Style-based abstractions for human motion classification

This paper presents an approach to motion analysis for robotics in which a quantitative definition of “style of motion” is used to classify movements. In particular, we present a method for generating a “best match” signal for empirical data via a two stage optimal control formulation. The first stage consists of the generation of trajectories that mimic empirical data. In the second stage, an inverse problem is solved in order to obtain the “stylistic parameters” that best recreate the empirical data. This method is amenable to human motion analysis in that it not only produces a matching trajectory but, in doing so, classifies its quality. This classification allows for the production of additional trajectories, between any two endpoints, in the same style as the empirical reference data. The method not only enables robotic mimicry of human style but can also provide insights into genres of stylized movement, equipping cyberphysical systems with a deeper interpretation of human movement.