A 3D Target Point Cloud Classification Method With Graph Convolutional Network

The research object of this topic is the 3D point cloud classification task.Our purpose is to verify the feasibility and effectiveness of the point cloud feature extraction algorithm based on graph convolution by implementing point cloud classification.As a basic task in the field of 3D perception,3D object classification can be implemented to verify the effectiveness of the 3D object feature extraction algorithm,and it can be applied to other tasks in the field of 3D perception,such as segmentation and reconstruction.Research significance and value.Considering that most current spatial information acquisition equipments provide spatial information of 3D objects in point cloud,this topic selects the point cloud data of 3D objects as the processing object of this task classification task.For 3D point cloud classification tasks,the traditional algorithm has the disadvantage that the point cloud category can only be determined by searching the feature library.The fitness of the selected features and the completeness of the feature library will directly affect the classification performance;the disadvantage of machine learning algorithms is that they still use The artificially designed point cloud features and fixed feature calculation methods limit the generalization of the algorithm;the current deep learning algorithms all extract the features of irregular point cloud data by modifying the convolution operation used to process regular data,and the algorithm can't fully fits the characteristics of point cloud data,and its generalization ability is not strong.We proposes an end-to-end point cloud classification network based on graph convolution by analyzing the point cloud characteristics to realize the classification task of a single model point cloud that only carries spatial location information.This topic proposes the following solutions for the three characteristics of point cloud : disorder,rotation invariance,and sparseness:(1)the overall network architecture draws on 2D image convolutional neural networks;(2)In order to solve the problem of disordered point cloud input,we replace regular convolution operations in convolutional neural networks by designing graph convolution operations based on graph theory;(3)Design a posture correction auxiliary network based on rigid body transformation and coordinate system transformation to solve point cloud data(4)In order to achieve the point cloud Multi-dimensional feature extraction and solve the sparseness problem of point clouds.we replace the pooling operation in the convolutional neural network by designing a pooling operation based on the sparse point cloud sampling method,spatial pooling method,and the physical meaning of the point cloud node attributes,,By testing on a standard point cloud dataset and comparing it with other point cloud classification algorithms,the experimental results show that the classification accuracy of our point cloud classification algorithm on a fixed pose point cloud dataset is not worse than that of other algorithms,and the classification accuracy on the random pose point cloud dataset is the highest.It is proved that the three-dimensional point cloud classification algorithm based on graph convolution proposed in this paper can achieve high accuracy point cloud classification,it is robust to the input point cloud attitude at the same time.And the feasibility and effectiveness of the point cloud feature extraction algorithm based on graph convolution are verified.