Deteksi Penggunaan Helm dan Rompi Pekerja Proyek Konstruksi dengan YOLOv8, YOLOv11, YOLOv26 dan K-Means Clustering Berbasis CRISP-DM

Abstract

This study developed a helmet and vest detection system for construction workers using YOLOv8, YOLOv11, and YOLOv26, combined with K-Means clustering based on CRISP-DM. The simulation involved 1,278 images (classes: person, helmet, vest, no-helmet, no-vest). Results showed that YOLOv8m achieved the highest mAP@0.5 (0.849), while YOLOv26m achieved the highest F1-Score (0.81) at a confidence threshold of 0.398 with an mAP@0.5 of 0.843, outperforming YOLOv11m, which recorded the lowest performance on both metrics (mAP 0.818, F1 0.76). Based on its precision–recall balance and stable performance across classes, YOLOv26m was selected as the system's primary model. Confusion matrix analysis revealed the highest accuracy for the helmet class (94%) and the lowest for no-vest (78%). This study demonstrates that integrating YOLO and K-Means clustering effectively detects and groups PPE compliance levels in real time, potentially improving the effectiveness of occupational safety supervision on construction projects.