Research On Key Technologies Of Large-Scale 3D Terrain Construction

The three-dimensional(3D)terrain simulates the undulating state and detail features of the terrain surface in a digital manner.It is widely used in flight vision simulation,battlefield situation simulation,simulation exercises,emergency command,disaster rescue and other application scenarios.The authenticity of 3D terrain and the real-time of visual rendering directly determine the fidelity,immersion and effectiveness of 3D applications.The key techniques of large-scale terrain 3D reconstruction based on sequence images collected by Unmanned Aerial Vehicles(UAVs)is studied in this paper,there are carried out from four aspects: sparse reconstruction,dense reconstruction,discrete point cloud gridding and large-scale terrain real-time rendering.Solving the key problems and giving specific solutions in the reconstruction process.In summary,this paper has done the following four innovative work:1.The image feature points extraction and the image feature matching are studied mainly when the terrain image is reconstructed by using sequence images.We investigated an improved SIFT algorithm.Under the premise of ensuring that the feature points are more and stable,by narrowing the selection range of near the feature point,the dimension of the feature description vectors are effectively reduced.Time performance is improved.Combining the properties of RANSAC and AC-RANSAC algorithm,an image registration method based on RANSAC and AC-RANSAC algorithm for essential matrix estimation is proposed.Different algorithms are used to eliminate mismatched pairs of points for different matching point sets.It is possible to eliminate more mismatches.2.The dense reconstruction method of sparse point cloud data is studied,and the dense reconstruction of PMVS was analyzed and summarized.When the sequence aerial image is densely reconstructed by PMVS,the algorithm has long calculation time and high space complexity.The improved multi-resolution layered diffusion reconstruction strategy is used to detect the plane curvature is detected on the reconstructed terrain,and different patch diffusion methods are selected to improve the overall speed of reconstruction and the accuracy of reconstruction results.3.The 3D reconstruction meshing method of discrete point cloud data is studied,combining with the characteristics of topographic point cloud data collected by UAVs sequence images,we proposed a fast meshing method based on Delaunay criterion and growth method for discrete terrain point cloud data.First,the space is split by a retractable bounding box,and spatial neighboring points(k-nearest neighbor points)of the point are find;then select a seed triangle with direction from the point cloud and design the search strategy for points and edges;finally,based on the existing outer edge,the outer candidate point set is determined,and one of the best points is selected from the outer candidate point set,and a new triangle is generated incrementally.4.During the rendering problems of large-scale terrain rule grids,this paper used standard image compression algorithm to convert regular grid data into 8-bit grayscale map,which effectively compresses the data volume of terrain block scheduling.The terrain data is horizontally meshed into blocks,and the vertical multi-resolution layered method is combined with the quadtree structure to reasonably organize the terrain data,and a two level quadtree index is established to improve the data retrieval efficiency;The multi-level buffer technology is used to realize the dynamic scheduling of large-scale terrain data between internal and external storage.When the terrain visualization is real-time rendering,the range and precision of the terrain data in the visible area will change with the change of the viewpoint.When the viewpoint changes greatly,a large number of terrain data blocks need to be scheduled,and a large number of disk I/O operations are performed.Not only does it bring delay to the rendering,it also affects the stability of the rendering frame rate.To ensure the smoothness and continuity of rendering,the terrain organization structure,scheduling strategy,node evaluation criteria,etc.based on quadtree structure are designed.We proposed a large-scale terrain real-time rendering algorithm based on CPU-GPU collaborative computing is proposed.The algorithm focuses on moving the batch LOD model from the CPU to the GPU.The CPU is mainly responsible for real-time scheduling of the terrain data blocks in the external memory into the memory,and loading the terrain block data into the memory of GPU,and the GPU is responsible for constructing the LOD model in parallel.