Huei‐Yung Lin, Liyu Chen
The localization of UAVs (unmanned aerial vehicles) is crucial for their safe, efficient, and accurate operation. Current developments are mostly based on GNSS/GPS signals for outdoor navigation. However, these technologies are vulnerable under the scenarios where the signals are interrupted or unavailable. This paper presents a GNSS-denied UAV localization technique using aerial and satellite image matching. In our proposed end-to-end network, the CLIP model is utilized to extract and correlate the geolocation of UAV acquired images and an existing satellite map. The image orientation is then incorporated for feature similarity computation to derive the flight heading information. To validate the model performance, an aerial image dataset is collected from the UAV flying at the altitude of 100 meters above the sea level. The evaluation conducted in a 2.23k m 2 region with the location and heading errors of 39.2m and 15.9 ∘ shows the feasibility of the proposed GNSS-denied UAV localization method. The codes and dataset are available at https://github.com/codebra721/CLIP-UAV-localization • A passive positioning technique for outdoor UAV navigation with aerial-satellite image matching is proposed. • The CLIP (contrastive language-image pretraining) model is adopted and improved for vision-based localization. • An end-to-end model is presented under the unified UAV positioning framework. • An aerial image dataset for outdoor UAV localization is created and made available publicly.