Estimating Pose of Object and Manipulator Grasping Control

Abstract

In this paper, we mainly study the pose estimation based on feature matching and manipulator grasping control. Obtaining the pose information of the object is an important part of autonomous grasp of the manipulator. In order to obtain the precise pose information of the object, an improved algorithm is proposed based on the GMS (Grid-Based Motion Statistics) algorithm. Firstly, we use the RANSAC algorithm to remove the point pairs whose the distance error is more than 1.5 pixels after transformation. Secondly, the Euclidean distance of the coefficients is calculated between original image and object image. Some point pairs with a larger distance are removed because of the affine invariant principle. The correct correspondences are transformed from 2D pixel coordinate frame to 3D camera coordinate frame with depth image. The least square method combined with SVD algorithm is used to solve the rotation and translation matrices of the object relative to the camera coordinate frame. These matrices are used to estimate the pose of the object. The high accuracy of feature matching in the improved GMS algorithm is verified. The estimated error of the position $(x, y, z)^{T}$ is within ±2.4mm, and the orientation $(\text{Roll}, \text{Pitch}, \text{Yaw})^{T}$ is within ±1°. Finally, the performance of the algorithm is verified through the grasping experiments with the manipulator.