Design of an improved graph-based model for real-time threat detection and dynamic evacuation management using crowd behaviour analysis

Design of an improved graph-based model for real-time threat detection and dynamic evacuation management using crowd behaviour analysis

연구 분야: Safety

학회: International Journal of Information Technology

Effective crowd management is highly significant in modern public spaces for safety and efficient use of resources. Conventional methods, based on static models and heuristic rules, are usually inadequate in the dynamic and unpredictable nature of crowds. This research work proposed an integrated framework combining agent-based modelling (ABM) with reinforcement learning (RL), convolutional neural networks (CNN) for real-time crowd density estimation, predictive analytics using Autoregressive Integrated Moving Average (ARIMA) and Long Short-Term Memory (LSTM) models, and multi-agent pathfinding to address all of these challenges. This framework would optimize the flow of crowds, predict the future behavior of the crowds, and intervene in a real-time sense to reduce congestion and operational inefficiencies. The key contributions of the work include real-time adaptability of behaviors, accurate crowd density prediction with optimally selected movement paths under high-density scenarios, thus providing a holistic approach and adaptive efficiency for handling crowds. Agent-Based Modeling (ABM) with Reinforcement Learning (RL) imitates individual agents with dynamic behavior rules to optimize the crowd flow and lessen congestion by up to 30%. This sets off an alert in less than 2 s whenever the crowd is densely packed. Predictive Analytics rendered by ARIMA and LSTM models with an accuracy of 85–90% in predicting future Crowd Densities and Transport Demands, thereby making Resource Allocations more efficient with a margin of 10–15%. Multi-Agent Pathfinding (MAPF) technology coordinates optimal paths of individuals together, enabling the reduction of travel times by 25–35% and congestion by 20–30%.