Novel AI-Driven Spatio-Temporal Crowd Monitoring for Enhanced Public Safety in Mass Gatherings

Abstract

Large-scale public events such as religious festivals and transportation hubs present significant crowd safety challenges, where traditional monitoring systems struggle with real-time risk assessment. To address this, we present STAR-Crowd, an AI-powered framework that integrates three key innovations: (1) a hybrid ViT-CNN encoder with occlusion-aware attention for improved density estimation in crowded scenes, achieving 96.3\% accuracy; (2) an LSTM-GNN decoder that models crowd movements as spatio-temporal graphs to predict dangerous conditions 1.8 seconds in advance; and (3) a Meta-RL module that dynamically adjusts risk thresholds, reducing false alarms by 22\% compared to static systems. Evaluated across five benchmarks including a new 10,000+ image Kumbh Mela dataset, STAR-Crowd demonstrates an 18.5\% higher F1-score than state-of-the-art methods. Real-world deployment at Delhi Railway Station reduced emergency response times by 37\% through proactive alerts, while maintaining efficient 47ms processing on edge devices. This work bridges the critical gap between offline crowd analysis and real-time intervention, offering a scalable solution for enhancing public safety in ultra-dense environments.