⚡ electrical engineering

Evaluation of Grid-based Uncertainty Propagation for Collaborative Self-Calibration in Indoor Positioning Systems

This paper presents an empirical evaluation of an enhanced grid-based Bayesian collaborative self-calibration algorithm for indoor UWB networks, demonstrating its ability to achieve sub-meter positioning accuracy and robustness in noisy, partially connected environments while significantly reducing the dependency on manually surveyed reference anchors.

Original authors: Paul Schwarzbach, Andrea Jung

Published 2026-04-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Paul Schwarzbach, Andrea Jung

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are in a giant, pitch-black warehouse with 12 friends. None of you know where you are standing, and there are no street signs or maps. The only thing you have is a walkie-talkie that can tell you exactly how far away your friends are, but sometimes the signal bounces off walls and gives you a slightly wrong distance.

Your goal? To figure out exactly where everyone is standing so you can navigate the warehouse.

This is the problem the paper solves. It's about collaborative self-calibration for indoor positioning systems (like Ultra-Wideband or UWB technology).

Here is the breakdown of the problem and their solution, using simple analogies:

The Old Way: The "Rigid Blueprint" (Closed-Form Method)

Traditionally, to map a room, you need three people to stand in known spots (anchors) to act as a reference.

The Flaw: If you guess the starting positions of those three people even slightly wrong, or if one of your distance measurements is a bit "noisy" (like a bad echo), the entire map collapses.
The Analogy: Imagine trying to draw a map of a city using a ruler. If you make a tiny mistake in the first inch of your drawing, the rest of the city gets stretched or squished until the whole map is useless. If the first three people you pick to stand in the "right" spots are actually standing in a spot with bad signal (Non-Line-of-Sight), the whole system crashes. It's like trying to build a house of cards on a shaky table; one wobble, and everything falls.

The New Way: The "Fuzzy Cloud" (Grid-Based Uncertainty Propagation)

The authors propose a smarter way. Instead of assuming a person is at one specific point, they treat their location as a cloud of possibilities.

The Analogy: Instead of saying, "My friend Bob is at exactly 5 meters," the system says, "Bob is probably at 5 meters, but there's a small chance he's at 4.8 or 5.2 meters."
How it works:
1. The Grid: Imagine the floor is covered in a giant chessboard. The system doesn't pick one square for a person; it puts a "probability weight" on many squares.
2. The Cloud: As people talk to each other, they don't just draw a single line to their friend. They draw a "fuzzy cloud" of possible locations based on the distance they measured.
3. The Magic: When these clouds overlap, the system looks for the area where the clouds agree the most. Even if one measurement is weird (due to a wall blocking the signal), the other measurements act like a safety net, keeping the "cloud" from drifting too far off course.

Why This Matters: The "Bad Weather" Test

The researchers tested this in a real industrial hall with 12 nodes (devices). They created two scenarios:

Clear Sky (Line-of-Sight): Everyone can see each other.
Stormy Weather (Non-Line-of-Sight): Some people are behind walls or machinery, causing signal echoes.

The Results:

The Old Way (Rigid Blueprint): When the "weather" got bad (NLOS), the old method panicked. It tried to force a perfect answer based on bad data, and the map fell apart completely. Errors jumped from less than half a meter to over 30 meters in some cases. It was like trying to drive a car with a broken GPS that insists you are in the middle of a lake.
The New Way (Fuzzy Cloud): When the weather got bad, the "cloud" just got a little wider and fuzzier. The system admitted, "I'm not 100% sure, but I'm pretty sure you're somewhere in this area." It gracefully degraded. Instead of crashing, it kept the error low (under 1 meter), even when the signal was bouncing off walls.

The Big Takeaway

This paper introduces a method that is resilient.

Old Method: "I need perfect data, or I give up."
New Method: "I can work with messy, imperfect data by keeping track of all the possibilities."

This is huge for real-world applications like warehouses, hospitals, or factories where you can't always guarantee a clear line of sight between devices. It means you can set up these positioning systems faster, cheaper, and without needing a team of surveyors to measure every single anchor point perfectly beforehand. The system can "calibrate itself" even when things get messy.

1. Problem Statement

Indoor positioning systems (IPS), particularly those using Ultra-Wideband (UWB), traditionally rely on stationary anchors with precisely surveyed coordinates. This manual calibration is time-consuming, costly, and scales poorly with network density.

Collaborative Self-Calibration: A solution where nodes estimate their own positions using peer-to-peer ranging without prior knowledge of the network topology.
The Core Challenge: Existing parametric methods (e.g., closed-form trilateration, covariance-based filters) fail in Non-Line-of-Sight (NLOS) environments. They approximate uncertainty using Gaussian distributions and covariance matrices, which discard spatial distribution information. When NLOS biases occur, these methods suffer from systematic error propagation, leading to catastrophic performance collapse and "overconfident" but incorrect estimates.
Connectivity Constraints: Conventional methods often require a fully connected network or specific geometric configurations (e.g., a perfect triangle of anchors) to initialize, making them fragile in industrial environments with partial connectivity.

2. Methodology

The authors propose a Grid-based Bayesian Uncertainty Propagation approach that replaces parametric approximations with discrete probability mass distributions.

A. Theoretical Framework

Instead of representing a node's position as a single point with a covariance matrix, the method discretizes the 2D space into a grid ( $M$ cells). It maintains a probability distribution over these cells, allowing for the representation of non-Gaussian and multimodal uncertainties (common in NLOS).

B. Algorithm Flow

The process follows a recursive Bayesian filter structure:

Initialization (Three-Anchor Setup):
- Anchor 0: Defined as the origin $(0,0)$ .
- Anchor 1: Estimated along the x-axis using a 1D histogram filter.
- Anchor 2: Estimated using a full 2D grid-based Bayesian update. This step incorporates three distance measurements ( $z_{01}, z_{02}, z_{12}$ ) to establish the coordinate frame while preserving uncertainty distributions.
Recursive Position Estimation:
- Motion Update: Predicts node movement based on velocity models (or static noise recovery).
- Measurement Update (Core Innovation):
  - Instead of using a single point estimate for reference anchors, the algorithm samples the top $\tau\%$ of probability mass from neighboring anchors as discrete hypotheses.
  - The likelihood function for a target node is computed by weighting the contributions of these multiple hypotheses (Eq. 10).
  - This creates an "aggregate likelihood field" that reflects the true spatial uncertainty of the network, rather than circular constraints derived from point estimates.
- Position Estimation: The final position is a weighted average of grid cells within a radius of the maximum likelihood estimate.

C. Experimental Setup

Hardware: 12 Qorvo MDEK1001 UWB transceivers in a fully meshed network (Direct Two-Way Ranging).
Environment: A $44 \times 30$ m² industrial hall with mixed Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) conditions.
Ground Truth: Established via a Leica TS16 total station.
Dataset: 132,000 ranging measurements across 2,000 epochs.
Configurations: Four initialization scenarios were tested:
- Conf. 1 & 2: Pure LOS initialization.
- Conf. 3 & 4: NLOS-contaminated initialization (2 out of 3 initial anchor links were NLOS).

3. Key Contributions

Self-Surveyed Dataset: Creation of a public dataset with 132k ranging measurements, ground truth, and LOS/NLOS classifications for a fully meshed UWB network.
Non-Parametric Uncertainty Propagation: Extension of a grid-based Bayesian formulation that preserves discrete probability mass across network nodes, avoiding the information loss inherent in covariance-based parametric methods.
Robustness to NLOS Initialization: Demonstration that the method maintains sub-meter accuracy even when the initial coordinate frame is established using NLOS-biased measurements, whereas closed-form methods fail.
Reduced Connectivity Requirements: The algorithm functions effectively with partial connectivity, as it can utilize available neighbors without requiring a complete distance matrix or specific geometric anchor arrangements.

4. Key Results

The study compared the proposed Probabilistic Grid-based Positioning (PGP) against a baseline Closed-Form (CF) method.

Ranging Performance:
- LOS Mean Error: 0.28 m.
- NLOS Mean Error: 3.78 m (reflecting realistic industrial partial occlusion).
- Overall Mean Error: 1.11 m.
Positioning Accuracy (LOS Initialization):
- Both PGP and CF performed similarly.
- PGP Median Error: 0.34–0.68 m.
- CF Median Error: 0.40–0.47 m.
Positioning Accuracy (NLOS-Contaminated Initialization):
- PGP (Robust): Showed "graceful degradation." Median errors increased slightly to 0.62–0.99 m. The 75th percentile remained under 1.66 m.
- CF (Fragile): Exhibited "performance collapse." Median errors jumped to 1.48–2.43 m. The 95th percentile exceeded 30 m in some cases, with extreme outliers.
- Availability: The CF method failed to position certain nodes (0% availability) in specific configurations due to geometric constraints, whereas PGP maintained 100% availability across all scenarios.

5. Significance and Conclusion

This paper demonstrates that probabilistic uncertainty propagation is critical for robust collaborative self-calibration in real-world industrial environments.

Resilience: By preserving the shape of uncertainty distributions (rather than collapsing them to a single point), the grid-based method can absorb NLOS biases during initialization without catastrophic failure.
Practicality: The approach reduces the dependency on manual anchor calibration and allows for deployment in scenarios with sparse connectivity or unreliable initial measurements.
Future Impact: The findings suggest that for automated UWB network initialization, non-parametric grid-based methods are superior to traditional closed-form or covariance-based approaches, particularly in environments where Line-of-Sight cannot be guaranteed.

The authors conclude that this method enables automated, infrastructure-independent UWB network deployment, significantly lowering installation costs and complexity.