Skip to content
Conference Open access

Geometry Over Depth: Multi-Node WiFi CSI Fusion for Robust Human Activity Recognition

Oct 2026 · 2026 22nd International Conference on Wireless and Mobile Computing, Networking and Communications (WiMob) · pp. 1-6 · 0 citations · 26 references

Abstract

WiFi Channel State Information (CSI) enables contactless, device-free Human Activity Recognition (HAR), but a single transmitter-receiver link observes motion from one fixed geometry and is sensitive to subject position. The common response is to absorb this fragility within increasingly deep models, at the cost of data requirements, interpretability, and a computational footprint ill-suited to low-power nodes. This paper takes a different position that the fragility of single-viewpoint sensing is better addressed through geometry than through model capacity. Using low-cost ESP32 nodes, we test whether spatial diversity from multiple observation points can substitute for model complexity in CSI-based HAR. We extract a set of 43 per-subcarrier features across statistical, nonlinear, communication-oriented, and spectral families, apply a two-stage subcarrier-then-feature selection scheme, and fuse the selected features of two spatially separated nodes for classification with conventional machine learning. Evaluation uses data from three subjects under a subject-activity-aware protocol that prevents train-test leakage, with results aggregated over ten randomized repetitions. Feature-level fusion of two nodes improves recognition accuracy from 75.56% for the best single-node configuration to 80%, providing initial evidence that interpretable fusion of spatially diverse lowcost nodes yields more robust HAR than either node alone under the tested conditions.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.