Partner und Internationale Organisationen
(Englisch)
|
MIRAlab University of Geneva (CH), University of Bath, England (UK), University of Glasgov, Scotland (UK), Ecole des Mines de Nantes (F), Ecole Polytechnique Fédérale de Lausanne (CH), Technical University of Vienna (A), Universite' Joseph Fourier, Grenoble (F), l.N.R.l.A., Rennes (F), Universitat de les Illes Balears, Mallorca (E), Limburgs Universitair Centrum, Diepenbeek (B)
|
Abstract
(Englisch)
|
The team at Ecole Polytechnique Federale de Lausanne (EPFL) have successfully extended and completed research on generative models from capture motion including skeleton models of human figures. By blending aligned data of captured motion from different people, our motion models can generate new motions while providing a set of meaningful, high-level control parameters including age, gender, weight and height.
Based on classical techniques of exact interpolation, conventional methods have to use the same number of basis functions as that of example motions and a global linear system has to be solved in determining the parameters of the interpolation functions. In addition, naive representations of motion data are used without taking advantage of redundancies among the example motions: for example, over a thousand numbers of real-valued interpolations have to be solved if captured motion data are used. Due to this high demand on storage and computation, their applications are limited to motion models whose control parameters are only those variant for a single individual such as emotional and physical parameters: in walking, happy-tired walks or walks at different radii of turn.
On the contrary, our work is based on a multi-step function approximation for the interpolation with motion data compressed by principal component analysis. Then, basis functions less than the number of examples can be used, no global linear system needs to be solved and only a few dozens of real-valued interpolations have to done even if densely sampled data of motion capture are used. This enables motion models constructed out of a rich set of examples and fast computed. Therefore, motion models can be controlled by attributes invariant for each individuals such as age, gender, height and weight, which require many of example motions from different individuals: for instance, male-female walks and young-old walks.
We demonstrated this by constructing a motion model for `punching' out of 68 captured motion data of 8 people varying in age, gender, height and weight.
The contributions made in 2001 can be roughly outlined in the following time line: (1st quarter) implementation and test of hierarchical radial basis function networks for the multi-step function approximation
(2nd quarter) implementation and test of pre-processing techniques for motion data including one to Euclidianize orientation data
(3nd quarter) motion capture of eight people in different age, gender, height and weight
(4th quarter) construction and demonstration of parametric motion models by intergrating the techniques on a set of the captured motion data
|
