Title: Enhancing Efficiency in WiFi CSI-Based Human Activity Recognition via Physics-Guided Attention in a Lightweight TCN

Abstract: The non-contact, cost-effective, and privacy-respecting attributes of Human Action Recognition (HAR) utilizing WiFi Channel State Information (CSI) have driven growing interest in this field. Nevertheless, current learning-driven methods predominantly depend on deep, resource-heavy architectures to implicitly deduce motion dynamics from CSI data, which exacerbates model complexity and hampers efficiency. We posit that embedding suitable inductive biases, designed around the physical properties of CSI signals, facilitates more efficient and potent learning. To this end, we present a streamlined temporal convolutional network (TCN) framework that explicitly embeds motion-aware inductive biases into the feature learning process. Our approach introduces a Doppler-energy-guided temporal attention mechanism within the feature space to highlight time segments characterized by significant motion, alongside a variance-driven channel attention module that adaptively assigns weights to informative subcarriers according to temporal motion statistics. By leveraging these domain-specific priors, the proposed model successfully captures motion dynamics without necessitating increased architectural depth. Comprehensive evaluations across various benchmark datasets reveal that our method outperforms deeper baseline models while substantially lowering both parameter counts and computational overhead.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC