Abstract

The advancements in Ambient Assisted Living (AAL) have been prompted by the growing population of elderly individuals facing diagnoses such as Dementia or Alzheimer’s, aiming to enhance their overall quality of life. To provide support it is important to know their daily activities and support them. A large portion of research in the field of Human Activity Recognition uses black box learning approaches such as deep learning, but there are cases where model based methods, such as Virtual Stochastic Sensors (VSSs) are competitive. This is possible because the model based methods can include system structure in the modeling process if it is known. VSS’s are derived from Hidden Markov Models (HMM) and applied to a CASAS single resident dataset, which is an apartment fitted with different types of am-bient sensors. For future applications a generalization of behavior, sensors or models is necessary so that models are not just trained and used for one specific apartment and setup. In this paper we analyze the effect of generalizing the residents behavior on the reconstruction accuracy. The generalization did lead to some improvements in the reconstruction accuracy, but the implications for the actual application need to be considered.