An Improved RT-DETR Model for Small Object Detection on Construction Sites

Traditional object detection methods often struggle with missed and false detections in the presence of complex backgrounds and small objects on construction sites, particularly with limited capability in detecting workers and protective equipment. This research proposes an enhanced RT-DETR (Real-Time Detection Transformer)based model that can detect small objects more accurately in complex scenarios in order to overcome these issues. A modified RepCAFPN encoder is proposed to optimize feature fusion and improve feature extraction capability. Additionally, the InnerCIoU loss function is employed to further enhance object localization accuracy and robustness. According to experimental findings, the enhanced model significantly raises $\mathrm{mAP} \text{@} 0.5$, with an enhancement of $2.7 \%$. At the same time, the improvement incurs minimal changes in computational complexity and parameter count. Notably, the proposed model achieves a $1.4 \%$ and $1.5 \%$ improvement in detection accuracy for small objects, such as workers and helmets, which occupy a small portion of the image. The study provides an applicable solution for object detection tasks on construction sites safety management.