Indoor Seat Occupancy Classification with Wi-Fi Channel State Information and Machine Learning Methods

Keeping a distance by monitoring the seat occupancy is an essential way to prevent the spread of virus inside a room. However, most current human sensing methods need customized devices, so a cheaper way of indoor seat occupancy classification is in need. Recent researches indicate that Wi-Fi channel state information (CSI) can be utilized for indoor human sensing without wearable sensors. This paper proposes a multi-person seat occupancy classification method based on machine learning and Wi-Fi CSI received by commercial network interface card. We designed an experimental scenario of 5 seats and 2 individuals, and use commercial Wi-Fi devices to build a multi-input multi-output (MIMO) system indoors to acquire an adequate dataset. Then a pipeline consists of phase calibration, linear interpolation, outlier removal and threshold de-noising was applied to preprocess the raw CSI amplitude and phase data. After sliding window feature extraction, convolutional neural network (CNN) and some conventional machine learning methods, such as naive Bayes (NB), decision tree (DT), support vector machine (SVM) and K-nearest neighbors (KNN), are used to classify seat occupancy, among which CNN performs the best, with a classification accuracy of 95%.