Аннотация
Introduction Manual monitoring of helmet use is impractical in dynamic, congested, and safety-critical construction environments, particularly when unmanned aerial vehicles (UAVs) capture workers from varying altitudes, viewpoints, illumination conditions, and occlusion levels. This study proposes a hybrid helmet-detection framework integrating YOLO-E and YOLOv12 to improve detection accuracy, small-object recognition, color fidelity, and open-set awareness for real-time construction safety monitoring. Methods The proposed framework combines the language-guided open-vocabulary capabilities of YOLO-E with the real-time feature extraction and object-detection capabilities of YOLOv12. Slicing-Aided Hyper Inference (SAHI), image preprocessing, synthetic-weather augmentation, and prompt-sensitivity analysis were incorporated to improve robustness under challenging operating conditions. The framework was evaluated using a publicly balanced PPE dataset and a newly developed UAV-based helmet dataset containing 6,471 annotated helmet instances captured across different altitudes, viewing angles, lighting conditions, weather scenarios, helmet colors, and occlusion levels. We conducted ablation experiments to measure the contribution of each component of the framework. Results The SAHI-enhanced hybrid framework increased mean average precision at the 50% IoU threshold from 89% to 97%, improved small-object recall from 0.41 to 0.80, and reduced localization error to 2.25 pixels. The proposed approach also achieved scores of 0.96 for weather robustness and 0.97 for color fidelity. Comparative experiments showed that the hybrid YOLO-E + YOLOv12 framework achieved superior precision, recall, and F1-score relative to the evaluated baseline and existing detection architectures, including Mask R-CNN, Faster R-CNN, RetinaNet, and SSD. Discussion The results demonstrate that combining open-vocabulary recognition, closed-set detection, SAHI-based small-object enhancement, and robustness-oriented preprocessing provides a scalable solution for UAV-based helmet monitoring. The framework is particularly suitable for construction safety auditing and large-scale surveillance scenarios requiring real-time detection under variable environmental conditions, while its open-set awareness improves adaptability to previously unrecognized visual conditions.