A Ground-Penetrating Radar Target Detection Method Using Dual-Attribute Feature Fusion

To minimize the disruption of detection activities on urban traffic and reduce interference from environmental factors, subsurface defect detection in urban roads using vehicle-mounted ground-penetrating radar (GPR) is typically conducted during nighttime hours, when traffic volume is low. This imposes stringent requirements on the accuracy and real-time interpretation of the detection results. Traditional manual interpretation methods, which rely on human expertise, struggle to meet the requirements for accurate and rapid interpretation of vehicle-mounted GPR data. Although existing automatic target detection approaches based on single-modal B-Scan images can effectively and rapidly identify prominent subsurface targets such as pipelines and voids, they often fail to detect or misidentify weak or ambiguous targets. To address these limitations, we propose a subsurface target detection framework Fuse-DETR, which is based on the RT-DETR architecture and integrates dual-modal image fusion. It uses B-Scan images and Instantaneous Amplitude (IA) images as two input branches, making use of their complementary features to improve detection of complex underground targets. Under the premise of ensuring annotation accuracy through on-site drilling verification, the first multimodal GPR dataset for urban roads was constructed, consisting of paired B-Scan and IA images. Comparative experiments and ablation studies were conducted on this dataset. The experimental results show that Fuse-DETR achieves a recall of 89.8%, an F1-Score of 85.6%, and an mAP of 81.16%, outperforming other single-modality methods.