소규모 건설 현장 안전모 탐지를 위한 엣지 기반 YOLO 모델의 성능 비교 및 개념 증명 연구 = Performance Comparison and Proof of Concept of Edge-based YOLO Models for Safety Helmet Detection in Small-scale Construction Sites

다국어 초록 (Multilingual Abstract)

Despite the technological advancements, small-scale construction sites continue to account for more than 70% of fatal accidents because of budget constraints and a lack of dedicated safety personnel. Existing high-cost server-based monitoring systems are difficult to deploy in infrastructure-poor environments. Therefore, this study examined the feasibility of a stand-alone real-time safety helmet-detection system based on Edge AI through a Proof of Concept (PoC). The performance of two state-of-the-art lightweight models, YOLOv11 Nano and YOLOv12 Nano, was compared, with YOLOv11 Nano ultimately selected for its superior stability and precision in edge environments. Trained on a dataset of 16,161 images, the model achieved a mAP@0.5 of 0.969 and an F1-Score of 0.94. It recorded an inference speed of 5.26 FPS on a Raspberry Pi 5 system. Overall, real-time practical monitoring is feasible by applying frame skipping and multi-frame decision logic, underscoring the validity of the proposed cost-effective technical solution to prevent serious accidents on small-scale construction sites.