Method for unsupervised non-rigid registration of motion capture markers on the surface of 3D mesh

Motion capture (MC) data are the most reliable source of information about human body motion. Such data are widely used in different kinds of applications, for example, movies creation, game development, physical simulation of human body etc. The task of markers registration on the surface of 3D mesh model, describing human body, is a necessary step of all mentioned applications. In publicly available MC databases, like Carnegie Mellon University (CMU) Graphics Lab Motion Capture Database, each marker has unique name and some semantic information about marker position on actor body; MC data contain time series of 3D marker positions and skeleton bones transformations. We present method for unsupervised search of correspondence between motion capture markers and 3D mesh vertices. Template mesh is a model of human in initial (T-shirt) pose, for registration any frame with arbitrary pose may be chosen. Correlated correspondence of markers to mesh vertices is reached by optimizing of joint probabilistic distribution over correspondence variables, represented as Markov Random Field (MRF). We define potentials, preserving geodesic distances between pairs of markers and correspondent vertices. Markers are added into consideration in portions by several steps. For each marker subset we create a Graphic Model as a minimum spanning tree, and execute Pearl inference procedure from Intel OpenSource Probabilistic Network Library (PNL). The obtaining results may be employed for pose deformation model learning in human animation and also in all mentioned applications.