Unsupervised probabilistic segmentation of motion data for mimesis modeling

Humanoid developments express the need for intelligent learning systems that can automatically realize behavior acquisition and symbol emergence. In the framework of mimesis model, we present an unsupervised dynamic HMM-based algorithm in order to analyze vectorial motion data. The efficiency of this algorithm is demonstrated by segmenting continuous sequence of real movements. We also propose to use it as the first level of an information treatment system by associating it with a recognition process. Unlike other existing segmentation-recognition system, our segmentation process does not need any learning of the parameters that increases the flexibility of the whole segmentation-recognition system and the range of its possible applications