A multi-frame UAV video inpainting and deep learning framework for vehicle-occlusion-free pavement distress detection

Unmanned Aerial Vehicles (UAVs) are increasingly used for non-contact pavement condition assessment because they enable rapid and high-resolution data collection over large road networks. However, during continuous UAV video acquisition under active traffic conditions, moving vehicles frequently occlude the pavement surface, resulting in incomplete visibility of distress features and reduced reliability of automated detection. To address this limitation, this study proposes a spatio-temporal vehicle-occlusion removal framework for UAV-based pavement distress assessment. The framework integrates frame differencing, motion-based filtering, morphological processing and temporal inpainting to reconstruct pavement regions temporarily hidden by moving vehicles, after which a YOLO-based model is applied for automated distress detection. The approach was validated using UAV video collected over a 5 km National Highway stretch and compared against manual field observations. Quantitative evaluation showed that the proposed preprocessing significantly improved the agreement between automated and manual distress counts. For cracks, the coefficient of determination (R 2) increased from 0.82 to 0.95 and the RMSE decreased from 2.14 to 0.70 after occlusion removal. Similar improvements were observed for potholes (R 2: 0.86 to 0.97; RMSE: 1.82 to 0.42) and patchwork (R 2: 0.86 to 0.98; RMSE: 1.57 to 0.27). These findings demonstrate that temporal inpainting can substantially improve the completeness and reliability of UAV-based pavement distress detection under real traffic conditions, thereby enhancing the practical applicability of UAV systems for pavement monitoring and infrastructure management.