UAVs Navigate Without GPS Using Buildings as Visual Landmarks

A Rotation and Scale Invariant Map Matching Method for UAV Visual Geolocalization

BLUF: Chinese researchers have developed a rotation and scale-invariant map matching system that enables UAVs to pinpoint their location in urban environments by matching building patterns from their cameras to digital maps, achieving accuracy within 11 meters even when GPS is unavailable or jammed.

---

When GPS signals fail—whether from deliberate jamming, urban canyons, or natural interference—unmanned aerial vehicles face a critical navigation challenge. Now, a research team from Xidian University and Hunan University has demonstrated a solution that leverages the one thing cities have in abundance: buildings.

Published in the IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing in December 2025, the Rotation and Scale Invariant Map Matching (RSIM) method represents a significant advance in GPS-denied navigation for urban UAVs. The system can determine a drone's position by analyzing the geometric patterns of buildings visible in its camera feed and matching them against vector map data.

How Buildings Become Navigation Beacons

The RSIM approach transforms urban landscapes into natural coordinate systems. Rather than requiring pre-collected aerial imagery—which demands massive onboard storage and struggles with seasonal or lighting changes—the system uses lightweight vector maps showing only building outlines and locations.

"Spatial scenes are most accurately characterized by spatial relationships among buildings rather than their geometric features," the researchers explain in their paper. The team, led by Yu Liu and colleagues, developed an algorithm that identifies buildings in UAV camera images, then analyzes the triangular patterns formed by any three building centers.

The innovation lies in using geometric relationships that remain constant regardless of how the image is rotated or scaled. When a UAV's orientation or altitude is unknown—common scenarios when inertial measurement units drift or fail—traditional matching methods collapse. RSIM sidesteps this by focusing on angular relationships and relative distances between buildings, features that persist regardless of camera perspective.

The system extracts several key characteristics: the angles formed by triangles connecting three building centers, the aspect ratios of each building's footprint, and the relative distances between structures. These measurements create a unique "fingerprint" for each location that can be matched against digital map databases.

Testing Against Deep Learning Competitors

To validate their approach, the researchers conducted extensive trials using simulated UAV imagery from Shijiazhuang and Xi'an, China. They compared RSIM against two prominent deep learning-based image matching systems: Local Feature Transformers (LoFTR) and Dense Kernel Feature Matching (DKM).

The results revealed stark performance differences. When tested on Xi'an datasets without image rotation, LoFTR achieved an average localization error of 6.92 meters, while DKM reached 6.77 meters. However, when the same images were randomly rotated—simulating real-world scenarios where UAV orientation is uncertain—LoFTR's error jumped to 122.19 meters and DKM's ballooned to 535.23 meters.

RSIM, by contrast, maintained consistent performance regardless of rotation, achieving average errors of 10.78 meters in Shijiazhuang and 11.23 meters in Xi'an across all test conditions. The system proved particularly robust compared to its predecessor, the Shape and Spatial Relationship Matching (SSRM) method, which achieved slightly better accuracy (7.38 meters) but only when image orientation and scale were known in advance.

"Current image-based matching methods are easily affected by significant feature differences between UAV images and georeferenced images," the researchers note, "and map-based matching methods are affected by image rotation and resolution differences."

The Building Extraction Challenge

The system's foundation rests on accurately identifying individual buildings from UAV camera footage—a computer vision challenge addressed using SOLOv2, an instance segmentation neural network. Trained on 3,000 aerial images, the model achieved 67.4% precision in correctly identifying and outlining buildings.

This extraction accuracy directly impacts navigation performance. The researchers conducted sensitivity analyses revealing that the system tolerates building omission rates up to 20% while maintaining reliable localization. When up to 10% of building boundaries contained pixel-level distortions, accuracy remained largely unaffected—a crucial resilience factor for real-world deployment where perfect building extraction is impossible.

Interestingly, false building detections had minimal impact. The triangular matching algorithm naturally filters out spurious detections because they fail to form consistent geometric relationships with actual structures in the reference map.

Processing Speed and Practical Deployment

Computational efficiency matters for airborne systems operating under power and weight constraints. Running on an Intel i7-7700 CPU with an Nvidia Tesla V100 GPU, RSIM processed each frame in an average of 1.55 seconds—0.46 seconds for building extraction and 1.09 seconds for scene matching.

This represents a significant speed advantage over SSRM (3.59 seconds per frame) while being slower than pure deep learning approaches like LoFTR (0.278 seconds) and DKM (1.236 seconds). However, the deep learning methods' catastrophic failure under rotation makes their speed advantage academic for GPS-denied navigation scenarios.

The system's map data requirements are remarkably modest. The Shijiazhuang dataset covering 786 square kilometers required only 21.1 megabytes of storage, while Xi'an's 520 square kilometers consumed 17.3 megabytes—a dramatic reduction compared to the 2.8 megabytes needed for each 1 square kilometer of georeferenced satellite imagery used by competing image-matching systems.

Limitations and Urban Applicability

RSIM's effectiveness depends on building density and pattern uniqueness. In areas with sparse buildings or highly repetitive layouts—such as suburban subdivisions with identical tract housing—the system may struggle to establish unique location signatures.

The researchers acknowledge this constraint but note that "large areas with highly repetitive building layouts are extremely rare" in Chinese cities based on their empirical analysis of satellite imagery. They suggest that flight planning can avoid such areas when selecting matching zones.

The system also requires reasonably accurate vector maps. Discrepancies between map data and actual building locations introduce errors, though the research demonstrates tolerance for moderate map inaccuracies.

Current testing used simulated UAV imagery from Google Earth rather than actual drone footage, leaving questions about performance with real-world image quality issues like motion blur. The researchers acknowledge this limitation, noting that image quality assessment algorithms could filter problematic frames before processing.

GPS-Denied Navigation Context

The drive to develop GPS-independent navigation systems reflects growing concerns about satellite signal vulnerability. Military operations increasingly confront sophisticated jamming and spoofing, while civilian applications face interference in urban canyons, tunnels, and indoor environments.

Traditional alternatives like inertial navigation systems accumulate error over time—position errors of 50 meters within two minutes and heading errors of 10 degrees within ten minutes for small UAV-mounted IMUs, according to the research. Visual localization methods avoid this error accumulation by establishing absolute position references at each measurement.

Previous GPS-denied UAV navigation research has explored various approaches, from road intersection matching to semantic scene understanding using deep neural networks. The trend toward map-based rather than image-based matching reflects practical constraints: smaller data storage requirements and inherent robustness to lighting variations that plague image-to-image matching.

Future Developments

The research team suggests several avenues for improvement. Enhanced building extraction accuracy through more sophisticated deep learning models could increase matching precision. In areas with sparse buildings, integrating additional features like road intersections might provide supplementary navigation references.

The triangular matching approach could potentially extend beyond buildings to any consistently identifiable urban features with stable spatial relationships. The fundamental principle—using geometric invariants rather than appearance features—offers a template for navigation solutions in diverse environments.

As UAV applications expand from military reconnaissance to commercial delivery and urban air mobility, reliable navigation in GPS-denied or GPS-degraded environments becomes increasingly critical. RSIM demonstrates that solutions need not always involve more sophisticated sensors or larger neural networks—sometimes, the answer lies in cleverly exploiting the geometric structure already present in the environment.

---

Verified Sources

1. Liu, Y., Bai, J., Xiao, Z., Lian, Y., & Jiao, L. (2026). A Rotation and Scale Invariant Map Matching Method for UAV Visual Geolocalization. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 19, 1616-1627. https://doi.org/10.1109/JSTARS.2025.3639900

2. Couturier, A., & Akhloufi, M.A. (2021). A review on absolute visual localization for UAV. Robotics and Autonomous Systems, 135, 103666. https://doi.org/10.1016/j.robot.2020.103666

3. Couturier, A., & Akhloufi, M.A. (2024). A review on deep learning for UAV absolute visual localization. Drones, 8(11), 622. https://doi.org/10.3390/drones8110622

4. Sun, J., Shen, Z., Wang, Y., Bao, H., & Zhou, X. (2021). LoFTR: Detector-free local feature matching with transformers. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8918-8927. https://doi.org/10.1109/CVPR46437.2021.00881

5. Edstedt, J., Athanasiadis, I., Wadenbäck, M., & Felsberg, M. (2023). DKM: Dense kernelized feature matching for geometry estimation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 17765-17775. https://doi.org/10.1109/CVPR52729.2023.01704

6. Wang, H., Zhou, F., & Wu, Q. (2024). Accurate vision-enabled UAV location using feature-enhanced transformer-driven image matching. IEEE Transactions on Instrumentation and Measurement, 73, 5502511. https://doi.org/10.1109/TIM.2024.3370777

7. Chen, Y., & Jiang, J. (2024). An oblique-robust absolute visual localization method for GPS-denied UAV with satellite imagery. IEEE Transactions on Geoscience and Remote Sensing, 62, 5601713. https://doi.org/10.1109/TGRS.2024.3367891

8. Wang, X., Zhang, R., Kong, T., Li, L., & Shen, C. (2020). SOLOv2: Dynamic and fast instance segmentation. Advances in Neural Information Processing Systems, 33, 17721-17732. https://proceedings.neurips.cc/paper/2020/hash/cd3afef9b8b89558cd56638c3631868a-Abstract.html