Source-linked AI summary
Map-Free Single-Anchor Position Localization Using Multipath Uncertainty at Upper Mid-Band
Xingchen Liu, Mingjun Ying, Xinquan Wang, Dipankar Shakya, Homa Nikbakht, Daniel Abraham, Theodore S. Rappaport
TL;DR
Map-free indoor localization is challenging because MPCs differ in geometric reliability and environmental maps may be unavailable or costly. This paper estimates per-MPC angular uncertainty from local directional observations and incorporates it into single-anchor geometric localization. Across 20 measured links, its covariance-weighted method outperforms the compared weighting variants and literature baselines, with a 3.86 m overall mean error.
Problem
Map-free methods may apply idealized geometry to MPCs affected by multibounce propagation, diffuse scattering, and path mixing, while environmental maps may be unavailable, incomplete, or costly to maintain.
Source: §I · p. 1Method
The framework estimates per-MPC angular covariance from local directional consistency and propagates it into a solver fusing point constraints with line fallbacks.
Source: §I · p. 1Results
3.86 m mean error, 2.66 m median error, and 70% of links within 5 m; CW achieves the best overall performance among the five methods.
Source: B. Localization performance comparison · p. 5Takeaways & Limitations
Across 20 measured links, local directional consistency provides MPC reliability information beyond received power alone.
Source: §V · p. 6Geometric model mismatch can still cause errors on challenging NLOS links despite covariance-based weighting.
Source: A. Localization results of the… · p. 5
Abstract
from arXiv · showhide
Future 5G-Advanced and 6G systems will exploit directional wideband channels for path-based localization, yet map-free indoor positioning remains challenging because extracted multipath components (MPCs) vary widely in localization reliability. This paper presents an uncertainty-aware map-free single-anchor localization framework using 16.95~GHz indoor directional channel measurements collected at NYU WIRELESS in Brooklyn, New York. Per-MPC angular covariance is estimated from the local consistency of neighboring pattern-de-embedded power delay profile (PDP) observations and propagated into geometric localization through weighted fusion of point-type and line-type constraints. Measurements across 20 links with TX--RX separations from 11~m to 97~m show that the proposed covariance-weighted method achieves a mean localization error of 3.86~m, a median of 2.66~m, and 70\% of links within 5~m using five retained MPCs. It outperforms the power-weighted and unweighted variants under identical point/line constraints, as well as the literature baselines. These results indicate that local directional consistency provides reliability information beyond received power alone and demonstrate a practical map-free approach for future indoor directional systems.
I. INTRODUCTION
Map-free localization needs to account for MPCs whose geometric reliability varies under multibounce propagation, diffuse scattering, and path mixing. The paper estimates path-specific reliability from directional measurements and evaluates the approach on measured indoor links.
I. INTRODUCTION
- Map-free methods can apply ideal LOS or single-bounce geometry to paths affected by multibounce propagation, diffuse scattering, and path mixing, motivating measurement-derived path reliability.Map-assisted methods can use environmental priors, but accurate maps may be unavailable, incomplete, or costly to acquire and maintain.Source: §I · p. 1
- The framework estimates per-MPC angular covariance from local directional consistency and propagates it into a solver that fuses LOS or feasible single-bounce point constraints with line fallbacks.It builds on antenna-de-embedded multipath extraction (ADME).Source: §I · p. 1
- The proposed framework outperforms power-weighted and unweighted variants using the same retained MPCs and constraints, as well as the planar NLOS and 3D ZF–JP baselines.The comparison is reported for the measurement campaign described in the paper.Source: §I · p. 2
A. Indoor Hotspot (InH) Measurement Campaign Overview
- The 16.95 GHz indoor measurement campaign contains 20 links, including 7 LOS and 13 NLOS links, with TX–RX separations from 11 m to 97 m.It spans enclosed rooms, open work areas, and connecting corridors.Source: A. Indoor Hotspot (InH) Measurement… · p. 2
B. MPC extraction and refinement overview
- ADME refines coarse MPC estimates using steering-pair observations within approximately one half-power beamwidth and selects the direction with the most consistent pattern-deembedded neighborhood powers.The refinement searches a local angular region for the smallest sum of squared neighborhood residuals.Source: B. MPC extraction and refinement… · p. 2
C. Local angular uncertainty modeling
The local uncertainty model estimates a separate angular covariance for each MPC from the consistency of its neighboring directional observations. Residual spread reflects local support, including disturbances beyond thermal noise.
C. Local angular uncertainty modeling
- The method estimates a local angular covariance for each MPC from neighborhood observations because different MPCs have different levels of directional support.The per-path uncertainty model is intended to capture that variation in reliability.Source: C. Local angular uncertainty modeling · p. 2
1) Local residual model and weighting:
- Neighborhood residuals measure how closely observations cluster around the estimated local power, with larger residuals indicating weaker support.Weak support may arise from low SNR, pattern mismatch, or MPC mixing within the delay window.Source: 1) Local residual model and weighting · p. 2
- The effective local residual scale is estimated using residual degrees of freedom for the five-parameter local model.The scale absorbs thermal noise and other difficult-to-separate disturbances, so it represents local uncertainty rather than purely thermal-noise variance.Source: 1) Local residual model and weighting · p. 2
2) Covariance approximation from neighborhood observations:
The paper approximates each MPC’s angular covariance by locally linearizing its neighborhood residual model. The approximation is conditional on adequate observations and a numerically well-defined local fit.
2) Covariance approximation from neighborhood observations:
- The covariance approximation uses a locally linearized residual model and retains the four-dimensional angular block of the five-parameter covariance.Its Jacobian uses finite differences for antenna-pattern angular columns and an analytic power derivative.Source: 2) Covariance approximation from… · p. 2
- MPCs with M_l ≤ 5 or nearly singular J_l^T J_l are excluded because their angular covariance cannot be reliably estimated.The covariance is used only when it is numerically well defined.Source: 2) Covariance approximation from… · p. 2
- Close antenna-pattern agreement yields a small angular covariance, while greater sample–pattern mismatch increases residual spread and covariance and may bias the refined AOA.The figure shows one-dimensional slices of the full four-dimensional neighborhoods.Source: 2) Covariance approximation from… · p. 3, Figure 1 · p. 3
- Delay is treated as fixed because synchronization-induced one-way path-length uncertainty is below 0.3 m; noise- or mixed-MPC delay errors remain for future joint angle–delay modeling.The stated synchronization accuracy is well below the observed localization errors.Source: 2) Covariance approximation from… · p. 3
A. Problem formulation
The framework estimates a horizontal RX position from refined multipath delays and angles with known TX position and fixed TX/RX heights, without an environment prior. It retains early paths and uses directional consistency to reduce the influence of less reliable MPCs.
A. Problem formulation
- The RX position is estimated from refined MPC delays and angles, with known TX position and fixed TX/RX heights and no map or surface prior for NLOS scatter points.The reflection or scatter point is unknown and is not associated with an environmental map or surface model.Source: A. Problem formulation · p. 3
B. MPC gating
- Only the K earliest arrivals are retained to favor LOS and low-order reflections, which are less likely than later paths to involve higher-order reflections, mixing, or diffuse scattering.Delay alone cannot reliably distinguish single-bounce from higher-order paths, and applying a single-bounce model to higher-order paths can cause large localization errors.Source: B. MPC gating · p. 3
- MPCs with inconsistent de-embedded neighborhood powers receive larger angular covariances and therefore less influence on the position estimate.Covariance weighting acts as a secondary soft gate.Source: B. MPC gating · p. 3
C. MPC-induced geometric constraints
- For each retained MPC, calibrated delay gives path length d_l = cτ̂_l, while refined departure and arrival angles define unit direction vectors.Residual delay uncertainty is neglected.Source: C. MPC-induced geometric constraints · p. 3
- The departure vector points from the TX along the departure direction, while the arrival vector points from the RX toward the arriving path’s source.The formulation also distinguishes horizontal projections and vertical components of the direction vectors and the TX horizontal position.Source: C. MPC-induced geometric constraints · p. 3
- Only the earliest-arriving retained MPC is tested for LOS, using approximate reciprocity between its departure and arrival directions.Later paths are treated as unlikely to be LOS.Source: 1) LOS point constraint · p. 3
1) LOS point constraint:
For an LOS path, the RX can be located from either departure or arrival direction together with delay. The method fuses both directions so that the more reliable side contributes more strongly.
1) LOS point constraint:
- The fused LOS direction weights the AOD and AOA vectors by inverse angular-uncertainty traces, allowing the less uncertain side to contribute more.The traces sum the azimuth and elevation variances in the corresponding covariance sub-blocks.Source: 1) LOS point constraint · p. 3
- The fused direction produces a 3D RX candidate that is retained only when its height is consistent with the RX height.Either direction alone, together with path delay, is sufficient for an LOS position; fusing both lets the more reliably estimated side dominate.Source: 1) LOS point constraint · p. 3
- The retained LOS candidate supplies a point-type geometric constraint, with a 2D residual for the unknown horizontal RX position.Source: 1) LOS point constraint · p. 3
2) Single-bounce NLOS point constraint and line fallback:
The method first seeks a feasible single-bounce NLOS point candidate. If that construction is infeasible or unstable, it uses a line constraint from the horizontal path geometry.
2) Single-bounce NLOS point constraint and line fallback:
- A single-bounce NLOS hypothesis relates the unknown scatter point, TX and RX segment lengths, and total path length, with the RX height fixed.The path length is split into TX-to-scatterer and scatterer-to-RX segments whose sum is the measured total.Source: 2) Single-bounce NLOS point constraint… · p. 3
- A single-bounce point candidate is used only when the segment lengths are nonnegative and the construction is well-conditioned.When feasible, it induces the same point-residual form as the LOS candidate.Source: 2) Single-bounce NLOS point constraint… · p. 4
- When the point construction is infeasible or numerically unstable, the fallback line joins endpoint hypotheses determined by AOD, AOA, and total path length.Any horizontal RX location consistent with a valid split of total path length lies on that line; the residual is signed perpendicular distance to it.Source: 2) Single-bounce NLOS point constraint… · p. 4, Figure 2 · p. 4
D. Weighted position estimation
Point and line constraints are combined in a type-aware weighted formulation. Angular uncertainty is propagated to point constraints, while line uncertainty is approximated at a pilot position estimate before the final position is solved.
D. Weighted position estimation
- Each retained MPC contributes a point or line constraint, and the method combines both constraint types in a unified weighted formulation.Point constraints have 2D vector residuals, while line constraints have scalar residuals.Source: D. Weighted position estimation · p. 4
- Point-residual uncertainty is obtained by first-order propagation of angular covariance through the point construction.The propagated quantity is also the covariance of the point residual.Source: D. Weighted position estimation · p. 4
- Because angular perturbations rotate an induced line, its residual variance depends on RX position and is approximated at an unweighted pilot estimate.The pilot estimate uses the same constraint set.Source: D. Weighted position estimation · p. 4
- With the propagated uncertainty quantities fixed, the final position estimate is obtained from a quadratic objective with a closed-form solution.A positive ε prevents numerical instability when a propagated covariance or variance is nearly singular.Source: D. Weighted position estimation · p. 4
- The processing chain extracts and refines MPCs, estimates angular covariance, retains early paths, selects point or line constraints, propagates uncertainty, and solves for the RX position.The algorithm computes line-residual variances when line constraints exist.Source: D. Weighted position estimation · p. 4
IV. NUMERICAL RESULTS
The evaluation uses 16.95 GHz indoor measurements and compares three weighting variants under shared point/line constraints with two literature baselines. Localization error is measured in 2D for LOS, NLOS, and all links.
IV. NUMERICAL RESULTS
- The dataset contains 7 LOS and 13 NLOS links, with TX and RX heights fixed at 2.4 m and 1.5 m, respectively.Error is the 2D Euclidean distance between estimated and ground-truth RX positions.Source: §IV · p. 4
- CW, PW, and UW use the same retained MPCs and point/line constraints, differing only in reliability weighting.CW uses propagated angular uncertainty, PW uses de-embedded MPC power, and UW assigns equal weight to each constraint.Source: §IV · pp. 4–5
- Table I reports localization errors separately for LOS, NLOS, and all links across the three internal variants and two literature baselines using K = 5.Source: Table I · p. 5
- The two literature baselines are Planar NLOS and 3D ZF–JP [10], evaluated using the same retained MPC set but their original single-bounce models.Planar NLOS omits elevation to match its 2D formulation; 3D ZF–JP uses azimuth and elevation angles with ToF.Source: §IV · pp. 4–5
A. Localization results of the proposed method
With K = 5, the proposed method generally estimates LOS RX positions with small errors, while significant errors are concentrated in a few challenging NLOS links. Covariance weighting reduces the influence of weakly supported paths but cannot eliminate geometric model-mismatch errors.
A. Localization results of the proposed method
- LOS links are generally estimated with small errors.Source: A. Localization results of the… · p. 5
- Significant errors are confined to a few challenging NLOS links, while most RX positions are well estimated.Larger errors occur for RXs deep inside rooms and for some long TX–RX separations, where retained MPCs may yield constraints inconsistent with the true RX location.Source: A. Localization results of the… · p. 5
- Covariance-based weighting reduces the influence of weakly supported paths but cannot fully eliminate errors caused by geometric model mismatch.Source: A. Localization results of the… · p. 5
B. Localization performance comparison
The covariance-weighted method achieves the strongest reported overall performance and leads the internal variants in error-distribution comparisons. Its error is lowest at K = 5 and remains stable for larger K, while the conclusion notes that too many MPCs can introduce unreliable late arrivals and that preferred MPC count depends on propagation conditions.
B. Localization performance comparison
- 3.86 m mean error, 2.66 m median error, and 70% within 5 m are CW’s overall results at K = 5; it outperforms PW, UW, Planar NLOS, and 3D ZF–JP.The literature baselines have overall mean errors of 15.35 m for Planar NLOS and 11.25 m for 3D ZF–JP, with 30% and 35% of links within 5 m, respectively.Source: B. Localization performance comparison · p. 5
- The K = 5 error CDF ranks CW ahead of PW and UW, while 3D ZF–JP and Planar NLOS show substantially heavier error tails.The three internal variants share the retained MPCs and point/line constraint construction.Source: B. Localization performance comparison · p. 5, Figure 4 · p. 6
C. Sensitivity to the Number of MPCs
- CW reaches its lowest overall mean error at K = 5 and remains near 4 m for larger K, whereas the preferred retained MPC count depends on path quality and propagation conditions.Errors are relatively large at K = 2 and decrease as additional MPCs are retained; retaining too many can introduce unreliable late-arriving paths.Source: C. Sensitivity to the Number of MPCs · pp. 5–6, Figure 5 · p. 6
V. CONCLUSION
- Across 20 measured links, local directional consistency provides reliability information beyond received power, and future work will examine adaptive MPC selection, cross-frequency generalization, and delay-uncertainty modeling.Source: §V · p. 6