Research On Semantic Segmentation Method Of SAR 3-D Point Clouds Based On Graph Signal Processing

In recent years,SAR(Synthetic Aperture Radar)3D imaging technology represented by synthetic aperture radar imaging technology has become one of the hotspots in the field of remote sensing.Tomographic SAR imaging technology can reverse the three-dimensional electromagnetic scattering structure of the target through the existing imaging data,and obtain the SAR three-dimensional point cloud of the imaging scene.It is of great significance in terrain mapping,urban target recognition and other applications.Compared with the point cloud data generated by optical detection system and RGBD(Red Green Blue Depth)camera,SAR 3D point cloud has the advantages of complete imaging scene and strong feature semantics,which makes SAR 3D point cloud have rich semantic features and great research value.However,it is difficult to segment 3D SAR data because of the number of unstructured points and unstructured points.Therefore,aiming at the segmentation of 3D point cloud of tomographic SAR,this paper studies the semantic segmentation algorithm based on graph convolution neural network,and verifies the semantic segmentation performance of the algorithm;At the same time,by studying the relevant knowledge of image signal processing,a joint target estimation method based on image signal processing is proposed,which increases the resolution of target imaging.The main research contents and innovations of this paper are as follows:1.The principle of tomographic SAR three-dimensional imaging and the principle of image signal processing are introduced.Firstly,the theoretical derivation of tomographic SAR 3D imaging is carried out,and the process of tomographic SAR 3D point cloud imaging is given;Secondly,the theory of graph signal processing method on unstructured data is described,and the result and hierarchical expression of graph convolution neural network are deduced.2.Aiming at the segmentation of 3D point cloud of tomographic SAR,a semantic segmentation algorithm of 3D point cloud of SAR Based on GCN(Graph Convolutional Network)is proposed.The spatial features and edge feature extraction ability of SAR 3D point cloud data are analyzed;Then,an improved Point Net semantic segmentation Neural Network Based on graph convolution MLP(Multi-Layer Perception)is proposed.The farthest point sampling method is used for point cloud downsampling,and the neighborhood graph of three-dimensional point cloud data is constructed.The features of these neighborhood points are constructed into graph signals,and the local features are learned by calculating the Laplace matrix of the neighborhood region;Finally,the pointnet graph convolution network model is trained and its performance in semantic segmentation on 3D SAR point cloud is verified.The performance improvement of graph convolution neural network in semantic segmentation is analyzed through comparative experiments.3.Aiming at the application of graph signal processing in radar array target estimation,a two-dimensional joint target estimation algorithm based on graph signal processing is proposed.Firstly,the principle of the traditional DoA(Direction of Arrival)target estimation algorithm is briefly described,and the disadvantage of the lack of phase information in the traditional target estimation algorithm is proposed;Secondly,an array phase data model based on graph signal algorithm is proposed,and the response graph signal structure is constructed according to the arrangement of different arrays and adjacent phases;Then,a general algorithm of the corresponding graph signal algorithm is proposed,and the target estimation results are obtained by traversing the target parameter domain through the construction of the search domain;Finally,the graph signal joint estimation algorithm is analyzed and verified by Monte Carlo experiment,which proves that the graph signal processing algorithm has better performance than the traditional algorithm in target joint estimation.