Research On Deep Network Semantic Segmentation Of LiDAR 3D Point Clouds

Semantic segmentation of 3D point clouds in remote sensing scenes has attracted more and more attention.Due to the complex background of remote sensing scenes,uncertain projection directions,and large target scales,the existing 3D point cloud semantic segmentation algorithms are applied to remote sensing scenes.1.It is difficult to identify the similarity of the target contour structure.In response to the above problems,the specific work of this paper includes the following two parts:(1)Aiming at the problems of complex target background and different target scales in remote sensing scenes,this paper proposes a multi-scale feature learning semantic segmentation algorithm based on multi-scale feature fusion network with serial structure.The algorithm fully combines the spatial information of objects and the contextual information of points to extract more representative features of 3D point clouds and solve the problems of complex target backgrounds and different target scales in remote sensing scenes.The method uses a hybrid feature extraction module that combines orientation encoding and density convolution,utilizes spatial information and density information in feature learning,extracts geometric attributes of objects,extracts features of salient regions in complex backgrounds,and improves the model in remote sensing scene backgrounds Robustness in complex situations.At the same time,the direction encoding and density convolution modules are combined for multi-scale feature learning,and the point self-attention module is used for multi-scale feature fusion and extraction of point cloud feature context information to improve local feature representation.In the case of different scale targets,improve the semantic segmentation accuracy of remote sensing scene point cloud.(2)Aiming at the problem that the target structure outline is similar and difficult to identify in the remote sensing scene,this paper proposes a multi-scale feature learning semantic segmentation algorithm guided by key point category based on the multi-scale feature fusion network of parallel structure.The algorithm first uses a multi-scale adaptive feature selection module to adapt to different degrees of sparsity in Li DAR 3D point cloud data,and re-weights its features by learning global context information to solve the problem of different target scales in remote sensing scenes.At the same time,the key point extraction module is introduced to infer the spatial key points through the learned features,and then generate the local detailed features of the key points and their spatial patterns; finally,by establishing the insensitivity to the point position consistency,but very sensitive to shape,to solve the problem of complex target background in remote sensing scenes; finally,by using the category-guided fusion module,different fusion strategies are used for the same category and between different categories to identify multiple adjacent structural shapes.It can solve the classification problem of objects with similar appearance structures in remote sensing scenes,and improve the accuracy of point cloud semantic segmentation.We evaluate our method on public datasets(Urban Semantic 3D(US3D)dataset,ISPRS Vaihingen 3D semantic labelling benchmark(ISPRS))and newly collected UAV Li DAR point cloud data on the campus of Guangdong University of Technology(GDUT Semantic3D)..Experiments show that the method proposed in this paper can well solve the problems existing in the application of 3D point cloud semantic segmentation algorithm in remote sensing scenes,improve the generalization ability of the network model,improve the performance of 3D point cloud semantic segmentation,and meet the requirements of practical applications.