Cross-Modal Transformer Fusion for UAV-Ground Integrated Road Surface Anomaly Detection and Severity Quantification

Growing demands on transportation infrastructure have increased the need for automated and reliable solutions for detecting and assessing road surface defects. This work introduces an AI-driven multimodal framework that integrates unmanned aerial vehicle (UAV) imagery with ground-level visual data to support accurate defect detection and quantitative severity estimation. The proposed architecture combines an enhanced YOLOv8 backbone with a Cross-Modal Fusion Transformer (CMFT), which enables effective alignment of aerial and ground-view features through attention-based spatial reasoning. In addition, a dedicated severity estimation head refines localized representations to categorize surface damage into minor, moderate, and severe levels. Experimental evaluation on a custom multimodal dataset shows that the proposed method consistently outperforms unimodal baselines in terms of detection accuracy and robustness.