Deep Transfer Learning Using Class Augmentation for Sensor-Based Human Activity Recognition

Abstract

Transfer learning (TL) is an essential technique for human activity recognition (HAR) since collecting a large amount of labeled sensor data is cost-intensive. In the image recognition field, parameter-based TL using ImageNet has become the defacto standard; however, no multipurpose TL method has been established in HAR thus far. Therefore, we proposed a simple and novel technique, i.e., class augmentation (CA), which augments the diversity of the class categories of source domains based on the hypothesis that HAR datasets used as source domains have low diversity in class categories. Data augmentation (DA) enhances the diversity of input data, whereas CA enhances the diversity of output labels. We evaluated three types of CA using DA and HAR-specific auxiliary information (gender and sensor position). Our experiments using public HAR datasets revealed that using CA seems to enhance the transfer performance. Moreover, we confirmed that CA closes the interdomain distance between the source and target domains.