One-Shot Accurate Cylinder Pose Estimation From Point Cloud Data With Density-Based Geometric Clustering

In recent years, the advent of machine learning technologies has significantly increased interest in factory automation (FA). Pose estimation, a crucial process in bin-picking within FA, has been explored extensively by researchers worldwide. This technology has applications in various fields including computer graphics (CG), virtual and augmented reality (VR/AR), and robotics. This paper specifically addresses the pose estimation of cylinders using a single point cloud (PC), a challenging problem due to potential ambiguities when the PC captures both the base and side of a cylinder, which can significantly impact the accuracy of pose estimation. To address this, we propose a geometric density-based clustering approach centered on the cylinder axis as the critical feature. Our method involves three steps: first, performing probability density estimation using two Gaussian spheres based on the normals and cross-products of the PC, applying directional kernel density estimation (DKDE). Second, choosing the dominant aspect either the base or the side through a point-to-point matching process to estimate the center point. Finally, conducting aspect clustering using an in-out circle created by cosine similarities to utilize the estimated cylinder axis. The center point is then determined either as the average of the PC or by the least square circle fitting, depending on the dominant aspect identified. Our approach demonstrates precise one-shot pose estimation results using a single PC.