Contactless Human Activity Recognition using Deep Learning with Flexible and Scalable Software Define Radio

TL;DR

This paper explores using Wi-Fi channel state information and deep learning models, especially LSTM, for contactless indoor human activity recognition, achieving over 95% accuracy with scalable software-defined radio technology.

Contribution

It introduces a novel Wi-Fi CSI-based HAR system employing deep learning models, demonstrating superior accuracy and scalability for contactless activity recognition.

Findings

LSTM outperforms CNN and hybrid models in accuracy.

Achieved 95.3% average accuracy in multi-activity classification.

Collected diverse CSI samples for multiple activities and directions.

Abstract

Ambient computing is gaining popularity as a major technological advancement for the future. The modern era has witnessed a surge in the advancement in healthcare systems, with viable radio frequency solutions proposed for remote and unobtrusive human activity recognition (HAR). Specifically, this study investigates the use of Wi-Fi channel state information (CSI) as a novel method of ambient sensing that can be employed as a contactless means of recognizing human activity in indoor environments. These methods avoid additional costly hardware required for vision-based systems, which are privacy-intrusive, by (re)using Wi-Fi CSI for various safety and security applications. During an experiment utilizing universal software-defined radio (USRP) to collect CSI samples, it was observed that a subject engaged in six distinct activities, which included no activity, standing, sitting, and leaning forward, across different areas of the room. Additionally, more CSI samples were collected when the subject walked in two different directions. This study presents a Wi-Fi CSI-based HAR system that assesses and contrasts deep learning approaches, namely convolutional neural network (CNN), long short-term memory (LSTM), and hybrid (LSTM+CNN), employed for accurate activity recognition. The experimental results indicate that LSTM surpasses current models and achieves an average accuracy of 95.3% in multi-activity classification when compared to CNN and hybrid techniques. In the future, research needs to study the significance of resilience in diverse and dynamic environments to identify the activity of multiple users.