An analysis of factors affecting point cloud registration for bin picking

Abstract

The robotic bin picking system is commonly used to automate processes in the manufacturing industry, by estimating the six degree-of-freedom (6-DoF) pose of an object. In particular, in vision-based systems, the pose of an object is estimated by registering a 3D point cloud acquired from a computer-aided design (CAD) model with a 2.5D point cloud acquired from a depth map. The registration process requires the correspondence points between 3D point cloud and 2.5D point cloud. Unfortunately, since the 3D point cloud and the 2.5D point cloud have different dimensions, performing registration is more challenging than with equivalent dimensions. In this paper, therefore, we analyze the process of 3D point cloud to 2.5D point cloud registration through the experiments to perform stable bin picking task. For the experiments, 2.5D point cloud is synthesized from 3D CAD model and uniformly adjusted for density and depth noise. By registering 3D point cloud to adjusted 2.5D point cloud, we quantitatively analyze how the adjusted density and depth noise affect the registration process.