AI Insight
Researchers developed YOLOv11s-R3, an enhanced object detection model for monitoring personal protective equipment (PPE) usage in construction settings. The model incorporates a refinement block that operates at higher resolution to better detect small-scale objects like helmets and vests, achieving precision of 0.9103 and maintaining real-time processing speed of approximately 190 frames per second. Testing on two datasets showed modest improvements over the baseline model, with statistically significant gains observed on the cross-dataset evaluation.
Why it matters
This technology could enable automated safety monitoring systems in construction sites, where over half of accidents are attributed to improper use of protective equipment. The model's ability to process images in real-time makes it suitable for practical deployment in active work environments.
by Alibek Barlybayev, Nurzada Amangeldy, Bijeesh Kozhikkodan Veettil, Bekbolat Kurmetbek, Akmaral Kassymova, Ramazan Sadvakassov
Automated monitoring of personal protective equipment usage is crucial due to the high incidence of accidents in the construction industry, with over half attributed to the improper use of protective equipment. Existing deep learning models show limited performance in detecting small-scale objects and reduced accuracy under challenging visual conditions. To address these limitations, this study introduces YOLOv11s-R3, an enhanced version of the YOLOv11s architecture with a P2-focused refinement block. The proposed model extends the standard feature pyramid network to stride-4 resolution and applies a single lightweight C2f block on the P2 feature map, targeting fine-scale objects. The model was trained on the Color Helmet and Vest dataset and evaluated on its held-out test split and on the GDUT Hardhat Wearing Detection dataset, with all experiments repeated across five independent random seeds. On the CHV test split, YOLOv11s-R3 achieved a precision of 0.9103 and mAP@50 of 0.9032, compared to 0.9035 and 0.9011 for the YOLOv11s baseline. The model maintained an inference speed of approximately 5 ms per image (190 FPS on an NVIDIA Quadro RTX 6000 GPU). On the CHV test split, the observed mAP@0.5 gain was not statistically distinguishable from run-to-run variance (paired t-test, p = 0.369). In contrast, the improvement observed on GDUT-HWD was more consistent relative to run-to-run variance, providing the primary evidence that the benefit of the proposed refinement extends across datasets. Together with an inference speed of approximately 5 ms per image, these findings support the applicability of YOLOv11s-R3 to real-time PPE monitoring.