A RGB-D based 6D Object Pose Estimation and Its Application in Robotic Grasping

Pose estimation of objects is critical to robotic grasping. Local optimization approach has been widely used to minimize the distance of the point pairs to estimate the 6D pose, which, however, is time-consuming and low-accuracy. To conquer this problem, a novel and time-efficient 6D object pose estimation neural network, PoseNet, is proposed in this paper. The input of PoseNet is the RGB-D image and a novel fusion network with channel attention mechanism is used to extract data. The random-sample-consensus-based voting method and rotation anchors are developed to predict, respectively, the translation of object and the rotation of object. The performance evaluation on the YCB-Video dataset show that the real-time inference and high accuracy are guaranteed. The proposed method is also demonstrated by a practical robotic grasping system, where the experiment video is avaliable at https://www.bilibili.com/video/BV1qf4y1s7in.