Human activity recognition using wearable devices

Human activity recognition (HAR) is a key application on wearable devices in the areas of fitness tracking, healthcare and elder care support. However, inaccurate recognition results may cause an adverse effect on users or even an unpredictable accident. Therefore, it is necessary to improve the accuracy of human activity recognition.This thesis aims to provide effective and efficient HAR methods to address main challenges of HAR, which can be divided into the following three contributions. The first contribution is a novel feature extraction and selection algorithm that addresses the interclass similarity problem in the confounding activity recognition. The second contribution is a novel approach of leveraging local and global features, which addresses both the intraclass variability and interclass similarity problems in HAR. The third contribution is a multiscale feature engineering approach, which leverages local and global features and addresses the negative effect on HAR caused by users' different habits. For the proposed approaches, extensive experiments have been conducted on real datasets or real scenarios. The experiments have demonstrated the proposed methods are superior to the state of the art.