Unsupervised Clustering of Free-Living Human Activities using Ambulatory Accelerometry

This paper investigates unsupervised pattern recognition approaches to segment raw accelerometer signals into a sequence of events, for activity monitoring in a free-living environment. Ambulatory devices, such as accelerometers, have made it possible to classify movement, measure long-term trends in activity, and detect unexpected events such as falls. While such technologies are gaining acceptance for use in controlled clinical settings, their use in varying and unrestricted environments is still problematic. This is principally due to the difficulty of obtaining sufficient annotated data to train supervised event classification models, exacerbated by the fact that the most significant events are likely to be extremely rare. To address these limitations, this paper researches two unsupervised event segmentation techniques to (1) coherently cluster free-living activities and (2) detect unspecified unusual events, without requiring labelled data for prior learning. Experiments are presented for the clustering of data collected from a subject in free-living conditions using a triaxial accelerometer attached to the waist. Results show high overall cluster purity performances of 0.81 for the coherent clustering of activities, and 'intuitive' clusters that suggest atypical activities for the unusual event experiments.