Publications

CONFERENCE (INTERNATIONAL) EAU-FM: Environmentally Adaptive UWB-Fingerprints Maintenance

Kotaro Ikunishi (Ritsumeikan University), Yuki Hirayama (Ritsumeikan University), Daichi Kawanishi (Ritsumeikan University), Kota Tsubouchi, Masamichi Shimosaka (Institute of Science Tokyo), Nobuhiko Nishio (Ritsumeikan University)

The 16th International Conference on Indoor Positioning and Indoor Navigation (IPIN 2026)

November 18, 2026

Ultra-wideband (UWB) fingerprint (FP) localization degrades when obstacles change multipath; however, rebuilding the entire FP database is costly. We address the unlabeled update problem: selecting which reference points (RPs) to re-survey using post-change observations. The proposed Environmentally Adaptive UWB-Fingerprints Maintenance (EAU-FM) pipeline combines CIR-derived multipath signatures with map-consistent geometric features. We train an autoencoder (AE) on pre‑change data and use the reconstruction error as a degradation score. A kernel density estimator (KDE) is then used to select a threshold that splits the samples into Good and Poor subsets. From Good samples, we estimate positions and detect unassigned RPs via density. Poor samples are assigned to candidate RPs in FP space (nearest‑neighbor). Gaussian smoothing on the RP graph produces a degradation map and a prioritized RP list. We evaluated EAU-FM across three practical scenarios: a standard environment with an introduced obstacle, uneven data distributions, and a large-scale space. In the standard setting, EAU-FM outperforms baseline methods by achieving the steepest error reduction during the initial update phase, rapidly decreasing the mean localization error from 73.46~cm to 29.85~cm with just two updates. Across all evaluated scenarios, the results demonstrate that EAU-FM is particularly effective during the early update phases, substantially minimizing the site survey effort required for maintenance.

Paper : EAU-FM: Environmentally Adaptive UWB-Fingerprints Maintenanceopen into new tab or window (external link)