Research Article
Human Activity Recognition Using MSHNet Based on Wi-Fi CSI
@INPROCEEDINGS{10.1007/978-3-030-64002-6_4, author={Fuchao Wang and Pengsong Duan and Yangjie Cao and Jinsheng Kong and Hao Li}, title={Human Activity Recognition Using MSHNet Based on Wi-Fi CSI}, proceedings={Mobile Networks and Management. 10th EAI International Conference, MONAMI 2020, Chiba, Japan, November 10--12, 2020, Proceedings}, proceedings_a={MONAMI}, year={2020}, month={12}, keywords={Human activity recognition Wi-Fi CSI MIMO Voting mechanism}, doi={10.1007/978-3-030-64002-6_4} }- Fuchao Wang
Pengsong Duan
Yangjie Cao
Jinsheng Kong
Hao Li
Year: 2020
Human Activity Recognition Using MSHNet Based on Wi-Fi CSI
MONAMI
Springer
DOI: 10.1007/978-3-030-64002-6_4
Abstract
In recent years, with the prominent population aging problem, health conditions of aged solitaries are inherently gaining more and more attentions. Among the techniques allowing real-time health monitoring, activity perception has become an important and promising eld in both academia and industry. In this paper, a human activity perception recognition model, named MSHNet (Multi-Stream-Hybrid-Network) based on Deep Learning is proposed to solve the problems of difficulty in extracting perceptual features of Wi-Fi signals and low recognition accuracy in traditional Machine Learning methods. MSHNet adopts passive wireless sensing technology, it uses commercial off-the-shelf Wi-Fi devices to collect Channel State Information (CSI) based on underlying physical equipment and automatically extracts human activity features characterized by amplitude in CSI. Then MSHNet aggregates the data streams of the same receiving antenna using the wireless signal transceiving characteristics of Multiple Input Multiple Output (MIMO) and trains the aggregated data streams respectively. At last, the voting mechanism is adopted to select the best training result. The experimental results demonstrate that MSHNet’s results on the public dataset have reached the state-of-the-art and on the datasets of four environments collected by ourselves the average recognition accuracy rate has reached 97.41%, satisfying the daily activity monitoring of the elderly, especially those living alone.
