Activity Recognition in Multi-User Environments Using Techniques of Multi-label Classification
Alaa Alhamoud, Vaidehi Muradi, Doreen Böhnstedt, Ralf Steinmetz · 2016
Activity recognition represents the cornerstone in realizing intelligent services such as energy conservation and ambient assisted living in smart environments. The problem statement of most activity recognition research assumes that only mutually exclusive activities occur in smart environments. The majority of research projects in this field focus on single-user environments where only one user performs a single activity at a given time. Such solutions are not applicable in real-world scenarios where multiple users reside in a home performing co-temporal activities. Our work addresses the problem of activity recognition in multi-user environments by utilizing the techniques of multi-label classification. It is based on a multi-label activity recognition dataset which we collected by deploying appliance-level power sensors as well as environmental sensors in a two-person apartment. In this dataset, a feature vector of sensor readings can have more than one label indicating the occurrence of more than one activity at a given time. In this work, we show that recognizing activities in smart environments can be achieved solely based on fine-granular power consumption data and without the need for installing any other sensing modality. Moreover, we prove that extracting and utilizing dependency relations between concurrent activities as well as temporal relations between subsequent activities provide a crucial enhancement of the predictive accuracy of activity recognition models.