6DoF Pose Estimation for Intricately-Shaped Object with Prior Knowledge for Robotic Picking

Abstract

Quick and accurate estimation for the 6-DoF pose of a randomly arranged object in intricate shape plays an important role in robotic picking applications. In this paper we propose an approach based on template matching by using the aligned RGB-D image with prior knowledge to recover the 6-DoF pose of a randomly arranged object. First, the object’s template database is generated with the help of a defined virtual imaging model and its CAD model. Then in the practical phase, we segment RGB-D image to get the mask representing the location of the object and then these data are modified into a comparable format with the characteristics of scale invariance. At last, a similar function with adjustable attention weight to color and depth data is defined to find Top-K matched templates. The selected matched templates are refined by ICP to generate the final answer. Experiments are conducted using an RGB-D camera and a robot arm to pick up given objects in intricate shape. The average recognition rate of the object in different poses is 97.826%. It also can work well with multiple objects randomly arranged with good masks representing the locations.