Unsupervised Domain Adaptation in Activity Recognition: A GAN-Based Approach

Sensor-based human activity recognition (HAR) is having a significant impact in a wide range of applications in smart city, smart home, and personal healthcare. Such wide deployment of HAR systems often faces the annotation-scarcity challenge; that is, most of the HAR techniques, especially the deep learning techniques, require a large number of training data while annotating sensor data is very time- and effort-consuming. Unsupervised domain adaptation has been successfully applied to tackle this challenge, where the activity knowledge from a well-annotated domain can be transferred to a new, unlabelled domain. However, these existing techniques do not perform well on highly heterogeneous domains. This article proposes shift-GAN that integrate bidirectional generative adversarial networks (Bi-GAN) and kernel mean matching (KMM) in an innovative way to learn intrinsic, robust feature transfer between two heterogeneous domains. Bi-GAN consists of two GANs that are bound by a cyclic constraint, which enables more effective feature transfer than a classic, single GAN model. KMM is a powerful non-parametric technique to correct covariate shift, which further improves feature space alignment. Through a series of comprehensive, empirical evaluations, shift-GAN has not only achieved its superior performance over 10 state-of-the-art domain adaptation techniques but also demonstrated its effectiveness in learning activity-independent, intrinsic feature mappings between two domains, robustness to sensor noise, and less sensitivity to training data.

Automated summary

Sensor-based human activity recognition (HAR) plays a significant role in various fields such as smart city, smart home, and personal healthcare. The shift-GAN proposed in this article integrates Bi-GAN and KMM to achieve robust feature transfer between two different domains. Experimental results demonstrate that shift-GAN outperforms over 10 existing domain adaptation techniques, learns intrinsic feature mappings independent of activities between two domains, is robust to sensor noise, and less sensitive to training data.