Exploiting Multiple Receivers for CSI-Based Activity Classification Using A Hybrid CNN-LSTM Model

Abstract

Channel State Information (CSI) has been used as an alternative sensing source for monitoring occupant's activities indoors. While various approaches have been proposed to extract features from the CSI and classify activities, those features fail to yield the spatial-temporal aspects of activities. In this context, this study presents new approach to extract appropriate features from multiple receivers. Time-series CSI data collected from a Wi-Fi receiver is converted into an image data by Short-Time Fourier Transform (STFT), and then such the image data from multiple receivers are combined into a large image data, so that it contains information about the spatial-temporal aspects and distinct patterns of the activities. Convolutional Neural Network (CNN) extracted miscellaneous features from the converted image data. The extracted features are then fed into Long Short-Term Memory (LSTM) to classify the activities. The proposed hybrid CNN-LSTM model offers over 95% accuracy for classifying the key activities in daily living (ADLs) (e.g., walking, eating, toileting, bathing, etc.). This approach also shows consistent performance in two different housing environments.