A novel dual-branch detection method for multiscale objects in construction sites
Huaifan Hang, Suhuan Bi, Liangliang Mu, Yan Xu, Yanli Bi, Hailong Yan · Architectural Engineering and Design Management · 2025
Accurate detection of multiscale objects is essential for effective safety management in complex and dynamic construction environments. This paper introduces a dual-branch framework to improve detection performance for objects with varying scales in cluttered settings. Each branch enhances an improved YOLOv8 model with Alterable Kernel Convolution (AKConv) and the C2f-TA module to improve adaptability and suppress background noise. One branch employs the improved YOLOv8 network tailored for large-scale objects detection on construction sites, where materials and structural components often dominate the visual field. The other branch focuses on small object detection by integrating Slicing Aided Hyper Inference (SAHI) with YOLOv8, enabling finer feature extraction and improved accuracy. Experimental results demonstrate that the YOLOv8-SAHI model reaches a mean Average Precision (mAP) of 82.36% on the SODA dataset, achieving a 2.65% improvement over the original model. Notably, the Average Precision (AP) for detecting helmets and vests increased significantly by 8% and 9%, respectively. These results demonstrate the effectiveness of YOLOv8-SAHI for enhancing site monitoring and safety management in construction environments.