Reliable autonomous localization is crucial for unmanned aerial vehicles (UAVs) to perform military tasks such as reconnaissance and surveillance. However, in battlefield environments where Global Navigation Satellite System (GNSS) signals are denied, the UAV's localization capability faces significant challenges. To address this, this paper proposes a novel visual localization method aimed at improving the robustness of UAVs under varying viewpoints, lighting, seasonal, and weather conditions. The core idea of the method is based on a multi-task learning framework combined with domain adaptation techniques, utilizing synthetic data containing depth and semantic information to transfer knowledge to real-world environments. By fusing geometric and semantic cues, we generate multi-scale feature representations with strong robustness to environmental changes, enabling precise image retrieval and position recognition. To mitigate the high cost of obtaining real-world data labels, high-quality pseudo-labels generated from synthetic data are used for supervised training. Extensive experimental results on public datasets and transferred real-world datasets demonstrate that the proposed method can effectively address visual localization challenges in complex battlefield environments. Furthermore, by integrating the pre-trained domain-adaptive features with the original baseline features, significant improvement in retrieval performance is achieved without the need for transfer training.
Liu et al. (Wed,) studied this question.