SD-YOLOv5: a rapid detection method for personal protective equipment on construction sites

Abstract

With the rapid growth of urbanization, construction sites are increasingly confronted with severe safety hazards. Personal protective equipment (PPE), such as helmets and safety vests, plays a critical role in mitigating these risks; however, ensuring proper usage remains challenging. This paper presents SD (Small object detection and DilateFormer attention mechanism)-YOLOv5s, an improved PPE detection algorithm based on YOLOv5s, designed to enhance the detection accuracy of small objects, such as helmets, in complex construction environments. The proposed model incorporates a dedicated feature layer for small target detection and integrates the DilateFormer attention mechanism to balance detection performance and computational efficiency. Experimental results on the CHV dataset demonstrate that SD-YOLOv5s achieves an average precision (AP) of 93.7%, representing an improvement of 2.8 percentage points over the baseline YOLOv5s (AP = 90.9%), while reducing the model’s parameter count by up to 14.6%. These quantitative improvements indicate that SD-YOLOv5s is a promising solution for real-time PPE monitoring on construction sites, although further validation on larger and more diverse datasets is warranted.