Building Roof Modeling And Optimization From Airborne LiDAR Point Clouds Based On Regular Constraints

Three-dimensional architectural models are an important part of digital city models,and have important applications in smart city construction,urban change detection,and urban spatial information analysis.Airborne LiDAR(Light Detection And Ranging)technology can quickly obtain 3D information on the surface,and has become one of the main data sources for the reconstruction of high-precision building 3D models.However,the original point clouds directly obtained has the characteristics of large data volume,uneven density,missing occlusion,and noise,which poses a huge challenge to the subsequent processing of point clouds data;at the same time,the existing building primitive patches The extraction method has problems such as poor robustness and low time efficiency;in addition,the existing airborne LiDAR building point clouds modeling results have unclear topological relationships and lack corresponding semantic information.In response to the above problems,this paper uses the airborne LiDAR building point clouds data to complete the entire process of 3D building roof model reconstruction,from building point clouds data preprocessing,building primitive patch extraction and building roof model reconstruction.On the one hand,discuss and study the theory and technology of efficiently constructing high-precision three-dimensional building roof models.The main research contents and work results include the following points:(1)Upsampling of point clouds data based on deep learning neural network.The idea of up-sampling the building point clouds data first and then down-sampling is used to obtain results with uniform density and moderate magnitude.Firstly,an adversarial neural network generator with sampling on the point clouds is designed,the feature layer is expanded and the discriminator is trained for confrontation,and a self-attention module is introduced to enhance the integration of features.(2)Extraction of building roof primitives based on regular constraints.In order to complete the fitting of the plane from the three-dimensional unstructured point clouds and maintain the geometric relationship between the reconstructed planes,the problem of building roof primitive patch fitting is transformed into a global regular constraint problem and a subset selection problem,so as to achieve multiple high-precision and high-efficiency extraction of building roof primitives.(3)Reconstruction of 3D building roof model.Use the segmented building roof primitive face pieces to extract the simplified contour lines of each face piece,and construct the roof topological map,and use the geometric structure rules of the roof face pieces to extract the roof structure points to complete the building roof model reconstruction.Combined with the research content,this paper uses the open source DALES airborne point clouds data set to experimentally verify the reconstruction process of the proposed building roof model.The test results show that the point clouds upsampling based on the adversarial neural network proposed in this paper can complete the upsampling of the building point clouds under the premise of ensuring the accuracy of the point clouds data,thereby achieving the acquisition of point clouds data with uniform density and moderate magnitude.Goal;the fitting of building roof primitives based on regular constraints achieves more efficient and accurate primitives extraction;the method of building roof model reconstruction follows the geometric structure rules,conforms to the basic structural characteristics of the building,and can Extract the key corners of the complete roof structure and complete the reconstruction of the 3D building roof model.