Human Activity Recognition Using FixMatch-based Semi-supervised Learning with CSI

Abstract

With the recent development of smart homes and IoT, sensing of daily activities has been attracting attention. Wi-Fi CSI-based methods have been studied as a method for activity recognition because they can reduce costs and protect privacy by not showing the identity of people. In a previous study it was reported that CSI can recognize simple human motions indoors. However, in the previous studies, the CSI-based methods were not able to recognize the various activities of people in their daily lives sufficiently. In addition, the training data is often manually annotated, and the training cost is high. Therefore, in this study, we performed action recognition using semi-supervised learning with a 1D-CNN on three receivers for everyday actions. The proposed method uses a 1D-CNN model to identify which room the subject is in when identifying actions such as global learning and then uses semi-supervised learning based on the FixMatch method as local learning to recognize more detailed actions. Although FixMatch was originally designed for image classification, the proposed method is implemented in such a way that it is effective even for time-series data format by using a data expansion method for time-series data. In basic experiments, we confirmed that the proposed method achieves an accuracy of 98% for global learning to determine rooms, and an accuracy higher than that of conventional supervised learning for local learning to recognize detailed actions.