Human Activity Recognition Based on On-Body Antennas and Semi-Supervised Learning

Abstract

Human activity recognition is of great importance in the field of health rehabilitation, man-machine interaction, physical training and so on. On-body propagation in wireless body area network provides a low-cost and power-efficient method to detect and recognize human activity. In this paper, radio propagation data related to human activities are collected by on-body antennas, and a semi-supervised learning method is proposed to recognize human activity. Specifically, transmission coefficient data of four activities are recorded. Short-time Fourier transform is carried out on the original measurement signal to get the time-frequency spectrograms. An autoencoder is employed to extract low-dimensional feature of the data and K-Means algorithm is used for clustering. The autoencoder and clustering are trained by a small number of labeled data to improve the clustering accuracy. After the semi-supervised learning model is optimized, the remaining large number of unlabeled data are clustered and different human activities are recognized.