Wi-Fi CSI-Based Outdoor Human Flow Prediction Using a Support Vector Machine

This paper proposes a channel state information (CSI)-based prediction method of a human flow that includes activity. The objective of the paper is to predict a human flow in an outdoor road. This human flow prediction is useful for the prediction of the number of passing people and their activity without privacy issues as a result of the absence of any camera systems. In this paper, we assume seven types of activities: one, two, and three people walking; one, two, and three people running; and one person cycling. Since the CSI can effectively express the effect of multipath fading in wireless signals, we expected the CSI to predict the various activities. In our proposed method, the amplitude and phase components are extracted from the measured CSI. The feature values for machine learning are the mean and variance of the maximum eigenvalue derived from the auto-correlation matrix and variance-covariance matrix composed of the amplitude or phase components and the passing time of flow. Using these feature values, we evaluated the prediction accuracy by leave-one-out cross-validation with a linear support vector machine (SVM). As a result, the proposed method achieved the maximum prediction accuracy of 100% for each direction and 99.5% for two directions.