An enhanced YOLOv10 framework for small-object safety helmet detection on construction sites.
TL;DR
Findings indicate that the YOLOv10 detector is a strong candidate for deployment-focused helmet-compliance monitoring in large-scale construction settings, although performance limitations remain most evident in far-field views and highly crowded scenes.
Abstract
Safety helmet compliance monitoring remains challenging because helmets often occupy only a few pixels in wide-area surveillance images. This study tackles these small-object difficulties in vision-based safety-helmet compliance monitoring (helmet vs. no-helmet) using wide-area surveillance imagery. An enhanced You Only Look Once version 10 (YOLOv10) detector is proposed by integrating Omni-Dimensional Dynamic Convolution (ODConv) into the backbone, an Efficient Multi-scale Attention-guided Bidirectional Feature Pyramid Network (EMA-BiFPN) for multi-scale feature fusion, a four-head detection scheme, and the Minimum Points Distance Intersection over Union (MPDIoU) regression loss. Across the experiments, the proposed detector reached an mAP50 of 94.28, with AP50 values of 96.55 for the helmet class and 92.01 for the no-helmet class, exceeding the performance of the YOLOv10 baseline and other benchmark detectors. The most notable gains appeared for extremely small and small targets ([Formula: see text] = 86.10, [Formula: see text] = 91.55), reflecting improved localization of helmets at long distances. Overall, these findings indicate that the method is a strong candidate for deployment-focused helmet-compliance monitoring in large-scale construction settings, although performance limitations remain most evident in far-field views and highly crowded scenes.