Structural Building Reconstruction From Point Clouds Under The Constraint Of Image Lines

The 21 st century is the era of big data,where cutting-edge technologies rapidly developed such as Internet of Things,cloud computing,artificial intelligence,and so forth.In this context,the acquisition and perception capabilities of spatiotemporal big data increased splendidly.In China,the construction of smart cities has risen as the national economic and technological strategy.The Ministry of Natural Resources announced the construction plan of Chinese Real Mesh Model in 2019,aimed at the city model reconstruction of the whole country.The spatial progress of urban construction requires the coverage of geographic data at different levels,from closeview micro scenes,to macro scenes at urban scale.To deal with diverse application requirements,smart city systems should be capable to provide different levels of technical services including visualization,model expression,planning,city design,etc.The three-dimensional models of the city provide powerful support for the construction of new smart cities,while building models stand out as the most valuable elements.The main goal of building model reconstruction is to express substantial semantic information with least data amount.The difficulty lies in how to abstract a parametric model that can accurately express the outlines and topologies of the building from complex and insufficient data sources.Although methods of data acquisition are increasing exponentially,the complexity of the environment and the lack of uniform standard for building structures make it ambiguous to form the modeling regulation.It is still difficult to obtain satisfactory data sources at low costs for fully automated building modeling.At present,the 3D reality mesh model can be automatically generated on a large scale,which contains huge data amount.The 3D triangular mesh model can provide basic functions such as visualization,environmental simulation.However,high-precision and compact models with semantic information still depend heavily on manual editing,which is inefficient,and the coverage is obviously inadequate.The smart services for urban planning and emergency response are limited to specific pilot scenes,which makes it difficult for scalable and extensible applications.A practical and reproducible building vector model should have accurate geometric positions,correct topological relationships,and clear texture maps,which can be expressed at multiple levels of details as required to meet different levels of applications.The construction of a novel smart city requires a comprehensive three-dimensional urban model with sufficient coverage and high quality to develop into more advanced forms such as smart city brains and digital twin cities and become a wisdom engine to promote social progress.In this paper,the complementary advantages of point clouds and optic images are fully utilized to reconstruct structural building models with texture.The main contributions consist of:(1)The Topo LAP algorithm is proposed,which is a polyhedral modeling method guided by linear and planar primitives.Polyhedral modeling based on space decompotion divides the bounding 3D space into convex polyhedron subsets,each of which belongs to either inside or outside of the model.The intersections of inside and outside polyhedrons conform the surface of the reconstructed polyhedral model.However,planar primitives extracted from point clouds are incomplete and nonuniform distributed,which makes the partition incomplete,especially in the urban environment.In this paper,linear primitives reconstructed from images are introduced as supplementary features which are immune with the poor image texture or walls of glass material.Image lines,boundary lines from planar point clouds,and intersection lines from planar pairs are combined to reconstruct structural outline of planes.And then lost planes are compensated by deducing the relationships between linear primitives and planar primitives.Finally,the complete planar primitives complex is used to partition the 3D space and the structural outlines are used to eliminate improper decompositions,which improves the efficiency to reconstruct polyhedral models.(2)The St PPP algorithm is proposed,which is a hierarchical Lo Ds generation method based on structural constrained triangulation.The border edges of a connected component of 2-manifold triangular mesh can shape arbitrary polygons.Based on this assumption,the St PPP algorithm reconstructs the polygon boundary by labeling the constrained Delaunay triangulation,which is involved in the following progresses during building reconstruction: 1)Footprints extraction,i.e.,select the facets inside the building from the triangulation of 2D point clouds;2)Extraction of rooftop boundaries,i.e.,partition the footprint into several polygons,each of which represents a rooftop facet,and the 3D rooftop model is reconstructed by reproject the 2D polygon boundaries to corresponding 3D rooftop planes.The intersection lines from plane pairs and the projected lines of facades are taken as structural features to constrain the boundary,and thus the topology of the rooftop planes can be reconstructed.Experimental results of the dense LiDAR point clouds from Dublin indicate that the proposed St PPP algorithm can handle the Lo Ds modeling of hyper sophisticated building structures.(3)The EATex algorithm is propsed,i.e.the edge-aware texture mapping of structural building models.The area of the surface of the building model can be too large to be imaged in a single image and the self-intersection and occlusion are common.To deal with these problems,an edge-aware texture mapping method is proposed focused on structural building models.The fusion models of wire-frame and triangular mesh are reconstructed from structural building models.Through visibility check,the facets which cannot be captured in one image are divided into sub-polygonal areas,each of which is totally visible or totally invisible.Then,the image lines are extracted and inserted to the mesh.The edges can be used as guidance during texture image selection:firstly,whether the edge is a shadow border is detected and used to guide the texture image avoiding the shadow area while prefer to clear area outside the shadow area;secondly,the edge is used to guide the border of texture patches to appear at the frame of the walls.The experimental results demonstrate that the texture mapping method is efficient for manually reconstructed regular building models as well as structural building models by automated reconstruction.Models can be represented with small amount of data and can be textured as fine as the models represented by dense mesh.(4)Introduction of the independently developed model reconstruction system integrated airborne LiDAR point clouds and multi-view images.The methods proposed in this paper are assembled in the system.The modeling strategy which reconstructs models after semantic segmentation is introduced.The efficiency and accuracy of modeling from different types of data are analyzed.Three test areas including LiDAR point clouds and multi-view images are involved in experiments,which are different in accuracy,density or structure characteristics.Modeling and texturing results indicate that the proposed methods can handle large-scale automated building reconstruction of massive complicated structures.In addition,the model representation used in this paper makes the model editable in the efficient interactive modeling system.In summary,under the constraints of image lines,the textured structural building models are reconstructed based on point clouds.Three practical algorithms,i.e.,Topo LAP,St PPP,and EATex are proposed,which are demonstrated by expriments.