A lightweight pose-guided fusion approach for accurate safety helmet compliance monitoring on edge devices

Abstract

This paper presents an AI framework employing automated visual analysis to verify personal protective equipment (PPE) compliance in complex construction and industrial settings. Within the framework of engineering automation, robust semantic modeling of worker safety configurations is essential for enabling real-time hazard monitoring and data-driven risk management. In such environments, lightweight architectures are crucial for low-latency inference and efficient deployment on resource-constrained edge computing devices. However, existing studies rely on standalone object detection, leading to unreliable PPE verification, or use computationally heavy models that limit real-time performance. The proposed framework performs helmet detection and human pose estimation within a unified YOLO11-based architecture using a shared backbone. A spatial consistency mechanism aligns head keypoints with helmet regions to verify correct PPE usage. Experimental results on the Construction Personal Protective Equipment (CPPE) dataset achieve an F1-score of 0.961 at 110.7 frames per second (FPS), maintaining both accuracy and efficiency. Further, the framework achieves up to a 67% reduction in false compliance alarms over standalone YOLO detectors and improved inference efficiency compared with previous multi-stage pipelines while reducing NVIDIA Jetson AGX Orin energy per frame, with practical deployment demonstrated through an integrated web platform supporting live multi-stream monitoring and automated violation reporting. Performance evaluations on edge devices highlight the framework’s potential for scalable safety deployment and its contribution to improving real-time compliance systems.