Source-linked AI summary
Direct Localization for Massive MIMO
Nil Garcia, Henk Wymeersch, Erik G. Larsson, Alexander M. Haimovich, Martial Coulon
TL;DR
Multipath makes conventional AOA-based localization unreliable because LOS paths can be difficult to identify, especially when arrivals are biased or obstructed. The paper proposes DiSouL, a compressed-sensing direct-localization method that jointly processes distributed massive-MIMO observations to distinguish LOS from NLOS paths. Numerical results report that DiSouL outperforms the compared localization techniques for most E/N0 values and can achieve sub-meter accuracy under the stated conditions.
Problem
In dense multipath environments, conventional AOA localization suffers biased estimates and difficult LOS-path identification, degrading localization accuracy.
Method
DiSouL jointly processes snapshots from distributed massive-MIMO base stations using compressive sensing to distinguish LOS components from NLOS components and directly estimate source position.
Results
DiSouL outperforms the compared techniques for most E/N0 values, while a theorem-validation experiment reports sub-meter accuracy with probability close to 1 when L=4 and w^2∈[3,4] at sufficiently high SNR.
Takeaways & Limitations
Joint processing of distributed massive-MIMO snapshots can estimate source location without explicitly estimating LOS AOAs, avoiding the challenging data-association problem.
Takeaways & Limitations
The gain in accuracy comes with higher computational complexity than previous techniques.
Abstract
from arXiv · showhide
Large-scale MIMO systems are well known for their advantages in communications, but they also have the potential for providing very accurate localization thanks to their high angular resolution. A difficult problem arising indoors and outdoors is localizing users over multipath channels. Localization based on angle of arrival (AOA) generally involves a two-step procedure, where signals are first processed to obtain a user's AOA at different base stations, followed by triangulation to determine the user's position. In the presence of multipath, the performance of these methods is greatly degraded due to the inability to correctly detect and/or estimate the AOA of the line-of-sight (LOS) paths. To counter the limitations of this two-step procedure which is inherently sub-optimal, we propose a direct localization approach in which the position of a user is localized by jointly processing the observations obtained at distributed massive MIMO base stations. Our approach is based on a novel compressed sensing framework that exploits channel properties to distinguish LOS from non-LOS signal paths, and leads to improved performance results compared to previous existing methods.
I. INTRODUCTION
Massive MIMO enables high-accuracy localization, but multipath biases AOA estimates and makes LOS-path identification difficult. DiSouL addresses this by jointly processing distributed base-station observations with compressed sensing to estimate source location directly.
- Motivation: Massive MIMO’s high angular resolution supports precise estimation of individual multipath AOAs, extending its use beyond communications.Its communication advantages include increased spectral efficiency, high directivity, and low complexity.
- Existing localization: Conventional AOA localization measures AOAs at multiple base stations and then triangulates the source position.This two-step procedure is effective in benign open-air settings but is degraded in dense multipath environments.
- Existing localization: Multipath biases AOA estimates, while identifying LOS paths is difficult because the strongest arrival may be NLOS and data association is NP-hard.These issues contribute to large localization errors in harsh multipath environments.
- Proposed direction: Direct localization estimates source position from data without first estimating intermediate LOS AOA parameters.Earlier direct-localization methods were designed for pure LOS environments, whereas multipath methods were not tailored to AOA information and massive arrays.
- Proposed direction: DiSouL jointly processes snapshots from distributed base stations using compressive sensing to distinguish LOS paths from NLOS paths.It exploits the common origin of LOS components and offers TOA-based search restriction and grid refinement variations.
- System model: The system model uses a known narrowband signal received through multipath, with unknown channel gains, AOAs, and TOAs, and relates LOS AOA and TOA to source position.The base stations are modeled as massive arrays whose response depends on the impinging angle.
III. PROPOSED METHOD
DiSouL selects sampling times and TOA upper bounds so that LOS energy is favored over NLOS interference, then jointly uses the resulting snapshots for direct localization. Its design also supports threshold-based delay estimation and computationally lighter grid refinement.
- Principle: DiSouL jointly processes matched-filter snapshots from all base stations to separate LOS paths from NLOS paths while estimating only the source position.The method avoids explicitly estimating LOS AOAs and uses TOA estimates as constraints rather than for precise localization.
- Practical considerations: Perfect knowledge of the transmitted pulse shape is required for matched filtering.With an entirely all-pass antenna and hardware response, the signal-to-noise ratio decreases but the number of multipath components remains unchanged.
- TOA Estimation: The method requires computing both matched-filter sampling instants and TOA estimates, which may also be obtained with alternative delay-estimation techniques.The desired sampling-time criterion differs from the criterion for estimating TOAs because NLOS components are treated as interference.
- Principle: Sampling times are chosen to maximize the LOS-to-NLOS energy ratio, while TOA estimates are required to be positively biased upper bounds on LOS TOAs.Consequently, sampling times are generally smaller than the corresponding TOA estimates.
- TOA Estimation: A multiple-antenna threshold matched filter estimates delays by aggregating observed antenna signals non-coherently and selecting a threshold crossing or peak.The threshold is selected so that the probability of an early noise-triggered false alarm is very low.
2) Sampling Time:
The localization stage discretizes candidate positions and arrival angles, then uses sparsity to identify a common source location and explain NLOS arrivals. Adaptive refinement can combine lower complexity with finer resolution, while three-dimensional extension increases computational cost.
- C. Localization: The localization problem introduces a uniform grid of Q candidate source locations and a uniform grid of M_l angles for each base station.These grids provide a discrete approximation to the continuous position-and-angle search.
- C. Localization: The LOS gain matrix is row-sparse because only one grid location represents the source, while NLOS variables are nonzero only at observed arrival angles.This structure encodes the common source location across base stations and the distinct NLOS arrivals at each array.
- C. Localization: The optimization promotes sparsity in source locations and NLOS paths while bounding the mismatch between observations and their reconstruction.It is formulated as a second-order cone program with parameters controlling allowed mismatch and the LOS/NLOS energy assignment.
- Remark: The current two-dimensional search can be generalized to three dimensions, but this increases computational complexity.The paper notes that three-dimensional geometry may also improve multipath robustness because distinct NLOS bearing lines generally do not intersect.
IV. PARAMETER SELECTION
The parameter ϵ controls the allowed mismatch between observations and reconstruction, while its value is selected probabilistically from the Gaussian noise model. In low SNR, the optimization can collapse to an all-zero solution, requiring a fallback location estimate.
- ϵ bounds the mismatch between observations and their reconstruction and is typically chosen as a noise bound.Because Gaussian noise is unbounded, the practical choice targets high-probability feasibility of the noiseless received signals.
- The feasibility condition is converted into a noise-only expression by substituting the observation model.The normalized aggregate noise energy follows a Chi-squared distribution with 2∑l=1^L Sl degrees of freedom.
- ϵ can be computed using the inverse cumulative distribution function of the relevant Chi-squared distribution.The probability parameter γ may be set to 0.99.
- Low SNR can make the aggregate snapshot energy no larger than ϵ, producing the trivial all-zeros solution.In that case, the method estimates the location whose LOS components correlate most strongly with the snapshots.
B. Setting the Parameter w
The weight w is selected to distinguish locations consistent with the LOS paths from alternatives under explicit assumptions about grid density, AOA recovery, and uniqueness. Lemmas establish conditions ensuring nonempty and consistent solutions, yielding a sufficient condition for correct source-location recovery.
- Definition: A location is consistent with L paths when the direct-path AOAs between that location and all base stations are true AOAs.The true source is consistent because its LOS components travel in straight lines.
- Assumptions: The weight analysis assumes unique consistency of the source, sufficiently dense location and angle grids, and correct recovery of the true AOAs.The AOA-recovery assumption is considered reasonable in high SNR conditions.
- Lemmas: If w is above the Lemma 1 threshold, every estimated location output by problem (13) is consistent with L paths.
- Lemmas: If w is below the Lemma 2 threshold, problem (13) outputs at least one location.
- Theorem: Under Assumptions A1–A3, Theorem 1 combines the two lemmas into a sufficient condition for correct recovery of the source location.The proof uses uniqueness: only the source location is consistent with L paths.
C. The Cases of Obstructed-LOS and Non-LOS
The method adapts its assumed number of LOS base stations when paths are obstructed or absent. It begins with ˆL=L and reduces this estimate when the optimization returns no location, while using a correlation fallback in low-energy cases.
- OLOS and NLOS: Attenuated or blocked LOS paths create OLOS or NLOS conditions that can prevent some LOS components from being detected.
- Weight adjustment: With one NLOS base station, the source is consistent with only L−1 paths, so a weight requiring L paths excludes the true source location.
- Weight adjustment: The adjusted weight requires an estimate ˆL of the number of base stations having LOS with the source.The true LOS count L∗ satisfies L∗≤L.
- Algorithm: Algorithm 1 initializes ˆL=L, sets ϵ, solves problem (13), and decreases ˆL when the solution matrix X is all zeros.When X is nonzero, the estimated position is the grid location corresponding to the row of X with maximum norm.
- Algorithm: If the aggregate snapshot energy does not exceed ϵ, Algorithm 1 estimates the position using the correlation-based fallback.
V. IMPROVING COMPUTATIONAL TIME AND PRECISION
The method improves precision and computational time through TOA-assisted grid trimming and adaptive refinement. Refinement starts from coarse location and angle grids, concentrates around estimates, and converges because each new problem retains the previous solution as feasible.
- Motivation: Performance depends on sampling-time quality and grid density, with grid density directly affecting computational complexity.
- TOA assistance: Sampling and TOA constraints trim the location grid; if the intersection is empty, a new grid can be generated in the feasible region.When needed, all TOA estimates are increased by v=1/B until a nonempty feasible region is obtained.
- Grid refinement: Dense grids provide fine resolution but increase computation, motivating refinement from coarse grids around estimated locations and angles.The computational complexity of problem (13) scales with the grid sizes, the number of base stations, and the number of angles.
- Grid refinement: The refinement process uses separate position and per-base-station angle grids, making it more complex than previous refinement approaches.
- Grid refinement: Each successive grid retains estimated points and neighboring points while also adding angles associated with estimated locations.The construction uses progressively finer resolutions for positions and angles.
- Stopping criterion: Refinement converges because each previous solution remains feasible and the objective is bounded below by zero.In practice, iterations stop when progress between consecutive steps becomes negligible.
C. The DiSouL Algorithm
DiSouL combines TOA assistance, grid construction, matched-filter snapshots, and Algorithm 1-based source estimation. The numerical setup uses a 100 m × 100 m area with four corner base stations and evaluates sub-meter accuracy as a function of the weight.
- Algorithm procedure: The algorithm sets η for the desired PFA, estimates TOAs, and constructs initial location and angle grids.It then trims the location grid through TOA assistance.
- Algorithm procedure: Matched filtering and sampling at computed times produce the snapshots used for source-location estimation.The sampling times are computed after the TOA-assisted grid step.
- Numerical setup: The numerical examples place a randomly located source in a 100 m × 100 m area with four base stations at the corners.The supplied setup passage begins listing the corner coordinates.
- Numerical evaluation: Figure 3 evaluates the probability of sub-meter accuracy versus the selected weight using Monte Carlo simulation with randomized signal strengths and phases.The figure concerns the scenario shown in Fig. 2.
A. Validation of Theorem 1
The experiments validate DiSouL under multipath and compare its localization performance with indirect and direct alternatives across signal, array, channel, and calibration conditions. DiSouL achieves high-precision localization, while its computational cost and matched-filter resolution impose practical boundaries.
- Validation of Theorem 1: A source–reflector experiment confirms sub-meter localization with probability close to 1 for w2 ∈ [3, 4] at sufficiently high SNR when L = 4.The condition tested is L^-1 ≤ w2 ≤ L.
- Localization accuracy: DiSouL achieves high-precision localization with high probability, followed by DPD and the two-step approaches, in the localization-error CDF.The CDF is evaluated at E/N0 = 20 dB and B = 30 MHz.
- Signal conditions: DiSouL outperforms the compared techniques for most E/N0 values, although its sub-meter accuracy saturates above 30 dB because of matched-filter time resolution.Higher E/N0 reduces the number of NLOS components entering the snapshots, but the benefit is limited by digital matched-filter resolution.
- Signal conditions: Increasing bandwidth improves all techniques, while shorter pulses include fewer NLOS components and decrease the risk of AOA-estimation errors.The bandwidth comparison fixes E/N0 = 20 dB.
- Array size: Increasing antennas improves DiSouL through higher angular resolution and better multipath separation, whereas indirect techniques remain approximately unchanged or improve very little.For IV and Stansfield, selecting the wrong path as LOS accounts for most errors.
VII. CONCLUSIONS
The paper addresses narrowband localization in massive MIMO under multipath using direct localization and compressive sensing. DiSouL jointly processes distributed-array snapshots to estimate source locations without explicitly estimating LOS AOAs, achieving high-probability sub-meter localization while requiring higher computational complexity.
- DiSouL jointly processes snapshots from widely distributed arrays to estimate source locations directly in narrowband multipath settings.The approach avoids explicitly estimating LOS AOAs and the associated data association problem.
- The method uses compressive sensing and massive arrays’ angular resolution to distinguish LOS from NLOS multipath components.It requires no statistical channel knowledge beyond the noise variance, supporting use across multipath environments.
- Coarse TOA estimates reduce execution time and enhance localization accuracy.
- Numerical simulations show high-probability sub-meter localization in dense multipath environments with narrow-band signals.
- The gain in accuracy comes with higher computational complexity than previous techniques.
APPENDIX A PROOF OF LEMMA 1
The proof establishes that, under the stated weight condition, an optimal reconstruction must assign the observed paths consistently with a common source location. It proceeds by comparing competing decompositions and their objective costs.
- The proof compares a reconstruction using one position with L associated LOS angles against alternatives using NLOS angles.
- The proof aims to show that w > L − 1 implies the estimated coefficient vector has ℓ0-norm L.This means all L paths are associated with the candidate location.
- The reconstruction is written as a location-associated component plus an NLOS component and a residual placeholder.
- When ∥x1∥0 = L, the candidate position π1 is consistent with L paths under Assumption A3).
- A vector indicator and the Cauchy–Schwarz inequality are used to derive the contradiction needed to establish the coefficient-support condition.
APPENDIX B PROOF OF LEMMA 2
The proof of Lemma 2 uses a contradiction argument to show that the optimization cannot omit all estimated locations when the weight satisfies the stated condition. It constructs a competing decomposition with a lower or equal cost.
- The proof assumes w < L and no estimated location, so every location coefficient is zero.
- The argument constructs a competing decomposition that transfers sufficient mass from NLOS observations to the LOS component.This construction is intended to preserve the observations while not increasing the objective cost.
- Because each true LOS angle lies in the admissible angle set, the observation admits a decomposition involving the true position p.
- The proof compares objective costs for the original and competing decompositions after discarding common terms.