Abstract
Real-time crowd monitoring has become an essential aspect of emergency
management in the context of increasing public gatherings and urban congestion. This study proposes
a deep learning-based modular framework for detecting and analyzing
crowd behavior in real-time to prevent stampedes, panic-induced dispersals, and
congestion-related hazards. The system integrates object detection (YOLOv8),
multi-person tracking (DeepSORT), pose estimation (OpenPose), density
estimation (CSRNet), and anomaly detection using LSTM Autoencoders and 3D Convolutional Neural Networks. Data were collected from open-source video datasets and controlled simulations to train and validate
the models. Evaluation showed detection accuracy
of up to 94.3%, F1-scores above 90%, and low latency
suitable for live deployment. The system effectively detects abnormal
motion patterns, congestion hotspots, and sudden crowd dispersal events.
Furthermore, ethical considerations were incorporated by ensuring data anonymization and non-intrusive monitoring. The study concludes
that the proposed approach offers a scalable,
real-time, and accurate solution for emergency response teams, city planners,
and public safety agencies. The framework enhances situational awareness and
supports timely decision-making, ultimately reducing the risk of crowd-related
disasters.