Concatenate feature extraction for robust 3D elliptic object localization

Abstract

Developing an efficient object localization system for complicated industrial objects is an important, yet difficult robotic task. To tackle this problem, we have developed a system consisting first of a vision model acquisition editor, where the object salient features are acquired through a human-in-the-loop approach. Subsequently, two feature extraction algorithms, region-growing and edge-grouping, are applied to the object scene. Finally, by Kalman filter estimation of a proper ellipse representation, our object localization system successfully generates ellipse hypotheses by grouping edge fragments in the scene. The proposed system is validated by experiments using actual industrial objects.