Scene Regions Guided Pose Estimation Using an Improved Voting Method in Cluttered Scenes

In the process of service robot performing home services, one of the key parts is robotic grasping. Meanwhile, accurate object pose estimation is essential for grasping. In home environments, estimating the poses of household textureless objects simply and effectively in cluttered and occluded scenes is challenging. This paper proposes a method by using the color information of object to extract the 3D scene regions where the object may exist. Point Pair Features voting approach is applied to obtained voting array in extracted 3D scenes. Then a novel votes adjustment method is proposed to recalculate the voting number to reduce the effects of occlusion. Our algorithm is evaluated on Linemod Occluded dataset and the experimental results show that the proposed algorithm can effectively improve the accuracy of pose estimation when there is object occlusion in a cluttered scene and improve the ranking of correct pose in candidate poses. Meanwhile, the average calculation time is shortened.