Home
Monitoring Personal Protective Equipment (PPE) compliance at construction sites in the Andean region of Apurímac, Peru, faces challenges due to extreme solar radiation (reported UV indices exceeding 11) and frequent physical occlusion. This study compares the performance of three architectures (YOLOv8n, YOLO11n, and RT-DETR-l) trained on the public PPE-CPXSZ v4 dataset (11 classes, including violations), which were additionally subjected to a synthetically degraded subset using light overexposure and simulated occlusion as a controlled approximation of Andean environmental stress. The models were evaluated using mAP50, mAP50-95, precision, recall, and real inference latency on an NVIDIA GPU. Under normal conditions, RT-DETR-l achieves the highest overall precision (92.66% mAP50 and 71.47% mAP50-95) thanks to its Transformer-type attention, with a latency of 34.8 ms per image (28.7 FPS), similar to that of YOLOv8n and YOLO11n (35.4–36.8 ms), despite its larger size (32.0 M parameters compared to 2.6–3.0 M parameters). YOLO11n maintains the lowest computational cost (2.6 M parameters, 5.5 MB), with an mAP50 of 89.21%. Under the adverse subset, all three models suffer a sharp drop in mAP50, but RT-DETR-l retains the highest relative value (46.40% compared to 24.75%–26.89%). It is concluded that when measuring real inference latency, the speed advantage attributed to YOLO diminishes compared to RT-DETR-l; therefore, model selection should prioritize precision and memory consumption over speed.