Real-time activity monitoring with a wristband and a smartphone

Activity monitoring is a very important task in lifestyle and health domains where physical activity of a person plays an important role in further reasoning or for providing personalized recommendations. To make such services available to a broader population, one should use devices that most users already have, such as smartphones. Since trends show an increasing popularity of wrist-worn wearables we also consider a sensor-rich wristband as an optional device in this research. We present a real-time activity monitoring algorithm which utilizes data from the smartphone sensors, wristband sensors or their fusion for activity recognition and estimation of energy expenditure of the user. The algorithm detects which devices are present and uses an interval of walking for gravity detection and normalization of the orientation of the devices. The normalized data is afterwards used for the detection of the location of the smartphone on the body, which serves as a context for the selection of location-specific classification model for activity recognition. The recognized activity is finally used for the selection of one or multiple regression models for the estimation of the user's energy expenditure. To develop the machine-learning models, which can be deployed on the smartphone, we optimized the number and type of extracted features via automatic feature selection. We evaluated each step of the algorithm and each device configuration, and compared the human energy expenditure estimation results against the Bodymedia armband and Microsoft Band 2. We also evaluated the benefit of decision fusion where appropriate. The results show that we achieve a 87% ± 5% average accuracy for activity recognition and that we outperformed both competing devices in the estimation of human energy expenditure by achieving the mean absolute error of 0.6 ± 0.1 MET on average.