Efficient Parameterized Methods for Physical Activity Detection using only Smartphone Sensors

Abstract

Detecting daily physical activities is very important in applications such as developing automated comfort scenarios for an individual. Motion smartphone sensors were previously used only as a complementary input whereas now, they are increasingly used as the primary data source for motion recognition. In this work, we use smartphone accelerometers to recognize online four daily human activities: sitting, walking, lying and running. We design two new hybrid protocols combining state of the art methods in a parameterized way. Then, we implement those protocols in the context of Android applications, which we develop. The first composition is more accurate and the second one is more energy efficient in terms of battery usage. Finally, we manage to personalize the model for online training of data sensors, which we create initially, to better adapt to the particular individual. According to our experimental performance evaluation, our hybrid methods achieve very high accuracy (even 99\%), while keeping battery dissipation at very satisfactory levels (average battery consumption 874mW).