Enhanced RT-DETR Architecture for UAV-Based Pavement Defect Detection in Complex Civil Infrastructure Scenarios

Uncrewed aerial vehicle (UAV) inspection increasingly is employed in pavement defect detection due to its high mobility and wide-area coverage. However, challenges such as complex backgrounds, small-scale defects, and limited computational resources hinder the effectiveness of existing detection models, often resulting in suboptimal accuracy, higher missed detection rates, and excessive computational costs. To address these issues, this paper proposes Real-Time Detection Transformer for UAV-Based Road Defects (RTDETR-URD), a defect detection model tailored for complex road environments. It integrates three key components—a high-efficiency feature balancing network (EFBNet), a semantic-preserving fusion module (RetBC3), and an adaptive multistage varifocal loss (AMVFL)—jointly enhancing both detection performance and inference efficiency. Experiments on a custom-built, multisource heterogeneous data set demonstrate that RTDETR-URD improves precision, recall, and mean average precision at 0.5 intersection-over-union ([email protected]) by 6.9%, 7.8%, and 6.4%, respectively, while reducing floating-point operation counts (GFLOPs) and parameter count by an average of 21.58%. Under challenging environmental conditions, the model maintains robust performance and precise boundary localization, outperforming mainstream counterparts. These results validate the effectiveness of the proposed architecture and suggest strong engineering adaptability, making it well-suited for UAV-based pavement inspection in real-world civil infrastructure scenarios.