A Method for Robust Object Recognition and Pose Estimation of Rigid Body Based on Point Cloud

Guiyu Zhao, Hongbin Ma, Ying Jin

Abstract: Object recognition and pose estimation of rigid body are important research directions in the field of both computer vision and machine vision, which has been widely used in robotic arm disorderly grasping, obstacle detection, augmented reality and so on. This paper introduces a method for object recognition and pose estimation of rigid body based on local features of 3D point cloud. A new 3D descriptor (MSG-SHOT) is proposed in the disordered grasping of robot, and only the depth information is used to complete the recognition and pose estimation of the object, which greatly improve the accuracy in the scenes full of clutters and occlusions. Firstly, the adaptive voxel filter based on local resolution is used to realize data reduction and keypoint extraction. Secondly, the MSG-SHOT descriptor is used to complete feature calculating and matching, and the preliminary object recognition and pose estimation of rigid body are completed. Finally, the fast non-maximum suppression algorithm based on point cloud is used to complete the screening of candidate objects. The experimental results show that our method has stability and accuracy, and has robustness to the scenes full of clutters and occlusions, which meets the standard of high-precision grasping of manipulator.