HARFormer: WiFi-Based Human Activity Recognition with Dynamic Tanh Transformer

Human Activity Recognition (HAR) is essential for applications in smart healthcare, ambient intelligence, and behavior monitoring. Recently, deep learning-based WiFi HAR approaches have achieved high accuracy, particularly those based on transformer architectures. Transformer-based HAR methods adopt layer normalization layers to improve model's stability and performance. However, it requires tuning the training hyperparameters and dampens learning dynamics. Targeting these challenges, we proposed a novel HARFormer architecture, a lightweight and efficient Transformer-based model specifically designed for sequential sensor data in HAR tasks. Different with traditional transformer, we first introduce Dynamic Tanh (DyT) to replace Layer Normalization, a minimalist alternative that directly suppressing extreme activations and introducing nonlinearity. This HARFormer not only simplifies the model structure but also enhances training stability and scalability. Extensive experiments on in-domain (Day 1) and cross-domain (Day 2 and Day 3) Channel State Information (CSI) datasets are conducted and show that the proposed HARFormer outperforms state-of-the-art methods across three scenarios, achieving accuracies of 98.93%, 87.73%, and 88.84% respectively.