⚡ electrical engineering

Single-Base-Station Indoor Localization via Super-Resolved Relative Power Delay Profiles

This paper proposes a super-resolution technique using expectation-maximization sparse Bayesian learning to reconstruct high-fidelity relative power delay profiles from noisy, finite pilot samples, thereby significantly enhancing single-base-station indoor localization accuracy without requiring line-of-sight paths or angle information.

Original authors: Fangqing Xiao, Dirk T. M. Slock

Published 2026-06-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Fangqing Xiao, Dirk T. M. Slock

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 trying to find a specific room inside a massive, windowless building. You can't see the walls, and you don't have a map. However, you have a special flashlight that sends out a quick "ping" of sound (or radio waves). When this ping hits the walls, furniture, and corners, it bounces back to you.

In a normal room, the sound bounces off many things at different times, creating a complex echo pattern. The paper argues that this echo pattern is unique to every single spot in the room. Even if you can't see the direct path to the source (the "Line of Sight"), the way the echoes bounce around the furniture creates a unique "fingerprint" for your location.

Here is the breakdown of the paper's idea, using simple analogies:

1. The Problem: The "Blurry" Photo

Usually, when a device tries to listen to these echoes, it takes a snapshot. But because the device only listens at specific, spaced-out moments (like taking photos with a camera that has a slow shutter speed), the resulting picture is blurry.

The Paper's View: The device doesn't see a perfect, sharp list of where every echo came from. Instead, it sees a smeared-out mess because of how the "camera" (the radio signal) works.
The Old Way: To fix this blur, people used to just "zoom in" on the picture by adding fake pixels (called "zero-padding"). This makes the image look smoother, but it doesn't actually reveal any new details; it just stretches the blur.

2. The Solution: The "Smart Detective" (EM-SBL)

The authors propose a new method called EM-SBL (Expectation-Maximization Sparse Bayesian Learning). Think of this not as a camera, but as a smart detective.

How it works: Instead of just stretching the blurry picture, the detective looks at the raw data and asks: "What is the simplest, most likely arrangement of echoes that could have created this specific blurry pattern?"
The Result: The detective reconstructs a much sharper picture of the echoes. It figures out that even though the signal is fuzzy, the underlying "shape" of the echoes is actually very distinct. It separates the real echoes from the noise and the blur.

3. The "Folded" Map

There is a tricky part: because the device listens at fixed intervals, very long echoes get "folded" back onto the map, like wrapping a long string around a small spool.

The Paper's Trick: Instead of trying to "unfold" the string (which is hard and often impossible), the paper says: "Let's just learn the folded string."
The Analogy: Imagine you are trying to recognize a person by their shadow. If the light source is weird, the shadow might look squashed or twisted. The paper says, "We don't need to fix the shadow to look like the person. We just need to memorize what the squashed shadow looks like for every room." As long as the "squashing" happens the same way every time, the squashed shadow is still a perfect ID card for that location.

4. The Results: Finding Your Way

The authors tested this in a computer simulation of a building (using a tool called QuaDRiGa). They compared three methods:

The Native Method: Looking at the raw, blurry echo.
The "Zoom" Method: The old way of stretching the blurry image.
The "Detective" Method (Their new way): Using the smart algorithm to reconstruct the sharp echo pattern.

The Outcome:

The "Detective" method was significantly better at guessing the correct room.
At a standard signal strength, the old methods were right about 76% of the time.
The new "Detective" method was right 93% of the time.
It also reduced the average distance of error from about 0.6 meters down to 0.4 meters (roughly the length of a large step).

Summary

This paper shows that you don't need expensive hardware or a clear view of the signal source to find your way indoors. You just need to stop treating the "messy" radio echoes as noise and start treating them as a unique fingerprint. By using a smart mathematical algorithm to clean up the "blur" caused by the device's limitations, you can pinpoint a location with much higher accuracy than before.

Key Takeaway: The paper claims that by using a specific mathematical "clean-up" tool on standard Wi-Fi-like signals, we can turn a blurry, confusing echo into a precise location ID, even in complex indoor environments.

Technical Summary: Single-Base-Station Indoor Localization via Super-Resolved Relative Power Delay Profiles

1. Problem Statement

The paper addresses the challenge of indoor localization using a single base station (BS) without relying on a Line-of-Sight (LoS) path, angle-of-arrival (AoA) information, or absolute time-of-arrival (ToA) ranging. While Power Delay Profiles (PDPs) have been explored as location fingerprints because multipath components are shaped by the surrounding geometry, a critical gap exists between theoretical PDPs and practical observations.

Communication receivers do not observe ideal, continuous-delay PDPs. Instead, they observe a finite number of noisy pilot-frequency samples. This limitation introduces three specific distortions:

Dirichlet Blur: Caused by the finite aperture of the pilot observation window.
Delay Folding: Caused by the uniform spacing of pilots, which creates an ambiguity period where physical delays separated by integer multiples of the ambiguity period are indistinguishable.
Off-Grid Mismatch: Arising when continuous physical delays do not align with the discrete delay dictionary used for reconstruction.

Existing approaches often either optimize channel reconstruction fidelity (ignoring localization discriminability) or assume a PDP has already been formed (ignoring how finite aperture and pilot spacing reshape the fingerprint). This paper aims to bridge the observation-to-localization gap by modeling these distortions directly from pilot-domain data.

2. Methodology

The proposed framework separates the localization chain into a PDP front end and a radio-map inference stage, ensuring that performance differences isolate the representation method rather than the downstream classifier.

A. Observation Model
The authors model the channel as a sum of multipath components. They explicitly characterize the effects of the pilot grid:

Folding: The observation is periodic with an ambiguity period $T_{amb} = 1/\Delta f_p$ . Paths are folded into this interval, and the phase is absorbed into complex gains. The folded PDP is treated as the location-dependent fingerprint.
Blur: The finite number of pilots $M$ results in a Dirichlet kernel point-spread function, broadening components and creating sidelobes.

B. Fingerprint Construction
Three PDP front ends are compared:

Native PDP: The squared magnitude of the inverse DFT of the pilot samples, sampled on the natural delay grid.
Zero-Padded (ZP) PDP: The native response interpolated via zero-padding (2x and 3x) to a finer grid. The authors note this adds no new observations or resolving power, only denser sampling of the blurred response.
EM-SBL PDP (Proposed): A sparse posterior-power reconstruction using Expectation-Maximization Sparse Bayesian Learning (EM-SBL).
- The model assumes a sparse latent vector $x_q$ on a fine grid ( $q \ge 1$ ).
- The algorithm infers the posterior second moment ( $|\mu_n|^2 + [\Sigma]_{n,n}$ ) to reconstruct the power profile.
- This approach treats the PDP as a sparse latent structure inferred from pilots, rather than a simple interpolation.

C. Feature Normalization and Inference

Normalization: To remove absolute power dependence, profiles are clipped at a level $B$ dB below the maximum (set to 50 dB) and normalized to $[0, 1]$ .
Radio Map: The service area is divided into cells. Offline, a map of normalized features is built. Online, a query feature is matched to the map using a weighted $k$ -Nearest Neighbor (kNN) classifier, though other classifiers (SVM, Random Forest) are also tested.

3. Key Contributions

Formulation of LoS-Free Localization: The paper formulates single-BS localization using relative PDPs as environmental fingerprints, explicitly treating multipath as information rather than impairment.
Explicit Modeling of Pilot-Induced Distortions: It derives the periodic Dirichlet point-spread function and the folded-delay representation induced by uniform pilot spacing, retaining the folded representation for both offline and online stages.
Sampling-Aware Reconstruction: It proposes using EM-SBL to reconstruct a sparse posterior-power profile directly from pilot samples, distinguishing sparse inference from simple zero-padding interpolation.
Controlled Evaluation: The study isolates the impact of the PDP front end by keeping the radio-map inference protocol, data splits, and noise realizations consistent across all methods.

4. Simulation Results

The evaluation utilizes spatially consistent channels generated via QuaDRiGa under a 3GPP indoor NLOS model (3.5 GHz carrier, 64 pilots, ~196 MHz bandwidth).

Accuracy: At an SNR of 20 dB, the proposed 2-fold EM-SBL method achieved a Top-1 cell accuracy of 93.27%. This represents a significant improvement over:
- Native PDP: 75.79%
- 3-fold Zero-Padding: 87.24%
Error Metrics: The mean two-dimensional localization error for 2-fold SBL was 0.392 m, compared to 0.587 m for the native PDP and 0.419 m for 3-fold zero-padding.
Robustness:
- SNR Sensitivity: SBL methods outperformed baselines across all tested SNRs (5–30 dB). 3-fold SBL showed slight advantages at very low SNRs, while 2-fold SBL offered the best performance-dimension tradeoff.
- Tail Suppression: Error Cumulative Distribution Functions (CDFs) showed that SBL significantly suppressed large localization errors (tail events) compared to native and zero-padded methods, particularly at 10 dB.
- Classifier Independence: The performance gains were consistent across five different localizers (Nearest Centroid, Weighted kNN, Linear SVM, RBF SVM, Random Forest), indicating the gain stems from the PDP representation, not the decision rule.

5. Significance and Claims

The paper claims that communication PDPs inherently contain sufficient environmental information for accurate indoor localization, provided that sampling-induced distortions (blur, folding, and grid mismatch) are properly handled.

The significance lies in demonstrating that super-resolution via sparse Bayesian learning on the pilot domain yields a more discriminative fingerprint than standard interpolation (zero-padding) or raw observation. The authors emphasize that this approach does not require physical delay unaliasing or LoS identification; instead, it leverages the consistent, folded, and blurred structure of the multipath environment as a unique signature for the user's location. The work establishes that with proper modeling of the observation process, a single base station can achieve high-precision indoor localization without specialized hardware or absolute ranging.