Wi-Fi CSI Based Human Activity Recognition

In this paper, we present a non-vision-based approach for Human Activity Recognition (HAR) aiming to address the privacy concerns associated with vision-based methods. In recent days, HAR finds applications in various areas, including fall detection, pose estimation in the healthcare sector and workplace safety. HAR tasks are predominantly performed by vision-based systems, leading to privacy concerns. Thus, we propose a system to leverage non-vision based technique using Wi-Fi Channel State Information (CSI) signals considering variations caused by human movements for detection and classification of activities including ‘idle’, ‘walking’ and ‘standing’. The proposed approach includes Long Short-Term Memory (LSTM) based architecture to predict human activity and serves as a way to monitor patient activities to ensure adequate rest and recovery. The contribution of the paper includes generation of CSI dataset for ‘idle’, ‘standing’, ‘walking’, ‘squatting’ and ‘jumping’ activities and demonstrate results using the LSTM model trained on the generated dataset thereby achieving an accuracy of $94.36 \%$.