Dataset and Approach for Helmet‐Wearing Detection With Edge Devices at Construction Sites

ABSTRACT To enhance construction site safety, we design HWD2024, a new benchmark dataset consisting of 5416 samples across nine classes, covering several on‐site construction scenarios. We propose a lightweight one‐stage detector, YOLO‐RPA, that incorporates the residual pyramid mechanism with average pooling attention to detect multi‐scale helmets. YOLO‐RPA identifies helmet colors and determines whether construction workers wear helmets properly. Additionally, a color transfer algorithm is proposed to address the issue of imbalanced sample distribution, while batch inference is introduced to improve the inference performance of helmet‐wearing detection on edge devices. Experiment results demonstrate that the accuracy of YOLO‐RPA in mean Average Precision (mAP) outperforms ten state‐of‐the‐art lightweight detectors, achieving 93.5% at mAP@0.50 and 73.2% at mAP@0.50:0.95. The color transfer method further improves mAP@0.50 by 0.5% and mAP@0.50:0.95 by 1.2% by balancing helmet instances. Using batch inference, YOLO‐RPA achieves inference speeds of 3.840.0 frames per second, with speedups ranging from 1.3 to 7.3 on ROCK 5B and NVIDIA Jetson AGX Orin. The HWD2024 dataset and an online evaluation system for its detection models are available at https://icnc‐fskd.fzu.edu.cn/hwd .