In this study, we investigate human activity recognition (HAR) using WiFi channel state information (CSI) signals, employing a single-layer gated recurrent unit (GRU) with an attention module. To overcome the limitations of existing state-of-the-art (SOTA) models, which, despite their good performance, have substantial model sizes, we propose a lightweight model that incorporates data augmentation and pruning techniques. Our primary goal is to maintain high performance while significantly reducing model complexity. The proposed method demonstrates promising results across four different datasets, in particular achieving an accuracy of about 98.92%, outperforming an SOTA model on the ARIL dataset while reducing the model size from 252.10 M to 0.0578 M parameters. Additionally, our method achieves a reduction in computational cost from 18.06 GFLOPs to 0.01 GFLOPs for the same dataset, making it highly suitable for practical HAR applications.
