Learning hierarchical non-parametric hidden Markov model of human motion

In this paper, we summarize the generative models for learning and automatic recreation of human motion as motion engine. We incarnate motion engine with a new model NPHHMM, a hierarchical hidden Markov model with non-parametric output densities. Our work contributes in three aspects: (1) NPHHMM models both temporal and spatial characteristics of human motion precisely so to support recreation of new motion; (2) compared with first-order hidden Markov model, NPHHMM has longer memory for more accurate prediction; (3) the EM learning algorithm of NPHHMM incorporates output densities in non-parametric form, which provides a compact representation of prototypical poses without requirements of explicit compression or losing accuracy of body model by extracting the most significant dependencies between body joints. NPHHMM learned from captured 3D human motion can be used to generate a variety of realistic new motions, thus is useful for data-driven motion editing and synthesis, which is recently an active research area to relieve animators from intensive labor of manual work and to automate the production of character animation.