Directional Fragmentation of Point Cloud for Model-Based Object Pose Estimation

Abstract

In the field of computer vision, six-degree-of-freedom (6DOF) object pose estimation has been researched continuously for a long time as one of the difficult problems to deal with. Recently, as learning-based problemsolving methods are in the spotlight, the paradigm of 6DOF object pose estimation is shifting from a traditional vision-based approach to a learningbased approach. The learning-based approach has powerful advantages in solving specific categories of problems. However, it is difficult to cope with a category of problems that deviate from predefined scenarios. This is significant when the model changes unexpectedly and frequently when performing model-based 6DOF object pose estimation. In this thesis, we propose a novel point cloud registration method for model-based object pose estimation that enhances the pose estimation method of the traditional vision techniques. We make several improvements to the pose estimation method using hand-crafted three-dimensional (3D) local feature descriptors, a traditional point cloud registration tool: (ⅰ) in order to efficiently register point clouds, only the essential elements required for registration are extracted from the model point cloud through resampling considering the geometric complexity of the model; (ⅱ) in order to estimate the pose robustly, the model is split into 6 pieces called fragments through a process of directional fragmentation; (ⅲ) in order to estimate the pose reliably, the fragments are properly switched according to the camera pose that changes with the camera movement. With these improvements, our method significantly improves the pose estimation performances. We measure the pose error, runtime, and recall for the RGBD image sequence to evaluate our method in experiments, and provide the results compared to the method without ours.