Semantic Segmentation And Scene Understanding Based On 3D Point Cloud

With the development of 3D measurement technology,the acquisition of 3D data has become more and more convenient,so that the research in 3D graphics and other related fields has received extensive attention.The point cloud is the basic expression form of 3D data and is applied to autonomous driving,robotics,and geographic remote sensing.This paper completes semantic segmentation and scene description of 3D scenes for point cloud data,which combines the method of deep learning.First of all,a multi-scale and multi-level feature extraction network(Di M-PCNet)based on point cloud is proposed in this paper.The Di M-PCNet network obtains uniform multi-scale point cloud information with the direction-consistent sampling method.Then the pyramid form is used in the feature extraction part to superimpose multi-level features layer by layer,which catches features of multi-scale and multi-level features.The Di M-PCNet network ensures the diversity and completeness of point cloud features,and solves the uneven density problem of point cloud.Moreover,the fusion of the local features and global features of the scene is used in the feature extraction of the Di M-PCNet network,which constructs the semantic segmentation network of point cloud scenes.And the feature extraction can obtain more key information of a 3D object outline with a method of interpolation up-sampling.The method in this paper is verified on the public data sets Model Net40,Shape Net,and S3 DIS,and the accuracy of classification and segmentation is improved.Finally,the indoor scene ontology is constructed through Protégé in this paper.In order to construct the ontology of the 3D indoor scene,the semantic template of the spatial division is defined based on Indoor GML which is the standard for Indoor Spatial Information.Then build the ontology classes and the relationships among classes.The concept of ontology joins into the scene understanding of point cloud,which sorts out the indoor scenes and broadens the depth and breadth of semantic understanding of point cloud scenes.