A 3D motion tracking method based on Nonparametric Belief Propagation

Most existing motion tracking methods works in specific predefined situations and requires large amount of a priori information about the target objects, such as, their shapes, appearances, kinematic structures, possible moves and physically valid poses. This work aims to investigate a generic motion tracking method that allows to reduce the amount of a priori knowledge employed. The proposed 3D tracking method mainly intends to allow the tracking of objects with distinct shapes, including cyclic dependencies between their different parts, and learning their representation models during the motion tracking process. To do so, the Nonparametric Belief Propagation (NBP) technique, the PArticle Message PASsing (PAMPAS) algorithm and the Loose-Limbed probabilistic graphical model are used into this novel approach. The proposed method is applied to distinct and previously unknown objects. The obtained results shown that the method is capable of deal adequately with such situations.