3D Object Classification And 6D Pose Estimation Based On Depth Image

In recent years,the application of deep neural networks in the field of two-dimensional images has achieved great success.With the popularity of three-dimensional sensors,robots can directly obtain depth data through depth cameras.How to use three-dimensional information to improve the robot's perception ability has also become A hot topic in the field of computer vision,this article focuses on machine grabbing tasks,using threedimensional information to achieve object recognition and pose estimation.This paper studies two core perception algorithms in object machine capture,including three-dimensional recognition and pose estimation.Among them,pose estimation is a relative pose estimation based on a standard model.During the grasping process,the position and category of the object need to be known first,so as to determine the grasping target.After obtaining the target,it is necessary to obtain its posture in order to control the robot to realize grasping.The relationship between 3D object recognition and pose estimation is tight.Since the final output pose is relatively standard model,it is necessary to first obtain a matching standard model from the model library through recognition.This paper addresses the need for high accuracy and robust 3D perception algorithms in machine grabbing tasks.In order to solve the problem of rotation invariance in 3D object recognition and noise interference in pose estimation,this paper proposes the use of rotation invariant encoder and weighted Huff The voting algorithm realizes the recognition of object rotation invariance and robust attitude estimation to noise.This solution first filters the original input depth data,and proposes different segmentation algorithms to segment the point cloud for two common scenes in the capture application.After segmentation,the point cloud on the surface of the target object is obtained.In the recognition phase,in view of the defect that the point cloud coordinates change with the change of the camera coordinate system,this paper researches and improves a rotationinvariant 3D object recognition method for 3D point cloud data,which includes a point-cloud rotation-invariant feature encoder,Point cloud downsampling method and classification network based on hierarchical graph convolution network.At the stage of three-dimensional object pose estimation,this paper extracts the point pair feature(PPF)feature to match the scene point cloud with the model point cloud,and calculates the pose through a weighted Huff voting method considering the noise.In addition,in order to improve the calculation efficiency,this paper improves a point cloud downsampling method based on meshing and local geometric change rate.