Synthesizing and Editing Human Motion from Sparse User Inputs

Abstract

An ideal 3D character animation system can easily synthesize and edit human motion and also will provide an efficient user interface for an animator. However, despite advancements of animation systems, building effective systems for synthesizing and editing realistic human motion still remains a difficult problem. In the case of a single character, the human body is a significantly complex structure because it consists of as many as hundreds of degrees of freedom. An animator should manually adjust many joints of the human body from user inputs. In a crowd scene, many individuals in a human crowd have to respond to user inputs when an animator wants a given crowd to fit a new environment. The main goal of this thesis is to improve interactions between a user and an animation system.

As 3D character animation systems are usually driven by low-dimensional inputs, there is no method for a user to directly generate a high-dimensional character animation. To address this problem, we propose a data-driven mapping model that is built by motion data obtained from a full-body motion capture system, crowd simulation, and data-driven motion synthesis algorithm. With the data-driven mapping model in hand, we can transform low-dimensional user inputs into character animation because motion data help to infer missing parts of system inputs. As motion capture data have many details and provide realism of the movement of a human, it is easier to generate a realistic character animation than without motion capture data.

To demonstrate the generality and strengths of our approach, we developed two animation systems that allow the user to synthesize a single character animation in realtime and edit crowd animation via low-dimensional user inputs interactively. The first system entails controlling a virtual avatar using a small set of three-dimensional (3D) motion sensors. The second system manipulates large-scale crowd animation that consists of hundreds of characters with a small number of user constraints. Examples show that our system is much less laborious and time-consuming than previous animation systems, and thus is much more suitable for a user to generate desired character animation.