Deep Learning Framework for Automated Worker Helmet Detection and Safety Compliance Monitoring Using the Yolo Object Detection Model

Abstract

Construction and industrial environments involve substantial occupational risks, particularly from falling objects, collisions, and unsafe work practices. Safety helmets are therefore a critical component of personal protective equipment, but conventional compliance monitoring relies heavily on manual inspection and periodic supervision. Recent advances in deep learning and computer vision have enabled automated identification of workers, helmets, and non-compliance conditions from images and video streams. Among available object-detection approaches, the You Only Look Once (YOLO) family has attracted considerable attention because of its favorable balance between detection accuracy, computational efficiency, and real-time inference capability. This review examines the development of YOLO-based worker helmet detection from YOLOv5 to more recent architectures, emphasizing small-object detection, complex backgrounds, occlusion, illumination variation, model lightweighting, temporal tracking, and deployment at construction sites. Existing studies demonstrate that YOLOv5 can provide high-speed detection, while improved YOLOv8 models address small and distant helmet targets through attention and feature-fusion mechanisms. More recent YOLOv9-based approaches integrate pose estimation and multiobject tracking to provide continuous helmet-wearing-state monitoring. The review further analyzes datasets, evaluation metrics, compliance decision mechanisms, alert generation, edge deployment, and practical limitations. Finally, research gaps are identified toward robust, privacy-aware, context-aware, and continuously operating safety-compliance systems.