Activity recognition in smart homes using absolute temporal information in dynamic graphical models

Human activity recognition (HAR) is a fundamental task in smart homes. In these environments residents´ data are collected via unobtrusive sensors, and human activities are inferred using machine learning mechanisms out of sensors´ data. Dynamic graphical models (DGMs) have been a widely used family of machine learning mechanisms for HAR. In DGM-based HAR methods relative temporal information and duration of activities are intrinsically hired for modelling activities, while neglecting absolute temporal information in their primitive forms. Some human activities in the home almost have certain temporal patterns i.e. they take place at certain points of time throughout a periodic time span. Such temporal information can improve the recognition efficiency. In this paper, a general method is proposed to incorporate absolute temporal information into a DGM-based HAR method of interest. To do this, a periodic time span is treated as a virtual sensor and added to the set of physical sensors. Then, a DGM-based HAR scheme can be hired to infer the activities using the data from the augmented set of sensors. To evaluate the proposed method, it is applied in four DGM-based activity recognition mechanisms, which are based on naive Bayes classifiers (NBCs), hidden Markov models (HMMs), hidden semi Markov models (HSMMs), and conditional random fields (CRFs). A well-known activity of daily life (ADL) dataset is used for the simulations. The results show that the proposed method improves the classification performance in terms of F-measure, and accuracy.