A Highly Efficient Human Activity Classification Method Using Mobile Data from Wearable Sensors

Abstract

Effectively recognising human activity from wearable sensors is a valuable yet challenging task due to the intrinsic data complexity and the inevitable dynamic nature of diverse application scenarios. Existing works are weak to address human activity dynamics and to capture the low level structural relationships between an activity and the context in a consecutive time window in a real-time manner. In this paper, we propose a novel bag-level human activity classification method that can efficiently recognise activities from sensing data. The proposed method hierarchically captures the low-level instance characteristics and the high-level intra-bag information. Meanwhile, this method leverages the learning speed by jointly using random projection and least squares, which inherits from extreme learning machine (ELM) with solid theoretical foundation. Experiments results over the USC-HAD dataset demonstrate that the proposed method consistently outperforms the state-of-the-art instance-level and bag-level detection methods in terms of both recognition accuracy and learning speed.