Abstract Estimating Six Degree-of-Freedom (6DoF) poses of known objects that are randomly placed in a cluttered bin is a fundamental task in computer vision and robotics, especially for mechanical parts, which are mostly metallic and texture-less. In this work, we focus on the mechanical parts 6DoF pose estimation, in which objects are always texture-less and occluded between each other. To tackle these problems, we propose a novel 3D descriptor, called BOLD3D, to detect and estimate the 6DoF pose in 3D point clouds. Our key observation is that the edge is one of the most important cues for the objects, especially for texture-less mechanical parts. Thus, we propose to utilize pairs of oriented 3D line segments, which are connected by the edge points in well organization, encoding the local geometric structure of the objects. Specifically, the edge points of the input objects are first extracted from 3D point clouds and then connected in order, after employing a discreetly downsample strategy. We then design an effective approach to normalize the 3D line segments orientation. The local geometric structure is represented by the BOLD3D features, each of which is a five-dimensional vector consisting of a pair of directed line segments. Our algorithm accelerates the poses estimation process, due to only the edges of objects are used. A variety of synthetic and real experiments show that our approach is capable of achieving satisfactory pose results with high accuracy and robustness for mechanical parts 6DoF pose estimation, even in the presence of a complex arrangement.
Paper
Full text
BOLD3D: A 3D BOLD descriptor for 6Dof pose estimation
Semantic Scholar · Engineering · 2020
Abstract
Abstract Estimating Six Degree-of-Freedom (6DoF) poses of known objects that are randomly placed in a cluttered bin is a fundamental task in computer vision and robotics, especially for mechanical parts, which are mostly metallic and texture-less. In this work, we focus on the mechanical parts 6DoF pose estimation, in which objects are always texture-less and occluded between each other. To tackle these problems, we propose a novel 3D descriptor, called BOLD3D, to detect and estimate the 6DoF pose in 3D point clouds. Our key observation is that the edge is one of the most important cues for the objects, especially for texture-less mechanical parts. Thus, we propose to utilize pairs of oriented 3D line segments, which are connected by the edge points in well organization, encoding the local geometric structure of the objects. Specifically, the edge points of the input objects are first extracted from 3D point clouds and then connected in order, after employing a discreetly downsample strategy. We then design an effective approach to normalize the 3D line segments orientation. The local geometric structure is represented by the BOLD3D features, each of which is a five-dimensional vector consisting of a pair of directed line segments. Our algorithm accelerates the poses estimation process, due to only the edges of objects are used. A variety of synthetic and real experiments show that our approach is capable of achieving satisfactory pose results with high accuracy and robustness for mechanical parts 6DoF pose estimation, even in the presence of a complex arrangement.