ALS point cloud facade geometry-based automatic reconstruction method for building three-dimensional model tending to LOD3 level

Abstract

The invention relates to the technical field of building three-dimensional modeling, and provides an ALS point cloud facade geometry-based LLO3-level-tending building three-dimensional model automatic reconstruction method, which comprises the steps of 1, building point cloud segmentation and bottom contour construction: segmenting a roof and facade area from an ALS point cloud, and constructing a bottom contour; automatically extracting the bottom contour of the actually grounded building through the vertical projection of the facade point cloud; step 2, door and window contour extraction based on facade geometry: projecting the facade point cloud into a two-dimensional image, identifying a door and window opening by using an Alpha Shape construction and geometry-semantic detection method, and extracting a door and window contour; and step 3, building a three-dimensional building model with enhanced facade details: inputting the bottom contour and the door and window contour into a modeling frame, and automatically embedding door and window components through polygonal surface patch combination and Boolean operation to generate a building model with a complete structure and semantic details. According to the method, the three-dimensional model of the building tending to the LOD3 level can be better and automatically reconstructed based on the ALS point cloud.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional modeling of buildings, in particular, to a LOD3-level building three-dimensional model automatic reconstruction method based on ALS point cloud facade geometry. BACKGROUND

[0002] Three-dimensional building models play a key role in many urban application scenarios such as digital twinning, urban planning, path navigation, energy analysis, and virtual reality visualization. The accurate reconstruction of geometric structures and the complete expression of semantic components in the model not only relate to the true restoration of spatial structures, but also directly affect the reliability and effectiveness of subsequent spatial analysis and system simulation. Among various three-dimensional spatial data acquisition methods, airborne LiDAR scanning (ALS) technology has become the mainstream data source for automatically generating large-scale urban building models due to its high resolution and wide coverage.

[0003] Current ALS-based building reconstruction methods mainly focus on the geometric modeling of roof structures. These methods usually project the roof boundary vertically to the ground to estimate the building outline. Although effective for regular structures, such methods often fail when buildings have overhanging eaves or inclined roofs, leading to outline expansion, position deviation, and geometric precision degradation.

[0004] To overcome these limitations, some studies attempt to introduce facade information. Methods based on facade point cloud projection and linear edge fitting can improve outline expansion, deviation, and precision deficiencies, but they struggle to generate closed outlines and cannot handle curved surface structures. Learning-based strategies explore the application of manually annotated facade datasets, but their performance highly depends on data quality and has limited generalization ability for different building types.

[0005] On the semantic level, windows, doors, and other facade openings are key components of functional description and model recognizability. Methods that fuse street view images or external contour data with other multi-source data can improve detection accuracy and detail reconstruction, but these methods require reliable external data sources, accurate cross-modal registration, and extensive preprocessing, severely limiting their scalability in large-scale ALS automatic reconstruction. SUMMARY

[0006] The present application provides a LOD3-level building three-dimensional model automatic reconstruction method based on ALS point cloud facade geometry, which directly reconstructs accurate building outlines and semantic elements such as windows and doors using facade geometric features, and integrates roof structures to generate complete models. This method effectively solves the geometric distortion problem of roof-based reconstruction methods, avoids excessive dependence on external data, improves the structural accuracy and semantic richness of building models, and thus promotes the refinement level and application value of automated large-scale three-dimensional reconstruction.

[0007] According to the application, a method for automatically reconstructing a three-dimensional model of a building of LOD3 level based on ALS point cloud facade geometry comprises the following steps:

[0008] Step 1: Building point cloud segmentation and bottom contour construction: roof and facade regions are segmented from the ALS point cloud by using normal analysis and plane fitting, and then the bottom contour of the building actually grounded is automatically extracted by vertical projection of the facade point cloud;

[0009] Step 2: Door and window contour extraction based on facade geometry: the facade point cloud is projected into a two-dimensional image, and the door and window contour is extracted by using AlphaShape construction and geometry-semantic detection method to identify the door and window opening, so as to realize the geometry acquisition of key facade semantic components;

[0010] Step 3: Three-dimensional building model construction with facade detail enhancement: the extracted bottom contour and door and window contour are input into the model-driven modeling framework, and the door and window components are automatically embedded by polygonal face combination and Boolean operation to generate a building model with complete structure and semantic details.

[0011] As preferred, in step 1, the planes of the ALS building point cloud are segmented by using RANSAC, and the specific steps are as follows:

[0012] 1.11) Plane point cloud segmentation: first, the plane of the building point cloud is segmented by using RANSAC algorithm; second, based on the normal parameters and center point of the fitted plane, the normal angle and center point distance between planes are calculated; finally, the plane point cloud with an angle and distance less than a threshold value is merged; finally, the merged plane parameters are re-estimated by using principal component analysis (PCA) method;

[0013] 1.12) Plane point cloud classification: first, the normal vector of each plane is calculated by using the plane point cloud parameters; then, the angle between the plane normal vector and the Z-axis is calculated, and if the normal vector is perpendicular to the Z-axis, it is classified as a facade point cloud, otherwise it is classified as a roof point cloud.

[0014] As preferred, in step 1, the specific steps of automatically extracting the bottom contour of the building are as follows:

[0015] 1.21) Projecting the facade point cloud to the ground: first, the building facade point cloud is orthogonally projected to the ground to generate a two-dimensional point set; second, the projected point set is sampled based on a regular grid to reduce the redundancy of the points and maintain the overall spatial distribution characteristics, i.e. the projected point set is divided into fixed-size grid cells, and only one representative point is retained in each grid cell; thereby obtaining a simplified two-dimensional scatter plot;

[0016] 1.22) Bottom contour extraction based on bisection approximation: First, construct the Delaunay triangulation of the two-dimensional ground projection point cloud; second, search for the optimal a value combined with the bisection recursion strategy, and extract the corresponding contour; if the contour is a multi-polygon, update a min , otherwise update a max ; until a max -a min > a ∈ , stop recursion and return the threshold and contour; a ∈ is the tolerance parameter;

[0017] 1.23) Contour simplification: Simplify the contour line using the RDP algorithm, replace the curve with a straight line segment whose error is less than a given threshold, thereby retaining key shape points; for rectangular shapes, introduce a quadratic programming algorithm QP to regularize the bottom contour line, use quadratic rules to globally adjust the vertex offset and angle offset, thereby ensuring the orthogonal relationship between the contour edges.

[0018] As preferred, in step 2, the complete steps of facade geometry-based door and window contour extraction are as follows:

[0019] 2.1) Facade planar image projection: First, project the segmented facade planar point cloud into a two-dimensional point set using formula (1); second, based on the range of the two-dimensional point set and the given resolution, construct a two-dimensional image; finally, based on formula (2), calculate the pixel coordinate position of each point, while expanding each point into a local neighborhood, and draw it into the two-dimensional image, thereby enhancing the density contrast between the opening and the wall;

[0020] P' = R · (P - P c ) (1)

[0021]

[0022] where R is a transformation matrix composed of three basis vectors obtained using principal component analysis, P' is the projected point coordinate, P is the point coordinate before projection, P c is the center coordinate of the facade planar point set, p is the pixel coordinate of the two-dimensional point, P' min is the minimum point coordinate of the two-dimensional point set, r is the image resolution, is the floor function;

[0023] 2.2) Data processing and model training: First, based on the projected two-dimensional image, construct a door and window bounding box detection dataset through manual annotation; second, use data augmentation operations to expand the dataset sample distribution; finally, use the dataset to train the YOLO target detection model;

[0024] 2.3) Door and window opening detection: Firstly, based on the trained model, predict the door and window opening bounding boxes of each facade image of the building; Secondly, use non-maximum suppression to remove the boxes with low confidence in the bounding boxes, and take the final bounding boxes as the door and window contours;

[0025] 2.4) Door and window contour re-projection: Calculate the pixel coordinates p of the four corner points of each door and window bounding box, and project the contour from the pixel coordinates back to the three-dimensional space based on formula (3);

[0026] P = R T ·(p*r+P' min )+P c (3).

[0027] As preferred, in step 3, specifically comprising: three-dimensional building body frame model reconstruction considering roof structure, and facade detail enhancement based on door and window contours.

[0028] As preferred, the three-dimensional building body frame model reconstruction considering roof structure comprises the following steps:

[0029] 3.11) Reconstruct the three-dimensional body frame model of the building using City3D: Firstly, input the extracted building facade bottom contour and segmented roof plane point cloud into the City3D system, and the system automatically generates vertical wall surface hypotheses and roof candidate facets; Secondly, construct an energy function, optimize the selection of facets and combine them with the wall to obtain an initial polyhedral body frame model;

[0030] 3.12) Water-tightness repair: Firstly, use edge stitching algorithm to merge the broken edge vertices of the simulated surface, and splice the surfaces together; Secondly, detect the holes existing after splicing, and use hole repair algorithm to fill the hole area; Finally, unify the normal direction and remove the non-manifold elements to obtain a water-tight building model that meets the component embedding requirements.

[0031] As preferred, the facade detail enhancement based on door and window contours comprises the following steps:

[0032] 3.21) Project the door and window contours to the wall surface: Firstly, calculate the normal vector of the plane where the door and window contours are located; Then, calculate the projection points of the contour corner vertices on the model wall surface along the normal vector direction, and remove the door and window contours that cannot be completely projected to the wall;

[0033] 3.22) Construct door and window entities: Extrude the door and window contours in the normal positive and negative directions by a certain thickness to generate door and window entities that conform to the wall thickness constraints;

[0034] 3.23) Door and window entity mosaic: doors and windows include two types of convex and concave; if the type is a concave door and window, the door and window entity is removed from the three-dimensional building model by using Boolean difference operation; if the type is a convex door and window, the three-dimensional building model is fused with the door and window entity by using Boolean union operation.

[0035] The beneficial effects of the present application are as follows:

[0036] The present application aims at the problems of insufficient utilization of facade information, distortion of bottom contour, and missing of facade details such as doors and windows in the existing ALS point cloud-based building three-dimensional modeling method, and proposes a LOD3 level building three-dimensional model automatic reconstruction method based on ALS point cloud facade geometry, researches and constructs the complete process from data preprocessing to model generation, realizes the automatic generation of building model from geometric contour to facade semantic details under the condition of ALS data, provides an efficient and stable solution for large-scale city three-dimensional modeling based on ALS point cloud, and lays a method foundation for facade level fine modeling in digital twin, city planning and virtual simulation and other application scenarios.

[0037] In terms of bottom contour construction, the present application overcomes the problems of contour expansion and distortion caused by traditional roof projection methods in dealing with overhanging eaves or complex walls through vertical projection of facade point cloud, makes up for the defects of insufficient fitting degree of complex contour based on facade projection method and excessive dependence on labeled data, and effectively improves the geometric accuracy of the bottom contour.

[0038] In terms of door and window contour extraction, the present application proposes a method combining point cloud local projection and YOLO target detection, realizes accurate identification and boundary recovery of door and window openings under the condition of ALS point cloud. Even in the case of sparse or local occlusion of point cloud, it still maintains good spatial consistency, and has high stability and reliability in facade detail enhancement.

[0039] In terms of three-dimensional model generation, the present application is based on City3D framework, fuses bottom contour, roof plane and door and window contour information, proposes a preprocessing strategy based on ray projection, and realizes automatic construction of building model with accurate geometric structure, complete topological structure and rich semantic details, and watertight repair by using energy optimization and Boolean operation.

[0040] In addition, the present application achieves good results in modeling efficiency, overall model quality and facade semantic component feature expression. Among them, the average modeling time is kept at the minute level, the automation efficiency is high; the average deviation of the generated model and the artificial modeling product is small (the average RMSE is 0.42 m), and the facade door and window opening recognition accuracy is high, the false detection rate is low and the geometric precision is high, compared with the traditional modeling idea relying on roof projection, the structural accuracy and semantic expression ability of the model are significantly improved, the real geometric shape and facade features of the building can be better restored, and the present application has high practical value and application potential. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The technical flowchart of the present application is shown in the figure;

[0042] Fig. 2(a) is a schematic diagram of building plane point cloud segmentation result;

[0043] Fig. 2(b) is a schematic diagram of building facade plane point cloud;

[0044] Fig. 2(c) is a schematic diagram of building roof plane point cloud;

[0045] Fig. 3(a) is a two-dimensional scatter plot of building facade point cloud;

[0046] Fig. 3(b) is a schematic diagram of optimal Alpha Shape building bottom contour;

[0047] Fig. 3(c) is a schematic diagram of simplified building bottom contour;

[0048] Fig. 4(a) is a schematic diagram of facade plane image projection;

[0049] Fig. 4(b) is a schematic diagram of door and window opening detection;

[0050] Fig. 4(c) is a schematic diagram of door and window contour re-projection;

[0051] Figure 5 Fig. 5 is a schematic diagram of three-dimensional building body frame model reconstruction considering roof structure;

[0052] Fig. 6(a) is a schematic diagram of door and window contour projection to wall surface;

[0053] Fig. 6(b) is a schematic diagram of constructing door and window entity;

[0054] Fig. 6(c) is a schematic diagram of door and window entity inlay;

[0055] Figure 7 Fig. 7 is a schematic diagram of generated model;

[0056] Figure 8 Fig. 8 is a schematic diagram of artificial modeling result;

[0057] Figure 9This is a schematic diagram showing the distribution of RMSE and MAE of model errors;

[0058] Figure 10 This is a diagram illustrating the error visualization results. Detailed Implementation

[0059] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0060] like Figure 1 As shown, this embodiment provides an automatic reconstruction method for a LOD3-level building 3D model based on ALS point cloud facade geometry, which includes the following steps:

[0061] Step 1: Building point cloud segmentation and bottom contour construction: Using normal analysis and plane fitting, the roof and facade areas are segmented from the ALS point cloud. Then, the actual grounded building bottom contour is automatically extracted by the vertical projection of the facade point cloud, avoiding the deviation caused by the roof projection in traditional methods.

[0062] Building plan point cloud segmentation and elevation extraction

[0063] Accurate segmentation of building rooftops and planar facades in ALS point clouds is a crucial prerequisite for subsequent tasks such as refined extraction of the building's base contour and identification of detailed structures like doors and windows. This mainly involves two steps: 1.11) planar point cloud segmentation and 1.12) planar point cloud classification. Accurate planar segmentation of the building within the ALS point cloud is the key.

[0064] Existing segmentation methods can be divided into two categories: data-driven methods and model-driven methods. Data-driven methods utilize the spatial relationships and topological similarity of point clouds to segment or cluster planes, but they are only applicable to dense, noise-free point clouds. Model-driven methods, on the other hand, find the optimal model parameters in the point cloud through voting or sampling, making them more adaptable to point cloud noise and non-uniformity. Random Consistency Sampling (RANSAC) is a typical model-driven method. It uses iterative random sampling and local model estimation to find the support plane containing the largest set of interior points in the point cloud. Compared to other model-driven methods such as Hough Transform, RANSAC has stronger noise resistance and robustness, and is therefore widely used in low-density or outlier-rich building point cloud processing. Therefore, this embodiment uses RANSAC to segment the planes of ALS building point clouds to improve the stability and accuracy of extracting geometric structures from noisy and structurally incomplete ALS building point cloud data. The specific steps are as follows:

[0065] 1.11) Plane point cloud segmentation: Firstly, RANSAC algorithm is used to segment the building point cloud; secondly, based on the normal parameters and center points of the fitted plane, the normal angle and center point distance between planes are calculated; finally, the plane point cloud with an angle and distance less than the threshold is merged; finally, the principal component analysis (PCA) method is used to re-estimate the parameters of the merged plane; the segmentation result is shown in Fig. 2(a).

[0066] 1.12) Plane point cloud classification: Firstly, the normal vector of each plane is calculated using the plane point cloud parameters; then, the angle between the plane normal vector and the Z axis is calculated, and if the normal vector is perpendicular to the Z axis, it is classified as a facade point cloud, otherwise it is classified as a roof point cloud. The final extracted facade plane and roof plane will be used for subsequent geometric modeling, and the recognition result is shown in Fig. 2(b) and Fig. 2(c).

[0067] Automatic construction of building bottom contour based on facade point cloud projection

[0068] The bottom contour of a building reflects its actual land occupation and overall geometric shape. Building bottom contour construction based on facade point cloud is beneficial to improve the accuracy of building three-dimensional expression. Based on the obtained facade point cloud, this embodiment proposes an automatic extraction method of building bottom contour based on ALS building facade point cloud, which mainly includes three steps of ground point projection, bottom contour automatic detection and contour simplification, and the contour detection is the most critical step.

[0069] Currently, traditional methods without deep learning are still the main method for extracting building bottom contour, which can be divided into two categories: structure detection-based method and edge tracking-based method. The structure detection-based method detects straight lines from the ground projection points and constructs closed shapes through hypothesis-selection strategy, but cannot handle non-straight line structures. The edge tracking-based method automatically extracts point cloud contours of arbitrary shape by sequentially tracking the edge point set of the ground projection point cloud. Among them, Alpha-Shape is the most commonly used edge tracking algorithm, which removes the long edges in the Delaunay triangulation of the projection point cloud by setting the edge length threshold α, thereby obtaining the contour consistent with the point cloud distribution. However, the problem of setting α value is still an important challenge faced by this method.

[0070] Since Alpha-Shape extracts building contour is essentially a convex hull approximation process, that is, when α changes from 0 to infinity, the edge contour changes from Multi-Polygon to Polygon and finally to Convex Hull. The ideal α is located between the change from Multi-Polygon to Polygon, and at this time the contour just contains all the points. Therefore, this embodiment proposes a binary approximation strategy based on the given range parameter (αmin ,α max ) and tolerance parameter a ∈ , a bisection recursion method is used to search a, so that when the parameter is a, a complete polygon is obtained, and when the parameter is a-a ∈ , a multi-polygon is obtained.

[0071] The specific steps of automatically extracting the building bottom contour are as follows:

[0072] 1.21) Project the facade point cloud to the ground: first, orthogonally project the building facade point cloud to the ground to generate a two-dimensional point set; second, sample the projected point set based on a regular grid to reduce the redundancy of the points and maintain the overall spatial distribution characteristics, that is, divide the projected point set according to fixed-size grid cells, and only keep one representative point (such as the centroid or a random point) in each grid cell. Thus, a simplified two-dimensional scatter plot is obtained, as shown in FIG. 3(a).

[0073] 1.22) Bottom contour extraction based on bisection approximation: first, construct the Delaunay triangulation of the two-dimensional ground projection point cloud; second, search for the optimal a value in combination with the bisection recursion strategy, and extract the corresponding contour. If the contour is a multi-polygon, update a min , otherwise update a max ; until a max -a min >a ∈ , stop recursion and return the threshold value and the contour, as shown in FIG. 3(b).

[0074] 1.23) Contour simplification: Since noisy point clouds can cause the contour to have jagged redundant features, it is necessary to simplify these redundant features. This embodiment uses the Ramer-Douglas-Peucker (RDP) algorithm to simplify the contour line. The core idea of this algorithm is to replace curves with small errors with straight line segments, thereby retaining key shape points. For rectangular shapes, to ensure orthogonality, a quadratic programming algorithm QP is introduced to regularize the bottom contour line. The basic idea of this algorithm is to use quadratic programming to globally adjust the vertex offset and angle offset, thereby ensuring the orthogonal relationship between the contour edges. The simplified result is shown in FIG. 3(c).

[0075] Step 2, facade geometry-based door and window contour extraction: project the facade point cloud into a two-dimensional image, use AlphaShape to construct and geometric-semantic detection method to identify door and window openings, extract door and window contours, and realize geometric acquisition of key facade semantic components.

[0076] Door and window extraction from building facade is one of the important steps to construct a fine-grained 3D building model. In ALS point cloud, doors and windows usually appear as hole regions on the facade. The main steps of door and window extraction include facade image projection, door and window contour recognition, and contour coordinate back-projection. The key to extracting door and window contour is to accurately identify the door and window hole.

[0077] The existing facade door and window extraction methods can be divided into traditional methods based on geometry and rules and learning-driven methods based on deep learning. The traditional method based on geometry and rules relies on the geometric features such as curvature of point cloud or the rules such as straight line edge of door and window to identify the door and window, but this kind of method is only designed for high-density complete point cloud. The learning-driven method directly learns geometric and semantic features from data through deep neural network to realize door and window target detection and semantic segmentation. However, the door and window point cloud segmentation model is usually designed for dense point cloud of ground scanning, and it is difficult to extract the boundary of window for incomplete door and window and rough edge of door and window hole. Therefore, the bounding box of target detection is used as the representative contour of door and window position and size in the embodiment.

[0078] For the problem of ALS point cloud facade door and window detection, the embodiment uses You Only Look Once (YOLO) model to realize the target detection of door and window bounding box. YOLO is a target detection and image segmentation model based on attention mechanism, which has the advantages of fast and high precision, and is widely used in various detection tasks. After projecting the facade point cloud to a binary image, the YOLO model is used to identify the hole region in the image. In the process of projecting the facade point cloud to the binary image, due to the sparsity and noise problem of ALS point cloud, direct single-point projection to the image will make the point pixel too sparse, which leads to the inability to distinguish the door and window from the background. To solve this problem, the embodiment proposes to "inflate" each point into a local neighborhood, that is, draw a circular point with a specified radius in the two-dimensional image, so as to enhance the density contrast between the opening and the wall and improve the separability of subsequent detection. The complete extraction steps are as follows:

[0079] 2.1) Facade plan image projection: first, project the segmented facade plan point cloud to a two-dimensional point set using formula (1); second, based on the range of two-dimensional point set and given resolution, construct a two-dimensional image; finally, calculate the pixel coordinate position of each point based on formula (2), and at the same time, to alleviate the problem of "point pixel" over-sparse caused by ALS sparsity, "inflate" each point into a local neighborhood and draw it to the two-dimensional image (Fig. 4(a)), so as to enhance the density contrast between the opening and the wall and improve the separability of subsequent detection;

[0080] P' = R · (P - P c ) (1)

[0081]

[0082] where R is a transformation matrix composed of three base vectors obtained using principal component analysis, P' is the projected point coordinate, P is the unprojected point coordinate, P c is the center coordinate of the facade plane point set, p is the pixel coordinate of the two-dimensional point, P' min is the minimum point coordinate of the two-dimensional point set, r is the image resolution, is the floor function.

[0083] 2.2) Data processing and model training: First, a door and window bounding box detection dataset is constructed based on the projected two-dimensional image through manual annotation; second, the dataset sample distribution is expanded using data augmentation operations; finally, the YOLO target detection model is trained using the dataset;

[0084] 2.3) Door and window opening detection: First, based on the trained model, the door and window opening detection frame of each facade planar image of the building is predicted; second, to avoid detection frame covering caused by false detection, non-maximum suppression processing is used to remove the frame with low confidence in the covered detection frame, and the final detection frame is taken as the door and window contour (Fig. 4(b));

[0085] 2.4) Door and window contour re-projection: Calculate the pixel coordinates p of the four corner points of each door and window bounding box, project the contour from the pixel coordinates back to the three-dimensional space based on formula (3), and the final result is shown in Fig. 4(c).

[0086] P = R T · (p * r + P' min ) + P c (3)

[0087] Step 3, facade detail enhanced three-dimensional building model construction: input the extracted bottom contour and door and window contour into the model-driven modeling framework, automatically embed door and window components through polygon patch combination and Boolean operation, and generate a building model with complete structure and semantic details.

[0088] This step is to unify the extracted bottom contour and door and window features with the roof plane obtained by segmentation to generate a watertight three-dimensional building model that can be used for component embedding and subsequent analysis. It mainly includes two steps: three-dimensional building body frame model reconstruction, and facade detail enhancement based on door and window contour.

[0089] Three-dimensional building body frame model reconstruction

[0090] The existing ALS point cloud-based building three-dimensional reconstruction methods are mainly divided into two categories: data-driven methods and model-driven methods. The data-driven methods focus on reconstructing the building surface by using the geometric and topological relationship between points, but they are sensitive to noise. In contrast, the model-driven methods are more robust to noise by identifying the structural primitives of point cloud. City3D is a model-driven method specifically for ALS point cloud. The method constructs candidate facet primitives based on the given building bottom contour and roof point cloud, and then proposes an energy optimization primitive selection framework with roof bias constraint based on the hypothesis-selection strategy, which can effectively reconstruct the roof details and flat walls, and finally obtain the polygonal mesh surface model of the building.

[0091] However, since the use of door and window contours enhances the facade details, the three-dimensional model needs to maintain water tightness. However, due to the floating-point error problem of City3D, the surface mesh of the model constructed by City3D has a fracture problem, and the fractured edges need to be stitched to ensure the water tightness of the model and facilitate the subsequent reconstruction of facade details. The edge stitching function and hole repair function of the computational geometry library (CGAL) are used to realize the water tightness repair of the mesh. CGAL is a high-precision computational geometry algorithm library that contains a large number of mesh-related algorithms. Edge stitching and hole repair are two important functions. The edge stitching is based on the similarity structure of the mesh edges to merge adjacent vertices, realizing the splicing of the polygonal mesh facets. The hole repair is aimed at the hole structure existing in the mesh surface, and the missing area is filled by triangulation.

[0092] Based on the above reasons, the method of City3D is used to construct a water-tight three-dimensional building block model, which mainly includes two steps: first, reconstructing a three-dimensional building block model using City3D; second, repairing the water tightness of the model. The specific steps are as follows:

[0093] 3.11) Reconstruct the three-dimensional building block model using City3D: first, input the extracted building facade bottom contour and segmented roof plane point cloud into the City3D system, and the system automatically generates vertical wall hypotheses and roof candidate facets; second, construct an energy function (point-to-plane consistency, normal / curvature smoothness, ridge / eave continuity, etc.), optimize the selection of facets and combine them with the wall to obtain an initial polyhedral block model;

[0094] 3.12) Water tightness repair: first, use the edge stitching algorithm to merge the fractured edge vertices of the simulated surface, and splice the surface together; second, detect the holes existing after splicing, and use the hole repair algorithm to fill the hole area; finally, unify the normal direction and remove the non-manifold elements to obtain a water-tight building model that meets the component embedding requirements. The final reconstruction result is shown in Figure 5 .

[0095] Facade detail enhancement based on door and window contour

[0096] Facade detail enhancement is a key step to promote the construction of 3D building model from LOD2 level to LOD3 level, which has strong practical value in various application scenarios. The core of facade detail enhancement is to embed door and window structure into the facade of building body frame model. The entity embedding method based on 3D Boolean operation is widely used in 3D modeling. The method splits entities by calculating the intersection boundary of the surface between entities, and selects the retained entity components according to the Boolean relationship between entities, and then composes a new entity. In this embodiment, the mesh extrusion and Boolean operation functions provided by CGAL are used to achieve the embedding task of door and window entity and building model entity through constructing door and window entity and Boolean operation.

[0097] However, due to the interference of ALS point cloud error, there may be deviations between the extracted door and window contour and the reconstructed facade surface structure of building body frame model, which will lead to the non-parallelism of embedded door and window and wall. Therefore, this embodiment proposes a door and window correction preprocessing step based on ray projection. The basic idea is to project the door and window contour along its normal vector to the model facade to realize the alignment of door and window and facade. Therefore, this embodiment designs a door and window embedding process based on 3D Boolean operation, including the following three steps:

[0098] 3.21) Project the door and window contour to the wall: first, calculate the normal vector of the plane where the door and window contour is located; then, calculate the projection point of the contour angular vertex on the model wall surface along the normal vector direction, and remove the door and window contour that cannot be completely projected to the wall. The schematic diagram is shown in Fig. 6(a).

[0099] 3.22) Construct door and window entity: extrude the door and window contour along the normal positive and negative directions by a certain thickness to generate a door and window entity that meets the wall thickness constraint (Fig. 6(b)). In this embodiment, the extrusion length refers to the "Code for Design of Masonry Structures" (GB50003-2011) to meet the actual building wall thickness standard.

[0100] 3.23) Door and window entity embedding: door and window include two types of external convex and internal concave; if the type is internal concave door and window, use Boolean difference operation to remove the door and window entity from the 3D building model; if the type is external convex door and window, use Boolean union operation to fuse the 3D building model and the door and window entity. Finally, generate the 3D building model of facade detail enhancement (Fig. 6(c)).

[0101] Experimental results

[0102] To comprehensively verify the effectiveness and applicability of the automatic reconstruction method proposed in this embodiment, this embodiment selects 10 representative urban buildings as experimental objects, covering regular shapes, multiple roof structures, and various types of buildings with door and window accessories. The average point cloud density of the ALS point cloud of the 10 buildings used in this embodiment is 132 points / m2 The facade enhanced 3D models of 10 buildings are reconstructed by the method proposed in this embodiment. Figure 7

[0103] To evaluate the reconstruction efficiency of the method, the average time consumption and total time consumption of the feature extraction stage and the enhanced modeling stage in the reconstruction of 10 buildings are counted, and the experimental environment is shown in Table 1. The statistical results are shown in Table 2. The statistical results show that the method exhibits good execution speed, and the average total time of modeling of 10 buildings is 91.16 seconds, in which the average total time consumption of the feature extraction stage is 2 seconds, and the modeling stage is about 89.16 seconds. This shows that the method has high efficiency in feature extraction, and the overall modeling speed is controlled in the minute level.

[0104] Table 1 Experimental environment table

[0105]

[0106] Table 2 Average time consumption table of each stage of modeling

[0107] Stage Feature extraction stage (s) Model production stage (s) Total (s) Average time cost 2 89.16 91.16

[0108] To evaluate the quality of the reconstructed model, this embodiment compares the artificial modeling results ( Figure 8 ) as a benchmark, analyzes from two aspects of overall structure quality and facade information expression, and quantitatively evaluates the geometric accuracy and semantic recognition ability of the model by using multiple indexes.

[0109] (1) Overall model quality analysis

[0110] To evaluate the overall model quality of the reconstructed model, this embodiment calculates the point-to-point distance error of the reconstructed model to the artificial model, and also counts the average absolute error (MAE) and the root mean square error (RMSE) of each building, and calculates the average MAE and the average RMSE of the two indexes under all 10 buildings, and the statistical results are shown in Figure 9 , and the error visualization results are shown in Figure 10 .

[0111] The statistical results ( Figure 9 ) show that the overall reconstruction quality of the 10 buildings shows certain difference. Overall, the RMSE of each building is higher than the MAE, in which the average RMSE is 0.42m, and the average MAE is 0.33m. The MAE reflects the overall offset of the model, and the value remains in the order of 0.3m, which shows that the reconstructed model is highly consistent with the artificial model in most positions, and only a small average geometric offset exists. The RMSE is more sensitive to local large error, and the average value is 0.42m, which is slightly higher than the MAE, which reveals that there is a larger local error in some complex areas.​

[0112] Specifically, the RMSE of most buildings is close to the MAE, and the error distribution is concentrated and stable; only a few buildings (such as Buildings 3, 5, and 9) have slightly higher RMSE, indicating that these models have deviations in local areas. From the rendering results of the ALS data, it can be seen that most of the building surfaces are dominated by blue areas, indicating that most of the vertex offsets are below 0.6 m. It is worth noting that these deviations mainly come from the extraction of complex roof shapes, and the reconstruction of bottom and facade details still maintains high accuracy and stability. This study only uses conventional segmentation and fitting methods for roof modeling and does not perform special optimization on complex roofs. Figure 9

[0113] (2) Facade information quality evaluation

[0114] To evaluate the recognition and reconstruction of door and window components, the following four commonly used indicators are used: Precision, Recall, F1, and mIoU, which reflect the correctness of classification and spatial geometric consistency, respectively. The experimental results are as follows:

[0115] The average Precision is 0.97, the Recall is 0.98, and the F1 Score is 0.98, with a standard deviation of about 0.03. This indicates that the method not only has high accuracy in door and window recognition, but also has strong stability. This means that false positives and false negatives are rare, and doors and windows can be almost completely and correctly identified;

[0116] In comparison, the average mIoU is 0.83, which is slightly lower than the classification indicators, but still at a high level, showing that the predicted results and the true contours have good consistency in geometric position and scale. Even in the case of sparse point clouds or incomplete RANSAC segmentation, the minimum mIoU remains at 0.74, indicating that even if the data quality decreases, the model can still maintain reliable door and window positioning capabilities. The standard deviation of about 0.03 further indicates that the method has strong robustness under different building types and point cloud conditions.

[0117] The above evaluation results can show that the method of the embodiment can effectively maintain geometric consistency with the artificial model at the overall structure level, with an average deviation of 0.3-0.4 m in a reasonable range; at the facade information level, the detection accuracy of door and window components is close to perfect, while maintaining high positioning consistency. This indicates that the method can achieve efficient and accurate three-dimensional reconstruction under ALS data conditions, and exhibits good reliability in detail expression.

[0118] ​This embodiment proposes a fine three-dimensional building model automatic reconstruction strategy based on facade geometric information. The strategy takes ALS facade point cloud as the core, and realizes the complete expression of the model from structure to detail by using its geometric and semantic characteristics. In this strategy, the complete process from data preprocessing to model generation is as follows: first, the accurate separation of roof and facade is realized based on RANSAC plane segmentation and classification, and the real bottom contour is extracted through facade point cloud orthogonal projection and Alpha Shape binary search; secondly, in the facade point cloud parallel projection image, combined with disc rendering and YOLO deep detection model, the robust recognition and back projection reconstruction of door and window contour are realized; finally, based on City3D framework and CGAL geometric operation library, the body frame model construction, watertight repair and door and window Boolean operation embedding are completed, and the building model closed and containing facade semantic details is generated. The above three-step strategy not only overcomes the geometric distortion problem based on roof projection, but also avoids the dependence on external image or database, providing a feasible path for large-scale automatic and fine modeling under ALS point cloud condition.

[0119] The above description of the present application and its embodiments is illustrative, not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the present application, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it, without departing from the purpose of the present application, similar structure and embodiments can be designed without creative design, which should belong to the protection scope of the present application.

Claims

1. A method for automatic reconstruction of 3D building models of LOD3 level based on the facade geometry of ALS point clouds, characterized in that: The method comprises the following steps: Step 1, building point cloud segmentation and bottom contour construction: the roof and facade area are segmented from the ALS point cloud by using normal analysis and plane fitting, and then the bottom contour of the real ground building is automatically extracted by vertical projection of the facade point cloud; Step 2, door and window contour extraction based on facade geometry: the facade point cloud is projected into a two-dimensional image, the door and window openings are identified by using Alpha Shape construction and geometric-semantic detection method, and the door and window contour is extracted to realize the geometric acquisition of key facade semantic components; Step 3, three-dimensional building model construction with facade detail enhancement: the extracted bottom contour and door and window contour are input into the model driven modeling framework, the door and window components are automatically embedded by polygonal face combination and Boolean operation, and the building model with complete structure and semantic details is generated.

2. The method of claim 1, wherein the method further comprises: determining a building footprint of the building based on the ALS point cloud; and determining a building height of the building based on the ALS point cloud. In step 1, the planes of the ALS building point cloud are segmented by RANSAC, and the specific steps are as follows: 1.11) Plane point cloud segmentation: firstly, the plane of the building point cloud is segmented by RANSAC algorithm; secondly, based on the normal parameters and center point of the fitted plane, the normal angle and center point distance between planes are calculated; finally, the plane point cloud with an angle and distance less than a threshold value is merged; finally, the principal component analysis (PCA) method is used to re-estimate the parameters of the merged plane; 1.12) Plane point cloud classification: firstly, the normal vector of each plane is calculated using the plane point cloud parameters; then, the angle between the plane normal vector and the Z-axis is calculated, and if the normal vector is perpendicular to the Z-axis, it is classified as a facade point cloud, otherwise it is classified as a roof point cloud.

3. The method of claim 2, wherein the method further comprises: determining a building footprint of the building based on the ALS point cloud; and determining a building height of the building based on the ALS point cloud. In step 1, the specific steps of automatically extracting the building bottom contour are as follows: 1.21) Project the facade point cloud onto the ground: firstly, the building facade point cloud is orthogonally projected onto the ground to generate a two-dimensional point set; secondly, based on the regular grid, the projected point set is sampled to reduce the redundancy of the points and maintain the overall spatial distribution characteristics, that is, the projected point set is divided into fixed size grid cells, and only one representative point is retained in each grid cell; thereby obtaining a simplified two-dimensional scatter plot; 1.22) Bottom contour extraction based on bisection approximation: First, construct the Delaunay triangulation of the 2D ground projection point cloud; second, search for the optimal α value combined with the bisection recursion strategy, and extract the corresponding contour; if the contour is a multi-polygon, update α min , otherwise update α max ; until α max -α min >α ∈ , stop recursion and return the threshold and contour; α ∈ is the tolerance parameter; 1.23) Contour simplification: the RDP algorithm is used to simplify the contour line, and the curve with an error less than a given threshold is replaced by a straight line segment, thereby retaining the key shape points; for rectangular shapes, a quadratic programming (QP) algorithm is introduced to regularize the bottom contour line, and a quadratic rule is used to globally adjust the vertex offset and angle offset, thereby ensuring the orthogonal relationship between the contour edges.

4. The method of claim 3, wherein the method further comprises: determining a building footprint of the building based on the ALS point cloud; and determining a building height of the building based on the ALS point cloud. In step 2, the complete steps of door and window contour extraction based on facade geometry are as follows: 2.1) Facade planar image projection: firstly, the segmented facade planar point cloud is projected into a two-dimensional point set using formula (1); secondly, based on the two-dimensional point set range and the given resolution, a two-dimensional image is constructed; finally, the pixel coordinate position of each point is calculated based on formula (2), and each point is inflated into a local neighborhood and drawn into the two-dimensional image, thereby enhancing the density contrast between the opening and the wall. P' = R - (P - P c ) (1) Wherein, R is a transformation matrix composed of three base vectors obtained by using principal component analysis method, P' is the coordinate of the projected point, P is the coordinate of the point before projection, P c is the center coordinate of the facade plane point set, p is the pixel coordinate of the two-dimensional point, P ' min is the minimum point coordinate of the two-dimensional point set, r is the image resolution, is the down rounding symbol; 2.2) Data processing and model training: Firstly, a door and window bounding box detection dataset is constructed based on the projected two-dimensional images through manual annotation. Secondly, the dataset sample distribution is expanded using data augmentation operations. Finally, the YOLO target detection model is trained using the dataset; 2.3) Door and window opening detection: Firstly, based on the trained model, the door and window opening detection box of each facade plan image of the building is predicted. Secondly, non-maximum suppression processing is used to remove the boxes with low confidence in the covered detection box, and the final detection box is taken as the door and window contour; 2.4) Door and window contour re-projection: Calculate the pixel coordinates p of the four corner points of each door and window bounding box, and project the contour from the pixel coordinates back to the three-dimensional space based on formula (3); P = R T • (p * r + P' (p * r + P' min ) + P c (3).

5. The method of claim 4, wherein the method further comprises: determining a building footprint of the building based on the ALS point cloud; and determining a building height of the building based on the ALS point cloud. In step 3, it specifically includes: three-dimensional building body frame model reconstruction considering roof structure, and facade detail enhancement based on door and window contour.

6. The method of claim 5, wherein the method further comprises: determining a building footprint of the building based on the ALS point cloud; and determining a building height of the building based on the ALS point cloud. The three-dimensional building body frame model reconstruction considering roof structure includes the following steps: 3.11) Reconstruct the three-dimensional body frame model of the building using City3D: Firstly, input the extracted building facade bottom contour and segmented roof plane point cloud into the City3D system, and the system automatically generates vertical wall surface hypotheses and roof candidate facets; Secondly, construct an energy function, optimize the selection of facets and combine them with the wall to obtain an initial polyhedral body frame model; 3.12) Water-tightness repair: Firstly, use the edge stitching algorithm to merge the broken edge vertices of the simulated surface to splice the surface together; Secondly, detect the holes existing after splicing and use the hole repair algorithm to fill the hole area; Finally, unify the normal direction and remove the non-manifold elements to obtain a water-tight building model that meets the component embedding requirements.

7. The method of claim 6, wherein the method further comprises: determining a building footprint of the building based on the ALS point cloud; and determining a building height of the building based on the ALS point cloud. The facade detail enhancement based on door and window contour includes the following steps: 3.21) Project the door and window contour to the wall: Firstly, calculate the normal vector of the plane where the door and window contour is located; Then, calculate the projection points of the contour corner vertices on the model wall surface along the normal vector direction, and remove the door and window contours that cannot be completely projected to the wall; 3.22) Construct door and window entity: Extrude the door and window contour in the normal positive and negative directions by a certain thickness to generate a door and window entity that conforms to the wall thickness constraint; 3.23) Door and window entity inlay: Doors and windows include two types: outward convex and inward concave. If the type is inward concave door and window, use Boolean difference operation to remove the door and window entity from the three-dimensional building model. If the type is outward convex door and window, use Boolean union operation to fuse the three-dimensional building model with the door and window entity.

Images