Unsupervised domain adaptation in sensor-based human activity recognition

Abstract

Sensor-based human activity recognition (HAR) is to recognise human daily activities through a collection of ambient and wearable sensors. Sensor-based human activity recognition 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, existing techniques do not perform well on highly heterogeneous domains.

To address this problem, this thesis proposes unsupervised domain adaptation models for human activity recognition. The first model presented is a new knowledge- and data-driven technique to achieve coarse- and fine-grained feature alignment using variational autoencoders. This proposed approach demonstrates high recognition accuracy and robustness against sensor noise, compared to the state-of-the-art domain adaptation techniques. However, the limitations with this approach are that knowledge-driven annotation can be inaccurate and also the model incurs extra knowledge engineering effort to map the source and target domain. This limits the application of the model.

To tackle the above limitation, we then present another two data-driven unsupervised domain adaptation techniques. The first method is based on bidirectional generative adversarial networks (Bi-GAN) to perform domain adaptation. In order to improve the matching between the source and target domain, we employ Kernel Mean Matching (KMM) to enable covariate shift correction between transformed source data and original target data so that they can be better aligned. This technique works well but it does not separate classes that have similar patterns. To tackle this problem, our second method includes contrastive learning during the adaptation process to minimise the intra-class discrepancy and maximise the inter-class margin. Both methods are validated with high accuracy results on various experiments using three HAR datasets and multiple transfer learning tasks in comparison with 12 state-of-the-art techniques.