Performance Comparison of YOLO Models for Safety Helmet Detection: Insights from YOLOv5 to YOLOv10 with Transfer Learning

Head injuries pose a significant risk in industries such as construction and manufacturing, making helmet use essential for safety. This paper presents a comparative analysis of six YOLO (You Only Look Once) models-YOLOv5mu, YOLOv6m, YOLOv7, YOLOv8m, YOLOv9 (m, GELAN-c, GELAN-e), and YOLOv10m-for helmet detection using a public dataset. The models are evaluated for mean Average Precision (mAP), frames per second (FPS), training time, and model size. YOLOv9 (GELAN-e) achieves the highest accuracy with 96.17% mAP at 0.1% confidence and 82.71% at 70% confidence. YOLOv8m is the fastest, achieving 45.37 FPS, while YOLOv10m has the highest mAP (85.10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>) at 70% confidence and is the most compact (33.5 MB). YOLOv9 (GELAN-c) offers the shortest training time (1.32 hours), while YOLOv9 strikes the best balance between accuracy, speed, and size. These findings suggest YOLOv9 as an excellent choice for applications requiring both accuracy and efficiency, while YOLOv8m is optimal for real-time scenarios.