The most attentive person selection using HMM with multiple sources

Abstract

We presented a novel HMM framework, Generative state model-based HMM (GHMM), treating the multiple sources whose outputs are simultaneously emitted while the conventional HMM is equipped for a single source. GHMM is designed for particular problems in which there is competition among sources (i.e., any GHMM state is a particular event when any source is more distinctive than others). The generative state model not only has ability to deal with the changes of the number of sources in runtimes but also forms the group relation in the sense of competition among sources, unlike the conventional HMM which is a predefined or fixed state model. We also applied the proposed method to the most attentive person selection. We have confirmed by preliminary experiments that the proposed method works well in the selection of the most attentive person to communicate with the robot.