Improvements in YOLO Architecture: A Novel Approach to High-Density Crowd Detection

Abstract

Detecting high-density crowds in dynamic environments presents challenges such as occlusion, scale variation, and real-time processing demands. Traditional object detection methods struggle with these issues due to overlapping individuals and variable visibility. YOLO (You Only Look Once), widely recognized for real-time object detection, underperforms in high- density scenarios. This paper proposes an enhanced YOLO architecture with multi-scale feature extraction using a Feature Pyramid Network (FPN), optimized anchor boxes for small and overlapping objects, and advanced data augmentation techniques to improve robustness. Evaluated on high-density datasets such as ShanghaiTech, UCF-QNRF, and WorldExpo'10, the proposed model achieves significant improvements in Mean Average Precision (mAP) while maintaining real-time processing speeds. This advancement supports applications in public safety, surveillance, and event management, marking a step forward in crowd detection technology.