Spectral domain graph convolution 3D point cloud classification method based on Fourier transform

Abstract

The invention discloses a spectral domain graph convolution 3D point cloud classification method based on Fourier transform, and the method comprises the steps: carrying out the geometric sampling ofan inputted original point cloud through a G-Point Net network model: setting an angle threshold value V, dividing points with neighborhood included angle values larger than an angle threshold V intoa geometric feature area G, and dividing the remaining points into other areas T; sampling to obtain point cloud of each area; introducing an expansion rate E based on a Dynamic KNN local graph construction method, and selectively establishing a local geometric graph every other E neighbor point clouds. Spectral domain graph convolution is carried out by using a spectral domain graph convolution method based on Fourier transform to obtain a plurality of pooled graph local features, global features are obtained through GPoint Net for classification, and a classification result is obtained. According to the method, the problem of non-uniform distribution of point cloud density is effectively solved, spatial geometric information is reserved, edge points of the point cloud can be efficientlydistinguished, meanwhile, noise points are separated, and the classification precision is improved.

Description

technical field [0001] The invention relates to a spectral domain map convolution 3D point cloud classification method based on Fourier transform, and belongs to the technical field of remote sensing image processing. Background technique [0002] With the contention of a hundred schools of thought in image processing technology, classification methods based on two-dimensional images emerge in an endless stream, and have made great achievements. However, the effect of deep learning processing methods based on 3D data is far inferior to that of 2D image classification. 3D data is typically represented as depth images, voxels, meshes, and point clouds. Compared with obtaining 3D data through RGB-D cameras or mainstream sensors, the 3D point cloud obtained by lidar can provide more reliable depth and contour information of 3D objects, and has been gradually applied to 3D object classification in recent years. [0003] In the previous work, most computer vision researchers use...

AI Technical Summary

Problems solved by technology

[0005] In order to overcome the shortcomings of the existing technology and effectively solve the problem that the traditional 3D point cloud classification method is affected by the spatial relationship and uneven distribution of the point cloud, the present invention provides a spectral domain graph convolution 3D point cloud classification method based on Fourier transform. On the premise of not changing the spatial information of the point cloud, a new form of representation-graph is proposed. The graph structure effectively solves the problem of the adjacency relationship between points and points in most point cloud deep learning models, and retains the spatial geometric information. The graph is very suitable for different arrangement. Regular non-European data; deep learning lacks a lot of research work in the spectral domain. This model combines the method of spectral domain graph convolution in the 3D point cloud framework for the first time. Spectral domain convolution has a solid mathematical theoretical foundation, and graph convolution is more focused. The adjacency relationship between key points; the G-PointNet of the present invention makes great improvements in feature point acquisition and local area division, proposes geometric sampling preprocessing and designs a dynamic K-nearest neighbor graph construction method Dynamic KNN, effectively solves the problem of The problem of uneven distribution of point cloud density

Images