Performance Comparison and Proof of Concept of Edge-based YOLO Models for Safety Helmet Detection in Small-scale Construction Sites

Key Points

• To validate the feasibility of an edge AI-based real-time safety helmet detection system in small-scale construction sites without servers or networks.
• Conducted concept proof (PoC) for a helmet detection system using edge AI technology.
• Compared performance of YOLOv11 Nano and YOLOv12 Nano models to determine the best fit for edge environments.
• Trained the optimal model on a dataset of 16,161 images.
• YOLOv11 Nano model achieved an mAP@0.5 of 0.969 and an F1-Score of 0.94, indicating high precision and reliability.
• System operated at an average inference speed of 5.26 FPS on a Raspberry Pi 5.
• Model performance can be enhanced through frame skipping and multi-frame decision logic.

Abstract