6D Pose Estimation Using Detection Networks and RANSAC-Based Global Registration

Abstract

6D pose estimation is one of the key techniques for object grasping. Recently, various deep learning-based techniques have been proposed and shown good estimation results. However, most of them are not accurate and robust enough for robot manipulation. In this paper, we propose 6D pose estimation method with deep learning network such as yolo-v3 and point cloud registration. First, it searches the region of interest (ROI) of the object using yolo-v3. Next, the pixels of ROI are replaced to point clouds corresponding to the ROI. Lastly, we register the point clouds with 3D model using the RANSAC based global registration. This RANSAC based global registration showed best estimation success rate compared to other registration techniques in the real experimental scenario where total 73 objects were used for test.