Research On Spatial Interpolation And Scene Semantic Segmentation Based On 3D Point Cloud Data

This paper aims at the non-uniformity of 3D point cloud data and the semantic segmentation of 3D scenes.Through the study of the imaging principles and related processing techniques of 3D point cloud data,and the principle of point cloud data processing by convolutional neural network and PointNet network models And the research of advantages,this paper proposes a two-dimensional image gray level interpolation algorithm and a three-dimensional point cloud semantic segmentation network model,LGF-Net(Local feature and Global feature Fusion Network).The specific work is as follows(1)Aiming at the non-uniformity of the point cloud,this paper uses a point cloud interpolation algorithm based on the uniformity of the point cloud and the gray value of the two-dimensional image.First,the point cloud is labeled with the result of the homogeneity calculation of the point cloud,the area to be interpolated is determined,and then the 3D point cloud data and the 2D image data are fused.The interpolation algorithm mainly adaptively weights the difference between the gray level of the two-dimensional image of the neighboring points and the gray level of the current interpolation point,and assigns a larger weight to the neighboring points with similar gray levels for depth prediction of the points to be interpolated.Compared with the traditional interpolation method,this algorithm improves the point cloud density while improving the uniformity of the point cloud(2)Aiming at the task of semantic segmentation of 3D point cloud scenes,this paper proposes an improved network model LGF-Net based on the PointNet network model.The convolutional layer of the network to extract deep features uses a point-by-point convolution method based on the furthest point sampling algorithm to extract the global information of the point cloud.The convolutional layer for extracting shallow features of the network applies directional coded convolution to extract local feature information of the point cloud.After merging local and global features,LGF-Net implements semantic segmentation of 3D point clouds.Experimental results show that LGF-Net significantly improves the accuracy of segmentation and recognition of small objects in point clouds.