Unmanned Aerial Vehicle (UAV)-Based Pavement Image Stitching Without Occlusion, Crack Semantic Segmentation, and Quantification

Unmanned Aerial Vehicle (UAV)-based pavement distress detection offers efficient and safe advantages. However, obstructions from road vehicles and the slender shape of cracks in UAV images challenge accuracy. To address this, this study established specific flight parameters, proposed the Historical Best Matching Image (HBMI) approach for data loss due to obstructions, and created the UAV-Crack500 dataset with 500 finely annotated crack images. Three algorithms with different loss functions were investigated, finding that the U-Net network combined with our Completely Asymmetric Loss (CAL) achieved the best performance, resolving the issue of class imbalance. Morphological analysis of the semantically segmented images provided precise crack morphology features. In complex scenarios, errors in features like crack area, length, mean width, and maximum width remained within 16%. This study establishes a comprehensive UAV-based pavement distress detection system, overcoming obstructions for accurate assessment.