Research On Deep Learning Network For Semantic Segmentation Of Indoor Scene Point Cloud Based On Directional Plane Projection Convolution

In the task of semantic understanding of indoor scenes,semantic segmentation of 3D point clouds is an important basis for realizing indoor scene understanding,which is widely used in the field of indoor intelligent 3D environment perception.With the deepening of deep learning technology in this field,the use of deep learning to solve point cloud scene understanding tasks represented by point cloud semantic segmentation has become a hot topic.However,because the original point cloud belongs to unstructured data,and the inherent sparseness and disorder of unstructured data,traditional convolutional neural networks cannot achieve effective extraction of point cloud features through conventional convolution operations.Initially,researchers tried to convert the sparse discrete point cloud into regular spatially arranged voxels,and then used 3D convolution to extract the voxelized point cloud features.However,the segmentation method based on the voxelization of point cloud data will bring huge memory occupation and computational overhead,while the lowresolution voxelized point cloud will lose local detail information.Therefore,in view of the problem that the irregular point cloud spatial distribution makes it difficult to effectively extract feature information with conventional convolution methods,this paper conducts research on the semantic segmentation of indoor scene point clouds based on the idea of partial plane regularization of point clouds and the idea of attention mechanism.The specific research content as follows:First,in view of the insufficient extraction of local feature information in the existing point cloud parameterized convolution operation when facing some flat objects in the indoor scene,based on the idea of regularization of the point cloud local plane,a local projection weight matrix is constructed to scatter the local space.The 3D point cloud is mapped into a regular grid plane,and then the local feature information of the point cloud is encoded by2 D convolution;in order to enhance the direction perception and spatial information expression ability of the local weight projection matrix,increase the orientation encoding neighborhood search,a orientation-encoding flattening projection convolution with directional perception capability is proposed;finally,according to the residual network idea,the point cloud semantic segmentation framework Point Resnet based on the orientationencoding flattening projection convolution is constructed.Secondly,in order to further improve the feature extraction ability of the local direction plane projection convolution in this paper,a three-dimensional point cloud interactive attention unit suitable for the point cloud direction plane projection convolution is proposed.The channel of the regular feature map obtained by the projection weight matrix Interactive calculation with spatial attention improves the abstract extraction ability of local spatial features of the convolution method in this paper,and further compensates for the loss of spatial information caused by local planarization processing;combined with the idea of residual network to construct interactive attention direction plane projection convolution Residual unit: Finally,combining the structural characteristics of Unet network with the residual unit in this chapter,we construct a three-dimensional point cloud semantic segmentation framework Point Res Unet network.Finally,in order to verify that the point cloud semantic segmentation framework of this paper is applicable to indoor scenes,the performance of the Point Resnet and Point Res Unet indoor scene point cloud semantic segmentation models proposed in this paper are tested and analyzed on the Scan Net dataset and the S3 DIS dataset.The three basic evaluation indicators in the field of point cloud semantic segmentation are used as the basis for network performance evaluation,and the superiority of the proposed method in indoor scene 3D point cloud semantic segmentation is verified through a fair comparison with other mainstream network models under the same experimental environment.