ActRec: A Wi-Fi-Based Human Activity Recognition System

Abstract

In this paper, we develop a Wi-Fi-based activity recognition system called ActRec, which can be used for the remote monitoring of elderly. ActRec comprises two parts: radio-frequency (RF) sensing and machine learning. In the RF sensing part, two laptops act as transmitter and receiver to record the channel transfer function of an indoor environment. This RF data is collected in the presence of seven human participants performing three activities: walking, falling, and sitting. The RF data containing the fingerprints of user activity is then pre-processed with various signal processing algorithms to reduce noise effects and to estimate the mean Doppler shift (MDS) of each data sample. We propose a feature extraction algorithm, which is applied to the MDS to obtain a feature vector used for activity classification. Moreover, we assess the activity recognition accuracy of three classification algorithms: K-nearest neighbors (KNN), naive Bayes, and decision tree. Our analysis reveals that the KNN, naive Bayes, and decision tree algorithms achieve an overall accuracy of 94%, 96.2%, and 98.9%, respectively.