Research On The Key Technology Of Urban Scene Perception And Building Reconstruction Based On Multi-source Spatial Data Fusion

The urban refined 3D model is the infrastructure of Digital China.The 3D urban building models are the spatial carrier for the association,convergence and integration of economic and social big data of cities.It has been widely used in fields such as smart city management,social comprehensive management,emergency decision-making and so on.At present,the 3D reconstruction of urban targets is still mainly based on multi-view images or point cloud data for interactive editing.There are problems of low efficiency,heavy workload and high cost with these kinds of method.The key technology of rapid automatic reconstruction of 3D urban building models needs breakthroughs: 1)The main modeling method based on 2D images needs to manually identify and extract the accurate three-dimensional vector information for scene reconstruction;2)The laser point cloud from a single platform lack the ability of obtaining the complete data of the scene,due to occlusion,limitation of view,etc.The lack of semantic information leads to great difficulty in point cloud-based target recognition and topology reconstruction.The key solution to solve the high-automatic reconstruction of refine models lies on effectively integration of images and point cloud data to provide a more complete target and structural feature description.However,the integration process mainly resides in texture mapping between images and high-precision cloud points of different platforms at present.The integration of feature level and even the decision making-level of two types of heterogeneous multi-source data need to be further studied.In view of the above problem,this paper focuses on the high-automatic and reliable reconstruction of urban targets supported by near-ground laser multi-sensor and multi-source information.The main data sources including multi-source data such as point cloud,image and trajectory information acquired by the near-surface multi-platform laser scanning system.By combining with the theory and method of feature expression and deep learning,researches to establish a high efficiency,high credibility and expandable three-dimensional reconstruction theory of urban architecture for new low-altitude multi-platform laser scanning multi-source data,the new system of the method can turn high-automated and fine reconstruction of urban construction into reality.The main research contents and achievements are as follows:In order to support the subsequent processing of efficient parallel primitive extraction from point cloud and images,target recognition and structure perception,the spatial distribution characteristics and processing requirements of multi-source data,such as near-surface multi-platform scanning point cloud,images and trajectory,are deeply studied and analyzed.Based on the analysis,combined with the distributed database,HBase,the multi-source data index construction and distributed storage method of Key-Value distributed database based on the geographic quadtree and the second-order random octree are proposed.Functions such as multi-source data dynamic scheduling,spatial query,dynamic update are implemented.The effectiveness of the multi-source data organization management method in this paper is verified by experiments,which can meet the requirements of subsequent multi-source combined data processing.Based on the analysis of existing point cloud registration and fusion methods,combined with the characteristics of a large number of planar features in urban point cloud data,this research suggests a fusion method.Based on low altitude UAV point cloud as the geographic basis of data fusion,this method utilizes multi-scale plane element constrained segmented vehicle-borne point cloud data,static scanning point cloud data and UAV-borne point cloud data.This paper presents the key steps of this method,including plane primitive extraction,homonym matching and multi-source point cloud registration strategy.Quantitative and qualitative analysis shows that the proposed algorithm can support high-precision registration and fusion of near-surface multi-platform scanning point cloud in urban area.On the basis of profound analysis and understanding of existing point cloud classification and image based target recognition methods,a multi-level classification network for ground,building,vegetation,vehicle,rod-shaped target and building detail window is constructed by combining rule set,random forest and Faster-RCNN deep learning.Experiments show that this method can effectively combine multi-dimensional features such as point cloud geometry and image texture,and can extract targets of interest reliably from multi-platform point cloud data and multi-view images.The feasibility of semantic structure perception in deep learning method for the urban scene is verified.As the center of this research,the building reconstruction in urban scene,based on the analysis of geometric characteristics of multi-source data,takes data driven reconstruction as the core.Combining the point cloud based building topological structure reconstruction with the imaged based of accurate boundary constraints,it develops a scientific method that integrates multi-source data,including airborne laser point cloud,ground laser point cloud and oblique aerial photography data,for 3D urban building reconstruction.Taking full advantage of geometrical characteristics of each data,it establishes a reliable solution for automatic building reconstruction.Moreover,it provides a high precision model framework for the next fine reconstruction interactive editing works.