Data-Driven Motion Reconstruction Using Local Regression Models

Reconstructing human motion data using a few input signals or trajectories is always challenging problem. This is due to the difficulty of reconstructing natural human motion since the low-dimensional control parameters cannot be directly used to reconstruct the high-dimensional human motion. Because of this limitation, a novel methodology is introduced in this paper that takes benefit of local dimensionality reduction techniques to reconstruct accurate and natural-looking full-body motion sequences using fewer number of input. In the proposed methodology, a group of local dynamic regression models is formed from pre-captured motion data to support the prior learning process that reconstructs the full-body motion of the character. The evaluation that held out has shown that such a methodology can reconstruct more accurate motion sequences than possible with other statistical models.