UAV coarse visual localization in large-scale continuous scenes

Visual localization in GNSS-denied environments is a critical technology for the autonomous navigation of UAVs. In practical missions, UAVs are typically required to execute continuous flights over large-scale areas. However, most existing studies construct small-scale satellite reference databases based on one-to-one aligned discrete sampling, which largely ignores the challenges introduced by continuous scenes. Although recent works have attempted to address this by constructing synthetic datasets with continuous sampling, a substantial domain gap remains between virtual and real-world scenarios. To address these issues, we propose a new UAV visual localization dataset, XIAN-Visloc, which contains over 80 km of continuous UAV flight sequences. Based on this dataset, we explicitly formulate the task of UAV coarse visual localization and present a comprehensive localization pipeline. At the data preprocessing stage, we investigate the construction of satellite image libraries for large-scale continuous scenes and propose an adaptive method that balances computational cost and localization accuracy. Furthermore, satellite positive samples are redefined based on geographic distance. At the model level, we propose a new baseline model based on a comprehensive study of training samples, sampling strategy, loss function, and retrieval method, aiming to enhance performance in continuous visual localization tasks. In addition, we evaluate the proposed method on the XIAN-Visloc and UAV-Visloc datasets using multiple retrieval and localization metrics. Comprehensive ablation studies are performed to assess the contribution of each component. Finally, real-world experiments demonstrate that, with pixel-level feature matching, the proposed method achieves a localization error of 15.07 m within a candidate area exceeding 50 km2[jls-end-space/], validating the practical effectiveness. Our code and dataset are available at https://github.com/VERYBC/UCVL.