Contextual pattern clustering for ontology based activity recognition in smart home

Abstract

Ambient Assisted Living (AAL) enabled in a smart home requires, the design of an activity recognition system. Generally, supervised machine learning strategies or knowledge engineering strategies are employed in the process of activity modeling. Supervised machine learning approaches incur overheads in annotating the dataset, while the knowledge modeling approaches incur overhead by being dependent on the domain expert for occupant specific knowledge. The proposed approach on the other hand, employs an unsupervised machine learning strategy to readily extract knowledge from unlabelled data and subsequently represents it as ontology activity model. The novelty in the proposed design is in the usage of Contextual Pattern Clustering (CPC) for activity modeling. The competence of the weighted Jaro Winkler similarity measure introduced in CPC lies in the utilization of contextual attributes for the composition of varied event patterns of an activity. Hierarchical strategy employed in CPC offers structured knowledge on activities and sub activities within a specific location. Additionally, the event organizer and habitual event generator subsystem introduced in the proposed framework derives knowledge related to event ordering and contextual description of an activity. The attained knowledge is later represented as a probabilistic ontology activity model to enable probabilistic reasoning over domain knowledge. An experimental analysis with a smart home dataset demonstrates the proficiency of the proposed unsupervised approach in activity modeling and recognition in comparison with that of the existing modeling strategies.