| Title |
Enhancing Drowsiness Detection in Construction Workers Through Generative Image Data and Computer Vision |
| Authors |
강경남(Kang, GyeongNam) ; 조선애(Cho, SunAe) ; 강민정(Kang, MinJeong) ; 김효상(Kim, HyoSang) ; 민종현(Min, JongHyeon) ; 전정호(Jeon, JungHo) |
| DOI |
https://doi.org/10.5659/JAIK.2026.42.3.349 |
| Keywords |
Construction Safety; Drowsiness Detection; Computer Vision; Data Augmentation; Generative AI |
| Abstract |
Drowsiness among construction workers presents significant safety risks by reducing attention and hazard awareness. This study investigates
the use of text-to-image (T2I) generative augmentation combined with computer vision models to improve automated drowsiness detection.
Four polygon-labeled datasets were prepared: 200 non-construction images, 200 construction-domain images, 400 synthetic images generated
using Stable Diffusion, and 100 construction-site test images. Two models, YOLOv8 and YOLO11, were trained and evaluated using
precision, recall, mAP@50, and mAP@50-95. Models trained on construction-domain data outperformed those trained on non-construction data,
achieving mAP@50 scores of 0.642 and 0.611 compared to 0.525 and 0.534. The addition of 400 synthetic images further improved
performance to 0.901 for YOLOv8 and 0.889 for YOLO11. Confusion matrix analysis demonstrated high accuracy in distinguishing AWAKE
and DROWSY states, although some background confusion remained. On-site validation confirmed improved generalization, with YOLOv8
detection confidence increasing from 0.51 to 0.72 for AWAKE and from 0.13 to 0.84 for DROWSY, alongside a reduction in false
detections. Overall, the findings emphasize the value of domain-aligned training data, the effectiveness of scalable synthetic augmentation, and
the robustness of YOLOv8 when applied to limited domain-specific datasets. The study provides practical evidence supporting the use of
T2I-based augmentation to enhance proactive safety management in construction environments. |