A LiDAR-Camera fusion framework assisted by geometric anomaly perception for real-time road subsidence detection

Road subsidence poses a significant threat to traffic safety, yet reliable detection remains challenging when relying solely on computer vision. To enable accurate and timely detection, this study developed a LiDAR-Camera real-time fusion framework and constructed three datasets for model training and evaluation: two road-subsidence image datasets (SubIMG1 and SubIMG2) and a LiDAR-Camera multimodal dataset (SubIMG-PCD). An integrated vehicle-mounted LiDAR–Camera acquisition system was developed, in which point-cloud data were used to support image annotation. For visual detection, we proposed DDS-YOLO, an enhanced model based on YOLOv13 that incorporates deformable convolution (DFC), dynamic sampling (DySample), and spatially adaptive feature modulation (SAFM). These modules improve the detection of subsidence targets characterized by irregular shapes, varying scales, and indistinct boundaries. To provide geometric verification, a road surface model incorporating road design elements was constructed, and a multiscale geometric anomaly detection (MGAD) algorithm was developed to extract physical evidence of subsidence, including depth and spatial continuity. The candidate bounding boxes generated by DDS-YOLO were subsequently matched with anomaly regions identified by MGAD, enabling decision-level confirmation of true subsidence cases while suppressing vision-induced false positives. Experimental results show that DDS-YOLO achieved F1 scores of 0.884 and 0.857 on SubIMG1 and SubIMG2, respectively, while maintaining an inference speed of 238 FPS and outperforming several mainstream detection models. On the SubIMG-PCD dataset, the proposed DDS-MGAD framework achieved the highest F1 score of 0.911 at 14.5 FPS, outperforming representative early- and feature-level fusion methods. The proposed framework provides an efficient, interpretable, and practical solution for rapid road-subsidence detection.