Passive Channel Charting: Locating Passive Targets using Wi-Fi Channel State Information

TL;DR

This work demonstrates that passive channel charting (PCC) can localize passive targets using Wi-Fi CSI without requiring active transmitters or explicit channel models. By casting passive localization as a dimensionality-reduction problem, PCC leverages a Siamese NN to learn a channel chart that preserves dissimilarity relationships derived from CSI, with optional augmentation from classical triangulation to place estimates in a global frame. Empirical results on indoor ESPARGOS-based data show PCC surpasses a classical triangulation baseline and approaches or matches supervised fingerprinting under certain conditions, though it can overfit to target type. The study underscores PCC’s potential for non-LoS sensing and multi-static scenarios, while highlighting open challenges in generalization and multi-target extension with future work addressing these limitations.

Abstract

We propose passive channel charting, an extension of channel charting to passive target localization. As in conventional channel charting, we follow a dimensionality reduction approach to reconstruct a physically interpretable map of target positions from similarities in high-dimensional channel state information. We show that algorithms and neural network architectures developed in the context of channel charting with active mobile transmitters can be straightforwardly applied to the passive case, where we assume a scenario with static transmitters and receivers and a mobile target. We evaluate our method on a channel state information dataset collected indoors with a distributed setup of ESPARGOS Wi-Fi sensing antenna arrays. This scenario can be interpreted as either a multi-static or passive radar system. We demonstrate that passive channel charting outperforms a baseline based on classical triangulation in terms of localization accuracy. We discuss our results and highlight some unsolved issues related to the proposed concept.

Figures (5)

• Figure 1: Concept of passive channel charting as developed in this work: Passive antenna arrays use signals from non-cooperative beacon transmitter to locate passive target (here: human) in environment, which perturbs clutter channel by scattering / reflecting or absorbing signal components.
• Figure 2: Photo of the environment with the four ESPARGOS arrays (one in foreground, two at the left / right edges and one in the background), the passive target (robot wrapped in aluminium foil) in the middle of the measurement area. Beacon transmitters are mounted to the ceiling and not visible. The dimensions of the measurement area are approximately $4.5\,\mathrm{m} \times 4.5\,\mathrm{m}$.
• Figure 3: Neural network structure: (a) Dense NN used for fingerprinting or as FCF and (b) FCF in Siamese configuration for channel charting training.
• Figure 4: Top view map of colorized position labels in $\mathcal{S}_\mathrm{rob,test}$ after clustering shown in (a). Datapoint colors are preserved for (b), (c), (d) and (e), which show (b) position estimates produced by the classical triangulation baseline (c) position estimates from the supervised NN, (d) the passive channel chart before the coordinate transform and (e) the augmented passive channel chart, with classical AoA estimates considered during training (all evaluated on $\mathcal{S}_\mathrm{rob,test}$).
• Figure 5: Empirical cumulative distribution functions of absolute localization errors for both baselines and PCC (augmented and un-augmented after coordinate transform $\mathcal{T}_\mathrm{opt}$)