Experiments with Dissimilarity Measures for Clustering Waveform Data from Wearable Sensors

摘要

Clustering waveform data is used in applications ranging from healthcare to economics and entertainment. In this paper, we present a study on clustering gestures enacted by subjects while wearing wrist-worn accelerometer sensors through different dissimilarity measures between individual components of multi-variate waveform data. We show how dissimilarity measures between different components of a multi-variate waveform database can measure the similarity, or the lack of it, between the motion of two hands in order to differentiate between different gestures, for applications in assistive technology and smart health-care. In doing so, we exploit a hierarchical clustering architecture and visualize it through single-linkage dendrograms and visual assessment of cluster tendency. Using annotations of the gestures, we describe the physical significance behind the formation of the hierarchy. We also discuss combining different dissimilarity measures by convex combination to improve clustering.