Research On 3D Point Cloud Recognition Based On Graph Convolutional Network

With the emergence of related algorithms and the continuous upgrading of hardware equipment,deep learning methods occupy an increasingly important part of computer vision.Convolutional neural networks can simulate the hierarchical processing of visual signals by the brain’s visual system,so it has achieved extremely high recognition accuracy in the field of 2D images.With the development of 3D sensors and related software,more and more image data is being presented in the form of 3D data.3D data has no fixed data organization,there are different data forms,there are different research direction,common 3D data forms are voxels,multi-views,and point clouds,etc.Among them,the 3D point cloud can better reflect the surface information and structure information of 3D objects,and is one of the important data for the research of 3D object recognition.but the representation of 3D point clouds is irregular and asymmetric,and its arrangement order is diverse,so deep neural networks that can be applied to 2D image data are not suitable for 3D point cloud data.Based on the data form of 3D point cloud data,this thesis proposes Graph PNet,a 3D point cloud recognition network based on graph convolution network for uniform point cloud density,and Graph PNet2,a 3D point cloud recognition network based on simplified graph convolution network and multi-scale feature fusion for non-uniform point cloud density.And then this thesis compares them with existing methods,and the effectiveness of both networks is verified.The specific work of this thesis are as follows:(1)This thesis introduces deep learning methods based on voxel and multi view,and then summarizes related research based on 3D point clouds in detail,analyzes the characteristics and deficiencies of existing methods(2)PointNet only extracts feature for isolated point,and ignores the neighborhood structure information among points,so this thesis proposes Graph PNet which is a 3D point cloud recognition network model based on graph convolutional network.The model assumes that the 3D point cloud is evenly distributed in space,combines the characteristics of disorder and invariance of the point cloud,the 3D point cloud is transformed into an undirected graph structure,and then the local structure information of the 3D point cloud is obtained by the structure,finally new point cloud features by fusing neighborhood information with single point information is obtained.At the same time,Graph PNet uses spatial transformation module and global pooling to solve the problem of disorder and invariance of point clouds.Compared with the existing deep learning methods based on 3D point clouds,Graph PNet improves the accuracy of recognition accuracy and achieves a more stable recognition effect.(3)Graph PNet2 model is proposed for the characteristics of uneven distribution of 3D point cloud data density,the model uses new method to divide neighborhoods,uses multi-scale feature fusion to adapt to 3D point cloud data of different densities.At the same time,Graph PNet2 model uses simplified graph convolution to reduce the calculation amount of multi-scale feature calculations.This thesis compares the difference between the recognition accuracy and calculation speed of the simplified graph convolution model and Graph PNet graph convolution model under the same architecture are compared through experimental.When the point cloud is unevenly distributed,the recognition accuracy of Graph PNet2 has been further improved compared to the existing methods,the recognition effect is relatively good...