Environment-Independent Wi-Fi Sensing Using Wavelet Scattering Transform

Human Activity Recognition (HAR) using Wi-Fi signals is a promising non-intrusive and device-free technique. Wi-Fi Channel Frequency Response (CFR) can be used to extract Doppler features corresponding to different types of human activities. The key challenge in extracting Doppler features from Wi-Fi CFR is that the Wi-Fi transmitters and receivers are not time and frequency synchronized, which causes phase ambiguity in the CFR. Hence, Doppler features for different types of human motion cannot be estimated accurately. In order to extract the Doppler features, CSI phase pre-processing technique is required. Another significant challenge in developing a practical Wi-Fi HAR system using deep learning techniques is the generality to different persons and environments. To address the aforementioned challenges, a novel CFR pre-processing technique for phase correction based on the Root-MUSIC algorithm is proposed. After phase correction, Doppler features for different human activities are extracted using the Wavelet Scattering Transform (WST). These features are then used to train a Convolutional Neural Network (CNN) for activity classification. The proposed HAR framework is evaluated using a IEEE 802.11ax CFR dataset across multiple scenarios with varying person, day, and environment. Experimental results demonstrate that the proposed HAR framework achieves a classification accuracy of 96% when tested for the same person, same environment on a different day. In the most challenging environment-independent scenario, the HAR framework achieves 74% accuracy.