Adaptive classification of arbitrary activities through hidden Markov modeling with automated optimal initialization

Abstract

An adaptive method for classification of arbitrary activities is presented that assesses continuously the activity in which a subject is engaged, thus providing contextual information facilitating the interpretation of any continuous data gathered from an (unsupervised) applied wearable robotics device and its bearer. Specifically the effect of a novel adaptive and fully automated initialization method using Potts energy functionals is discussed. Exemplary data suggests that this method very likely improves overall performance equally or better than more traditional methods. This includes state of the art methods based on segmental k-means initialization that do require substantial recurrent manual intervention.