Movement Classes from Human Motion Data

Abstract

We present a new method for identifying a set of movement types from unlabelled human motion data. One typical approach first segments input motion into a series of intervals, and then clusters those into a set of groups. Unfortunately, the dependency between segmentation and clustering causes trouble in alternate tuning of parameters. Instead, we unify those two tasks in a single optimization framework that searches for the optimal segmentation maximizing the quality of clustering. The genetic algorithm is employed to address this combinatorial problem with our own genetic representation and fitness function. As the primary benefit, the user is able to obtain a repertoir of major movements just by selecting the number of classses to be identified. We demonstrate the usefulness of our approach by providing visual descriptions of motion data, and an intuitive animation authoring interface based on movement collections.