Large-Scale Geometric Map-Based Localization of UAVs in GNSS-Denied Urban Environments
In the authors' words
Unmanned aerial vehicles (UAVs) operating in GNSS-denied urban environments require alternative methods for position estimation. Existing approaches based on satellite image retrieval or learned descriptors are sensitive to appearance variation and degrade rapidly as the search area grows. We present a vision-based localization system that matches building patterns observed from a downward-facing UAV camera against a reference building footprint database. Our approach detects buildings in aerial imagery, accumulates observations across frames into a unified map, and matches local building arrangements against reference footprints using a novel geometry-driven descriptor that augments local triangle structure with per-building shape features. By encoding spatial relationships between nearby buildings rather than visual appearance, the system is robust to appearance variations and remains discriminative over large search areas. Evaluations on seven flights across four municipalities in a large metropolitan area demonstrate 100% Recall@1 at search areas of approximately 113 km and 254 km, and 71.4% Recall@1 when expanded to approximately 452 km, encompassing up to 277,000 buildings. In contrast, baseline methods degrade rapidly and achieve 0% Recall@1 at 254 km and 452 km.
Appeared: Thursday, September 24. arXiv. Preprint, not yet peer-reviewed.
Authors' comment: Accepted IROS 2026