UAV Imagery for Ground Vehicle Visual Localization: Reference Database Construction versus Query-Stage Matching

Accurate localization of unmanned ground vehicles (UGVs) is difficult in global navigation satellite system (GNSS)-denied urban environments, and high-definition maps involve substantial construction and maintenance costs.Image-based reference maps have been studied as a cost-effective alternative, and the integration of unmanned aerial vehicle (UAV) imagery with ground-level imagery has been reported to improve localization performance.However, it remains unclear whether this improvement originates from the construction of a geometrically stable reference database or from direct matching between aerial references and ground-level query images at the localization stage.This study quantitatively decomposed the contribution of UAV imagery into these two stages.Four reference-database configurations were evaluated using an identical query-localization pipeline: aerial-only (A), ground-only (G), integrated reconstruction queried with ground-level references only (G*), and fully integrated (GA).Experiments were conducted along an approximately 2 km urban route using 813 UAV images, 498 ground-level reference images, and 273 query images acquired in an independent session under different seasonal conditions.Including aerial imagery in the construction of the reference database reduced the checkpoint root mean square error (RMSE) by 35% (from 0.82 m to 0.53 m).Recall@1 m increased from 58.2% for G to 96.3% for G*, while adding aerial references at the query stage produced no further improvement under the default 10 m ground-reference spacing (96.3% for GA).Thus, the 38.1-percentage-point gain was attributable almost entirely to aerial-supported reference database construction, whereas direct querystage matching with aerial references provided only a secondary benefit under extremely sparse groundreference conditions.Multi-frame refinement substantially reduced coverage-related failures but did not eliminate positional errors associated with distortions in the reference-database geometry.These results indicate that UAV imagery is more valuable for constructing geometrically stable reference databases than as a direct matching source for ground-level queries, potentially reducing the density of ground-level references required for UGV localization.