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

TL;DR

The paper tackles indoor localization in GNSS-denied environments by proposing a topology-aware hybrid Wi‑Fi/BLE fingerprinting pipeline that combines physically consistent RSS normalization, Bayesian denoising, dual regressors (RF and wKNN), and principled fusion via Dempster–Shafer theory with Choquet integration, augmented by per-sample persistent-homology descriptors. It delivers both point estimates and interpretable belief maps, designed for embedded deployment on microcontrollers. Across two heterogeneous datasets, including a new ESP32-based collection, the full pipeline consistently improves over a strong PF+RF baseline, achieving 3.40 m and 2.45 m RMSE under 10% synthetic RSS noise, and 0.44 m / 0.32 m in noise-free conditions, while providing uncertainty quantification. The approach emphasizes robustness to noise, interpretability, and real-time feasibility, offering competitive performance with lower computational demands than learning-heavy methods and paving the way for multi-radio, topology-informed, on-device localization.

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-scoring or dBm -> linear mW -> 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) estimates and interpretable belief maps, and is engineered for microcontroller-class deployment with per-update cost O(T log M + log M + Mp + 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 m versus 6.292 +/- 0.13 m (20.6% relative reduction; 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.

Figures (12)

• 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 $\mathcal{O}(T\log M+\log M+M_p+S)$.
• Figure 2: Real picture of Test area for Dataset 1 on the third floor of the Faculty of Electrical and Computer Engineering, University of Tehran.
• Figure 3: Map of Test area for Dataset 1 on the third floor of the Faculty of Electrical and Computer Engineering, University of Tehran. Fifteen reference points are distributed across a $6\,\mathrm{m}\times14\,\mathrm{m}$ area ($50$ cm each lattice).
• Figure 4: Regressor comparison (noise-free). RF attains the best overall accuracy; wKNN is second due to strong local interpolation. Linear models and single trees underperform; the ranking is consistent across both datasets.
• Figure 5: Filter accuracy/latency trade-offs. Particle filtering (PF) yields the lowest error but at a higher computational cost; Kalman filtering (KF) offers the best real-time compromise. Trends are consistent across both datasets.