Sensor Event and Activity Prediction using Binary Sensors in Real Homes with Older Adults

Abstract

A fair amount of research on smart home functions has aimed at assisting older adults in their everyday life. The implementation of these functions relies on the performance of activity recognition and prediction algorithms. The research literature contains a number of well-performing algorithms for activity recognition and prediction. However, most of this work uses data collected in controlled environments (e.g. lab environments) using scripted activities. This thesis applied, evaluated, and compared the performance of state-of-the-art prediction algorithms in real-world scenarios. The project conducted a field trial including eight smart homes with one resident each, located in a care facility in Oslo for older adults over 65 years old. The homes have about 15 sensors, including motion, magnetic, and power sensors. Probabilistic methods and neural networks were applied to predict the next sensor event and activity of daily living in the home, as well as its time of occurrence. The work shows that it is possible to achieve acceptable prediction accuracy with few sensors and about three weeks of collected data. Transfer learning between apartments has also shown to be well performing. In addition, the study can be useful for deciding which prediction methods to use depending on the case, and in accordance with project constraints (e.g. the number of sensors, user privacy).