Perception subsystem for object recognition and pose estimation in RGB-D images

Abstract

RGB-D sensors have become key components of all kind of robotic systems. In this paper we present a perception subsystem for object recognition and pose estimation in RGB-D images. The system is able to recognize many objects at once, disregarding whether they belong to one or many classes. Next to the detailed description of the principle of the system operation we present several off-line and on-line experiments validating the system, including verification in the task of picking up recognized objects with IRp-6 manipulator.