Unsupervised Learning Motion Models Using Dynamic Time Warping
Marek Kulbacki, Artur Bąk · 2002
This paper concerns essential, practical problem in automatic animation human-like figures with the support of informatics technologies connected with motion capture domain. The main problem we want to solve is partition set of primitive motions into appropriate groups according to similarity between motions. Up to now, experiments in systems of this kind, appeared be not too adequate to needs. In this situation, we had been faced with the necessity of creating new methods for supporting process of managing motion data. We construct motion models to easier extract features of given motions. Using these models we propose measure of discrepancy between motions. It shows how two motions are similar to each other, normalizes length of motions and decreases high dimension of considered motion data, so clustering may take place in dimensionally reduced space.