Generation and implementation of grasping rectangle based on hierarchical shape context and kernel density estimation towards robotic grasping unknown objects

This paper focuses on the perception step of robotic grasping unknown objects in order to get a stable grasping hypothesis. At first, hierarchical shape context feature is proposed to depict the local and global shape character of a sample point along the edges of the object. Moreover a kind of random forests classifier is adopted to recognize the grasping candidates in the image from vision system so that a 2D grasping rectangle can be generated through kernel density estimation. Finally, by means of stereo matching, the grasping rectangle can be mapped into the 3D space. Thus, the center of the grasping rectangle can be applied as the center of the gripper. The approaching vector and the grasping rectangle direction can be employed to determine the pose of the gripper. Simulated experiments showed that a reasonable and stable grasping rectangle can be generated for various unknown objects.