Learning Statistical Models of Human Motion
Richard Bowden · 2000
: Non-linear statistical models of deformation provide methods to learn a priori shape and deformation for an object or class of objects by example. This paper extends these models of deformation to that of motion by augmenting the discrete representation of piecewise nonlinear principle component analysis of shape with a markov chain which represents the temporal dynamics of the model. In this manner, mean trajectories can be learnt and reproduced for either the simulation of movement or for object tracking. This paper demonstrates the use of these techniques in learning human motion from capture data. 1: Introduction Mathematical models of how an object or class of object deform and move with time are important in both computer vision and in graphical simulation. By maintaining an internal representation of shape and motion, object location and tracking can be simplified, or realistic motion can be generated. The key to this approach is the ability to learn what is shape...