3D Map-Aided Visual Localization for UAVs in Mid-to-Low Altitude Environment

Accurate 6-DoF pose estimation for Unmanned Aerial Vehicles (UAVs) in Global Navigation Satellite System (GNSS)-denied environment is paramount, and visual localization algorithms have emerged as a highly effective and widely adopted paradigm. However, this approach is proved ineffective in low-altitude scenario due to the repetitive structures (e.g., buildings, grasslands), which introduces visual ambiguities. To overcome it, we propose a novel hierarchical deep learning-based visual localization system for robust 6-DoF pose estimation in low-altitude scenario. Our approach first constructs a comprehensive multi-view database from high-precision 3D maps. This database is then utilized by a coarse localization module for efficient scene retrieval. Subsequently, a dedicated module performs fine-grained visual matching, leveraging the wider scene context to accurately estimate the full 6-DoF pose. The hierarchical framework significantly enhances localization accuracy and robustness, demonstrating superior performance in challenging low-altitude environment where traditional methods struggle.