| Three-dimensional(3D)modeling of buildings is a basic task in remote sensing and plays a significant role in urban modernization.Airborne LiDAR can collect the 3D surface information of ground objects accurately,so aerial scanning point clouds are widely applied in3 D modeling of urban areas.However,limited by the positioning type and moving trajectory of the airborne scanner and occlusion of the ground objects,the building point clouds are sparse,incomplete and uneven in density,which makes it difficult to model watertight and complete building models from airborne LiDAR point clouds.At present,most traditional building modeling methods based on LiDAR point clouds need to set empirical parameters,and the ratio of length,width and height of obtained building models is inaccurate,which makes it worth to automatically reconstruct building models with accurate ratio from LiDAR point clouds.Meanwhile,buildings often have sharp edges and plane structures,and downstream tasks have explicit requirements on the memory usage of the model.Therefore,the application scene of the smooth dense grid model is limited.Aiming at these problems,this paper studies the 3D shape modeling and model simplification of buildings.1.Aiming at the sparsity,incompleteness and uneven density of LiDAR point clouds,this paper proposes an implicit modeling framework—GEOP-Net,which embeds with highdimensional geometric features.This method uses the geometric encoding module based on orthogonal basis to map features into high-dimensional polynomial space,so as to enhance the ability of the network to obtain detailed information.Then,the point-voxel encoding module is combined to enhance the extraction ability of local features and global features.Finally,the occupancy probability value of the query point is inferred by the occupancy probability decoder,and then complete and watertight building models are modeled by extracting the iso-surface.In this paper,the feasibility of the proposed method is verified by experiments on the building dataset in Zurich,and the effectiveness of the method is verified by comparing with seven existing methods.2.Aiming at the fact that the surface of dense meshes obtained by the implicit modeling method tends to be round and large memory consumption caused by excessive number of model faces,this paper proposes a model simplification method based on rolling guided normal filtering—RNFMS.This method introduces the rolling guided normal filter to smooth building models while retaining sharp features to ensure that accurate planes are extracted by region growing and combined with the mesh simplification method on this basis,then lightweight mesh model of the building is constructed.In this paper,experiments are carried out on dense grid models obtained by the proposed implicit modeling method,compared with three model simplification methods,the proposed method is verified to be effective. |
