Motion synthesis

We investigate the problem of synthesizing natural looking motions for virtual characters. Animating virtual characters is an important problem for many applications ranging from training simulations to computer games and movie production. In this thesis, we present motion synthesis methods that generate character motions by cutting and pasting motion capture data. Selecting a collection of clips from a motion database, that yields an acceptable motion is a combinatorial problem. Human motion is dynamically constrained by muscle strength and physics. This means we can not change our motion arbitrarily quickly. We take advantage of this observation and present efficient motion synthesis algorithms for generating motions that satisfy a variety of constraints automatically. These constraints include geometric constraints that prescribe the spatial structure of the motion, such as "Go from point A to point B in a specific amount of time" as well as annotation constraints that prescribe the types of motions to be performed, such as "Walk, pickup an object and then run". Human looking motions that satisfy the constraints are synthesized interactively. Complex motions can be synthesized by imposing a combination of multiple constraints.; We also present a semi-supervised motion classification method which is useful for annotating large sets of motion. This method has been successfully used to learn instances of particular types of motion for our synthesis algorithms and is also useful for action recognition.