Methodologies for continuous activity classification of user through wearable devices: Feasibility and preliminary investigation

Abstract

Nowadays, the possibility to use wearable devices at low-cost and with measuring capabilities is widely spread. They are able to collect multiple physiological and physical quantities, but the information provided is generally of low level (series of raw data) for the end-user, caregiver or the medical staff. A smart processing and an intelligent combination of these data may lead to the identification of more refined indicators (e.g. the kind of activity performed by the user) and so easier to understand information. In this work, the authors have performed a preliminary investigation on how to discriminate different levels of physical activity (resting, walking, running) conducted by a user. The analysis has been carried out with different trials and the use of a commercial wearable sensor (BioHarness 3.0), which measures five physiological and physical quantities. Different approaches, based on classification and clustering techniques, have been tested to prove their ability in discriminating the tasks performed. Results show that a threshold-based approach, applied to the physical quantities measured, is able to identify the activities conducted by the user, with an average accuracy of 98%. Moreover, a novel indicator based on the combination of both physiological and physical quantities is proposed and discussed in the paper. The integration of physiological data (e.g. Heart Rate) could lead to more high level information, which should be better investigated in future works.