A progressive multi-level detail monomer reconstruction method for real-life 3D building models

Abstract

The present application provides a method for progressive multi-level detailed monomer reconstruction of a real-life three-dimensional building model, which relates to the field of monomer reconstruction of inclined models. The method comprises: sequentially constructing a LOD0-level vector element model, a LOD1.0-level building white model, a LOD2.0-level building white model and a LOD2.0-level building standard model through DEM data and DSM raster data converted from the real-life three-dimensional model, completing the progressive multi-level detailed monomer reconstruction of the real-life three-dimensional model, and fully automatically performing monomer reconstruction of the building without manual intervention, thereby greatly reducing the cost of monomer reconstruction and improving the efficiency of monomer reconstruction of the real-life three-dimensional building.

Description

Technical Field

[0001] The present application relates to the field of individual reconstruction of tilted models, and in particular to a method for progressive multi-level individual reconstruction of real-scene three-dimensional architectural models. Background Art

[0002] A real-life 3D model refers to a 3D model that expresses the real world and is obtained and generated using modern surveying and mapping technologies such as photogrammetry, remote sensing, 3D laser scanning, and drone mapping. Among them, the oblique photography model is currently the most common real-life 3D model and plays an important role in urban-level real-life 3D construction. However, the oblique photography model is usually a continuous triangulated network structure (a triangulated network is a mesh-like plane control graphic composed of a series of continuous triangles). The model has a large amount of data, is difficult to update, lacks semantic information about scene objects, and cannot perform object-based management of geographic entities such as buildings, which limits the further development of real-life 3D applications.

[0003] In response to the problem of building singulation in real-life three-dimensional models, relevant researchers have proposed a dynamic singulation technology based on cutting singulation based on vector contours and superimposing vector data during three-dimensional rendering. However, this technology only performs singulation processing on the basis of the original inclined model, and does not perform lightweight reconstruction of the complex triangulated network and discrete textures of the inclined model. It cannot fundamentally solve the inherent limitations of the inclined model, such as large data volume, uneven surface, presence of holes and suspended objects, and difficulty in applying three-dimensional spatial analysis.

[0004] At present, the individual reconstruction of real-life 3D building models, especially the individual modeling of buildings better than LOD1.3 and LOD2.0, is mainly based on manual modeling and semi-automation. This technical solution has low reconstruction efficiency, high cost, and is prone to geometric structure errors, making it difficult to meet the requirements of large-scale real-life 3D applications.

[0005] In addition, different real-life 3D applications have different requirements for the accuracy of reconstruction of individual building models and the maximum load of model rendering, and often require the production of model data with multiple levels of details such as LOD1.0, LOD1.3, and LOD2.0. At present, when performing multi-level LOD modeling of real-life 3D building models, the method of separate independent construction or step-by-step extraction from high-level models to low-level models is generally adopted, which greatly limits the production efficiency of multi-level individual building models. Summary of the invention

[0006] The purpose of the present invention is to provide a method for progressive multi-level detail singulation reconstruction of real-life three-dimensional building models in order to solve the problems of high cost of manual singulation modeling, difficulty in ensuring the correctness of model geometric structure, and low efficiency in building multi-level LOD models in the existing real-life three-dimensional model building singulation process.

[0007] The above-mentioned purpose of the present application is achieved through the following technical solutions:

[0008] S1: Based on the real-life 3D model of the building, a deep learning model based on RandLA-Net and PointNet++ is used to obtain classified point cloud data;

[0009] S2: Mesh the point cloud data to obtain each point cloud block;

[0010] Use point cloud blocks that only retain ground classification to generate DEM data that express terrain features grid by grid;

[0011] Using point cloud blocks that retain building classification and ground classification, DSM raster data expressing building height characteristics is generated grid by grid;

[0012] This market

[0013] S3: According to the pixel height value, the DSM raster data is rendered into a rasterized three-dimensional model and cut horizontally step by step to obtain the multi-layer contour surface data of the building; the multi-layer contour surface data of the building is subjected to building contour surface regularization processing, and the building contour surface after regularization processing is output;

[0014] S4: Perform spatial merging operation on the building outline surface to obtain a LOD0 level vector element model;

[0015] S5: Based on the LOD0 vector feature model, combined with DSM raster data and DEM data, determine the roof height and bottom height of the building;

[0016] S6: Based on the building outline surface, roof height value and bottom height value of the LOD0 level vector element model, a model mesh body is generated to obtain a LOD1.0 level building white model of the second level of detail;

[0017] S7: Obtain LOD1.3 block white mold according to the building multi-layer contour surface data and LOD1.0 level building white mold;

[0018] S8: Mapping a universal texture to the LOD1.3 block white model to obtain a LOD1.3 building block model;

[0019] S9: Perform grid division and plane fitting processing on the DSM grid data to determine the roof cover structure; generate the LOD2.0 level building white mold according to the roof cover structure and the LOD1.3 block white mold;

[0020] S10: Obtain the first wall texture and the first roof texture of each building through the LOD2.0 level building white model;

[0021] S11: Map the first roof texture and the first wall texture to the LOD2.0 building white model to generate a LOD2.0 building standard model, and complete the progressive multi-level detail monomer reconstruction of the real-scene 3D model.

[0022] By adopting the above-mentioned technical scheme, the present application is different from other existing technologies that can only output models of one level of precision in one reconstruction. The present invention realizes a one-time individual reconstruction based on the original tilted model through the innovation of the multi-level building individual reconstruction mode, and progressively outputs LOD0, LOD1.0, LOD1.3, LOD2.0 multi-level detailed building individual models. The high-level model construction can make full use of the reconstruction results of the low-level model, saving the individual modeling time and reducing the modeling cost.

[0023] This application adopts the above technical solution. The existing technology based on the reconstruction of real-life 3D models generally directly cuts and reconstructs based on the tilted model Mesh triangulation, which has a huge amount of calculation and cannot solve the geometric structure anomalies of the reconstruction results caused by mesh triangulation holes and suspended objects. The present invention first generates a color point cloud based on the tilted model, uses deep learning technology to classify the point cloud, and then uses the point cloud that only retains ground points and building points to generate DSM raster data for building reconstruction, which greatly improves the model reconstruction efficiency and the robustness of reconstruction for abnormal tilted data, and has stronger adaptability to abnormal Mesh data.

[0024] Step S1 includes:

[0025] Obtain a real-life 3D model;

[0026] Perform interval sampling on the real-scene 3D model to generate a 3D color point cloud;

[0027] A deep learning model based on the combination of RandLA-Net and PointNet++ is used to perform point cloud semantic segmentation on the three-dimensional color point cloud to obtain classified point cloud data; point cloud semantic segmentation includes: building segmentation, ground segmentation and vegetation segmentation.

[0028] This application adopts the above-mentioned technical solution, uses RandLA-Net as the main framework of point cloud semantic segmentation and integrates the point cloud sampling algorithm of PointNet++ to ensure the efficiency, accuracy and adaptability of point cloud classification to different density areas of point clouds.

[0029] Step S2 includes:

[0030] The point cloud data is meshed according to the preset size, and the overlapping areas with preset distances are set between the meshes.

[0031] This application uses the above technical solution to grid the point cloud data to avoid the problem of memory overflow caused by excessive point cloud data volume and reduce the requirements for the computer hardware environment. The division of multiple grids can also be parallelized to further improve the efficiency of later building reconstruction. In addition, in order to solve the problem of buildings being divided in the grid boundary area, resulting in half of the building, a 10-meter overlap area is maintained between grids.

[0032] Step S3 includes:

[0033] The step of performing building contour surface regularization processing on the building multi-layer contour surface data and outputting the building contour surface after regularization processing comprises:

[0034] S31: Construct building polygons through multi-layer building contour surface data;

[0035] Determine the main direction of the building through the building polygon, the specific steps include: statistically weighting each edge of the building polygon to obtain a candidate direction; calculate the angle deviation value and length ratio coefficient of each edge of the building relative to the candidate direction; determine the contribution value according to the angle deviation value and the length ratio coefficient; determine the candidate direction with the largest contribution value as the main direction of the building;

[0036] S32: Filter and group the characteristic edges of the building polygons according to the characteristic edge screening conditions, including:

[0037] Set the edge length threshold determination coefficient and , ;

[0038] The feature edge screening conditions include: the length of the feature edge is greater than , the characteristic edge is perpendicular or parallel to the main direction of the building and the edge length is greater than , or there is an adjacent edge perpendicular to the feature edge and the edge length is greater than ;

[0039] S33: performing a right-angle processing on the boundary of the building polygon, the specific steps include: performing a 90-degree right-angle processing on two adjacent sides of the building polygon, taking the main direction of the building as a reference;

[0040] S34: Use the Douglas compression algorithm to simplify the building polygons, remove redundant points and non-feature nodes, and output the regularized building contour surface.

[0041] Step S5 includes:

[0042] Based on the LOD0 level vector feature model, the DSM raster data expressing the building is superimposed to calculate the pixel height value. The specific steps include: using the quartile method to remove the extreme values ​​of each roof height value within the building vector range;

[0043] Calculate the average of the roof height values ​​after removing the extreme values, and use the average as the roof height value of the building;

[0044] Based on the LOD0 level vector feature model, DEM data is superimposed and the pixel value of the center point of the building vector outline is used as the bottom height value of the building.

[0045] This application adopts the above technical solution, and the method of using the quartile method to remove extreme values ​​and then calculate the average height can calculate a relatively accurate height value that expresses the vertical position of the building roof, avoiding the influence of a few outliers on the height of the building top. The height of the building bottom uses the DEM pixel value of the center point of the vector contour to make the building bottom fit the ground better.

[0046] Step S7 includes:

[0047] The overlapping area ratio algorithm is used to calculate the similarity of the multi-story building contour surface data, and the multi-story building contour surfaces with similarity greater than a preset threshold are clustered and merged to obtain a preset number of contour surfaces of different heights;

[0048] The model mesh is generated for the preset number of contour surfaces of different heights after clustering and merging, forming a multi-level stacked closed prism, and the LOD1.3 block white model that distinguishes blocks of different heights is output.

[0049] Step S8 includes:

[0050] Common textures include: mapped roof textures and wall textures;

[0051] For the top and wall of the LOD1.3 block white model, map the roof texture and wall texture respectively by repeated mapping. The steps include:

[0052] On the wall of the LOD1.3 block white mold, place single window images in sequence in the horizontal direction from left to right, the vertical direction from top to bottom and the preset spacing. Stop placing when the horizontal or vertical position is not enough to place a window; merge the wall and window images to complete the mapping of the wall texture of the LOD1.3 block white mold;

[0053] The LOD1.3 building block model is generated by mapping the LOD1.3 block white mold after the wall texture and roof texture.

[0054] Step S9 includes:

[0055] Dividing the DSM grid data expressing the building height characteristics according to a preset size of grids to obtain a preset number of DSM grids;

[0056] The RANSAC random sampling consensus algorithm is used to perform plane fitting on the three-dimensional pixel point set of the DSM grid. The main implementation steps are as follows:

[0057] S91: Read the plane coordinates and height values ​​of all pixel points within the preset range of the DSM grid, and store them in an array in the form of three-dimensional coordinates;

[0058] S92: Select three non-collinear points from an array and calculate the plane equation ,in is the normal vector of the plane, Represents the vertical distance from the plane to the origin of the coordinate system along the normal direction;

[0059] S93: Calculate each point within the preset range Distance to the plane;

[0060] If the distance from the point to the plane is less than the preset threshold, the point is considered an interior point;

[0061] If the distance from the point to the plane is greater than or equal to the preset threshold, the point is considered an outlier;

[0062] S94: repeating steps S92 to S93 until a preset number of times is reached, obtaining the plane with the largest number of internal points as the current best model of the DSM grid, and outputting the best fitting plane for each DSM grid;

[0063] Calculate the normal vector angle between the best fitting plane and its adjacent small mesh planes, merge the small mesh planes whose normal vector angle is less than a preset threshold degree, and obtain the roof feature surface; the roof feature surface includes: slope feature structure and plane feature structure;

[0064] Filtering the plane feature structures whose area is smaller than a preset threshold and the slope feature structures whose area is smaller than a preset threshold in the roof feature surfaces;

[0065] According to the edge of the roof feature surface, the filtered slope feature structure and plane feature structure are buffered and extended to form a roof cover structure that is interconnected and seamless;

[0066] On the basis of the LOD1.3 block white mold, the roof cover structure is superimposed and the edges and seams are blended to generate the LOD2.0 level building white mold.

[0067] The present application adopts the above-mentioned technical scheme to generate DSM data that only reflects the characteristic structure of the building roof based on the point cloud classification results, divides it into small grids, performs feature surface fitting, and then recursively merges adjacent feature surfaces with a coplanar trend to obtain the plane feature surface and the inclined feature surface of the building roof, and extends and merges the feature surfaces to generate a closed, continuous, high-precision, and topologically good roof cover structure, which retains the main characteristic structure of the roof while reducing the amount of model data.

[0068] Step S10 includes:

[0069] On the basis of LOD2.0 building white model, the RGB color value of the inclined texture image pixel by pixel is read according to the top-view orthographic projection method and the side-view horizontal projection method, and a single roof texture image and a preset number of wall texture images are generated respectively; the preset number of wall texture images are baked and merged into a single wall texture image; the single roof texture image is the first roof map of the building; the single wall texture image is the first wall map of the building.

[0070] A computer-readable storage medium stores instructions. When the instructions are executed, a method for progressive multi-level detail monomer reconstruction of a real-scene three-dimensional building model is executed.

[0071] The beneficial effects of the technical solution provided by this application are:

[0072] The method of first converting the inclined model into DSM raster data that expresses the undulating structure of the building and then performing multi-level contour extraction and roof feature surface extraction is adopted to better ensure the correctness of the geometric structure and the adaptability to abnormal Mesh data; the idea of ​​progressive construction from low-level models to high-level models is adopted. The construction process of high-level models can make full use of the results of low-level models. One-time construction can generate multi-level detailed models, reducing the time-consuming multi-level repeated modeling and LOD chain generation of building fine models, providing technical support for the construction of multi-level LOD models of building data in ultra-large scene ranges and multi-scale expression of three-dimensional scene building data.

[0073] The individual reconstruction process is truly fully automated. From the segmentation of buildings to the individual reconstruction of white models, from the texture mapping of the model to the output of multi-level LOD models, all are automatically realized using deep learning technology and three-dimensional reconstruction algorithms. The output of individual building models with the highest LOD2.0 level can be achieved without human intervention, reducing the time and cost of individual building modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] The present application will be further described below with reference to the accompanying drawings and embodiments, in which:

[0075] Figure 1 is a step diagram in an embodiment of the present application;

[0076] Figure 2 This is a schematic diagram of point cloud sampling, segmentation and classification based on a tilt model in an embodiment of the present application;

[0077] Figure 3 is a schematic diagram of point cloud block processing in an embodiment of the present application;

[0078] Figure 4 It is a DEM and DSM data map based on the classified point cloud output in the embodiment of the present application;

[0079] Figure 5 It is a contour surface diagram generated by multi-layer horizontal sectioning of a DSM three-dimensional model expressing a building in an embodiment of the present application;

[0080] Figure 6 is a regularized processing diagram of the building outline in the embodiment of the present application;

[0081] Figure 7 It is the LOD0 building vector model diagram in the embodiment of the present application;

[0082] Figure 8 This is a schematic diagram of calculating the building height and top height according to DEM and DSM in an embodiment of the present application;

[0083] Fig. 9 It is the LOD1.0 level building white model drawing in the embodiment of the present application;

[0084] Fig.10 It is the LOD1.3 block white model in the embodiment of the present application;

[0085] Fig.11 It is a LOD1.3 building block model diagram of the universal texture mapped in the embodiment of the present application;

[0086] Fig.12 is a characteristic surface diagram of the roof in the embodiment of the present application;

[0087] Fig.13 is a structural diagram of a roof top cover in an embodiment of the present application;

[0088] Fig.14 It is the LOD2.0 level building white model drawing in the embodiment of the present application;

[0089] Fig.15 It is a single rooftop map and a wall map generated based on oblique pixel-by-pixel sampling in the embodiment of the present application;

[0090] Fig.16 It is the LOD2.0 level building standard model diagram in the embodiment of this application. DETAILED DESCRIPTION

[0091] In order to have a clearer understanding of the technical features, purposes and effects of the present application, the specific implementation methods of the present application are now described in detail with reference to the accompanying drawings.

[0092] The embodiments of the present application provide a method for progressive multi-level detail monomer reconstruction of a real-life three-dimensional building model.

[0093] Please refer to Figure 1 , Figure 1 This is a step diagram of a method for progressive multi-level detail monomer reconstruction of a real-scene three-dimensional building model in an embodiment of the present application, comprising:

[0094] S1: Based on the real-life 3D model of the building, a deep learning model based on RandLA-Net and PointNet++ is used to obtain classified point cloud data;

[0095] Specifically, RandLA-Net is a network structure for efficient semantic segmentation of large-scale point cloud data. It achieves semantic segmentation of point clouds through multiple steps such as Local Spatial Encoding (LocSE), Point Feature Enhancement, Attentive Pooling, Weighted Sum, Dilated Residual Block, Random Sampling (RS) and Local Feature Aggregation (LFA). Compared with the PointNet++ deep learning model, RandLA-Net is more efficient in semantic segmentation of large-scale point clouds and has lower requirements for computer video memory configuration. However, PointCNN++ uses the multi-scale method MSG and the multi-level method MRG to solve the problem of uneven sampling of sparse point clouds, and PointCNN++ has better sampling effect. The three-dimensional color point cloud and point cloud classification results are shown in the figure. Figure 2 shown.

[0096] S2: Mesh the point cloud data to obtain each point cloud block;

[0097] Use point cloud blocks that only retain ground classification to generate DEM data that express terrain features grid by grid;

[0098] Using point cloud blocks that retain building classification and ground classification, DSM raster data expressing building height characteristics is generated grid by grid;

[0099] Specifically, the schematic diagram of grid division is as follows Figure 3The DEM and DSM results based on the point cloud classification results are shown in Figure 4 shown.

[0100] S3: According to the pixel height value, the DSM raster data is rendered into a rasterized three-dimensional model and cut horizontally step by step to obtain the multi-layer contour surface data of the building; the multi-layer contour surface data of the building is subjected to building contour surface regularization processing, and the building contour surface after regularization processing is output;

[0101] S4: Perform spatial merging operation on the building outline surface to obtain a LOD0 level vector element model;

[0102] Specifically, a spatial merging operation is performed on the building contour surface, and the merged building contour surface is used as the building base surface, and the LOD0 level vector element model expressing the first level detail of the single building base contour is output. The LOD0 level vector element model of the first level detail is as follows: Figure 7 shown.

[0103] S5: Based on the LOD0 vector feature model, combined with DSM raster data and DEM data, determine the roof height and bottom height of the building;

[0104] S6: Based on the building outline surface, roof height value and bottom height value of the LOD0 level vector element model, a model mesh body is generated to obtain a LOD1.0 level building white model of the second level of detail;

[0105] Specifically, based on the building outline surface of the LOD0 vector element model and the roof height value and bottom height value of each building unit, the model mesh is generated to form a closed building prism geometry structure, and the LOD1.0 building white model of the second level of detail is obtained, as shown below Fig. 9 shown.

[0106] S7: Obtain LOD1.3 block white mold according to the building multi-layer contour surface data and LOD1.0 level building white mold;

[0107] S8: Mapping a universal texture to the LOD1.3 block white model to obtain a LOD1.3 building block model;

[0108] S9: Perform grid division and plane fitting processing on the DSM grid data to determine the roof cover structure; generate the LOD2.0 level building white mold according to the roof cover structure and the LOD1.3 block white mold;

[0109] S10: Obtain the first wall texture and the first roof texture of each building through the LOD2.0 level building white model;

[0110] S11: Map the first roof texture and the first wall texture to the LOD2.0 building white model to generate a LOD2.0 building standard model, and complete the progressive multi-level detail monomer reconstruction of the real-scene 3D model.

[0111] Specifically, the first roof texture and the first wall texture are mapped to the LOD2.0 building white model, and the model texture and UV texture coordinates (U coordinate corresponds to the horizontal direction of the texture, and V coordinate corresponds to the vertical direction of the texture) are calculated to generate the fourth level of detail LOD2.0 building standard model, such as Fig.16 shown.

[0112] Step S1 includes:

[0113] Obtain a real-life 3D model;

[0114] Perform interval sampling on the real-scene 3D model to generate a 3D color point cloud;

[0115] A deep learning model based on the combination of RandLA-Net and PointNet++ is used to perform point cloud semantic segmentation on the three-dimensional color point cloud to obtain classified point cloud data; point cloud semantic segmentation includes: building segmentation, ground segmentation and vegetation segmentation.

[0116] Step S2 includes:

[0117] Point cloud data Grid division, with an overlap area of ​​10 meters between grids.

[0118] Step S3 includes:

[0119] The step of performing building contour surface regularization processing on the building multi-layer contour surface data and outputting the building contour surface after regularization processing comprises:

[0120] S31: Construct building polygons through multi-layer building contour surface data;

[0121] Determine the main direction of the building through the building polygon, the specific steps include: statistically weighting each edge of the building polygon to obtain a candidate direction; calculate the angle deviation value and length ratio coefficient of each edge of the building relative to the candidate direction; determine the contribution value according to the angle deviation value and the length ratio coefficient; determine the candidate direction with the largest contribution value as the main direction of the building;

[0122] S32: Filter and group the characteristic edges of the building polygons according to the characteristic edge screening conditions, including:

[0123] Set the edge length threshold determination coefficient and , ;

[0124] The feature edge screening conditions include: the length of the feature edge is greater than , the characteristic edge is perpendicular or parallel to the main direction of the building and the edge length is greater than , or there is an adjacent edge perpendicular to the feature edge and the edge length is greater than ;

[0125] Specifically, the characteristic edge must meet one of the following conditions: 1) The edge length is greater than ; 2) perpendicular or parallel to the main direction, and the side length is greater than ; 3) There is an adjacent edge perpendicular to it, and the length of the edge is greater than .

[0126] S33: performing a right-angle processing on the boundary of the building polygon, the specific steps include: performing a 90-degree right-angle processing on two adjacent sides of the building polygon, taking the main direction of the building as a reference;

[0127] S34: Use the Douglas compression algorithm to simplify the building polygons, remove redundant points and non-feature nodes, and output the regularized building contour surface.

[0128] Specifically, the three-dimensional model is cut horizontally from the lowest point to the highest point at a vertical interval of 0.3 meters to obtain the multi-story building contour surface data with height information. The contour surface generation result is as follows Figure 5 The results before and after the building outline is regularized are shown in Figure 6 shown.

[0129] Step S5 includes:

[0130] Based on the LOD0 level vector feature model, the DSM raster data expressing the building is superimposed to calculate the pixel height value. The specific steps include: using the quartile method to remove the extreme values ​​of each roof height value within the building vector range;

[0131] Calculate the average of the roof height values ​​after removing the extreme values, and use the average as the roof height value of the building;

[0132] Based on the LOD0 level vector feature model, DEM data is superimposed and the pixel value of the center point of the building vector outline is used as the bottom height value of the building.

[0133] Specifically, the schematic diagram of the top and bottom elevation assignment results of the building outline is as follows: Figure 8 shown.

[0134] Step S7 includes:

[0135] The overlapping area ratio algorithm is used to calculate the similarity of the multi-story building contour surface data, and the multi-story building contour surfaces with similarity greater than a preset threshold are clustered and merged to obtain a preset number of contour surfaces of different heights;

[0136] The model mesh is generated for the preset number of contour surfaces of different heights after clustering and merging, forming a multi-level stacked closed prism, and the LOD1.3 block white model that distinguishes blocks of different heights is output.

[0137] Specifically, the calculation steps of the overlapping area ratio algorithm are as follows: in the building multi-layer contour surface data, such as adjacent contour surface A and contour surface B, first perform spatial geometric intersection of surface A and surface B to obtain the intersection C surface, and then divide the area of ​​surface C by the area of ​​surface A and surface B respectively. If both are higher than 95%, it is considered that contour surface A and contour surface B are similar contours and can be clustered and merged. Otherwise, they are considered to be different feature contour surfaces, and so on. Then, the model mesh body is generated for one or more contour surfaces of different heights after clustering and merging to form a multi-level stacked closed prism, and the LOD1.3 block white model that distinguishes blocks of different heights is output, such as Fig.10 shown.

[0138] Step S8 includes:

[0139] Common textures include: mapped roof textures and wall textures;

[0140] For the top and wall of the LOD1.3 block white model, map the roof texture and wall texture respectively by repeated mapping. The steps include:

[0141] On the wall of the LOD1.3 block white mold, place single window images in sequence in the horizontal direction from left to right, the vertical direction from top to bottom and the preset spacing. Stop placing when the horizontal or vertical position is not enough to place a window; merge the wall and window images to complete the mapping of the wall texture of the LOD1.3 block white mold;

[0142] The LOD1.3 building block model is generated by mapping the LOD1.3 block white mold after the wall texture and roof texture.

[0143] Specifically, the steps of mapping the roof texture on the LOD1.3 block white model by repeated mapping include: mapping the roof texture and the wall texture onto the LOD1.3 white model to obtain the LOD1.3 building block model of the third level of detail, such as Fig.11 shown.

[0144] Step S9 includes:

[0145] The DSM grid data expressing the building height characteristics is The grid size is divided to obtain a preset number of DSM grids;

[0146] The RANSAC random sampling consensus algorithm is used to perform plane fitting on the three-dimensional pixel point set of the DSM grid. The main implementation steps are as follows:

[0147] S91: Read DSM grid The plane coordinates and height values ​​of all pixel points within the range are stored in an array in the form of three-dimensional coordinates;

[0148] S92: Select three non-collinear points from an array and calculate the plane equation ,in is the normal vector of the plane, Represents the vertical distance from the plane to the origin of the coordinate system along the normal direction;

[0149] S93: Calculation Every point in the range Distance to the plane;

[0150] If the distance from the point to the plane is less than the preset threshold, the point is considered an interior point;

[0151] If the distance from the point to the plane is greater than or equal to the preset threshold, the point is considered an outlier;

[0152] S94: repeating steps S92 to S93 until a preset number of times is reached, obtaining the plane with the largest number of internal points as the current best model of the DSM grid, and outputting the best fitting plane for each DSM grid;

[0153] Calculate the normal vector angle between the best fitting plane and its adjacent small mesh planes, merge the small mesh planes whose normal vector angle is less than a preset threshold degree, and obtain the roof feature surface; the roof feature surface includes: slope feature structure and plane feature structure;

[0154] Filtering the plane feature structures whose area is smaller than a preset threshold and the slope feature structures whose area is smaller than a preset threshold in the roof feature surfaces;

[0155] According to the edge of the roof feature surface, the filtered slope feature structure and plane feature structure are buffered and extended to form a roof cover structure that is interconnected and seamless;

[0156] On the basis of the LOD1.3 block white mold, the roof cover structure is superimposed and the edges and seams are blended to generate the LOD2.0 level building white mold.

[0157] Specifically, the roof features are as follows: Fig.12 The interconnected and seamless roof structure, such as Fig.13 As shown. Fig.14As shown, the geometric structures of the building roof and wall reconstructed by the present invention are clear, and the typical plane feature structure and slope surface feature structure of the roof are retained. The maximum degree of simplification is carried out while ensuring the main body and local form of the building. Compared with the original inclined model data, the simplification degree of the triangulated network can reach more than 90%.

[0158] Step S10 includes:

[0159] On the basis of LOD2.0 building white model, the RGB color value of the inclined texture image pixel by pixel is read according to the top-view orthographic projection method and the side-view horizontal projection method, and a single roof texture image and a preset number of wall texture images are generated respectively; the preset number of wall texture images are baked and merged into a single wall texture image; the single roof texture image is the first roof map of the building; the single wall texture image is the first wall map of the building.

[0160] Specifically, wall maps and roof maps, such as Fig.15 shown.

[0161] The time consumption comparison between the progressive multi-level detail monomer reconstruction method proposed in the present invention and the existing monomer modeling method mainly based on manual modeling is shown in Table 1 below.

[0162] Table 1 Comparison of time consumption of progressive multi-level detail monomer reconstruction and manual modeling

[0163]

[0164] According to the workload of building monomer modeling in a city of 1 square kilometer and one modeler working 8 hours a day, the method of the present invention is used to build the LOD2.0 building standard model, which is about 25 times more efficient than manual modeling. Since the present invention adopts a progressive multi-level detailed monomer reconstruction method, the LOD2.0 model is constructed while outputting LOD1.3, LOD1.0 and LOD0 models, saving the time of constructing low-level LOD models separately. Therefore, the efficiency of LOD0-LOD2.0 multi-level LOD model construction is increased by about 43 times, which greatly improves the efficiency of building monomer reconstruction based on tilted models.

[0165] The present application also discloses a computer-readable storage medium storing a plurality of instructions suitable for loading by a processor to execute the above-mentioned method for progressive multi-level detail monomer reconstruction of a real-scene three-dimensional building model.

[0166] The above are only exemplary embodiments of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification and the truth of practice, those skilled in the art will easily think of other embodiments of the present disclosure.

[0167] This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art not described in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A method for progressive multi-level detail monomer reconstruction of a real-life 3D building model, characterized in that: The method comprises the following steps: S1: Based on the real-life 3D model of the building, a deep learning model based on RandLA-Net and PointNet++ is used to obtain classified point cloud data; S2: Mesh the point cloud data to obtain each point cloud block; Use point cloud blocks that only retain ground classification to generate DEM data that express terrain features grid by grid; Using point cloud blocks that retain building classification and ground classification, DSM raster data expressing building height characteristics is generated grid by grid; S3: According to the pixel height value, the DSM raster data is rendered into a rasterized three-dimensional model and cut horizontally step by step to obtain the multi-layer contour surface data of the building; Performing building outline regularization processing on the building multi-layer outline surface data, and outputting the building outline surface after regularization processing; S4: Perform spatial merging operation on the building outline surface to obtain a LOD0 level vector element model; S5: Based on the LOD0 vector feature model, combined with DSM raster data and DEM data, determine the roof height and bottom height of the building; S6: Based on the building outline surface, roof height value and bottom height value of the LOD0 level vector element model, a model mesh body is generated to obtain a LOD1.0 level building white model of the second level of detail; S7: Obtain LOD1.3 block white mold according to the building multi-layer contour surface data and LOD1.0 level building white mold; S8: Mapping a universal texture to the LOD1.3 block white model to obtain a LOD1.3 building block model; S9: Perform grid division and plane fitting processing on the DSM grid data to determine the roof cover structure; generate the LOD2.0 level building white mold according to the roof cover structure and the LOD1.3 block white mold; S10: Obtain the first wall texture and the first roof texture of each building through the LOD2.0 level building white model; S11: Map the first roof texture and the first wall texture to the LOD2.0 building white model to generate a LOD2.0 building standard model, and complete the progressive multi-level detail monomer reconstruction of the real-scene 3D model.

2. A method for progressive multi-level detail monomer reconstruction of a real-scene three-dimensional building model as claimed in claim 1, characterized in that: Step S1 includes: Obtaining a real-life 3D model; Perform interval sampling on the real-scene 3D model to generate a 3D color point cloud; A deep learning model based on the combination of RandLA-Net and PointNet++ is used to perform point cloud semantic segmentation on the three-dimensional color point cloud to obtain classified point cloud data; point cloud semantic segmentation includes: building segmentation, ground segmentation and vegetation segmentation.

3. The method for progressive multi-level detail monomer reconstruction of a real-scene three-dimensional building model according to claim 1, characterized in that: Step S2 includes: The point cloud data is meshed according to the preset size; overlapping areas with preset distances are set between the meshes.

4. The method for progressive multi-level detail monomer reconstruction of a real-scene three-dimensional building model according to claim 1, characterized in that: Step S3 includes: The step of performing building contour surface regularization processing on the building multi-layer contour surface data and outputting the building contour surface after regularization processing comprises: S31: Construct building polygons through multi-layer building contour surface data; Determine the main direction of the building through the building polygon, the specific steps include: statistically weighting each edge of the building polygon to obtain a candidate direction; calculate the angle deviation value and length ratio coefficient of each edge of the building relative to the candidate direction; determine the contribution value according to the angle deviation value and the length ratio coefficient; determine the candidate direction with the largest contribution value as the main direction of the building; S32: Filter and group the characteristic edges of the building polygons according to the characteristic edge screening conditions, including: Set the edge length threshold determination coefficient and , ; The feature edge screening conditions include: the length of the feature edge is greater than , the characteristic edge is perpendicular or parallel to the main direction of the building and the edge length is greater than , or there is an adjacent edge perpendicular to the feature edge and the edge length is greater than ; S33: performing a right-angle processing on the boundary of the building polygon, the specific steps include: performing a 90-degree right-angle processing on two adjacent sides of the building polygon, taking the main direction of the building as a reference; S34: Use the Douglas compression algorithm to simplify the building polygons, remove redundant points and non-feature nodes, and output the regularized building contour surface.

5. The method for progressive multi-level detail monomer reconstruction of a real-scene three-dimensional building model according to claim 1, characterized in that: Step S5 includes: Based on the LOD0 vector feature model, the DSM raster data expressing the building is superimposed to calculate the pixel height value. The specific steps include: using the quartile method to remove the extreme values ​​of each roof height value within the building vector range; calculating the average value of the roof height value after removing the extreme values, and taking the average value as the roof height value of the building; Based on the LOD0 level vector feature model, DEM data is superimposed and the pixel value of the center point of the building vector outline is used as the bottom height value of the building.

6. The method for progressive multi-level detail monomer reconstruction of a real-scene three-dimensional building model according to claim 1, characterized in that: Step S7 includes: The overlapping area ratio algorithm is used to calculate the similarity of the multi-story building contour surface data, and the multi-story building contour surfaces with similarity greater than a preset threshold are clustered and merged to obtain a preset number of contour surfaces of different heights; The model mesh is generated for the preset number of contour surfaces of different heights after clustering and merging, forming a multi-level stacked closed prism, and the LOD1.3 block white model that distinguishes blocks of different heights is output.

7. The method for progressive multi-level detail monomer reconstruction of a real-scene three-dimensional building model according to claim 1, characterized in that: Step S8 includes: Common textures include: mapped roof textures and wall textures; For the top and wall of the LOD1.3 block white model, map the roof texture and wall texture respectively by repeating the mapping method. The steps include: On the wall of the LOD1.3 block white mold, place single window images in sequence in the horizontal direction from left to right, the vertical direction from top to bottom and the preset spacing. Stop placing when the horizontal or vertical position is not enough to place a window; merge the wall and window images to complete the mapping of the wall texture of the LOD1.3 block white mold; The LOD1.3 building block model is generated by mapping the LOD1.3 block white mold after the wall texture and roof texture.

8. The method for progressive multi-level detail monomer reconstruction of a real-scene three-dimensional building model according to claim 1, characterized in that: Step S9 includes: Dividing the DSM grid data expressing the building height characteristics according to a preset size of grids to obtain a preset number of DSM grids; The RANSAC random sampling consensus algorithm is used to perform plane fitting on the three-dimensional pixel point set of the DSM grid. The main implementation steps are as follows: S91: Read the plane coordinates and height values ​​of all pixel points within the preset range of the DSM grid, and store them in an array in the form of three-dimensional coordinates; S92: Select three non-collinear points from an array and calculate the plane equation ,in is the normal vector of the plane, Represents the vertical distance from the plane to the origin of the coordinate system along the normal direction; S93: Calculate each point within the preset range Distance to the plane; If the distance from the point to the plane is less than the preset threshold, the point is considered an interior point; If the distance from the point to the plane is greater than or equal to the preset threshold, the point is considered an outlier; S94: repeating steps S92 to S93 until a preset number of times is reached, obtaining the plane with the largest number of internal points as the current best model of the DSM grid, and outputting the best fitting plane for each DSM grid; Calculate the normal vector angle between the best fitting plane and its adjacent small mesh planes, merge the small mesh planes whose normal vector angle is less than a preset threshold degree, and obtain the roof feature surface; the roof feature surface includes: slope feature structure and plane feature structure; Filtering the plane feature structures whose area is smaller than a preset threshold and the slope feature structures whose area is smaller than a preset threshold in the roof feature surfaces; According to the edge of the roof feature surface, the filtered slope feature structure and plane feature structure are buffered and extended to form a roof cover structure that is interconnected and seamless; On the basis of the LOD1.3 block white mold, the roof cover structure is superimposed and the edges and seams are blended to generate the LOD2.0 level building white mold.

9. The method for progressive multi-level detail monomer reconstruction of a real-scene three-dimensional building model according to claim 1, characterized in that: Step S10 includes: On the basis of LOD2.0 building white model, the RGB color value of the inclined texture image pixel by pixel is read according to the top-view orthographic projection method and the side-view horizontal projection method, and a single roof texture image and a preset number of wall texture images are generated respectively; the preset number of wall texture images are baked and merged into a single wall texture image; the single roof texture image is the first roof map of the building; the single wall texture image is the first wall map of the building.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed by a computer, the method according to any one of claims 1 to 9 is executed.

Images