Fast template matching and pose estimation in 3D point clouds

In: Computers & Graphics (Apr. 2019), 79(36-45)
 

Abstract

Template matching for 3D shapes in point cloud data is an essential prerequisite for a multitude of applications such as bin picking tasks for known objects, detection and completion of redundant object instances during scanning endeavors, and verification of industrial assemblies. Building on existing approaches for template matching, especially on methods utilizing point tuple features for the quick generation of transformation guesses in a RANdom SAmple Consensus (RANSAC) setting, we introduce an improved, targeted sampling strategy as well as an efficient hypothesis validation approach to drastically improve the overall runtime. In our experiments the proposed optimizations lead to a performance increase by two orders of magnitude in comparison to an unoptimized implementation. Several experiments on diverse real-world and simulated datasets demonstrate the robustness of our proposed approach.

Bilder

Bibtex

@ARTICLE{vock2019,
    author = {Vock, Richard and Dieckmann, Alexander and Ochmann, Sebastian and Klein, Reinhard},
     pages = {36--45},
     title = {Fast template matching and pose estimation in 3D point clouds},
   journal = {Computers {\&} Graphics},
    volume = {79},
      year = {2019},
     month = apr,
  abstract = {Template matching for 3D shapes in point cloud data is an essential prerequisite for a multitude of
              applications such as bin picking tasks for known objects, detection and completion of redundant
              object instances during scanning endeavors, and verification of industrial assemblies. Building on
              existing approaches for template matching, especially on methods utilizing point tuple features for
              the quick generation of transformation guesses in a RANdom SAmple Consensus (RANSAC) setting, we
              introduce an improved, targeted sampling strategy as well as an efficient hypothesis validation
              approach to drastically improve the overall runtime. In our experiments the proposed optimizations
              lead to a performance increase by two orders of magnitude in comparison to an unoptimized
              implementation. Several experiments on diverse real-world and simulated datasets demonstrate the
              robustness of our proposed approach.},
       doi = {10.1016/j.cag.2018.12.007}
}