A point cloud building component modeling method and system based on feature extraction

Abstract

The present invention relates to the technical field of three-dimensional modeling in engineering measurement, and particularly to a point cloud building component modeling method and system based on feature extraction. The method comprises the following steps: obtaining three-dimensional point cloud data of building components, constructing a multi-scale local neighborhood with each point as the center, calculating a covariance matrix, decomposing eigenvalues, generating an edge probability map, semantically outputting corresponding component categories, and screening corresponding key feature point sets; constructing a three-level B-spline surface model with the feature point cloud as control points, and adjusting the corresponding spatial distribution density; forming a weighted complete graph through the key feature point set, identifying and processing non-planar areas using the Kuratowski theorem, applying a QEM algorithm to lightweight the three-dimensional model, automatically extracting key feature points of the building components for screening and optimization, and constructing a concise and accurate lightweight model based on feature point topology optimization and graph theory algorithms, thereby improving the automation and efficiency of modeling and effectively resolving the contradiction between model lightweighting and accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional modeling in engineering measurement, and in particular to a point cloud building component modeling method and system based on feature extraction. Background Art

[0002] Currently, laser point cloud technology has been widely used in fields such as building component modeling and three-dimensional reconstruction. It can quickly obtain spatial geometric information of building entities and has become a core data source for engineering applications such as building information modeling (BIM), digital twins, and urban modeling. The high-density three-dimensional point cloud obtained by laser scanning provides rich data support for the digital representation of building components. However, in the actual modeling process, most existing methods rely on manual modeling or semi-automatic modeling methods, and still have many problems. For example, the modeling process is highly dependent on manual operation, the operation process is cumbersome, the overall efficiency is low, and the accuracy and consistency of the component results are difficult to guarantee. At the same time, the lack of standardized feature point extraction and recognition mechanisms makes the modeling results prone to structural misjudgment, blurred boundaries and other problems.

[0003] In recent years, to improve modeling automation, academia and industry have gradually introduced deep learning-based point cloud recognition technology. By fusing multi-source data (such as RGB-D images and laser point clouds) and derived features (such as normal vectors and semantic labels), this technology significantly improves the classification and recognition accuracy of building components. Although these methods have made some progress in component segmentation and semantic labeling, key bottlenecks remain in the structural modeling of feature points: ① The lack of systematic analysis and optimization of the topological structure between feature points leads to incomplete modeling structures and imprecise geometric representation; ② The presence of a large amount of redundant data or invalid point connections in the model affects the lightweight representation of the overall model; the inability to achieve differentiated modeling control based on component complexity and semantic importance not only affects reconstruction efficiency but also increases subsequent processing costs; ③ The lack of a universal lightweight modeling mechanism makes the model difficult to directly apply in practical scenarios such as BIM systems, GIS platforms, 3D visualization, and operation and maintenance analysis. Summary of the Invention

[0004] In response to the problems existing in the existing technology, such as incomplete topological structure, serious model redundancy and lack of lightweight control, the present invention proposes a point cloud building component modeling method and system based on feature extraction. It integrates deep learning and geometric analysis to realize the automatic extraction of key feature points of building components, and combines adaptive reconstruction and multi-scale detail enhancement technology to adjust the model granularity and structural accuracy as needed; further, through the topological optimization of feature points and graph theory algorithms, spatial connection relationships are constructed to generate a lightweight model with simple structure and accurate expression, effectively improving the degree of automation and application efficiency of building component modeling, and is suitable for engineering scenarios such as BIM, GIS and 3D visualization.

[0005] The present invention provides a point cloud building component modeling method based on feature extraction, comprising the following steps:

[0006] Perform laser scanning on the target building component area to obtain 3D point cloud data of the building components and perform data preprocessing;

[0007] Based on the preprocessed 3D point cloud data, a multi-scale local neighborhood is constructed with each point as the center. The covariance matrix of the point set in the local neighborhood at each scale is calculated and the eigenvalues ​​are decomposed. The normalized curvature index is extracted and normalized using the Sigmoid function to generate an edge probability map for screening candidate edge feature points. The improved PointNet++ network is used for semantic segmentation, and the candidate edge feature points are output to the corresponding component category. The corresponding key feature point set is then screened based on the component category.

[0008] Using the feature point cloud of key feature points as control points, a three-level B-spline surface model is constructed to fit the geometric shape of the target component. The density adjustment function is set according to the importance of each feature point cloud in a preset local neighborhood to dynamically adjust the corresponding spatial distribution density. This is used to adaptively reconstruct the three-dimensional model of the target component and introduce the Laplace deformation algorithm for detail enhancement.

[0009] Based on the adaptively reconstructed 3D model, a weighted complete graph is formed through a set of key feature points. The improved Kruskal algorithm is used to generate a minimum spanning tree to determine the preliminary connection relationship. The simulated annealing algorithm is introduced to iteratively optimize the connection relationship through the energy function. The scalability of the connection results is verified by the Kuratowski theorem. Non-planar areas are identified and processed. The QEM algorithm is applied to lightweight the 3D model to complete the 3D modeling of the target component.

[0010] Furthermore, based on the preprocessed 3D point cloud data, a multi-scale local neighborhood is constructed with each point as the center. The covariance matrix of the point set in the local neighborhood of each scale is calculated and the eigenvalues ​​are decomposed. The normalized curvature index is extracted and normalized by the Sigmoid function to generate an edge probability map for screening candidate edge feature points, including:

[0011] Calculate the covariance matrix of each point in the multi-scale local neighborhood and decompose it to obtain the eigenvalues ​​in the three-axis directions;

[0012] The normalized curvature index is calculated based on the minimum eigenvalue of the three-axis direction;

[0013] Based on the normalized curvature index of each point in the multi-scale local neighborhood, a multi-scale edge response function is constructed using weighted fusion, and normalized mapping is performed through the Sigmoid function to generate an edge probability map;

[0014] The points whose edge response values ​​in the edge probability map are greater than the preset threshold are screened to obtain candidate edge feature points.

[0015] Furthermore, the semantic segmentation candidate edge feature points are outputted through the improved PointNet++ network to obtain the corresponding component categories, and the corresponding key feature point sets are screened based on the component categories, including:

[0016] The improved PointNet++ network is used to segment the candidate edge feature points, taking the spatial coordinates, normal vectors, and RGB color information of the candidate edge feature points as input. The candidate edge feature points with semantic confidence higher than the set confidence threshold are screened, labeled with the corresponding component categories, and then output.

[0017] According to the output component category, corresponding feature enhancement is performed to extract the corresponding key feature point set.

[0018] Furthermore, corresponding feature enhancement is performed according to the output component category, and the corresponding key feature point set is extracted;

[0019] When the component type is a plane component, a plane fitting residual function is constructed based on the extracted candidate edge feature points. The optimal plane parameters are iteratively solved by the LM algorithm and the corner point detection is performed in combination with the convex hull vertex spacing of the plane projection point set. The candidate edge feature points at the structural turning position are screened to form a key feature point set.

[0020] When the component category is cylindrical, the column axis energy function is constructed, the endpoint position is optimized using the conjugate gradient method, the geometric deviation, semantic confidence and point spacing are introduced to construct a multi-constraint scoring function, and the corresponding candidate edge feature points are screened out through the branch and bound algorithm to form a key feature point subset.

[0021] Furthermore, a three-level B-spline surface model is constructed to fit the geometric shape of the target component using the feature point cloud in the key feature point set as the control point, including:

[0022] The initial control grid is constructed based on the key feature point set, and the layer-by-layer refinement mechanism is used to reconstruct the B-spline surface model. The layers include coarse-grained layer, medium-grained layer and fine-grained layer. The B-spline surface model is:

[0023] ;

[0024] in, For the model level, For the Layer, through parameter coordinates The generated B-spline surface, , is the B-spline basis function along the u and v directions, used to construct the shape control function of the surface, For the The coordinates of the control points corresponding to the i,j control grid nodes in the layer, 、 For the The number of control points in the layer along the u and v directions.

[0025] Furthermore, the density adjustment function is ;in, The number of points contained in the sphere neighborhood with a preset radius centered at (x, y, z); is the volume of the sphere neighborhood; is the curvature weight; is the semantic weight of the corresponding component category.

[0026] Furthermore, based on the adaptively reconstructed three-dimensional model, a weighted complete graph is formed through a set of key feature points. The improved Kruskal algorithm is used to generate a minimum spanning tree to determine the preliminary connection relationship. The simulated annealing algorithm is introduced to iteratively optimize the connection relationship through the energy function, including:

[0027] Map the key feature point set to a graph node, construct a weighted connection edge between any two points, and generate a weighted complete graph. The weight of the connection edge The calculation formula is as follows:

[0028]

[0029] in, 、 is the three-dimensional coordinate of the key feature point; Global average point ; is the angle between the normal vectors of adjacent facets; Semantic Differentiation Factor , is the component confidence; 、 is the weight coefficient;

[0030] The improved Kruskal algorithm is used to sort the edge sets of the weighted complete graph in ascending order, and the union-find data structure is used to monitor the connected components, generate the minimum spanning tree and retain the redundant edge sets.

[0031] A simulated annealing algorithm is introduced for the minimum spanning tree structure to iteratively optimize the connection path and output the corresponding topological connection structure.

[0032] Furthermore, the Kuratowski theorem is used to verify the scalability of the connection results and identify and handle non-planar regions, including:

[0033] Based on the optimized topological connection structure, the expandability is verified by Kuratowski theorem to identify the non-expandable non-planar areas in the topological connection structure.

[0034] The connection structure of the non-planar area is topologically segmented and reconstructed into the corresponding developable surface subgraph using NURBS surface.

[0035] Furthermore, the QEM algorithm is applied to perform lightweight processing on the 3D model to complete the 3D modeling of the target component, including:

[0036] Based on edge collapse cost function Compress non-critical areas and critical areas accordingly, and use BC7 texture compression and JPEG-XL encoding to optimize the multi-level texture storage structure; where v is the position of the vertex after edge folding, 、 is the quadratic error matrix of adjacent triangles;

[0037] The lightweight 3D model is double-verified for reconstruction accuracy and geometric integrity.

[0038] Based on the same inventive concept, the present invention also provides a point cloud building component modeling system based on feature extraction, which adopts the above-mentioned point cloud building component modeling method, including:

[0039] The data acquisition module is used to perform laser scanning on the target building component area, obtain the three-dimensional point cloud data of the building component and perform data preprocessing;

[0040] The feature extraction module is used to construct a multi-scale local neighborhood centered on each point based on the preprocessed 3D point cloud data. The covariance matrix of the point set in the local neighborhood at each scale is calculated and the eigenvalues ​​are decomposed. The normalized curvature index is extracted and normalized using the Sigmoid function to generate an edge probability map for screening candidate edge feature points. The improved PointNet++ network is used to perform semantic segmentation on the candidate edge feature points, output the corresponding component category, and then the corresponding key feature point set is screened based on the component category.

[0041] The adaptive reconstruction module is used to construct a three-level B-spline surface model to fit the geometric shape of the target component using the feature point cloud concentrated in the key feature point as the control point. The density adjustment function is set according to the importance of each feature point cloud in the preset local neighborhood to dynamically adjust the corresponding spatial distribution density. It is used to adaptively reconstruct the three-dimensional model of the target component and introduce the Laplace deformation algorithm for detail enhancement.

[0042] The lightweight processing module is used for three-dimensional models based on adaptive reconstruction. It forms a weighted complete graph through a set of key feature points, uses an improved Kruskal algorithm to generate a minimum spanning tree to determine the preliminary connection relationship, introduces a simulated annealing algorithm to iteratively optimize the connection relationship through an energy function, verifies the scalability of the connection results through the Kuratowski theorem, identifies and processes non-planar areas, and applies the QEM algorithm to lightweight the three-dimensional model to complete the three-dimensional modeling of the target component.

[0043] Compared with the prior art, the present invention has at least one of the following technical effects:

[0044] This invention utilizes building component recognition technology that integrates deep learning with geometric features to automatically extract key feature points of building components. It then utilizes adaptive reconstruction and multi-scale detail enhancement techniques to screen and optimize these feature points, dynamically adjusting the model's level of detail based on actual project needs. Finally, based on feature point topology optimization and graph theory algorithms, a concise and accurate lightweight model is constructed. This invention improves the automation and efficiency of building component modeling, effectively resolving the conflict between lightweight and accurate models, and providing strong technical support for the digital and intelligent transformation of the construction industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments:

[0046] Figure 1 It is a flowchart of the steps of the point cloud building component modeling method of the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. The specific implementation methods of the present invention are described below in conjunction with the drawings and embodiments.

[0048] First embodiment

[0049] While laser point cloud technology is currently widely adopted in the 3D modeling of building components, it still has significant shortcomings in identifying and efficiently modeling complex components. Existing modeling methods mostly rely on manual modeling or semi-automated operations, which are not only labor-intensive and time-consuming, but also lack effective means for detail expression, geometric accuracy, and model redundancy control. For example, when scanning and modeling the facades of large buildings, the point cloud data often contains a large number of invalid and redundant points, and feature extraction relies on manual screening, which severely restricts modeling efficiency and the lightweighting effect of the model.

[0050] Furthermore, although existing point cloud recognition methods based on deep learning can achieve preliminary semantic segmentation and classification, they are limited in their ability to extract key feature points within components, especially in the lack of modeling and analysis of the geometric topological relationships between feature points, resulting in discontinuous modeling structures or missing connection relationships. At the same time, most existing solutions are single-scale processing, failing to dynamically adjust modeling details based on component complexity and semantic information, making it difficult to balance model accuracy and computing performance, and unable to adapt to the needs of BIM system integration, operation and maintenance analysis, and even expandable structure identification in actual engineering scenarios. For example,

[0051] During a large-scale commercial complex construction project in a certain city, the project team needed to digitally model the building's main structure for subsequent operations and maintenance management and BIM system integration. However, after acquiring point cloud data using laser scanning equipment, the team still relied heavily on manual annotation and modeling of components such as walls, beams, and columns. This process was not only inefficient and time-consuming, but also, due to the lack of a unified component feature extraction mechanism, often resulted in unstable modeling accuracy, loss of details, and repeated modeling.

[0052] Furthermore, when faced with multi-layered, intersecting structural components (such as arched beams and diagonal walls), traditional methods are unable to automatically identify their geometric categories and topological relationships. This results in a large amount of redundant data, connection errors, or missing key structures in the model, seriously affecting the integrity of the 3D model and its subsequent application effectiveness. Furthermore, most current point cloud modeling tools lack a mechanism to balance model complexity and semantic importance, and are unable to automatically adjust modeling accuracy and point cloud density for different components, resulting in increased storage and rendering burdens.

[0053] To this end, this embodiment proposes a point cloud building component modeling method based on feature extraction, which is mainly a laser point cloud building component lightweight modeling method based on feature extraction and geometric topology analysis. This method integrates the improved PointNet++ semantic segmentation network and the multi-scale curvature feature extraction mechanism, which can automatically identify edge feature points and component categories; by constructing a topological connection graph and introducing the Kruskal minimum spanning tree and simulated annealing optimization, the automatic simplification of the structural relationship of feature points is achieved; by combining scalability verification and Laplace deformation methods, the rationality of the model structure and the accuracy of geometric expression are improved; finally, the QEM algorithm is used to achieve lightweight compression of the three-dimensional model, while ensuring the integrity of the component semantics and geometric features, the modeling efficiency and model applicability are greatly improved, and it is adapted to the integration requirements of various engineering platforms. The specific implementation methods are as follows:

[0054] The present invention provides a point cloud building component modeling method based on feature extraction, comprising the following steps:

[0055] Use a 3D laser scanner to perform laser scanning on the target building component area, obtain 3D point cloud data of the building components and perform data preprocessing;

[0056] Specifically, laser scanners can collect data at high density on construction sites to obtain three-dimensional point cloud data, and perform pre-processing operations such as noise removal, point cloud registration, and missing data filling on the original point cloud data to obtain the point cloud set to be processed. ;

[0057] Based on the preprocessed 3D point cloud data, a multi-scale local neighborhood is constructed with each point as the center. The covariance matrix of the point set in the local neighborhood at each scale is calculated and the eigenvalues ​​are decomposed. The normalized curvature index is extracted and normalized using the Sigmoid function to generate an edge probability map for screening candidate edge feature points. The improved PointNet++ network is used for semantic segmentation, and the candidate edge feature points are output to the corresponding component category. The corresponding key feature point set is then screened based on the component category.

[0058] Specifically, an adaptive feature point extraction method based on geometric shape and semantic information automatically extracts key feature points of building components from laser point clouds. Based on the component's geometric characteristics, the method calculates characteristic values ​​such as curvature at each point in the point cloud to identify multiple candidate edge feature points. Combining the component's semantic information with its geometric features, the candidate edge feature points are screened and optimized to extract a set of key feature points that have a decisive impact on model construction.

[0059] Using the feature point cloud of key feature points as control points, a three-level B-spline surface model is constructed to fit the geometric shape of the target component. The density adjustment function is set according to the importance of each feature point cloud in a preset local neighborhood to dynamically adjust the corresponding spatial distribution density. This is used to adaptively reconstruct the three-dimensional model of the target component and introduce the Laplace deformation algorithm for detail enhancement.

[0060] Specifically, this feature point cloud-driven adaptive reconstruction technology mainly adjusts the number and positional relationship of feature points based on the local geometric features and global positional relationship of the feature point cloud through multi-level and multi-scale refined processing methods such as interpolation, smoothing, and texture mapping. This ensures that the model accurately restores the detailed information of building components while maintaining its lightweight characteristics, thereby enhancing the realism of the model.

[0061] Based on the adaptively reconstructed 3D model, a weighted complete graph is formed through a set of key feature points. The improved Kruskal algorithm is used to generate a minimum spanning tree to determine the preliminary connection relationship. The simulated annealing algorithm is introduced to iteratively optimize the connection relationship through the energy function. The scalability of the connection results is verified by the Kuratowski theorem. Non-planar areas are identified and processed. The QEM algorithm is applied to lightweight the 3D model to complete the 3D modeling of the target component.

[0062] Specifically, during the model reconstruction process, graph theory algorithms are used to perform topological optimization on feature points. Feature points are considered nodes in a graph, and by calculating spatial positional relationships such as distances and angles between nodes, a connection graph between feature points is constructed. Simulated annealing is used to optimize the algorithm, adjusting the connection relationships between feature points and simplifying the model structure. The connection paths between feature points are optimized, and the optimal or near-optimal solution for the connection paths between feature points is found. Planar graph determination algorithms such as the Kuratowski theorem in graph theory are used to ensure that the constructed model is composed of non-intersecting planes or surfaces, thereby avoiding incorrect model structures.

[0063] Furthermore, based on the preprocessed 3D point cloud data, a multi-scale local neighborhood is constructed with each point as the center. The covariance matrix of the point set in the local neighborhood of each scale is calculated and the eigenvalues ​​are decomposed. The normalized curvature index is extracted and normalized by the Sigmoid function to generate an edge probability map for screening candidate edge feature points, including:

[0064] Calculate the covariance matrix of each point in the multi-scale local neighborhood and decompose it to obtain the eigenvalues ​​in the three-axis directions;

[0065] The normalized curvature index is calculated based on the minimum eigenvalue of the three-axis direction;

[0066] Based on the normalized curvature index of each point in the multi-scale local neighborhood, a multi-scale edge response function is constructed using weighted fusion, and normalized mapping is performed through the Sigmoid function to generate an edge probability map;

[0067] The points whose edge response values ​​in the edge probability map are greater than the preset threshold are screened to obtain candidate edge feature points.

[0068] Specifically, the covariance matrix of each point in the multi-scale local neighborhood is calculated and decomposed to obtain the eigenvalues ​​of the three-axis directions, including,

[0069] (1) Each point As the center, three scales of local neighborhoods are constructed, and the neighborhood radius is set to 5mm, 15mm and 30mm respectively. At each scale, the covariance matrix of the neighborhood point set is calculated, and the eigenvalue decomposition is performed to obtain the eigenvalues ​​in the three main axis directions. , used to measure the local curvature characteristics of the point, the calculation formula of the local neighborhood covariance matrix C is,

[0070] ;

[0071] Among them, k is the number of neighborhood points; Neighborhood point coordinates is the neighborhood centroid coordinate ;

[0072] Extract the minimum eigenvalue Calculate the normalized curvature index:

[0073] ;

[0074] is the eigenvalue of the covariance matrix C; m is the neighborhood radius, in millimeters.

[0075] (2) Construct edge probability map by fusion of three-scale features, where the edge response function is,

[0076]

[0077] in, is the dynamic threshold, is the global mean curvature, is the variance of the curvature at this size; is a probability function used for probability normalization. Therefore, It is an adaptive edge judgment baseline;

[0078] When the preset threshold is 0.7, When the value is greater than 0.7, it is determined as an edge point, and the edge points that meet the requirements are extracted as candidate edge feature points.

[0079] Furthermore, the semantic segmentation candidate edge feature points are outputted through the improved PointNet++ network to obtain the corresponding component categories, and the corresponding key feature point sets are screened based on the component categories, including:

[0080] The improved PointNet++ network is used to segment the candidate edge feature points, taking the spatial coordinates, normal vectors, and RGB color information of the candidate edge feature points as input. The candidate edge feature points with semantic confidence higher than the set confidence threshold are screened, labeled with the corresponding component categories, and then output.

[0081] According to the output component category, corresponding feature enhancement is performed to extract the corresponding key feature point set.

[0082] Furthermore, corresponding feature enhancement is performed according to the output component category, and the corresponding key feature point set is extracted;

[0083] When the component type is a plane component, a plane fitting residual function is constructed based on the extracted candidate edge feature points. The optimal plane parameters are iteratively solved by the LM algorithm and the corner point detection is performed in combination with the convex hull vertex spacing of the plane projection point set. The candidate edge feature points at the structural turning position are screened to form a key feature point set.

[0084] When the component category is cylindrical, the column axis energy function is constructed, the endpoint position is optimized using the conjugate gradient method, the geometric deviation, semantic confidence and point spacing are introduced to construct a multi-constraint scoring function, and the corresponding candidate edge feature points are screened out through the branch and bound algorithm to form a key feature point subset.

[0085] Specifically, based on the extracted candidate edge feature points, an improved PointNet++ network is used to perform semantic segmentation on point cloud data. The network input includes a 9-dimensional feature vector consisting of the spatial coordinates, normal vector, and RGB color information of each point, and the output includes the component category label and semantic confidence value of each point.

[0086] Perform differential geometric analysis and optimization based on component categories and extract key feature point sets.

[0087] (1) When the component type is a planar component, such as a wall, floor or other building component:

[0088] Construct a plane fitting residual function:

[0089]

[0090] in, , , is the three-dimensional coordinate of the i-th point, a, b, c, d are the number of plane equations, To represent the absolute value of the distance from the i-th point to the plane, that is, the vertical distance, is the modulus of the plane normal vector, used to normalize the distance, It represents the sum of the distances from all points to the fitting plane, that is, the "fitting residual". The smaller the value, the better the fit.

[0091] The optimal plane equation is solved iteratively using the LM (Levenberg-Marquardt) algorithm: , the iterative solution of the optimal plane equation is the fitting solution goal.

[0092] Calculate the convex hull vertex spacing of the plane projection point set for corner detection and filter out points that meet the following conditions:

[0093] ;

[0094] , are two adjacent convex hull corner points, is the distance between points, It is the diagonal length of the model bounding box. If the distance between adjacent corner points is greater than 15% of the overall component length, it is considered as a valid corner point at the structural corner and retained as a feature point.

[0095] (2) When the component type is a cylindrical component, such as a beam, column or other building component:

[0096] Construct the geometric expression of cylinder fitting,

[0097] ,

[0098] in, , is the coordinate of the center of the cylinder cross section, r is the cylinder radius, Used to define the vertical height range of the column.

[0099] Construct the column axis energy function to optimize the endpoints:

[0100]

[0101] Among them, α and β are the orthogonality and continuity weight coefficients, α is 0.7, and β is 0.3. Orthogonality means that the column axis should be orthogonal to the cross section, and continuity means that the column axis should be smooth and continuous in the vertical direction. They are used to optimize the direction of the column centerline, that is, whether the centerline of the column is straight and symmetrical.

[0102] Solving the axis parameters by conjugate gradient method

[0103] Construct a multi-constraint optimization function to optimize and screen feature points:

[0104]

[0105] Among them, N is the number of candidate feature points, is the i-th candidate feature point, The geometric feature deviation is used to measure whether a part is a turning point by the curvature difference; is the semantic confidence; is the point spacing constraint; the weight coefficients are set to: ω1=0.6, ω2=0.3, ω3=0.1

[0106] Use the branch and bound algorithm for global optimization and set the termination condition:

[0107] ;

[0108] is the objective function of the kth iteration,

[0109] Furthermore, a three-level B-spline surface model is constructed to fit the geometric shape of the target component using the feature point cloud in the key feature point set as the control point, including:

[0110] Based on the key feature point set, the initial control grid is constructed, and the layer-by-layer refinement mechanism is used to reconstruct the B-spline surface model. The layers include coarse-grained layer, medium-grained layer and fine-grained layer. The three-level progressive B-spline surface reconstruction method is mainly used to construct a multi-resolution model from coarse to fine. The B-spline surface model dynamically balances the model accuracy and lightweight requirements.

[0111] ;

[0112] in, For the model level, For the Layer, through parameter coordinates The generated B-spline surface, , is the B-spline basis function along the u and v directions, used to construct the shape control function of the surface, For the The coordinates of the control points corresponding to the i,j control grid nodes in the layer, 、 For the The number of control points in the layer along the u and v directions.

[0113] Node vector configuration:

[0114] First floor 50mm, suitable for overall contour construction;

[0115] Second floor 20mm, to add main structural details;

[0116] Third floor 5mm, used for depicting high-precision feature edges.

[0117] Furthermore, the density of feature points is dynamically adjusted based on local geometric complexity and semantic importance to achieve multi-scale modeling with high fidelity in important areas and lightweight in secondary areas. A density adjustment function is established, which is: ;in, The number of points contained in the sphere neighborhood with a preset radius centered at (x, y, z); is the volume of the sphere neighborhood; is the curvature weight; is the semantic weight of the corresponding component category, where cylindrical components =1.5, plane component =0.7. For example, columns, beams, and other cylindrical components should be built more clearly, while flat components such as wall decorations and plate patterns can be built thicker to save data.

[0118] The Laplace deformation algorithm is used to enhance the sharpness of feature edges while maintaining the smooth continuity of the surface, and the deformation equation is constructed:

[0119]

[0120] in, is the Laplace matrix, element ; It is a hard constraint weight to ensure the displacement deviation of feature points ; is a diagonal matrix, when When it is a feature point ,otherwise is the deformation constraint vector, which is calculated by the difference between the initial surface and the target position.

[0121] Furthermore, based on the adaptively reconstructed three-dimensional model, a weighted complete graph is formed through a set of key feature points. The improved Kruskal algorithm is used to generate a minimum spanning tree to determine the preliminary connection relationship. The simulated annealing algorithm is introduced to iteratively optimize the connection relationship through the energy function, including:

[0122] Map the key feature point set to a graph node, construct a weighted connection edge between any two points, and generate a weighted complete graph. The weight of the connection edge The calculation formula is as follows:

[0123]

[0124] in, 、 is the three-dimensional coordinate of the key feature point; Global average point ; is the angle between the normal vectors of adjacent facets; Semantic Differentiation Factor , is the component confidence; 、 is the weight coefficient;

[0125] The improved Kruskal algorithm is used to sort the edge sets of the weighted complete graph in ascending order, and the union-find data structure is used to monitor the connected components, generate the minimum spanning tree and retain the redundant edge sets.

[0126] A simulated annealing algorithm is introduced for the minimum spanning tree structure to iteratively optimize the connection path and output the corresponding topological connection structure.

[0127] Furthermore, the Kuratowski theorem is used to verify the scalability of the connection results and identify and handle non-planar regions, including:

[0128] Based on the optimized topological connection structure, the expandability is verified by Kuratowski theorem to identify the non-expandable non-planar areas in the topological connection structure.

[0129] The connection structure of the non-planar area is topologically segmented and reconstructed into the corresponding developable surface subgraph using NURBS surface.

[0130] Furthermore, the QEM algorithm is applied to perform lightweight processing on the 3D model to complete the 3D modeling of the target component, including:

[0131] Based on edge collapse cost function Compress non-critical areas and critical areas accordingly, and use BC7 texture compression and JPEG-XL encoding to optimize the multi-level texture storage structure; where v is the position of the vertex after edge folding, 、 is the quadratic error matrix of adjacent triangles;

[0132] The lightweight 3D model is double-verified for reconstruction accuracy and geometric integrity.

[0133] Specifically, during the model reconstruction process, graph theory algorithms are used to perform topological optimization on feature points. Feature points are considered nodes in a graph, and a connection graph between them is constructed by calculating spatial relationships such as distances and angles between nodes. Simulated annealing is used to optimize the connection between feature points, simplifying the model structure and optimizing the connection paths between them. The optimal or near-optimal solution for these paths is found. This ensures that the constructed model consists of non-intersecting planes or curved surfaces, thus avoiding incorrect model structures.

[0134] The feature point set is mapped into graph nodes, and a weighted complete graph is constructed through geometric relationships and semantic constraints. The improved graph theory algorithm is used to optimize the connection relationship and generate a lightweight topological structure.

[0135]

[0136] in, 、 is the three-dimensional coordinate of the feature point; Global average point ; is the angle between the normal vectors of adjacent facets; Semantic Differentiation Factor , is the component confidence; 、 is the weight coefficient, =0.7.

[0137] The improved Kruskal algorithm is used to generate the minimum spanning tree, retaining redundant edge sets. First, the edge sets are sorted in ascending order. Then, a union-find data structure is used to monitor the connected components and generate the minimum spanning tree (MST), retaining redundant edge sets.

[0138] Simulated annealing optimization, building an energy function to optimize the connection relationship:

[0139]

[0140] Where, is the node degree; is the number of crossing edges in the projection plane.

[0141] Set the initial temperature T0 = 1000, the annealing coefficient α = 0.95, and the Markov chain length L = 1000. Perform state transition operations: edge swap (randomly replace two edges with new non-crossing edges), node perturbation (apply Gaussian noise displacement), and topology mutation (reverse local subgraph connections with a 10% probability).

[0142] Applying Kuratowski's theorem to detect non-planar structures: Constructing an adjacency matrix , if exists or The subgraph is judged as non-planar; the non-planar area is segmented to generate the Euler characteristic Surface patch; use NURBS surface to reconstruct the non-developable area:

[0143]

[0144] in, is the inverse distance weighted control point weight, is the B-spline basis function.

[0145] The geometry simplification uses the QEM algorithm for lightweight processing and defines the edge collapse cost:

[0146]

[0147] 、 is the quadratic error matrix between adjacent triangles. Low-cost edges are iteratively collapsed, with a simplification rate of 70% for less important areas and 30% for more important areas. Texture optimization constructs anisotropic mipmaps, using BC7 compression and JPEG-XL encoding.

[0148] Verify the reconstruction quality and perform double verification of the reconstruction accuracy and geometric integrity of the lightweight model, where the geometric error meets , topological integrity evaluation index: .

[0149] Second embodiment

[0150] Based on the same inventive concept, the present invention also provides a point cloud building component modeling system based on feature extraction, which adopts the above-mentioned point cloud building component modeling method, including:

[0151] The data acquisition module is used to perform laser scanning on the target building component area, obtain the three-dimensional point cloud data of the building component and perform data preprocessing;

[0152] The feature extraction module is used to construct a multi-scale local neighborhood centered on each point based on the preprocessed 3D point cloud data. The covariance matrix of the point set in the local neighborhood at each scale is calculated and the eigenvalues ​​are decomposed. The normalized curvature index is extracted and normalized using the Sigmoid function to generate an edge probability map for screening candidate edge feature points. The improved PointNet++ network is used to perform semantic segmentation on the candidate edge feature points, output the corresponding component category, and then the corresponding key feature point set is screened based on the component category.

[0153] The adaptive reconstruction module is used to construct a three-level B-spline surface model to fit the geometric shape of the target component using the feature point cloud concentrated in the key feature point as the control point. The density adjustment function is set according to the importance of each feature point cloud in the preset local neighborhood to dynamically adjust the corresponding spatial distribution density. It is used to adaptively reconstruct the three-dimensional model of the target component and introduce the Laplace deformation algorithm for detail enhancement.

[0154] The lightweight processing module is used for three-dimensional models based on adaptive reconstruction. It forms a weighted complete graph through a set of key feature points, uses an improved Kruskal algorithm to generate a minimum spanning tree to determine the preliminary connection relationship, introduces a simulated annealing algorithm to iteratively optimize the connection relationship through an energy function, verifies the scalability of the connection results through the Kuratowski theorem, identifies and processes non-planar areas, and applies the QEM algorithm to lightweight the three-dimensional model to complete the three-dimensional modeling of the target component.

[0155] Although the present invention has been disclosed above in terms of preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications to the technical solutions of the present invention by using the methods and technical contents disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the scope of protection of the technical solutions of the present invention.

Claims

1. A point cloud building component modeling method based on feature extraction, characterized in that the steps include: Perform laser scanning on the target building component area to obtain 3D point cloud data of the building components and perform data preprocessing; Based on the preprocessed three-dimensional point cloud data, a multi-scale local neighborhood is constructed with each point as the center. The covariance matrix of the point set in the local neighborhood of each scale is calculated and the eigenvalues ​​are decomposed. The normalized curvature index is extracted and normalized by the Sigmoid function to generate an edge probability map for screening candidate edge feature points. The improved PointNet++ network is used to perform semantic segmentation on the candidate edge feature points to output the corresponding component category, and the corresponding key feature point set is screened based on the component category. Using the feature point cloud in the key feature point set as the control point, a three-level B-spline surface model is constructed to fit the geometric shape of the target component, and a density adjustment function is set according to the importance of each feature point cloud in a preset local neighborhood to dynamically adjust the corresponding spatial distribution density, so as to adaptively reconstruct the three-dimensional model of the target component and introduce a Laplace deformation algorithm for detail enhancement; Based on the adaptively reconstructed three-dimensional model, a weighted complete graph is formed through the key feature point set, an improved Kruskal algorithm is used to generate a minimum spanning tree for determining a preliminary connection relationship, a simulated annealing algorithm is introduced to iteratively optimize the connection relationship through an energy function, the scalability of the connection result is verified by the Kuratowski theorem, non-planar areas are identified and processed, and the QEM algorithm is applied to perform lightweight processing on the three-dimensional model to complete the three-dimensional modeling of the target component.

2. The point cloud building component modeling method according to claim 1, characterized in that: Based on the preprocessed three-dimensional point cloud data, a multi-scale local neighborhood is constructed with each point as the center. The covariance matrix of the point set in each scale local neighborhood is calculated and the eigenvalue is decomposed. The normalized curvature index is extracted and normalized by the Sigmoid function to generate an edge probability map for screening candidate edge feature points, including: Calculating the covariance matrix of each point in the local neighborhood of multiple scales, and decomposing it to obtain the eigenvalues ​​in three-axis directions; Calculating the normalized curvature index based on the minimum eigenvalues ​​in the three-axis directions; Based on the normalized curvature index of each point in the multi-scale local neighborhood, a multi-scale edge response function is constructed by weighted fusion, and normalized mapping is performed by a Sigmoid function to generate the edge probability map; The points whose edge response values ​​in the edge probability map are greater than a preset threshold are screened to obtain the candidate edge feature points.

3. The point cloud building component modeling method according to claim 2, characterized in that: The improved PointNet++ network is used to perform semantic segmentation on the candidate edge feature points to output the corresponding component category, and the corresponding key feature point set is filtered based on the component category. include, Taking the 9-dimensional feature vector composed of the spatial coordinates, normal vector, and RGB color information of the candidate edge feature points as input, the improved PointNet++ network is used for segmentation, the candidate edge feature points with semantic confidence higher than the set confidence threshold are screened, and the corresponding component categories are labeled and output; Corresponding feature enhancement is performed according to the output component category, and a corresponding key feature point set is extracted.

4. The point cloud building component modeling method according to claim 3, characterized in that: Perform corresponding feature enhancement according to the output component category and extract the corresponding key feature point set; When the component category is a plane component, a plane fitting residual function is constructed based on the extracted candidate edge feature points, the optimal plane parameters are iteratively solved by the LM algorithm, and corner point detection is performed in combination with the convex hull vertex spacing of the plane projection point set, and the candidate edge feature points at the structural turning position are screened to form the key feature point set; When the component category is a cylindrical component, a column axis energy function is constructed, the endpoint position is optimized using the conjugate gradient method, geometric deviation, semantic confidence and point spacing are introduced to construct a multi-constraint scoring function, and the corresponding candidate edge feature points are screened out through the branch and bound algorithm to form the key feature point subset.

5. The point cloud building component modeling method according to claim 4, characterized in that: Using the feature point cloud in the key feature point set as control points, a three-level B-spline surface model is constructed to fit the geometric shape of the target component, including: An initial control grid is constructed based on the key feature point set, and a layer-by-layer refinement mechanism is adopted to reconstruct the B-spline surface model, wherein the layers include a coarse-grained layer, a medium-grained layer, and a fine-grained layer. The B-spline surface model is: ； in, For the model level, For the Layer, through parameter coordinates The generated B-spline surface, , is the B-spline basis function along the u and v directions, used to construct the shape control function of the surface, For the said The coordinates of the control points corresponding to the i,j control grid nodes in the layer, 、 For the said The number of control points in the level along the u and v directions.

6. The point cloud building component modeling method according to claim 5, characterized in that: The density adjustment function is ;in, The number of points contained in the sphere neighborhood with a preset radius centered at (x, y, z); is the volume of the neighborhood of the sphere; is the curvature weight; is the semantic weight corresponding to the component category, is the normalized curvature index of the minimum eigenvalue.

7. The point cloud building component modeling method according to claim 6, characterized in that: Based on the adaptively reconstructed three-dimensional model, a weighted complete graph is formed through the key feature point set, an improved Kruskal algorithm is used to generate a minimum spanning tree for determining a preliminary connection relationship, and a simulated annealing algorithm is introduced to iteratively optimize the connection relationship through an energy function, including: The key feature point set is mapped to a graph node, and a weighted connecting edge is constructed between any two points to generate the weighted complete graph. The weight of the connecting edge is The calculation formula is as follows: in, 、 is the three-dimensional coordinate of the key feature point; Global average point ; is the angle between the normal vectors of adjacent facets; Semantic Differentiation Factor , is the component confidence; 、 is the weight coefficient; An improved Kruskal algorithm is used to sort the edge sets of the weighted complete graph in ascending order, a union-find data structure is used to monitor connected components, the minimum spanning tree is generated, and redundant edge sets are retained; A simulated annealing algorithm is introduced for the minimum spanning tree structure to iteratively optimize the connection path and output the corresponding topological connection structure.

8. The point cloud building component modeling method according to claim 7, characterized in that: The scalability of the connection results is verified by Kuratowski's theorem, and non-planar areas are identified and processed. include, Based on the optimized topological connection structure, the expandability is verified by Kuratowski's theorem to identify non-expandable non-planar regions in the topological connection structure; The connection structure of the non-planar area is topologically segmented and reconstructed into a corresponding developable surface subgraph using a NURBS surface.

9. The point cloud building component modeling method according to claim 4, characterized in that: Apply the QEM algorithm to perform lightweight processing on the three-dimensional model to complete the three-dimensional modeling of the target component. include, Based on edge collapse cost function Compress non-critical areas and critical areas accordingly, and use BC7 texture compression and JPEG-XL encoding to optimize the multi-level texture storage structure; where v is the position of the vertex after edge folding, 、 is the quadratic error matrix of adjacent triangles; The lightweight three-dimensional model is double-verified for reconstruction accuracy and geometric integrity.

10. A point cloud building component modeling system based on feature extraction, using the point cloud building component modeling method according to any one of claims 1 to 9, characterized in that: include, The data acquisition module is used to perform laser scanning on the target building component area, obtain the three-dimensional point cloud data of the building component and perform data preprocessing; A feature extraction module is configured to construct a multi-scale local neighborhood centered on each point based on the preprocessed three-dimensional point cloud data, calculate the covariance matrix of the point set within the local neighborhood of each scale and decompose the eigenvalues, extract the normalized curvature index, perform normalized mapping using a Sigmoid function, generate an edge probability map for screening candidate edge feature points, perform semantic segmentation using an improved PointNet++ network, output the corresponding component category of the candidate edge feature points, and screen the corresponding key feature point set based on the component category; An adaptive reconstruction module is used to construct a three-level B-spline surface model to fit the geometric shape of the target component using the feature point cloud in the key feature point set as control points, and dynamically adjust the corresponding spatial distribution density by setting a density adjustment function based on the importance of each feature point cloud in a preset local neighborhood, so as to adaptively reconstruct the three-dimensional model of the target component and introduce a Laplace deformation algorithm for detail enhancement; A lightweight processing module is used to form a weighted complete graph based on the adaptively reconstructed three-dimensional model through the key feature point set, use an improved Kruskal algorithm to generate a minimum spanning tree for determining a preliminary connection relationship, introduce a simulated annealing algorithm to iteratively optimize the connection relationship through an energy function, verify the scalability of the connection result through the Kuratowski theorem, identify and process non-planar areas, and apply a QEM algorithm to perform lightweight processing on the three-dimensional model to complete the three-dimensional modeling of the target component.

Images