Human Activity Recognition based on Wi-Fi CSI Data -A Deep Neural Network Approach
Andrii Zhuravchak, Oleg Kapshii, Evangelos Pournaras
Abstract
Using Wi-Fi Channel State Information (CSI) is a novel way of environmental sensing and human activity recognition (HAR). These methods can be used for several safety and security applications by (re)using Wi-Fi routers without the need for additional costly hardware required for vision-based approaches, known also to be particularly privacy-intrusive. This work introduces a full pipeline of a Wi-Fi CSI-based system for human activity recognition that assesses and compares two deep learning methods. We analyze how different hardware configurations affect WiFi CSI signals. We contribute a novel and more realistic data collection process, in which human activity recognition is seamlessly integrated in real-life, resulting in more reliable assessments of the model classification performance. We analyze how InceptionTime and LSTM-based classification models perform in human activity recognition. The source code and collected dataset are made publicly available for reproducibility and encouraging further research in the community.