Urban Localization Without Satellite Signals

299 concepts257 papers1 proposal

Global Navigation Satellite Systems (GNSS), such as GPS, are the bedrock of modern positioning. Yet, they possess a fundamental vulnerability: they rely on an unobstructed line of sight to satellites orbiting over 20,000 kilometers away. In dense city centers—often called urban canyons—this geometry collapses. Tall buildings, tunnels, and dense foliage block these weak space-borne signals entirely or reflect them off concrete and glass surfaces. This reflection causes multipath interference, where a receiver mistakes a bounced signal for a direct one, throwing off distance calculations by tens or hundreds of meters.

To achieve the centimeter-level precision required by autonomous vehicles and mobile robotics, a major research direction has shifted away from space-borne trilateration. Instead, modern systems treat the urban environment not as an obstacle, but as a rich, deterministic map of landmarks. By matching local sensory data against pre-existing structural, topological, or terrestrial radio templates, these systems can pinpoint a vehicle's location entirely independent of satellite availability.


Ground-to-Satellite Cross-View Alignment

One of the most elegant ways to bypass a live GPS connection is to use a single ground-level camera and match its feed directly against historical satellite or aerial imagery. The core challenge here is a drastic perspective mismatch: a vehicle's camera captures a horizontal, perspective view of a street, while satellite imagery provides an orthographic, top-down view.

To bridge this gap, researchers design networks that project ground features into a shared embedding space or a unified Bird's-Eye View (BEV) plane. In FG²: Fine-Grained Cross-View Localization by Fine-Grained Feature Matching, Zimin Xia et al. demonstrate that isolating fine-grained features along the vertical height dimension allows a system to align sparse ground points precisely with aerial maps, significantly reducing localization errors.

Rather than treating this alignment as a single-shot classification problem, recent frameworks frame it as an iterative optimization process. For example, in PIDLoc: Cross-View Pose Optimization Network Inspired by PID Controllers, Wooju Lee et al. apply the logic of Proportional-Integral-Derivative (PID) control loops to cross-view matching. By treating the spatial discrepancy between the ground camera's estimated pose and the satellite map as an error signal, the network smoothly steers the estimated coordinates toward the true global position.


Semantic Landmarks as Geometric Anchors

While matching raw pixels or dense 3D point clouds can provide high accuracy, these methods are brittle. A city street looks entirely different at midnight in the rain than it does at noon in the summer. Furthermore, streaming dense 3D maps requires immense memory and bandwidth.

The solution is semantic abstraction. Instead of memorizing every pixel, localization algorithms extract permanent, sparse geometric structures—such as lamp posts, traffic signs, and curbs—that remain invariant across seasons and weather conditions.

[Raw Sensor Data] ---> [Semantic Segmentation] ---> [Sparse Landmark Map]
(LiDAR / Camera)       (Isolate Poles/Curbs)        (Match with HD Map)

In Online Pole Segmentation on Range Images for Long-term LiDAR Localization in Urban Environments, Hao Dong et al. bypass heavy 3D point cloud processing by converting LiDAR data into 2D range images. They train a network to isolate vertical poles (like streetlights and signs) as lightweight, permanent anchors.

However, dense urban traffic frequently occludes these landmarks from a vehicle's line of sight. To combat this, Tarlan Suleymanov et al. show in LiDAR Lateral Localisation Despite Challenging Occlusion from Traffic that deep neural networks can infer the positions of hidden road boundaries and curbs behind surrounding vehicles. By reconstructing these occluded boundaries, the system maintains stable lateral positioning even when the physical landmarks are completely blocked by surrounding traffic.


Trajectory Fingerprinting on Topological Maps

If a vehicle is dropped into an unfamiliar city street without GPS, a single visual frame might be ambiguous—many intersections look identical. However, as the vehicle moves, its sequence of actions creates a unique spatial signature. A trajectory consisting of a 100-meter straight drive, a sharp 90-degree left turn, and a subsequent 50-meter acceleration can typically match only a handful of locations on a city map.

By treating a sequence of turns and distances (visual or inertial odometry) as a "fingerprint," systems can query lightweight, non-visual topological maps like OpenStreetMap to resolve ambiguity over time.

  • Recurrent Trajectory Matching: In Map-Based Temporally Consistent Geolocalization through Learning Motion Trajectories, Bing Zha et al. use Recurrent Neural Networks (RNNs) to process sequential motion patterns. The network tracks the temporal consistency of a vehicle's path, gradually eliminating false candidate locations on a topological map until only the true path remains.
  • Probabilistic Self-Localization: This concept is mirrored in the crowd-sourced framework proposed in Lost! Leveraging the Crowd for Probabilistic Visual Self-Localization. The system treats noisy visual odometry as a continuous probabilistic distribution over a road network, rapidly converging on the correct coordinates within a few seconds of driving without any initial position prior.

Terrestrial Radio Infrastructure

When visibility is severely degraded by extreme weather, fog, or darkness, optical and geometric sensors degrade. In these scenarios, researchers exploit existing terrestrial radio infrastructure, such as 5G and LTE cellular networks, as a substitute for satellite constellations.

Terrestrial cellular base stations offer two distinct advantages over satellites: they are much closer to the ground, providing significantly higher signal-to-noise ratios, and their positions are fixed and known with millimeter precision.

Feature Satellite Navigation (GNSS) Terrestrial Radio (5G/LTE)
Signal Source Space-borne (20,000 km away) Ground-based cell towers
Signal Strength Weak, easily blocked by buildings Strong, penetrates urban structures
Primary Failure Mode Multipath reflections in canyons Non-line-of-sight (NLOS) delays

To utilize these signals for precise positioning, systems must map how radio waves propagate through complex urban geometry. In LocUNet: Fast Urban Positioning Using Radio Maps and Deep Learning, Çağkan Yapar et al. utilize a convolutional neural network to predict Received Signal Strength (RSS) radio maps. By comparing a mobile device's live signal profile against this neural radio map, the system can instantly deduce its location.

To push terrestrial positioning to sub-meter accuracy, researchers are moving away from simple signal strength indicators toward direct signal processing. In Cramér-Rao Bound for Direct Position Estimation in OFDM Based Cellular Systems, Sijia Li et al. demonstrate the theoretical limits of Direct Position Estimation (DPE) using cellular signals. By processing the raw radio waveforms directly at the transmission level rather than using a traditional two-step arrival-time calculation, DPE mitigates the severe multipath distortions that typically cause terrestrial radio positioning to fail in deep urban canyons.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 247 papers
  2. Version 1 · 2026-10-04 · 254 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_satellite_signal_occlusion_d246239e,
  title = {satellite signal occlusion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/satellite-signal-occlusion/rdr_d246239ecaa14b669e4313ab5fc49522},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_satellite_signal_occlusion_d246239e_v2,
  title = {satellite signal occlusion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/satellite-signal-occlusion/rdr_d246239ecaa14b669e4313ab5fc49522/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake