Simulation-Based Stampede Prevention with Machine Learning and Explainable AI

Large scale events held outdoors present a significant threat to public safety owing to excessive crowd density and insufficient pre-event risk assessment mechanisms. In this study, we propose a simulation-assisted crowd safety framework using pedestrian dynamics modeling, ensemble learning, and Explainable Artificial Intelligence (XAI). We have built digital event venues where pedestrian movement has been simulated considering various arrival rates and other environmental conditions. Features related to crowd dynamics such as density, average velocity, turbulence, and crowd pressure are analyzed from simulations based on around 20,000 scenarios in order to predict risks. XGBoost and LightGBM algorithms are utilized to calculate the continuous score representing the risk of stampedes while SHAP explanations give insightful understanding of the underlying cause of prediction outcomes. Our experiment proves good predictive performance with extremely high R² values (>0.998) and negligible prediction errors. The result of the explainability analysis suggests that crowd density and crowd pressure play the most important role in the rising risks. The proposed framework can help authorities to identify hot spots and test out new event venue plans.