Topology-Aware Hybrid Wi-Fi/BLE Fingerprinting via Evidence-Theoretic Fusion and Persistent Homology

Behrad Mousaei Shir-Mohammad, Behzad Moshiri, Abolfazl Yaghmaei · 2025

Read Topology-Aware Hybrid Wi-Fi/BLE Fingerprinting via Evidence-Theoretic Fusion and Persistent Homology by Behrad Mousaei Shir-Mohammad, Behzad Moshiri, Abolfazl Yaghmaei (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-16557.txt.

Introduction

Behrad Mousaei Shir-Mohammad, Behzad Moshiri, and Abolfazl Yaghmaei place indoor users in a 2025 study. GPS fails inside, so they use Wi-Fi and Bluetooth signal prints instead. A print is the set of signal strengths heard at one spot. Mixed phones, echoes, and noise make such prints hard to read. Their frame cleans prints, then blends two reader tools.

Shape traits from persistent homology, a tool that maps holes and loops, join each print. A proof-based fusion step merges the readers' calls with doubt scores. Tests on two data sets, one with 1,200 fresh samples, show clear gains. Error fell by about a fifth on mean across splits. The full frame runs on tiny chips for live use.

Catalog record

CREATORBehrad Mousaei Shir-Mohammad, Behzad Moshiri, Abolfazl Yaghmaei
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-16557

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Topology-Aware Hybrid Wi-Fi/BLE Fingerprinting via Evidence-Theoretic Fusion and Persistent Homology Behrad Mousaei Shir-Mohammad, Behzad Moshiri, Abolfazl Yaghmaei Abstract Indoor localization remains challenging in GNSS-denied environments due to multipath, device heterogeneity, and volatile radio conditions. We propose a topology-aware, hybrid Wi-Fi/BLE fingerprinting framework that (i) applies physically consistent RSS normalization (dBm z z -scoring or dBm → \!\to linear mW → z \!\to z -score), (ii) denoises streams with classical Bayesian filters (KF/UKF/PF), (iii) combines complementary regressors (Random Forest and weighted kNN with a diagonal Mahalanobis metric), (iv) performs evidence-theoretic fusion via Dempster–Shafer theory (DST), and (v) augments each sample with persistent-homology (PH) descriptors. The system outputs both ( x , y ) (x,y) estimates and interpretable belief maps, and is engineered for microcontroller-class deployment with per-update cost O ⁡ ( T ​ log ⁡ M + log ⁡ M + M p + S ) O(T\log M+\log M+M_{p}+S) . We evaluate on two heterogeneous datasets—including a new 1,200-sample ESP32 survey—and report ablations, robustness to test-only noise, and significance across 10 stratified splits. Under 10% synthetic RSS noise , the full pipeline attains 3.40 m (Dataset 1) and 2.45 m (Dataset 2) RMSE, improving a strong PF + RF baseline by about 37% . Averaged across splits, it yields 4.993 ± 0.15 4.993\pm 0.15 m versus 6.292 ± 0.13 6.292\pm 0.13 m (20.6% relative reduction; p < 0.001 p<0.001 ). In noise-free tests, accuracy tightens to 0.44 m and 0.32 m (up to 56% better). Compared with recent learning-heavy approaches that assume large site-specific datasets and GPU inference, our method delivers competitive accuracy with formal uncertainty quantification and low computational cost suitable for real-time deployment. Keywords Indoor localization, Wi-Fi fingerprinting, BLE beacons, Data Fusion, Discrete calculus, Kalman filter, particle filter, sensor fusion, Dempster–Shafer theory, Choquet integral, persistent homology, uncertainty quantification, embedded systems. 1 Introduction Global Navigation Satellite Systems (GNSS) degrade or fail indoors, yet most buildings already host dense Wi-Fi access points and, increasingly, Bluetooth Low Energy (BLE) beacons. RSS fingerprinting leverages these ambient radios by matching received-signal-strength (RSS) vectors to a surveyed database of reference points (RPs) [ 1 , 2 ] . In practice, however, multipath, device heterogeneity, and transient interference inject substantial noise into RSS, which destabilises distance metrics and degrades position estimates. Our goal is a lightweight, interpretable pipeline that explicitly addresses these realities while remaining deployable on commodity microcontrollers. RSS acquisition Wi-Fi (7 APs), BLE (3 beacons) Raw \mathbf{f}\in\mathbb{R}^{d} per scan Normalisation \tilde{\mathbf{f}}=(\mathbf{f}-\boldsymbol{\mu})\oslash\boldsymbol{\sigma} or \mathbf{p}=10^{\mathbf{f}/10} mW, \tilde{\mathbf{f}}=(\mathbf{p}-\boldsymbol{\mu}_{p})\oslash\boldsymbol{\sigma}_{p} RSS denoising (KF/UKF/PF) KF: \mathbf{x}_{t}=\mathbf{x}_{t-1}+\mathbf{v}_{t-1} , \tilde{\mathbf{z}}_{t}=\mathbf{x}_{t}+\mathbf{n}_{t} PF: w_{t}^{(m)}\!\propto\!p(\tilde{\mathbf{z}}_{t}\!\mid\!\mathbf{x}_{t}^{(m)}) ; resample if M_{\mathrm{eff}}<0.3M_{p} Feature assembly Fingerprint \tilde{\mathbf{z}}\in\mathbb{R}^{d} Per-sample PH on \{(i,\tilde{f}_{i})\}_{i=1}^{d} \phi_{\mathrm{PH}}=[\mathrm{NoP}_{0},\mathrm{PE}_{0},\mathrm{NoP}_{1},\mathrm{PE}_{1}] with \mathrm{PE}_{k}=-\sum_{i}\tilde{\ell}_{i}\log\tilde{\ell}_{i} , \tilde{\ell}_{i}=\dfrac{d_{i}-b_{i}}{\sum_{j}(d_{j}-b_{j})} Augment: \mathbf{x}=[\tilde{\mathbf{z}};\,\phi_{\mathrm{PH}}]\in\mathbb{R}^{d+4} Random Forest (RF) \hat{\mathbf{y}}_{\mathrm{RF}}=\dfrac{1}{T}\sum_{t=1}^{T}h_{t}(\mathbf{x}) Weighted k NN (wKNN) \delta(\mathbf{x},\mathbf{f}_{r})=(\mathbf{x}-\mathbf{f}_{r})^{\!\top}\boldsymbol{\Sigma}^{-1}(\mathbf{x}-\mathbf{f}_{r}) w_{r}\propto 1/(\delta+\epsilon) , \hat{\mathbf{y}}_{\mathrm{KNN}}=\dfrac{\sum_{r}w_{r}\,\mathbf{y}_{r}}{\sum_{r}w_{r}} Dempster–Shafer (DST) fusion Cells \Theta=\{\theta_{j}\}_{j=1}^{S} , m_{s}(\{\theta_{j}\})\propto e^{-\alpha\,\delta_{s}(j)} m(A)=\dfrac{\sum_{B\cap C=A}m_{1}(B)m_{2}(C)}{1-\sum_{B\cap C=\varnothing}m_{1}(B)m_{2}(C)} Belief map & estimate Belief m(\cdot) over cells; \theta^{\star}=\arg\max_{A\subseteq\Theta}m(A) Point \hat{\mathbf{y}} from cell centroid or fused regression Uncertainty: highest-belief regions Offline (normalisation stats, \boldsymbol{\Sigma} , RF training) Online (per-scan; cost \mathcal{O}(T\log M+\log M+M_{p}+S) ) Figure 1: End-to-end hybrid Wi-Fi/BLE fingerprinting pipeline. RSS is normalised, optionally denoised (KF/UKF/PF), and augmented with per-sample persistent-homology descriptors. Two complementary regressors (RF and wKNN) produce predictions that are fused by Dempster–Shafer (DST), yielding a point estimate and an interpretable belief map. Offline: learn normalisation statistics, \boldsymbol{\Sigma} , and RF. Online: per-scan operation with overall cost \mathcal{O}(T\log M+\log M+M_{p}+S) . In the broader context of sensor fusion and state estimation, prior work has demonstrated the effectiveness of Kalman-family filters and nonlinear blending for noisy, heterogeneous sensors: GPS/INS integration with nonlinear blending filters [ 3 ] , MEMS-IMU calibration with Kalman filtering [ 4 ] , and robust fault handling in microgrids using the Unscented Kalman filter [ 5 ] . These results motivate our choice of KF/UKF/PF to stabilise RSS before regression, providing accuracy gains with modest computational cost. Prior indoor-localisation work spans three strands. (i) Lightweight Wi-Fi–only methods (e.g., selective/range-limited k NN) are easy to deploy but typically plateau at several metres and rarely quantify uncertainty; for example, SLWKNN reports \sim 4.8 m and DumbLoc \sim 6.2 m in noise-free conditions, without explicit RSS noise modelling or multi-radio fusion [ 6 , 7 ] . (ii) Classical learning methods improve regressors or features (e.g., RF-assisted AP selection with kernel ELM) and achieve roughly 3.9 m, but still rely on a single radio and lack uncertainty handling [ 8 ] . (iii) Learning-heavy models learn complex radio–environment mappings and can reach 1–2 m using CNN–LSTM or high-order GNNs [ 9 , 14 ] ; extensions with data augmentation and temporal modelling further push accuracy, e.g., CSI trajectory fingerprinting with GAN-augmented CNN–LSTM [ 10 ] and generative hidden-Markov Wi-Fi fingerprinting attaining sub-meter accuracy under controlled conditions [ 11 ] . In parallel, Xia et al. optimise RSSI fingerprint matching in wireless sensor networks by refining the matching scheme for efficiency and accuracy [ 12 ] . While effective, these approaches are single-modality and typically do not expose uncertainty or perform evidence-theoretic fusion; by contrast, our hybrid Wi-Fi+BLE design integrates DST/Choquet fusion and PH features, and targets microcontroller-class deployment. Hybrid sensing has also been explored: SPOTTER fuses Wi-Fi and BLE in a particle-filter framework and reports \sim 2.7 m, but with comparatively heavy computation and without demonstration on resource-constrained hardware [ 13 ] . Complementary studies of multi-sensor fusion show that Dempster–Shafer evidence theory is effective under conflicting/uncertain measurements, and fuzzy aggregation can capture non-linear complementarity [ 22 , 23 ] . These gaps motivate a solution that pairs multi-modal sensing and principled fusion with efficient, classical components. We therefore propose a lightweight, real-time hybrid Wi-Fi/BLE fingerprinting framework that integrates four components: (i) RSS filtering via Kalman, Unscented Kalman, or Particle filters to attenuate fast fading [ 18 ] ; (ii) dual regressors—Random Forest (RF) and weighted k -nearest neighbours (wKNN)—to produce complementary coordinate predictions; (iii) model fusion using Dempster–Shafer Theory (DST) for principled evidence combination and a two-source Choquet fuzzy integral to capture non-linear complementarity [ 19 , 20 ] ; and (iv) topology-aware features computed per sample via persistent homology (PH) to summarise channel-structure geometry beyond variance-based statistics [ 21 ] . DST yields spatial belief maps visualised as highest-belief regions rather than Gaussian ellipses, improving operator interpretability. RF and wKNN respond differently to noise and RP sparsity, so fusing them stabilises predictions across sites; filtering reduces measurement noise at the source; and PH descriptors capture coarse geometric structure that PCA-like features may miss. All components are commodity, require no GPUs, and suit embedded deployment. Our design is informed by successes of nonlinear Kalman filtering and fuzzy aggregation in related sensing domains [ 3 , 4 , 5 , 22 , 23 ] . Problem Statement and Notation We consider indoor fingerprinting on a bounded floor plan. At time t , the sensor acquires a Wi-Fi/BLE RSS vector \mathbf{f}_{t}\in\mathbb{R}^{d} across N_{W} Wi-Fi and N_{B} BLE channels ( d=N_{W}+N_{B} ). Let \mathcal{D}_{\text{train}}=\{(\mathbf{f}^{(n)},\mathbf{r}^{(n)})\}_{n=1}^{M} be the labelled fingerprint–location pairs at surveyed reference points (RPs), with \mathbf{r}^{(n)}=(x^{(n)},y^{(n)})\in\mathbb{R}^{2} . The goal is to estimate \hat{\mathbf{r}}_{t}=(\hat{x}_{t},\hat{y}_{t}) for a new \mathbf{f}_{t} , while also exposing spatial uncertainty via a belief map over discretised cells \Theta=\{\theta_{j}\}_{j=1}^{S} . We normalise fingerprints either in dBm (direct z -score) or via dBm \to linear

Read the complete text