, and contribute to better event management strategies in stadiums, pilgrimage sites, and transit hubs.