Automated Methods for Data-Driven Synthesis of Realistic and Controllable Human Motion


Human motion is difficult to animate convincingly --- not only is the motion itself intrinsically complicated, but human observers have a lifetime of familiarity with human movement, which makes it easy to detect even minor flaws in animated motion. To create high-fidelity animations of humans, there has been growing interest in motion capture, a technology that obtains strikingly realistic 3D recordings of the movement of a live performer. However, by itself motion capture offers little control to an animator, as it only allows one to play back what has been recorded. This dissertation shows how to use motion capture data to build generative models that can synthesize new, realistic motion while providing animators with high-level control over the properties of this motion. In contrast to previous work, this dissertation focuses on automated algorithms that make it feasible to work with the large data sets that are needed to construct expressive motion models. Two models in particular are considered. The first is the motion graph, which allows one to rearrange and seamlessly attach short motion segments into longer streams of movement. The second is motion blending, which creates motions ``in between'' a set of examples and can be used to create continuous and intuitively parameterized spaces of related actions. Automated methods are presented for building these models and for using them to generate realistic motion that satisfies high-level requirements.


My dissertation may be downloaded either in its entirety or as individual chapters.