YOLO-MSD: An Accurate Detection Method for Small Target on Construction Sites
Xiuyi Guo, Hongbin Liu, Yongze Zhao, Peng Dong, Yilin Wang, Xingyu Li, Qixin Zhuang · 2025
The construction site environment is complex, with many small targets like construction tools, building materials, and safety equipment. Accurate small target detection can boost site work efficiency. Existing object detection algorithms (e.g., the YOLO series) do well in general object detection, but have issues such as low recognition accuracy and inaccurate positioning in small object detection. To address these problems, this paper presents an improved detection model, YOLO-MSD (YOLO with Multi Scale Detection), for more accurate small object detection on construction sites. The YOLO-MSD model uses multi scale training and feature fusion to gain rich edge information. The multi scale training strategy dynamically adjusts input image sizes, enabling the model to handle targets of different scales, especially small ones. Through feature fusion, YOLOMSD integrates features at different network levels. By fusing ViT (Vision Transformer) features with multi level features, edge feature details are enhanced, improving the detector's robustness. Experimental results on the public dataset SODA show that the YOLO-MSD detector outperforms state of the art methods, especially for small object detection in construction site scenes.