Abstract: Human activity recognition is a prerequisite for many innovative applications including elderly activity monitoring and support in order to enable elderly people to live longer independently in their own homes. Over the past decade, a diversity of different activity recognition approaches has been developed from which the majority focuses on the processing of data from one sensor modality only (e.g., vision). Nonetheless, a merging of data from multiple disparate sources has the potential of offering more accurate, robust, descriptive, intuitive, and meaningful results due to the availability of complementary and partially redundant information. This article (1) gives a review of existing multimodal approaches for elderly activity recognition in home settings, (2) introduces a powerful activity recognition model based on brain-inspired multimodal data mining methods, (3) employs this model for the purpose of daily activity recognition in a home setting using a publically available real world dataset, and (4) quantitatively compares the obtained results with state of the art multimodal activity recognition methods including hidden Markov models, conditional random fields, decision trees, a Bayes approach, and a context lattice.
