基于HG-RTDETR的建筑工地小目标检测方法OA

To address the challenges of limited pixel information,unstable illumination,large-scale variations,and occlusion interference in small-object detection within construction sites,this paper proposes an improved Transformer-based detection model,HG-RTDETR.A Single-Head Gated feature extraction module(SHG)was constructed,employing a single-head attention mechanism to jointly model global context and local features,with a gating mechanism to optimize feature selection and enhance discrimination under occlusion.A Cross-Scale Channel and Context Feature Fusion(CS-CCFF)was designed,incorporating Reconfigurable-Field Attention Convolution(RFAConv)and Hierarchical Feature Integrator(HFI).RFA adaptively adjusts receptive fields for multi-scale fea-ture extraction while suppressing noise,and HFI integrates high-level semantics,mid-level details,and large-scale features via a feature pyramid architecture to establish cross-level feature interaction.In addition,the MPDIoU loss function was introduced during bounding-box regression to refine the alignment between predicted and ground-truth boxes,improving small-object localization accuracy and robustness.Experiments on the MOCS construction dataset demonstrate improvements of 2.8%in mAP50 and 1.8%in mAP50:95.