Single View-Based Pose Estimation of Unknown Circular Object With RGB-D Camera

Zhaoxi Li, Cai Meng, Dingzhe Li, Limin Liu, Weimin Yuan, Zeweng Huang

Abstract: The conventional single view-based circular object pose estimation requires that the radius is known in advance, but it is not always possible for an uncooperative target. Furthermore, the pose error is highly dependent on the extracted elliptical pixels, and there is an ambiguity pose solution that is hard to distinguish without other information. To address these problems, we propose a method to estimate the unambiguous pose of an unknown circular object with a single view from red-green-blue-depth (RGB-D) camera. At first, a purification filter is introduced to select the elliptic-arc pixels of the projected circle and segment the point cloud of the spatial circle from the depth map to get its support plane. Then, these pixels are backprojected onto the support plane and a spatial circle is fitted to get an initial pose solution. Finally, the initial pose is optimized by correcting the orientation based on the Sampson distance, and by moving the support plane to an optimal place from the original circular cloud. In addition, the radius is also estimated as a byproduct. To verify the feasibility and robustness of our method, both synthetic and physical datasets are generated by Blender software and a precise optical tracking system. Experiments show that the proposed method is not only feasible but also outperforms other methods about the robustness and the accuracy of pose and radius.