A framework for object training and 6 DoF pose estimation

Abstract

A high potential future robot vision addresses personal robots able to fulfill useful tasks in everyday environments like offices, hospitals, care centers or at home. When operating in such highly unstructured and dynamic environments, perceptual capabilities play a central role towards the final objective of a fully autonomously acting robot. This paper presents a framework that targets the complete perceptual process including data acquisition, object training and object localization to create a system that is capable of easily teaching a robot new objects to enable its recognition within unknown and unstructured environments. The system operates on single images and is designed to handle occlusions and multi-occurrences of objects.