Data-Driven Multiple Character Motion Synthesis

Abstract

Research on multiple characters has recently emerged. It is difficult to apply existing data-driven techniques such as editing, blending, or stitching to the motion of multiple characters because their behavior is completely different from that of single characters. The main reason for the difficulty is the interaction between the characters. In multiple character scenes, interactions such as hand shaking, pushing, and carrying a heavy object collaboratively may occur between characters. Most of the interaction necessitates that appropriate actions occur in a carefully coordinated manner in both space and time. This precise coordination of interaction is very fragile during the motion synthesis process. Another reason why it is difficult to apply existing techniques to multiple character motion is the increase in the system's complexity caused by interaction. As the number of characters grows, synthesizing coordinated action sequences simultaneously for all interacting characters causes an exponential increase in the computation time and memory storage usage.

In this thesis, we propose data-driven techniques for synthesizing multiple character motion, which can solve the difficulties caused by the interaction. Our approach is to formulate interactions between characters as mathematical constraints and effectively integrate the formulated interactions into a motion synthesis system without significantly increasing the system's complexity. A wide range of interactions is formulated as combinations of various types of manipulation handles and interpersonal constraints to specify absolute/relative spatial location, direction, temporal duration, and timing of interactions. We can create multiple character motions while maintaining the interaction if the interaction's formulation is incorporated into the motion synthesis system. The integration should be handled in different ways according to the characteristics of each motion synthesis field. We effectively integrate the interaction with two typical motion synthesis techniques for multiple characters. The first application is multiple character motion editing. The interaction formulated as a collection of linear constraints is integrated into our motion editing system as a linear system with linear equality constraints. Therefore, our motion editing technique allows the user to manipulate synchronized multiple character motions at interactive rates using various manipulation handles. The second application is highly complex multiple character motion synthesis. We use motion patches as a way to integrate formulated interactions. By tiling the patches spatially and temporally, our algorithm automatically generates highly complex animation of multiple interacting characters.