A Symbol Grounding Problem of Gesture Motion through a Self-organizing Network of Time-varying Motion Images

Abstract

We applied a method for self-organizing network to gesture motion images. We extract both so-called common parts and singular parts of gesture by analysis of the network topology. We call these parts to elemental motion units. In this paper, we propose a method to extract elemental motion units from gesture motion images and a gesture recognition method of a large vocabulary on the basis of recognition of elemental motion units. When recognizing gestures in time-varying motion images, the usual adopted method has been to distinguish different kind of gestures from a series of movie or images. However when a gesture targeted for recognition has an extensive vocabulary, the recognition of that gesture with an extended version of the usual method is difficult. So we use elemental motion unit to recognize gestures with large vocabulary. We have already proposed the incremental path method that is one of self-organizing network method. This method is constructing an incremental network structure given an input sequence. We applied this method to gesture motion images, so that we extract common parts in some gestures and singular parts. We show detail of incremental path method in section 11.5 and Automated extraction method in section 11.6. At last we describe new gesture recognition method using this elemental motion units and conclusion.