Wireframe-guided 6-DoF aerial visual localization with Level-of-Detail 3D city maps

Aerial visual localization (AVL) is a fundamental yet challenging problem for the autonomy of unmanned aerial vehicles (UAVs) when GNSS signals are unreliable. However, current AVL research typically depends on texture-rich 3D representations, resulting in high storage requirements, considerable maintenance costs, and privacy concerns. To address these challenges, we propose a new method named WGVL for visual localization in the air. Unlike existing AVL algorithms, WGVL estimates the UAV’s 6-degree-of-freedom (6-DoF) pose using only a Level-of-Detail (LoD) 3D map by aligning wireframes projected from the LoD map with those predicted by a neural network. Specifically, given a query image, a prior pose, and a LoD city map, WGVL consists of three modules: feature extraction, coarse pose estimation, and fine pose refinement. The feature extraction module extracts wireframe features from the query image and feeds them into the next module. The coarse pose estimation module hierarchically constructs a cost volume over uniformly sampled pose hypotheses to approximate the pose likelihood distribution and selects the most probable pose. Finally, the fine pose refinement module further improves the selected pose via a differentiable 6-DoF optimization based on the Gauss–Newton method. Due to the lack of public datasets for this task, we construct two datasets with LoD 3.0 and LoD 2.0 maps, respectively, each containing real-world RGB and thermal infrared images, along with ground-truth pose annotations. Extensive experiments demonstrate that WGVL achieves strong localization accuracy and offers a practical solution for global aerial visual localization.