Drone-Based Vehicle-Aware Spatiotemporal Reliability Fusion for Bridge Deck Pavement Inspection Under Normal Traffic

Vehicle occlusion disrupts bridge deck pavement inspection when unmanned aerial vehicle (UAV) video is acquired under normal traffic. This study addresses how temporally repeated UAV observations can be reliably integrated for vehicle-occluded pavement reconstruction and introduces a novel Vehicle-Aware Spatiotemporal Reliability Fusion (VSRF) framework that preserves directly observed surface details for quantitative inspection. Vehicle detections are converted into dilated occlusion masks and combined with temporal appearance stability; after panorama-coordinate registration, the highest-reliability unoccluded observations are transferred across the reconstructed pavement surface and blended locally at source-transition boundaries. The reconstructed mosaics provide the image domain for crack segmentation and reference-scaled crack-length estimation. Evaluation used 4020 sampled frames acquired from five bridges. The resulting 177.5–212.0-megapixel panoramas maintained continuous lane-scale coverage while retaining visible pavement texture, markings, joints, and crack-related details. Across 792 held-out cross-frame recovery pairs, transferred observations achieved a mean PSNR of 30.15 dB and a mean SSIM of 0.9612 against independently retained clear references. Field-referenced crack-length estimates obtained from the VSRF mosaics yielded a mean absolute percentage error of 12.6%, compared with 24.0% for median fusion. These findings demonstrate the potential of reliability-guided observation transfer for efficient, traffic-tolerant bridge deck inspection and establish a foundation for extending observation-grounded reconstruction to broader infrastructure monitoring scenarios.