Fast and Accurate Pose Estimation for Industrial Workpieces Robotic Picking

Abstract

6DoF Pose estimation plays an important role in industrial robotic picking applications, and it is particularly challenging when dealing with complex-shaped workpieces, often with little texture. This paper proposes a complete approach to customize a fast and accurate workpiece picking system, based on dense reconstruction, object detection and point cloud registration schemes. For any target object, the required input is its CAD model. First, we use a depth camera and an eye-in-hand robot to capture the scene in RGB-D image form. Then, we align the CAD models to some reconstructed point clouds, and automatically generate datasets of annotated images with the help of projective rendering. The data is used to train a neural network object detector, in order to detect a region of interest in color images. Next, as the detected 2D region is projected into a 3D space, the depth information inside this space is extracted to conduct point cloud registration with the object's model for pose estimation, and its result guides the system to carry out an optical picking action. Moreover, our method is accelerated with the parallelized computation in GPU to raise the efficiency of the system.