An adaptive UAV-based daily highway inspection method based on vision foundation models and image matching technique

As a maintenance practice, daily highway inspection (DHI) is essential to maintaining pavement performance, ensuring traffic safety, and preserving service quality. In recent years, a combination of unmanned aerial vehicles (UAV) and computer vision techniques has offered an efficient and objective way to realize road inspection. However, UAV-based DHI methods remain limited by training efficiency on diverse inspection objects, classification accuracy under severe class imbalance and adaptability to dynamic datasets. This study proposes a UAV-based DHI method to address the aforementioned technical challenges in real-world engineering scenarios. First, we employ Distillation with No Labels version 3 to extract not only global context but also local features from UAV image tiles. Second, an interest-driven annotation strategy is proposed to address severe class imbalance and annotation scarcity problems. Third, we adopt Hierarchical Navigable Small World for efficient nearest-neighbor retrieval in DHI object classification. Fourth, dense matching utilizes local features to provide a quantitative pavement distress assessment without pixel-level annotations. Our method is validated on UAV data collected from Guyuan, China, over the last two years. Experimental results indicate that the proposed method not only achieves competitive classification performance over all baseline models but also exhibits cross-dataset adaptability. The interest-driven annotation strategy improves training efficiency for all models. Furthermore, the dense matching program provides a reliable quantitative assessment of pavement distress and supports practical DHI. Our adaptive method completes the full DHI workflow for a highway segment at approximately 17.21 min/km, covering UAV data acquisition, annotation, and classification. This efficiency provides a practical way for long-term DHI deployment under real-world highway maintenance scenarios.