Optimization Estimating 3D Object Pose Using Levenberg-Marquardt Method

Abstract

The accuracy of estimating object pose is a crucial task in the field of industrial robotic vision. As the further work will be related to the robot control movement of grabbing the object. Another consideration is regarding the cost. To keep the low cost camera, that is using mono vision camera, this paper presents the methods to optimize the object pose accuracy using Levenberg-Marquardt Method. The input is the coarse pose result which applies Directional Chamfer Matching Method and Line Segment Detector Method. The optimization method resulting that the translational (or Cartesian position) average error of synthetic camera image is 0.99 mm and the rotational average error is 0.32°. Meanwhile, for the real camera image case, the translational (or Cartesian position) average error is 0.37 mm and the rotational average error is 1.65°.