Style-based human motion segmentation

This paper presents a method for segmenting human motion based on a notion of quality and the movement of a user such that the exact segmentation is tailored for different subjects. The problem is solved via an inverse optimal control problem where the parameter of optimization is a time along the movement trajectory that splits the longer trajectory into distinct “moves.” First, trajectories are generated using a “forward” optimal control problem; then, the match of these generated trajectories is optimized via a second, “inverse” optimization, which determines the appropriate point of segmentation. An analytical solution to this set up, its numerical implementation, and an application to real data are presented. A key novel contribution of this paper is the analytical derivation of first order necessary conditions for optimality. The segmented movements may populate a library of movement primitives in order for robots and automated systems to perform and interpret novel tasks.