Dirichlet Process Gaussian Mixture Model for Activity Discovery in Smart Homes with Ambient Sensors

Most of the existing approaches to activity recognition in smart homes rely on supervised learning with well annotated sensor data. However obtaining such labeled data is not only challenging but sometimes also an unobtainable task, especially for senior citizens who may suffer various mental health disorders. Other unsupervised learning approaches to activity discovery are based on fixed complexity models that require the number of activities to be specified in advance. Such models may not be suitable for smart home setting as the activity space may change over time. In this paper, we propose to use a Bayesian nonparametric clustering method to discover the activities from ambient sensors deployed in smart homes. Our model can automatically infer the number of activities from observed data, thus can be widely applicable in smart home environment. We test our method on two smart home datasets, including a public dataset and a dataset collected in our project. The experiment results demonstrate the efficiency of our method in activity discovery in smart home environment.