Self-Attention Networks for Human Activity Recognition Using Wearable Devices

Abstract

Human activity recognition has gained a lot of attention in recent years as it has many potential applications, such as in smart homes, healthcare and sport monitoring. Sensors in wearable devices and smartphones are widely used not only because they are low cost but also because they are not invasive to users and are easy to deploy. However, accurately predicting human activities using wearable devices is challenging as the generated data provides only indirect information about the activities being performed. In this study, we propose a self-attention network that processes data from inertial measurement unit sensors of smartphones to classify common human activities. Self-attention networks are able to extract useful information from time-dependent signals by carefully allocating their focus among relevant input features. Our method was tested along several popular human activity recognition algorithms using two datasets, including a new human activity dataset that is publicly released in this study. Our method consistently obtains state-of-the-art results predicting the activities of the tested datasets with an average accuracy of 97%.