Segmentation of Human Behavior Using Singular Spectrum Transformation

Abstract

Recently, needs of supporting human activities of daily life are growing, where understanding their intentions from observed behaviors in real time is of importance. Because of the too many variations of the behaviors, conventional pattern matching methodologies for signals are not adequate, and time series data of the behaviors should be interpreted in terms of actor's intention and/or feelings as well as of their contexts. From this point of view, a "semiotic approach" is adopted, and as a first step towards intention learning from observation, a new method for segmenting human continuous behaviors into meaningful chunks corresponding to the actor's varied intentions by usage of Singular Spectrum Transformation is proposed.