TF-Mamba: A Lightweight State-Space Model for Wi-Fi-Based Human Activity Recognition

Human activity recognition (HAR) using Wi-Fi-based systems is essential for applications such as smart homes, healthcare, and security. Traditional HAR methods, which rely on channel state information (CSI) to detect human movements, often struggle to balance recognition accuracy with computational efficiency. In this article, we introduce time-frequency Mamba (TF-Mamba), a lightweight state-space model (SSM) optimized for HAR through time-frequency analysis of CSI data. TF-Mamba features a dual-stream architecture comprising Time-Mamba and Frequency-Mamba blocks, efficiently capturing both the temporal (time-domain) and spectral (frequency-domain) features. The incorporation of a 2-D discrete wavelet transform (DWT) enhances the model’s capability to extract comprehensive features from CSI data. Extensive experiments on benchmark datasets, including UT-HAR, NTU-Fi, and HUST-HAR, confirm TF-Mamba’s superiority over existing methods. Notably, TF-Mamba achieves an accuracy of 99.72% on the HUST-HAR dataset while reducing computational complexity by up to 98.9% in multiply-accumulate operations (MACs) compared with previous state-of-the-art (SOTA) models. These results highlight TF-Mamba’s potential as an efficient and highly accurate solution for HAR, especially in resource-constrained environments. Source code and datasets are available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/Junshuo-Lau/HUST_HAR</uri>.