Recognizing Activities Using a Kinect Skeleton Tracking and Hidden Markov Models

Knowing in which activities users are involved is an essential part of their context, which become more and more important in modern context-aware applications, but determining these activities could be a daunting task. Many sensors have been used as information source for guessing human activity, such as accelerometers, video cameras, etc., but recently the availability of a sophisticated sensor designed specifically for tracking humans, as is the Microsoft Kinect has opened new opportunities. The aim of this paper is to determine some human activities, such as eating, reading, drinking, etc., while the person is seated, using the Kinect skeleton structure as input. In this paper we take an unsupervised approach based on K-means for clustering activities, and Hidden Markov Models (HMM) to recognize the activities captured with the Microsoft Kinect's skeleton tracking feature. We show also how the number of clusters affects the performance of the HMM, and that after reaching a certain number of clusters, the performance of the HMM models to recognize activities does not improve anymore.