Visual Servoing of Textured Free-Form Objects in 6 Degrees of Freedom

Abstract

This thesis presents and examines methods for visual tracking of textured free-form surfaces as well as appropriate setups for catching these objects with a robot manipulator. Faced with limited computational resources, the object is firstly localised in 2 degrees of freedom (DoF). Successively, this hypothesis is refined to a pose estimate in 6 DoF using an universal object description by means of three-dimensional textured point clouds. In particular, an analytic formula is devised that allows for the efficient prediction of intensity changes respective to object motion. Known approaches of illumination compensation are considered and adapted to the particular models of surface shape and appearance matching. The presented methods are evaluated systematically with respect to convergence. Finally, the system is experimentally validated for the desired human-robot interaction by grasping moving objects from the users hand.