Abstract
To reduce the accident rate in the construction industry, an improved YOLOv5n-based hazard recognition system for construction sites is proposed. By incorporating optimization mechanisms such as the ECA attention module, ghost module, SIoU loss, and EIoU-NMS into YOLOv5n, the system achieves both lightweight acceleration and improved accuracy. Two ultrasmall models (approximately 2.5 MBs each) were trained on a self-built dataset to detect "unsafe human behaviors" and "unsafe object conditions," achieving mAP@0.5 scores of 93.6% and 99.5%, respectively. After deployment on the Jetson Nano B01 edge platform, the system was constructed, and its high efficiency in onsite hazard detection was validated.