CMADS v1.0: Crowd Monitoring and Anomaly Detection System Using YOLOv8, ByteTrack, and Vision Transformer

Abstract

Managing crowds and noticing unusual behaviors are important for ensuring public safety at large events, particularly in areas with multiple gates where movement can be crowded and unpredictable. Traditional surveillance systems often fall short in accurately identifying individuals, gauging crowd density, or detecting abnormal activity. To tackle these issues, we propose an AI-based automated crowd monitoring and anomaly detection system, CMADS v1.0, that integrates YOLOv8 for people detection, ByteTrack for robust tracking, and Vision Transformers (ViT) for gender classification. This system can estimate the number of people present, including gender distribution, and detect unusual behaviors such as loitering, sudden surges, or slow movement—all of which are useful for monitoring gate zones at large events. The system achieves 13-15 frames per second and 89% mean Average Precision (mAP) in people detection, with 98.7% accuracy in gender classification. It also successfully identifies loitering behavior with 87% accuracy, abnormal speeds with 90% accuracy, and crowd surges with 92% accuracy. These preliminary results are promising, demonstrating that the system can perform the tasks effectively in real time. The system aims to support security teams by providing detailed information on crowd movements and possible hazards, finally enhancing crowd control efforts.