Research On CUDA Based 3D Model Reconstruction Method For Oblique Images

Realistic 3D model data,as the basis and carrier of many applications,has been widely used in urban construction and management,public life and travel and other fields.Among the many model reconstruction methods,3D model reconstruction based on UAV oblique images has the advantages of large range,high accuracy,high clarity,low cost,high efficiency,etc.,and has now developed into the mainstream 3D reconstruction method.However,the increasing complexity of model structure and the gradual increase of reconstruction scale pose new challenges for realistic 3D model reconstruction based on oblique images.There are still many problems to be solved,such as:(1)The current 3D model reconstruction method based on oblique images has the problem of missing reconstruction details,which cannot guarantee the accuracy and integrity of the model.(2)The 3D model reconstruction speed based on oblique image dense point cloud is slow,which affects the production efficiency of oblique image 3D model.(3)The efficiency of levels of detail(LOD)model construction method is not high,which restricts the promotion and application of oblique image 3D model.In the context of rapid development of general computing of graphic processing unit(GPU),the introduction of compute unified device architecture(CUDA)provides a powerful tool for researchers and improves the parallel computing efficiency of GPU.Therefore,based on the CUDA parallel computing platform,this thesis focuses on the optimization of 3D model reconstruction effect based on oblique images,the optimization of 3D model reconstruction efficiency,and the optimization of LOD model construction efficiency,aiming at optimizing the 3D reconstruction effect of oblique images and improving the efficiency of 3D reconstruction.The main work and innovation points of this thesis are as follows:(1)In terms of effect optimization of 3D model reconstruction based on oblique images,aiming at the problem of missing reconstruction details in existing 3D reconstruction methods,an adaptive 3D reconstruction method based on local features of the model is proposed.First,an adaptive energy calculation strategy based on local features of the model is designed to enhance the noise filtering ability at the detail parts of the model.Second,an adaptive end tetrahedron finding method is proposed to solve the problem of wrong label assignment for end tetrahedron.Finally,experimental verification is carried out on several groups of oblique image dense point cloud data with different features.The experimental results show that the adaptive local feature-based oblique image 3D model reconstruction method proposed in this thesis can improve the accuracy and integrity of the reconstructed 3D model without reducing the noise filtering ability,and better preserve the scene details.(2)In the aspect of 3D model reconstruction efficiency optimization of oblique images,aiming at the problem of low efficiency of 3D reconstruction method based on Delaunay triangulation,a Delaunay triangulation construction method based on CUDA is proposed.First,Delaunay triangulation of oblique image dense point cloud is calculated by point-by-point insertion method.Secondly,a CUDA-based Delaunay triangulation energy calculation parallel strategy is designed.Based on block data,a data and index structure suitable for parallel processing is proposed.The parallel computing ability of GPU is used to improve the efficiency of 3D reconstruction.Finally,several optimization strategies are used to further improve the performance of parallel method.The experimental results show that compared with traditional methods,the CUDA-based parallel 3D reconstruction method based on Delaunay triangulation proposed in this thesis achieves about 4 to 7 times efficiency improvement,significantly improving the efficiency of 3D reconstruction.(3)In terms of efficiency optimization of three-dimensional LOD model construction,aiming at the problem of long computation time of traditional LOD model construction method,a CUDA-based LOD model construction method is proposed.First,based on the principle of edge folding simplification method,the three-dimensional model is simplified.Second,a CUDA-based LOD clipping parallel strategy is proposed,which decomposes the clipping process in the traditional LOD model construction method into three parts: topology type judgment,intersection calculation and polygon triangulation.Under the premise of keeping thread load balance,geometric data is mapped to threads for parallel processing,which improves computation efficiency.Finally,several optimization strategies are proposed to further improve the performance of LOD construction method.Experiments show that according to different model scales,the CUDA-based LOD model construction method proposed in this thesis can achieve 2 to 3 times efficiency improvement compared with traditional methods,greatly improving the LOD construction efficiency.