An enhanced framework for real-time dense crowd abnormal behavior detection using YOLOv8

Abstract

Abnormal behavior detection in dense crowd, during the Hajj pilgrimage is vital to public security. Existing approaches face challenges due to factors like occlusions, illumination variations, and uniform attire. This research introduces the Crowd Anomaly Detection Framework (CADF), an improved YOLOv8-based model, integrating Soft-NMS to improve detection accuracy under complex conditions. CADF extensively evaluated on the Hajjv2 dataset, delivering an AUC of 88.27%, a 13.09% improvement over YOLOv2 and 12.19% over YOLOv5, with an Accuracy of 91.6%. To validate its generalizability, the framework is also tested on UCSD and ShanghaiTech datasets. Comparisons with state-of-the-art models, including VGG19 and EfficientDet, demonstrated CADF’s superiority in accuracy, AUC, precision, recall, and mAP metrics. By addressing the unique challenges of Hajj crowd and achieving strong performance across diverse datasets, CADF highlights its potential for real-time crowd anomaly detection, contributing to enhanced safety in large-scale public gatherings and aligning with Sustainable Development Goals 3 and 11.