Joint Transfer Strategy for Cross-Domain Human Activity Recognition

Abstract

Human activity recognition based on wearable sensors has been considered as an important work in the increasingly developed artificial intelligence and Internet of Things. Since acquiring enough activity labels is often expensive and time-consuming, an intuitive way is to leverage existing data from one sensor (source domain) to recognize target data from another sensor (target domain). Unfortunately, the data from different domains may have a large discrepancy. Existing approaches typically consider reducing the domain discrepancy to transfer knowledge. However, these approaches do not take full advantage of data structure, which is important for activity recognition. In this paper, we propose an effective method, named Joint Transfer Strategy (JTS). The method has high accuracy and robustness for cross-domain human activity recognition. Specifically, JTS first obtains pseudo labels for the target domain via a majority voting technique. Then, it leverages a joint transfer strategy to map the source and target domain data into a shared domain-invariant subspace and preserve the data structure information. Finally, an adaptive classifier is learned to label target data. Comprehensive experiments on three large public activity recognition datasets demonstrate that JTS outperforms other state-of-the-art methods in terms of classification accuracy.