Comparative Study of Deep Learning Algorithms for Wi-Fi Channel State Information Based Human Activity Recognition

Human Activity Recognition (HAR) using Wi-Fi Channel State Information (CSI) is a promising device-free, noninvasive sensing approach. This study compares deep learning architectures, including Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), Long Short Term Memory (LSTM), Convolutional Neural Network - Bidirectional LSTM(CNNBiLSTM), Multi-Head Self-Attention - Bidirectional LSTM (MHA-BiLSTM), and Transformer encoders, for HAR using amplitude-based CSI data. Data were preprocessed with denoising filters, including Butterworth and Wavelet filters, followed by frequency domain feature extraction. Models were tested across diverse indoor environments to assess generalizability. Results show hybrid architectures like CNN-BiLSTM, and MHABiLSTM outperform simpler models, especially in varying environments. Effective preprocessing, such as denoising and sequence generation, significantly enhances accuracy. This study highlights optimal model and preprocessing strategies for robust, environment-agnostic HAR systems