Radio-frequency (RF) sensing for deep awareness of human physical status

Abstract

Over the last few years, there have been a variety of human sensing applications developed through Radiofrequency (RF) sensing engaged in multiple different sectors. Traditional human activity recognition (HAR) methods have involved the use of sensors, which can be inconvenient and invade the user’s privacy. As such, Wi-Fi sensing, a type of RF sensing, provides a contactless yet effective way to achieve similar effects as traditional sensors. In this project, a Bidirection Long Short-Term Memory (BiLSTM) model was used to train channel state information (CSI) data from Wi-Fi signals collected in an indoor environment, which achieved an accuracy of 83.65% when classifying between static and dynamic actions. Alongside the use of the wavelet denoising method, the results indicate that while human activities can be classified with CSI information with high accuracy, further research is still necessary for improved accuracy and applicability of the model to the real-world environment.