PaperlayerView exact revision on arXiv

Source-linked AI summary

LocAttMamba: A Low-Complexity Mamba Framework with Attention-Based Multi-AP Fusion for Indoor Localization

MohammadMahdi Ghadaksaz, Mohammad Farzanullah, Melike Erol-Kantarci

arXiv:2610.05673v1cs.NI

TL;DR

Indoor localization must balance accuracy and robustness with the computational limits of positioning devices, while existing approaches face NLoS and complexity challenges. LocAttMamba processes AP measurements with Mamba encoders and attention fusion, jointly predicting location and uncertainty. On real-world 5G and UWB data, it reports lower mean errors than the second-best benchmarks while requiring the fewest multiply-accumulate operations.

Abstract

from arXiv · show

Accurate and low-complexity indoor localization is important for location-based services in fifth generation (5G) and sixth generation (6G) networks, where positioning devices operate under limited computational budgets and non conditions. Indoor localization has been studied widely using traditional signal-level localization approaches. However, these techniques often show degraded performance in on-line-of-sight (NLoS) scenarios. Recently, artificial intelligence (AI)-based techniques, including transformers, have been applied to address these challenges. While transformer-based architectures can capture the dependencies within the measurements collected from distributed access points (APs), their high computational complexity results in a large number of multiply-accumulate operations and long inference time. In contrast, lightweight recurrent and convolutional models trade this cost for degraded accuracy. In this paper, we propose LocAttMamba, a low-complexity localization framework in which the channel impulse response (CIR) and time-based features of each AP are processed by a separate Mamba encoder with near-linear complexity, and the resulting per-AP embeddings are fused through a multi-head attention layer that weights each AP according to its importance at every time step. The framework jointly predicts the user location and its per-axis uncertainty, which is refined through a post-hoc calibration step. We evaluate the proposed framework using two real-world 5G and ultra-wideband (UWB) measurement datasets. Our numerical results reveal that LocAttMamba obtains a mean two-dimensional (2-D) positioning error of 1.004 m and 0.599 m, respectively, on 5G and UWB datasets, outperforming the second-best benchmark by 9.79% and 5.82%, while requiring the fewest multiply-accumulate operations among all evaluated models and being 4.4-16 times faster than the transformer-based benchmarks.

I. INTRODUCTION

The introduction frames indoor localization as a balance of accuracy, robustness, and computational cost, then presents LocAttMamba as an attention-fused Mamba approach evaluated on real-world 5G and UWB datasets.

I. INTRODUCTION

  • Conventional signal-level methods degrade in NLoS-rich settings, while transformer inference can impose high computational and memory costs.Devices also have limited computing power, and measurements are affected by noise and changing surroundings.Source: §I · p. 1

A. Related Works

  • Prior work spans MLP, recurrent, convolutional, attention, transformer, and emerging Mamba approaches, but the cited Mamba studies do not adaptively weight individual APs.The paper identifies this gap for degraded measurements and real-world 5G and UWB data.Source: B. Motivations & Contributions · pp. 2–3, A. Related Works · p. 2

B. Motivations & Contributions

  • LocAttMamba processes each AP with a dedicated bidirectional Mamba encoder and fuses per-AP embeddings with multi-head attention.The contributions also describe robustness evaluation under single- and multiple-AP CIR distortions and analysis of learned attention weights.Source: B. Motivations & Contributions · p. 3
  • LocAttMamba achieves mean errors of 1.004 m on 5G and 0.599 m on UWB, beating the second-best benchmarks by 9.79% and 5.82%, respectively.It requires the fewest multiply-accumulate operations among evaluated models and is 4.4–16× faster than transformer-based benchmarks.Source: B. Motivations & Contributions · p. 3
  • The framework predicts uncertainty alongside location and applies post-hoc calibration; the reported normalized squared error statistic falls from 34.6 to 2.15.The paper describes 2 as the ideal value for a two-dimensional Gaussian.Source: §I · p. 2, B. Motivations & Contributions · p. 3

II. PROBLEM DESCRIPTION

The problem formulation treats localization as mapping synchronized measurements from multiple APs over a time window to the user's two-dimensional locations.

II. PROBLEM DESCRIPTION

  • Each AP's real-valued feature vector combines the real and imaginary CIR parts with two time-related features.The resulting vector has dimension 2ν+2, and each observation matrix contains W time steps.Source: §II · p. 3
  • The formulation constructs synchronized per-AP observation matrices for each time step and aggregates them across time to form the training dataset.The dataset uses windows of W time steps and a stride S.Source: §II · pp. 3–4
  • A trainable mapping f converts observation matrices from all APs into predicted locations, with parameters optimized against ground-truth locations over the training data.The trained parameters are then used to predict locations from test measurements.Source: §II · p. 4

III. PROPOSED SCHEME

LocAttMamba uses per-AP bidirectional Mamba processing followed by attention fusion to predict location and per-axis uncertainty; the section introduces SSMs as a sequence-modeling foundation.

III. PROPOSED SCHEME

  • The proposed scheme processes each AP with bidirectional Mamba blocks, then fuses the per-AP embeddings with multi-head attention to predict location and per-axis uncertainty.This is the framework's stated overall architecture.Source: §III · p. 4

A. Preliminaries: SSMs & Mamba

  • An SSM transforms input sequences to output sequences through an implicit latent state, and can be computed as a recurrence or a convolution.The formulation uses input and output sequences and model matrices to specify a continuous-time linear system.Source: A. Preliminaries: SSMs & Mamba · p. 4
  • Discretizing the system with a step size produces the sequence representation; constant S4 matrices enable a global convolution but limit input-dependent modeling.The text identifies this input-independent structure as a limitation for more complex problems.Source: A. Preliminaries: SSMs & Mamba · p. 4

B. Mamba Block

Mamba makes its state-space recurrence input-dependent, allowing time-varying parameters while retaining efficient selective-scan computation. The block combines an SSM branch with a gating branch.

B. Mamba Block

  • Input-dependent B, C, and δ make Mamba’s recurrence time-varying, addressing the limited expressiveness of S4’s input-independent parameters.The scalar SSM is applied independently to each input channel, and δ_k is constrained to be positive.Source: B. Mamba Block · pp. 4–5
  • The block projects its input, then applies depthwise 1-D convolution and SiLU before forming the SSM-branch sequence.The SSM branch’s projected sequence supplies the input-dependent parameters and recurrence.Source: B. Mamba Block · p. 5
  • The SSM state is a per-channel matrix, and its outputs are collected across the window as a sequence.The recurrence’s time variation comes through δ_k, B_k, and c_k while A remains static.Source: B. Mamba Block · p. 5
  • A separate gating branch is combined elementwise with the SSM output, and the result is linearly projected to produce the block output.Source: B. Mamba Block · p. 5
  • Because input-dependent parameters prevent global-convolution evaluation, Mamba uses a hardware-aware selective scan to compute the recurrence efficiently.Source: B. Mamba Block · p. 5

C. Proposed LocAttMamba Framework

LocAttMamba processes each access point’s sequence with its own encoder built from stacked Mamba blocks. The encoder forms per-AP embeddings through forward and reverse processing, residual addition, projection, and normalization.

C. Proposed LocAttMamba Framework

MLP MLP

The framework stacks per-AP embeddings by time step and uses multi-head attention to weight AP contributions. It then predicts location and per-axis uncertainty for each time step.

MLP MLP

  • At each time step, a shared query and the stacked AP embeddings are projected into queries, keys, and values for attention-based fusion.Source: MLP MLP · p. 6
  • The resulting multi-head attention weights determine each AP’s contribution to the fused embedding at that time step.The AP importance measure is the average normalized attention weight across heads.Source: MLP MLP · pp. 6–7
  • The location head predicts x- and y-axis coordinates, while the uncertainty head predicts uncertainty along both axes.Predictions are collected over all time steps.Source: MLP MLP · p. 7

D. Loss Functions

Training begins with mean square error loss and switches after warm-up to β-negative log likelihood. The latter jointly optimizes location and the uncertainty head.

D. Loss Functions

  • Mean square error loss is used during warm-up training before the model switches to β-negative log likelihood.Source: D. Loss Functions · p. 7
  • After warm-up, β-negative log likelihood is used to optimize location and uncertainty-head log σ jointly.The loss formulation uses a stop-gradient operator.Source: D. Loss Functions · p. 7
  • The algorithm processes each AP through forward and reverse directions, merges the results with a residual, and updates them through a normalized feed-forward block.Source: D. Loss Functions · p. 7

E. Post-hoc Uncertainty Calibration

The framework calibrates predicted per-axis uncertainty after training because β-NLL-trained models tend to be overconfident. It fits separate axis scales on validation data and applies them unchanged at test time.

E. Post-hoc Uncertainty Calibration

  • Because β-NLL-trained models tend to be overconfident, the predicted per-axis variance is calibrated post hoc.Calibration quality is measured by mean normalized squared error; χ²=2 indicates a well-calibrated two-dimensional Gaussian, while larger values indicate overconfidence.Source: E. Post-hoc Uncertainty Calibration · p. 7
  • Per-axis temperature scaling fits separate scales on the validation set to correct miscalibration.The resulting scaled standard deviations are applied unchanged to the test set.Source: E. Post-hoc Uncertainty Calibration · pp. 7–8
  • The closed-form scale in (53) exactly minimizes Gaussian negative log-likelihood over a per-axis scale.Source: E. Post-hoc Uncertainty Calibration · p. 8

IV. NUMERICAL RESULTS

The numerical evaluation uses a real-world 5G indoor-localization dataset with a held-out test set. Performance is assessed using validation distance, with metrics averaged over four independent runs.

IV. NUMERICAL RESULTS

  • The 5G dataset contains six APs, 128 CIR elements per AP, and TDoA and ToA-offset features, with separate training and test sets.It was collected at 100 MHz bandwidth and 3.7 GHz carrier frequency; the held-out test set contains 15,722 time steps.Source: §IV · p. 8
  • The last 20% of training time steps form the validation set, split before windowing to prevent windows crossing the train–validation boundary.The training and test sets contain 18,867 and 15,722 time steps, respectively.Source: §IV · p. 8
  • Performance metrics are averaged across four independent runs, with validation distance used as the key evaluation metric.The distance compares ground-truth and predicted locations along the x- and y-axes.Source: §IV · p. 8
  • Training uses AdamW for 15 epochs with six warm-up epochs, Z-score normalization, and windows of 25 time steps with stride five.Validation and test data use normalization statistics obtained from the training set.Source: §IV · p. 8

A. Fine-Tuning & Analysis

The analysis selects hyperparameters based on validation behavior, examines the accuracy–runtime trade-off from adding APs, and compares model and feature ablations against the full design.

A. Fine-Tuning & Analysis

  • η=0.003 and η=0.01 achieved the lowest validation distance, with η=0.003 showing lower variance; BS=16 had slightly lower variance than BS=8.dstate=16 achieved the lowest validation loss and distance, while λ=0.05 obtained slightly lower error overall; increasing dstate to 32 increased complexity without improvement.Source: A. Fine-Tuning & Analysis · p. 8, Figure 3 · p. 9
  • 29.12% lower validation distance from one to two APs came with 94.99% longer training and 77.61% longer validation time.From five to six APs, validation distance improved by 10.28%, while training and validation time increased by 17.59% and 17.47%; selective scan reduced six-AP training and validation time by 50.08% and 18.97%.Source: A. Fine-Tuning & Analysis · p. 8, Figure 4 · p. 9

B. Ablation Study

  • 51.76% performance degradation with No Fusion and 9.43% with Average Fusion show the advantage of attention fusion over these alternatives.Dropping TDoA and ToA-offset caused 26.15% performance loss, while dropping only ToA-offset caused 1.03% degradation; unidirectional Mamba, 1-D CNN, 1-D RNN, and MSE-only training degraded performance by 15.79%, 12.73%, 12.73%, and 4.00%.Source: B. Ablation Study · pp. 8–9

C. Performance on The Test Set

LocAttMamba achieves the best mean error on both test datasets while using the fewest multiply-accumulate operations among the benchmarks. Its test-set analysis also reports calibrated uncertainty and region-dependent AP attention weights.

C. Performance on The Test Set

  • The test-set mean error is 0.999 m and median error is 0.865 m, based on a single run.Fig. 5 maps prediction error by color: darker points indicate lower errors and brighter, yellow-near points higher errors.Source: C. Performance on The Test Set · p. 9, Figure 5 · p. 10
  • Per-axis temperature scaling reduces the test-set uncertainty statistic from χ2 = 34.6 to χ2 = 2.15, close to the ideal value of 2.Validation-fitted scales (sx = 3.26, sy = 4.87) match test-optimal scales within approximately 10%; calibration changes only the uncertainty head, not the location head.Source: C. Performance on The Test Set · p. 9
  • Average attention weights vary by region: AP 5 weighs more in lower-left regions, while obstructed AP 3 has the lowest average weight.The reported pattern is associated with AP proximity and line-of-sight availability.Source: C. Performance on The Test Set · pp. 9–10, Figure 6 · p. 10

D. Benchmark Comparison

  • LocAttMamba has the fewest multiply-accumulate operations and third-lowest inference time among the benchmarks.On 5G, it has 37.10% lower inference time and nearly 63 times fewer multiply-accumulate operations than Sec5GLoc; inference times were measured on a T4 GPU at batch size 1.Source: D. Benchmark Comparison · p. 10
  • On the 5G error CDF, LocAttMamba has better P75 and P90 errors, while Sec5GLoc has a lower median and SVM has the worst curve.Fig. 7 presents the benchmark error CDF on the 5G test set.Source: D. Benchmark Comparison · pp. 10–11, Figure 7 · p. 11
  • LocAttMamba beats the second-best mean-error benchmark by 9.79% on 5G and 5.82% on UWB, averaged over four runs.Sec5GLoc has lower median error on 5G, while L-SwiGLU-T has lower median and P75 error on UWB; LocAttMamba has better P75 and P90 errors on 5G.Source: D. Benchmark Comparison · p. 10, Table III · p. 11

E. Noise Robustness

Across the 5G and UWB datasets, LocAttMamba has the best or near-best accuracy, robustness, and complexity in the reported evaluations. Under CIR degradation, its attention weights decrease for the distorted AP, though noisy UWB results at severe degradation favor other models.

E. Noise Robustness

  • The evaluation considers single-AP CIR degradation over a range of SNRs and multiple-AP CIR degradation for APs {1, 3, 5} at a single SNR of 5 dB.The same AP is degraded for all algorithms.Source: E. Noise Robustness · pp. 11–12
  • With multiple APs distorted, LocAttMamba maintains superiority on both datasets in most reported settings, but Bi-LSTM has 5.48% lower mean error on 5G with five APs distorted.At three-AP distortion on 5G, its mean error is lower than MambaLoc by 37.85%, Sec5GLoc by 74.80%, L-SwiGLU-T by 67.76%, and Bi-LSTM by 9.64%.Source: E. Noise Robustness · p. 12, Table IV · p. 12
  • With one AP degraded to -10 dB, LocAttMamba has lower error than L-SwiGLU-T and Bi-LSTM on 5G, but L-SwiGLU-T and ACNN outperform it on UWB below -2.5 dB.At -10 dB on UWB, L-SwiGLU-T and ACNN have 50.50% and 16.85% lower error, respectively.Source: Figure 8 · p. 11, E. Noise Robustness · p. 12
  • For noisy AP 0, its average attention weight falls from 0.2267 with clean CIR to 0.2057 at 0 dB and 0.1784 at -10 dB.The distorted AP's weight does not fall to zero, with clean TDoA and ToA-offset features still present.Source: E. Noise Robustness · p. 12, Figure 9 · p. 12

V. CONCLUSIONS

  • LocAttMamba achieves mean errors of 1.004 m on 5G and 0.599 m on UWB, beating the second-best benchmarks by 9.79% and 5.82%, respectively.It has the fewest multiply-accumulate operations and third-lowest inference time among evaluated models.Source: §V · p. 12

Related papers

Loading 2610.05673v1…