Multilinear Motion Synthesis Using Geostatistics

The trend in interactive entertainment is towards scenes with massive numbers of characters, and requiring huge amounts of motion data, which must be compactly and efficiently stored without sacrificing quality or controllability of the motions. Multilinear algebra is a powerful tool for efficiently representing multivariate data, including human motion data, through the analysis of multimodal correlations. The multilinear model, however, often suffers from undesirable artifacts when motion data are sparsely and non-uniformly sampled in a high-dimensional control space. For overcoming this defect, we introduce a geostatistical interpolation to the multilinear model by formulating it to fit into the motion representation with tensor approximation. The advantages of this approach are demonstrated by the motion synthesis in a high-dimensional control space and by a level of detail control. This technique provides practical tools for implementing interactive animations of many characters while ensuring accurate and flexible controls with a small amount of storage.