Research on vision-based autonomous positioning of UAV in low-altitude environment

Abstract

Unmanned Aerial Vehicles (UAVs) commonly rely on GNSS for absolute positioning, but GNSS signals may be unreliable in low-altitude urban, campus, or interference-prone environments. This thesis studies vision-based UAV self-positioning using UAV-satellite cross-view geo-localization. The task is formulated as image retrieval, which means a UAV-view image is matched with satellite-view or reference images. And a method named DinoUAV-Loc is proposed based on DINOv2 method. It uses a shared Siamese structure, combines CLS token global features with patch-level local features, and applies cosine similarity for matching. Experiments are conducted on DenseUAV, DenseUAV-Hui, and a self-built field dataset. The method achieves 90.00% R@1 on DenseUAV, 87.95% R@1 on DenseUAV-Hui, and 96.35% R@1 on the self-built dataset. The results indicate that the method can support retrieval-based UAV positioning in low-altitude environments.