Large-Scale Geometric Map-Based Localization of UAVs in GNSS-Denied Urban Environments

Garth Terlizzi; Kaveh Fathian

cs.ROSep 23, 2026

Large-Scale Geometric Map-Based Localization of UAVs in GNSS-Denied Urban Environments

Authors: Garth Terlizzi, Kaveh Fathian

Organizations: Department of Computer Science, Colorado School of Mines

Abstract

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 $^2$ and 254 km $^2$ , and 71.4% Recall@1 when expanded to approximately 452 km $^2$ , encompassing up to 277,000 buildings. In contrast, baseline methods degrade rapidly and achieve 0% Recall@1 at 254 km $^2$ and 452 km $^2$ .

Figures & tables

Fig. 1 : Successful localization example using our proposed approach, across areas spanning ${\approx}113$ km 2 (blue circle), ${\approx}254$ km 2 (orange circle), and ${\approx}452$ km 2 (green circle). Our novel star descriptors, generated from observed buildings, are matched across the UAV and reference maps for localization. The UAV map contains only $104$ observed buildings, yet is successfully matched against up to $272{,}878$ reference buildings, demonstrating the large-scale capability of our method.

Fig. 2 : Overview of the proposed localization pipeline. Aerial keyframes are segmented for buildings, stitched into an internal map, triangulated, and matched against a reference building footprint database via star descriptors and consensus voting.

Panel of Figure 3 — Fig. 3 : Effect of a missed detection on the Delaunay triangulation. Left: ground truth from Microsoft Building Footprints. Right: UAV -segmented map with minor error (jitter and distortion) and a missed detection. Red triangles are significantly altered by the missed detection, while green triangles retain their shape, highlighting Delaunay’s neighborhood-preservation behavior under local segmentation noise.

Feature	Definition	Dim
Triangle geometry
Sorted angles	$\alpha_{1}\leq\alpha_{2}\leq\alpha_{3}$	$3$
Perimeter	$P=\sum\\|e_{i}\\|$	$1$
Building shape (avg.)
Compactness	$4\pi A\,/\,L^{2}$	$1$
Elongation	$w^{\min}/w^{\max}$	$1$

TABLE I : Star descriptor features. Each triangle (central and three neighbors) is described by eight features: sorted interior angles and perimeter from triangle geometry, plus four building-shape metrics averaged over the three vertex buildings.

Fig. 4 : Example star descriptors. The central triangle (consisting of the dark buildings) and its adjacent neighbors form a local geometric pattern that encodes both spatial arrangement and per-building shape features.

Flight #	Length (km)	Alt. (m)	Keyframes	# Buildings Detected	Environment
$1$	$0.39$	$108$	$20$	$30$	C
$2$	$1.68$	$116$	$67$	$78$	C+S
$3$	$2.19$	$105$	$101$	$133$	S
$4$	$3.18$	$115$	$256$	$104$	R
$5$	$2.26$	$98$	$76$	$87$	C+S
$6$	$2.85$	$106$	$181$	$121$	C+S

TABLE II : Summary of UAV flight trajectories used for evaluation. In our environment labels, suburban refers to single-family homes, while residential refers to multi-family homes and apartment complexes.

Panel of Figure 7 — Fig. 5 : Flight trajectories overlaid on reference building footprint maps. (a)–(g) show flights $1$ – $7$ in Table II .

Flight #	# Reference Buildings in Search Radius			Coarse-Stage Error (m)			Final Localization Error (m)
Flight #	$R{=}6$ km	$R{=}9$ km	$R{=}12$ km	$R{=}6$ km	$R{=}9$ km	$R{=}12$ km	$R{=}6$ km	$R{=}9$ km	$R{=}12$ km
$1$	$18{,}711$	$51{,}452$	$99{,}484$	$22.9$	$22.9$	$22.8$	$7.3$	$7.3$	$7.6$
$2$	$49{,}962$	$120{,}434$	$201{,}628$	$45.0$	$47.4$	$47.4$	$25.4$	$32.9$	$31.3$
$3$	$17{,}987$	$52{,}290$	$101{,}521$	$99.9$	$102.5$	$105.3$	$40.0$	$38.2$	$40.7$
$4$	$84{,}318$	$178{,}164$	$272{,}878$	$27.6$	$27.4$	$27.3$	$20.5$	$17.7$	$17.7$
$5$	$78{,}849$	$159{,}946$	$260{,}750$	$153.7$	$156.1$	–	$22.1$	$30.3$	–

TABLE III : Reference-map building count (within each search radius), coarse-stage error, and final localization error across search radii. A trial is counted as successful if the final estimate lies within

100

m of ground truth. “–” indicates localization failure at flight conclusion.

Fig. 6 : Qualitative localization example for Flight $7$ at the $6$ km search radius ( ${\approx}113$ km 2 ), covering $64,662$ buildings. For each method, the top- $3$ ranked candidates are shown. Our method correctly identifies the true location as its top candidate, while BRM and NetVLAD do not recover the correct location in their top- $3$ candidates. The geometry-only ablation (w/o shape) ranks the correct location lower, illustrating how shape features improve candidate discrimination.

Method	Metric	$R{=}$ $1$ km	$R{=}$ $3$ km	$R{=}$ $6$ km	$R{=}$ $9$ km	$R{=}$ $12$ km
NetVLAD	R@ $1$	$5/7$	$2/7$	$0/7$	$0/7$	$0/7$
NetVLAD	R@ $5$	$\mathbf{7/7}$	$3/7$	$0/7$	$0/7$	$0/7$
NetVLAD	R@ $10$	$\mathbf{7/7}$	$3/7$	$0/7$	$0/7$	$0/7$
BRM	R@ $1$	$6/7$	$1/7$	$1/7$	$0/7$	$0/7$
BRM	R@ $5$	$6/7$	$2/7$	$1/7$	$0/7$	$0/7$
BRM	R@ $10$	$6/7$	$3/7$	$1/7$	$0/7$	$0/7$

TABLE IV : Comparison with baselines and ablation across search radii with a

100

m correctness threshold. Entries are Recall@

K

values reported as successful flights out of

7

. Best results per column are shown in bold.

Fig. 7 : Progressive coarse localization at $R{=}6$ km. Each trace shows the best rank among the ground-truth bin and its 8-connected neighboring bins among $11{,}310$ candidates as keyframes accumulate from $1$ to $N$ . Only rank- $1$ constitutes coarse-stage success. Star markers denote each flight’s first stable rank- $1$ point. Fluctuations arise from false-positive or missed building detections that perturb the internal Delaunay topology and alter scoring.

GNSS-Denied Localization

Global Navigation Satellite System (GNSS)-denied localization estimates position when satellite signals are unavailable or unreliable.

See all 57 papers

Unifying UAV Cross-View Geo-Localization via 3D Geometric Perception
Apr 2, 2026Haoyuan Li, Wen Yang, Fang Xu +4Cross-View Geo-Localization GNSS-Denied Localization
School of Electronic Information, Wuhan University, Wuhan 430072, China · School of Artificial Intelligence, Wuhan University, Wuhan 430072, China · Technology and Engineering Center for Space Utilization and the Key Laboratory of Space Utilization, Chinese Academy of Sciences, Beijing 100094, China +1
AeroMap3D: Anchoring Monocular UAV 6-DoF Localization to Visual-Geometric-Semantic Map Priors
Jul 15, 2026Zhiyun Deng, Luis SentisCross-View Geo-Localization Image Matching
The University of Texas at Austin · AIVE AI Systems
SASGeo: Stability-Aware Semantic Map Localization for GNSS-Denied UAVs -- A Framework and Synthetic Proof of Concept
Jul 7, 2026Natalia Trukhina, Vadim VashkelisCross-View Geo-Localization UAV Navigation
Embedded Intelligence Lab (emilab.org)

Apr 2, 2026J/K move · Enter open · S save

Unifying UAV Cross-View Geo-Localization via 3D Geometric Perception

Haoyuan Li, Wen Yang, Fang Xu, Hong Tan+3

School of Electronic Information, Wuhan University, Wuhan 430072, China · School of Artificial Intelligence, Wuhan University, Wuhan 430072, China · Technology and Engineering Center for Space Utilization and the Key Laboratory of Space Utilization, Chinese Academy of Sciences, Beijing 100094, China+1

Cross-View Geo-Localization GNSS-Denied Localization UAV Navigation Camera Pose Estimation

Open paper AskPDF arXiv

Cross-view geo-localization for Unmanned Aerial Vehicles (UAVs) operating in GNSS-denied environments remains challenging due to the severe geometric discrepancy between oblique UAV imagery and orthogonal satellite maps. Most existing methods address this problem through a decoupled pipeline of place retrieval and pose estimation, implicitly treating perspective distortion as appearance noise rather than an explicit geometric transformation. In this work, we propose a geometry-aware UAV geo-localization framework that explicitly models the 3D scene geometry to unify coarse place recognition and fine-grained pose estimation within a single inference pipeline. Our approach reconstructs a local 3D scene from multi-view UAV image sequences using a Visual Geometry Grounded Transformer (VGGT), and renders a virtual Bird's-Eye View (BEV) representation that orthorectifies the UAV perspective to align with satellite imagery. This BEV serves as a geometric intermediary that enables robust cross-view retrieval and provides spatial priors for accurate 3 Degrees of Freedom (3-DoF) pose regression. To efficiently handle multiple location hypotheses, we introduce a Satellite-wise Attention Block that isolates the interaction between each satellite candidate and the reconstructed UAV scene, preventing inter-candidate interference while maintaining linear computational complexity. In addition, we release a recalibrated version of the University-1652 dataset with precise coordinate annotations and spatial overlap analysis, enabling rigorous evaluation of end-to-end localization accuracy. Extensive experiments on the refined University-1652 benchmark and SUES-200 demonstrate that our method significantly outperforms state-of-the-art baselines, achieving robust meter-level localization accuracy and improved generalization in complex urban environments.

Abstract

GNSS-Denied Localization

Explore similar work

Unifying UAV Cross-View Geo-Localization via 3D Geometric Perception

AeroMap3D: Anchoring Monocular UAV 6-DoF Localization to Visual-Geometric-Semantic Map Priors

SASGeo: Stability-Aware Semantic Map Localization for GNSS-Denied UAVs -- A Framework and Synthetic Proof of Concept