Human Activity Recognition Using Wi-Fi and Machine Learning

EasyChair Preprint 6268

Abstract

In the era of Internet of things, access points (APs) will be deployed everywhere. The wireless signals offered by these APs can be used for more than just Internet connectivity. In fact, the human movement causes Doppler shifts in the received wireless signals. By combining signal processing techniques and machine learning, it is possible to recognize human activity from Wi-Fi signals. This paper builds on these ideas and develops a human activity recognition system that comprises two parts: RF sensing and machine learning. In the RF sensing part, we record the channel transfer function of an indoor environment in the presence of a participant performing three activities: walking, falling, and picking up an object. Using signal processing techniques, we estimate the mean Doppler shift (MDS) of the channel, which contains the fingerprint of the user activity. The MDS is used by a classifier to determine the type of performed activity. We assess the activity recognition performance of three classification algorithms: cubic support vector machine (CSVM), K-nearest neighbor (KNN), and linear discriminant analysis (LDA). Our analysis shows that the CSVM, LDA, and KNN algorithms achieve an overall accuracy of 99.5%, 97.3%, and 95.1%, respectively.

Keyphrases: Channel State Information, Channel transfer function, Human Activity Recognition, Mean Doppler shift, machine learning

@booklet{EasyChair:6268,
  author    = {Ali Chelli and Muhammad Muaaz and Ahmed Abdelgawwad and Matthias Pätzold},
  title     = {Human Activity Recognition Using Wi-Fi and Machine Learning},
  howpublished = {EasyChair Preprint 6268},
  year      = {EasyChair, 2021}}