Efficient machine learning for Wi-Fi CSI-based human activity recognition using fast Monte Carlo based feature extraction

Wi-Fi sensing is gaining attention to recognize human activities without intruding on people's privacy. Using existing wireless infrastructure, human activity can be detected and classified without wearables or cameras. Once there is a movement within a WI-Fi channel, the phase and amplitude are altered and are distinct for each activity. The channel state information (CSI) contains detailed information about the phase and amplitude within the channel and picks up subtle changes in the signal caused by a person's movement through space. While CSI offers rich temporal and spatial information, its high dimensionality poses significant challenges for real-time processing, storage, and machine learning. This thesis as presented in chapter 2, presents an efficient feature extraction framework for CSI-based human activity recognition (HAR) that uses a fast Monte Carlo-based algorithm grounded in the Frieze-Kannan-Vempala (FKV) low-rank approximation technique, which is employed to extract a compact and discriminative subset of features from the Doppler-transformed CSI data. The algorithm utilizes length-squared sampling to identify high-energy rows and columns of the data matrix and applies singular value decomposition (SVD) on a smaller submatrix to uncover latent activity patterns. Rejection sampling subsequently selects features that align with learned directions of maximum relevance, guided by label information. The selected features significantly reduce the input dimensionality, enabling efficient training of convolutional neural networks (CNNs), long-short-term-model (LSTM), CNN-LSTM, and decision tree for activity classification. Experimental results demonstrate that the proposed feature extraction method helps preserve classification accuracy while substantially reducing computational complexity. This makes the approach suitable for edge-device deployment and scalable HAR systems. This work advances low-cost, privacy-preserving, and computationally efficient HAR by integrating signal processing, randomized algorithms, and machine learning.