Research On 3D Point Cloud Classification And Segmentation Algorithm Based On Graph Neural Network

With the rapid development of science and technology,3D point cloud data has been widely used in fields such as autonomous driving,robot perception,and virtual reality.However,due to its high-dimensional,disordered and sparse characteristics,it is difficult for traditional methods to effectively classify and segment it,which brings great challenges to the understanding and analysis of 3D scenes and objects.The method based on graph neural network is introduced into 3D point cloud data processing,which can directly perform endto-end training on point cloud data,and can effectively use local geometric information for classification and segmentation by constructing a graph structure.Therefore,the 3D point cloud classification and segmentation method based on graph neural network can be well adapted to discrete 3D point clouds.In order to enrich the local detail features and semantic features of 3D point cloud,this thesis proposes two 3D point cloud classification and segmentation networks based on graph neural network.The main research contents include:(1)Aiming at the problem that the difference in the density of point cloud data in different regions affects the learning of local features,a 3D point cloud classification and segmentation algorithm AN-GCNet based on an adaptive neighborhood graph convolutional network is proposed.First,the farthest point sampling algorithm is used to down-sample the original point cloud data,and then the optimal neighborhood search of each point is obtained using the neighborhood entropy function to increase their uniqueness,and then the dynamic graph convolution operation is used to learn each The neighborhood information features of each node are aggregated to dynamically obtain the local structure of the point cloud data,and finally the local structure features are extracted in an end-to-end manner.Experimental results on publicly available datasets demonstrate the effectiveness of the proposed algorithm in typical 3D point cloud classification and segmentation tasks.(2)In order to further enrich semantic features and improve computational efficiency,a classification and segmentation algorithm PN-GANet based on graph attention and pyramid network is proposed.The algorithm proposes two different operators,a graph attention module and a pyramid network.The graph attention module constructs the point cloud on the graph structure,and calculates the similarity of adjacent points through the self-attention mechanism,and obtains a new feature vector by weighted summation,thereby improving the local feature expression ability of the model.The pyramid network uses four different sizes of convolution kernels to downsample the feature map,and then uses bilinear interpolation for upsampling to restore the original feature map,expand the receptive field,and retain as many geometric features as possible.Experimental results show that compared with some existing models,the proposed network model has achieved good results.(3)Conduct comparative experiments on the two graph neural network-based algorithms proposed above,and analyze their respective advantages and complementarities.The ANGCNet algorithm has higher classification accuracy due to its flexible local structure,and is suitable for 3D point cloud data requiring high-precision classification and tasks requiring global information;while the PN-GANet algorithm directly uses graph attention to weight different points,In addition,the pyramid network is used to fuse multiple scale information,which performs better in segmentation performance and has certain advantages in efficiency.It is suitable for larger 3D point cloud data and segmentation tasks that require faster convergence speed.