Automatic Construction of Compact Motion Graphs

Motion capture data often requires substantial processing before it becomes useful. We propose a technique that automatically distills a compact motion graph from an arbitrary collection of motion capture data. At its heart, the process identifies clusters of similar motions which we call “motion bundles”. Motion bundles and their encompassing motion graph provide a readily understandable structuring of the motion data. They can serve as a shared tool in support of common types of motion processing, including motion segmentation, motion compression, the creation of blend spaces, and the identification of connectivity to support motion resequencing. We use a novel string-based representation of motions to help find motion bundles. Users can specify a preference for longduration bundles or bundles containing many motions. We demonstrate results using data for boxing, walking and various exercise motions, and we show that meaningful partitions are retained in the face of noise.