A city digital twin scene static LOD processing method and device and a storage medium

Abstract

To address the problems of semantic structure destruction, geometric feature distortion, and difficulty in controlling multi-level errors in the LOD simplification process of existing 3D models, this application discloses a static LOD processing method, device, and storage medium for urban digital twin scenes, belonging to the field of computer graphics and 3D modeling technology. The method includes: semantic recognition and region division of the urban building 3D mesh model to construct a semantically labeled mesh; establishing a simplification control model based on the semantically labeled mesh and determining simplification constraints; performing mesh simplification processing under constraints; preserving geometric features of the simplification results; evaluating errors in each LOD level model and performing adaptive correction; and replacing semantic components that cannot be preserved to generate multi-level LOD model data. This application reduces model complexity while achieving the coordinated preservation of semantic consistency and geometric accuracy, making it suitable for efficient rendering and visualization applications of urban-level 3D scenes.

Description

Technical Field

[0001] This application relates to the fields of computer graphics and 3D modeling technology, and more specifically, to a method, device and storage medium for static LOD processing of urban digital twin scenes. Background Technology

[0002] With the rapid development of City Information Modeling (CIM) and digital twin technologies, city-level 3D scenes are widely used in planning and design, operation and maintenance management, emergency command, and visualization. However, due to the massive scale and complex geometry of urban building model data, problems such as high storage pressure, low loading efficiency, and insufficient rendering performance often arise during network transmission and real-time rendering. Therefore, it is usually necessary to use Level of Detail (LOD) technology to represent the model hierarchically in order to achieve a balance between accuracy and performance in different application scenarios.

[0003] Existing static LOD processing methods are mostly based on pure geometric simplification algorithms, performing vertex folding or face merging on the mesh. While this can reduce model complexity, it does not fully consider the semantic attributes and structural differences of building components, easily leading to distortion of key components, blurred contour features, and destruction of semantic information. Furthermore, in the multi-level generation process, the lack of a unified semantic constraint mechanism and error control methods makes it difficult to maintain geometric continuity and visual consistency between different LOD levels, affecting the realism and usability of digital twin scenarios.

[0004] Furthermore, existing technologies lack effective closed-loop evaluation and adaptive correction mechanisms to address the issues of local error accumulation and loss of important components during the simplification process. They also struggle to reasonably represent and replace semantic components that cannot retain their geometric structure, thereby affecting the integrity and expressive power of multi-level model data.

[0005] In summary, how to balance semantic structure protection, geometric feature preservation, and multidimensional error control during the simplification of urban building 3D meshes, and achieve high-quality, multi-level, and sustainably optimized static LOD model generation, has become an urgent technical problem to be solved. Summary of the Invention

[0006] To overcome a series of shortcomings in existing technologies, the purpose of this application is to provide a static LOD processing method for urban digital twin scenes, comprising the following steps:

[0007] Semantic recognition and region division are performed on the building components in the three-dimensional mesh model of the city buildings to obtain a semantically labeled mesh containing multiple semantic regions;

[0008] Based on the semantic annotation grid, a semantically related simplified control model is established, so that each semantic region corresponds to a corresponding simplified constraint condition.

[0009] Based on the simplified control model, semantic constraint mesh simplification processing is performed on the three-dimensional mesh model of urban buildings to obtain a simplified mesh model;

[0010] Perform geometric feature preservation processing on the simplified mesh model to stabilize the building outline features in the simplified mesh model;

[0011] Based on the urban building 3D mesh model, the simplified mesh model of each LOD level is evaluated for error, and the local area where the error exceeds the preset threshold is corrected and controlled according to the evaluation results.

[0012] In some embodiments, semantic recognition and region division are performed on building components in the urban building 3D mesh model to obtain a semantically labeled mesh containing multiple semantic regions, including the following steps:

[0013] Using triangular facets as basic units, local geometric descriptors including face normals, Gaussian curvature, mean curvature, and facet area are extracted.

[0014] Construct an adjacency graph of adjacent faces, using faces as nodes and shared edges as edges to express local geometric topological relationships;

[0015] A multi-layer graph convolutional network is used to perform message passing and feature aggregation on the face adjacency graph, and each face is assigned a building component category label to obtain a set of face semantic category labels.

[0016] Perform connectivity analysis on the set of semantic category labels for facets, merge spatially adjacent and class-consistent facets to form preliminary semantic regions, and obtain a preliminary semantic annotation grid.

[0017] For the boundary transition region, boundary segments are extracted along the shared edges of adjacent semantic regions. Hard and soft boundaries are determined based on the angle between the normal vectors of the patches and the curvature gradient. Sharp cutting and boundary alignment are performed on the hard boundaries, and gradual transition assignment is performed on the soft boundaries. Patches are assigned to adjacent regions, and finally a semantic annotation mesh with clear spatial ownership and continuous boundaries is obtained.

[0018] In some embodiments, based on the semantic annotation grid, a semantically related simplified control model is established, such that each semantic region corresponds to a corresponding simplified constraint condition, including the following steps:

[0019] For each semantic region and its corresponding building component in the semantic annotation grid, feature parameters are extracted from four dimensions: relative area ratio of components, average observation frequency, structural recognition information content, and contour saliency. The weights of each dimension are determined by combining large-scale urban scene user visual attention data, and then the comprehensive visual importance score of each semantic region is calculated.

[0020] Based on the comprehensive visual importance score, the semantic region is divided into three simplification constraint levels: high fidelity, standard simplification, and radical simplification.

[0021] Set corresponding simplified constraints for the semantic regions of each simplified constraint level;

[0022] Based on the simplified constraint levels and their simplified constraints, a semantically relevant simplified control model is established.

[0023] In some embodiments, based on a simplified control model, a semantic constraint-based mesh simplification process is performed on the three-dimensional mesh model of urban buildings to obtain a simplified mesh model, including the following steps:

[0024] Candidate folded edges in the 3D mesh model of urban buildings are identified under the guidance of a simplified control model, and the geometric quadratic error cost of each folded edge is calculated as the basic cost component.

[0025] Based on the simplified control model, the semantic regions to which the two vertices of each candidate folded edge belong are checked. If they belong to different regions, a semantic cross-region penalty coefficient is added to the basic cost.

[0026] For candidate folded edges within high-fidelity semantic regions, a sparse penalty term related to the density of the remaining vertices in the region is added based on a simplified control model, so that the high-fidelity region adaptively reduces the folding rate during the simplification process.

[0027] After each folding operation is completed, the cost of all candidate folding edges in the neighborhood of the affected vertex is updated according to the simplified control model, and their order is adjusted in the priority queue to ensure that the next folding operation proceeds according to the current optimal strategy.

[0028] Driven by a simplified control model and a priority queue, the folding operation is executed cyclically to simplify the 3D mesh model of urban buildings while ensuring both geometric accuracy and semantic structure protection, until the preset simplification goal is achieved and a simplified mesh model is obtained.

[0029] In some embodiments, the method for performing geometric feature preservation processing on a simplified mesh model is as follows:

[0030] Extracting building outline feature lines from a 3D mesh model of urban buildings;

[0031] The feature lines of the building's outer contour are segmented and fitted with polyline to obtain polyline segments, and the endpoints of the polyline segments are used as anchor points to construct a three-dimensional feature constraint framework.

[0032] During the mesh simplification process, displacement constraints are imposed on vertices falling within the influence range of the 3D feature constraint framework, and the building volume perception and outline continuity are adaptively maintained at different LOD levels.

[0033] When a region containing a polyline segment loses vertex support due to simplification, a retained vertex is inserted on the polyline segment to ensure that the contour features in the simplified mesh continue to exist, while maintaining visual and structural continuity at different LOD levels.

[0034] In some embodiments, the method for error evaluation of simplified mesh models at each LOD level based on a 3D mesh model of urban buildings is as follows:

[0035] Several sampling points are uniformly determined on the surface of the simplified mesh model at each LOD level;

[0036] Calculate the shortest undirected point-to-surface distance from each sampling point to the surface of the 3D mesh model of the city buildings, and express the geometric distance error as the root mean square of the absolute values ​​of the distances to all sampling points.

[0037] Compare the unit normal vectors of the simplified mesh model and the 3D mesh model of the city building at the same sampling point, and take the average deviation of the angle to represent the normal deflection error;

[0038] The closed volume of each semantic region in the three-dimensional mesh model and the simplified model of urban buildings is calculated separately, and the volume change rate is calculated by weighting according to the simplification constraint level of each semantic region to obtain the semantic volume error;

[0039] The geometric distance error, normal deflection error, and semantic volume error are normalized and weighted to obtain a comprehensive error score.

[0040] In some embodiments, the simplified mesh model at each LOD level is evaluated based on the urban building 3D mesh model, and local areas where the error exceeds a preset threshold are corrected and controlled according to the evaluation results, including the following steps:

[0041] Local areas where the error exceeds a preset threshold are identified, all non-compliant faces are marked in units of triangular faces, and a preset buffer range is expanded based on the spatial bounding box of the non-compliant faces to determine the correction domain.

[0042] Within the correction domain, a Loop subdivision is performed on each non-compliant triangular facet by one level, so that the midpoint of each edge is uniformly subdivided into four sub-triangular faces, thereby increasing vertex density and improving the degree of freedom of local geometric adjustment.

[0043] Using the surface of a local area corresponding to the 3D mesh model of urban buildings as the optimization target, the Laplacian surface deformation method is used to iteratively optimize the vertex positions within the correction domain, so that the vertex projection positions onto the surface of the 3D mesh model of urban buildings gradually move closer.

[0044] When the local area error still exceeds the preset threshold, repeat the process of illegal area localization, adaptive mesh refinement and local surface optimization until the local area error converges to below the preset threshold.

[0045] In some embodiments, the static LOD processing method for urban digital twin scenes further includes the following steps:

[0046] When the LOD level reaches the preset simplification level, determine whether there are semantic components in the current simplified mesh model that cannot retain the geometric structure;

[0047] If there are semantic components that cannot preserve the geometric structure, then the semantic components are represented by alternative expressions, and the simplified mesh models of each LOD level are combined to finally obtain multi-level LOD model data;

[0048] If there are no semantic components that cannot preserve the geometric structure, the simplified mesh models of each LOD level are directly integrated to obtain multi-level LOD model data.

[0049] In addition, to achieve the above objectives, this application also proposes a static LOD processing device for urban digital twin scenes. The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the static LOD processing method for urban digital twin scenes as described above.

[0050] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the static LOD processing method for urban digital twin scenes as described above.

[0051] Compared with the prior art, this application has the following beneficial effects:

[0052] This application constructs a semantically related simplified control model by performing semantic recognition and region division on the three-dimensional mesh model of urban buildings. In the process of mesh simplification, semantic constraint regulation, geometric feature preservation and multi-dimensional error closed-loop correction mechanism are introduced to achieve the generation of multi-level LOD models that take into account geometric accuracy, semantic structural integrity and contour stability. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating a static LOD processing method for a city digital twin scene disclosed in an embodiment of this application. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some embodiments of this invention, but not all embodiments.

[0055] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] The embodiments and directional terms described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0057] like Figure 1 As shown, a static LOD processing method for a city digital twin scene includes the following steps:

[0058] A three-dimensional mesh model of urban buildings is obtained, and semantic recognition and region division are performed on the building components in the three-dimensional mesh model of urban buildings to obtain a semantically labeled mesh containing multiple semantic regions.

[0059] Based on the semantic annotation grid, a semantically related simplified control model is established, so that each semantic region corresponds to a corresponding simplified constraint condition.

[0060] Based on the simplified control model, semantic constraint mesh simplification processing is performed on the three-dimensional mesh model of urban buildings to obtain a simplified mesh model;

[0061] Perform geometric feature preservation processing on the simplified mesh model to stabilize the building outline features in the simplified mesh model;

[0062] Based on the urban building 3D mesh model, the simplified mesh model of each LOD level is evaluated for error, and the local area where the error exceeds the preset threshold is corrected and controlled according to the evaluation results.

[0063] When the LOD level reaches the preset simplification level, determine whether there are semantic components in the current simplified mesh model that cannot retain the geometric structure;

[0064] If there are semantic components that cannot preserve the geometric structure, then the semantic components are represented by alternative expressions, and the simplified mesh models of each LOD level are combined to finally obtain multi-level LOD model data;

[0065] If there are no semantic components that cannot preserve the geometric structure, the simplified mesh models of each LOD level are directly integrated to obtain multi-level LOD model data.

[0066] The static LOD processing method for urban digital twin scenarios described in this application performs semantic recognition and region division on the 3D mesh model of urban buildings, constructs a semantically labeled mesh, and establishes a simplified control model corresponding to the semantic regions to achieve hierarchical constraint control of different components during the simplification process. During the mesh simplification stage, semantic cross-domain constraints and adaptive folding strategies are introduced to reduce model complexity while protecting key semantic structures. Simultaneously, geometric feature preservation processing is performed on the simplified mesh, and constraints are applied to the building outline and salient features to ensure stable volume perception and outline continuity at different LOD levels. Furthermore, based on the original 3D mesh model, multi-dimensional error evaluation is performed on the simplification results of each LOD level, and refinement and optimization corrections are performed on local areas where the error exceeds the threshold, forming a closed-loop control mechanism. When the LOD level reaches the preset simplification level, semantic components that cannot retain their geometric structure are replaced with alternative expressions, and the model data of each level are integrated to finally generate multi-level LOD model data with complete structure, consistent semantics, and controllable accuracy. This achieves high-quality, high-performance hierarchical expression and stable output of 3D models in urban digital twin scenarios.

[0067] In some embodiments, the method for obtaining a three-dimensional mesh model of urban buildings is as follows:

[0068] Oblique aerial photography of the target urban area is used to obtain multi-view aerial images with sufficient overlap.

[0069] The structure-reconstruction-motion algorithm is applied to multi-view aerial images to generate sparse point clouds and solve for camera pose.

[0070] Generate dense point clouds containing texture coordinates based on a multi-view stereo matching algorithm;

[0071] Extract building information model data of the target urban area and complete the registration of dense point cloud and building information model data through iterative nearest point algorithm;

[0072] The registered and fused point cloud is reconstructed using Poisson surfaces to generate triangular meshes, and topological defects are repaired to obtain a complete and closed 3D mesh model of urban buildings.

[0073] The method for obtaining a 3D mesh model of urban buildings described in this application acquires multi-view image data with high overlap through oblique photogrammetry aerial photography. A sparse point cloud is generated using a structure-reconstruction-motion algorithm, and the camera pose is calculated to establish initial spatial geometric relationships. Based on this, a multi-view stereo matching algorithm is used to generate a dense point cloud containing texture coordinate information, achieving high-precision reconstruction of urban building details. Simultaneously, building information model data of the target area is extracted, and an iterative nearest-point algorithm is used to achieve precise registration between the dense point cloud and the building information model data, realizing the fusion of geometric data and structural semantic information. Subsequently, Poisson surface reconstruction is performed on the fused point cloud to generate a triangular mesh, and the generated mesh undergoes topology repair and closure processing. Finally, a 3D mesh model of urban buildings with complete structure, geometric continuity, and semantic consistency is obtained, providing high-quality basic data support for subsequent digital twin scene construction and LOD processing.

[0074] In some embodiments, semantic recognition and region division are performed on building components in the urban building 3D mesh model to obtain a semantically labeled mesh containing multiple semantic regions, including the following steps:

[0075] Using triangular facets as basic units, local geometric descriptors including face normals, Gaussian curvature, mean curvature, and facet area are extracted.

[0076] Construct an adjacency graph of adjacent faces, using faces as nodes and shared edges as edges to express local geometric topological relationships;

[0077] A multi-layer graph convolutional network is used to perform message passing and feature aggregation on the face adjacency graph, and each face is assigned a building component category label to obtain a set of face semantic category labels.

[0078] Perform connectivity analysis on the set of semantic category labels for facets, merge spatially adjacent and class-consistent facets to form preliminary semantic regions, and obtain a preliminary semantic annotation grid.

[0079] For the boundary transition region, boundary segments are extracted along the shared edges of adjacent semantic regions. Hard and soft boundaries are determined based on the angle between the normal vectors of the patches and the curvature gradient. Sharp cutting and boundary alignment are performed on hard boundaries, and gradual transition assignment is performed on soft boundaries. Patches are reasonably assigned to adjacent regions, and finally a semantic annotation mesh with clear spatial ownership and continuous boundaries is obtained.

[0080] Furthermore, the method for determining hard and soft boundaries based on the angle between surface normal vectors and curvature gradient includes the following steps:

[0081] Extract each shared edge that constitutes the boundary, and determine the first and second triangular facets that are adjacent to each other on both sides of the shared edge;

[0082] Calculate the unit normal vectors of the first and second triangular facets respectively, and calculate the angle θ between the normal vectors based on the dot product of the two unit normal vectors;

[0083] The curvature parameters of the first and second triangular facets are calculated respectively, and the absolute value of the difference between their curvatures is used as the curvature gradient Δk, wherein the curvature parameter is the mean curvature or Gaussian curvature.

[0084] The angle θ between the normal vectors and the curvature gradient Δk are normalized to obtain the normalized angle parameter θ′ and the normalized curvature parameter Δk′.

[0085] The boundary type is determined based on preset judgment rules: when θ′≥T1 and Δk′≥T2, the boundary is judged as a hard boundary; when θ′≤T3 and Δk′≤T4, the boundary is judged as a soft boundary; when θ′ and Δk′ are between the above intervals, a comprehensive judgment is made based on the weighted discriminant function F=α·θ′+β·Δk′, where when F≥T5, it is judged as a hard boundary, otherwise it is judged as a soft boundary;

[0086] Among them, T1, T2, T3, T4, and T5 are preset judgment thresholds, and α and β are weighting coefficients. T1, T2, T3, T4, T5, α, and β are set through experience or determined based on sample data statistics. F represents the judgment index used to comprehensively characterize the degree of geometric abrupt change at the boundary. It is obtained by weighted summation of the normalized normal vector angle parameter and the curvature gradient parameter, and is used to reflect the strength of geometric discontinuity at the boundary.

[0087] The semantic recognition and region division method for urban building components described in this application extracts geometric descriptors such as face normals, curvature, and area using triangular facets as basic units. It constructs a face adjacency graph to express local topological relationships and uses a multi-layer graph convolutional network for feature aggregation and category determination to achieve facet-level semantic annotation. On this basis, it merges adjacent facets of the same type through connected component analysis to form a preliminary semantic region. It distinguishes between hard and soft boundaries based on the normal angle and curvature gradient of the region boundary and performs sharp cutting alignment or gradual transition assignment processing respectively. Finally, it obtains a semantic annotation mesh with clear spatial ownership and continuous and stable boundaries, providing a structured foundation for subsequent semantic constraint simplification.

[0088] In some embodiments, based on the semantic annotation grid, a semantically related simplified control model is established, such that each semantic region corresponds to a corresponding simplified constraint condition, including the following steps:

[0089] For each semantic region and its corresponding building component in the semantic annotation grid, feature parameters are extracted from four dimensions: relative area ratio of components, average observation frequency, structural recognition information content, and contour saliency. The weights of each dimension are determined by combining large-scale urban scene user visual attention data, and then the comprehensive visual importance score of each semantic region is calculated.

[0090] Based on the comprehensive visual importance score, the semantic region is divided into three simplification constraint levels: high fidelity, standard simplification, and radical simplification.

[0091] Set corresponding simplification constraints for semantic regions of each simplification constraint level to ensure the structural information and visual fidelity of key components, while allowing secondary components to be simplified efficiently;

[0092] Based on the simplified constraint level and its simplified constraints, a semantically related simplified control model is established to hierarchically control the simplification of the 3D mesh, thereby achieving fidelity of key components and optimization and simplification of decorative details.

[0093] Furthermore, the method for determining the weights of each dimension by combining user visual attention data in large-scale urban scenarios includes the following steps:

[0094] Acquire visual attention data of users in urban scenes, including fixation point location, fixation duration, and fixation frequency information;

[0095] Visual attention data is mapped onto the surface of a 3D mesh model of urban buildings, and attention intensity indices for each semantic region are statistically analyzed. The attention intensity indices are obtained by weighted summation of the number of gaze points falling into the semantic region, the duration of gaze, and the number of gazes.

[0096] For each semantic region, a feature vector is constructed consisting of the relative area ratio of components, average observation frequency, structural recognition information, and contour saliency. The attention intensity index is used as the target response value to form a sample dataset.

[0097] Based on the sample dataset, a multiple linear regression model is used to establish the mapping relationship between each feature dimension and the attention intensity index, and the regression coefficients are solved by the least squares method.

[0098] The regression coefficients are normalized to obtain the weight coefficients corresponding to each feature dimension, which are used to characterize the degree of contribution of different dimensions to visual importance.

[0099] Furthermore, the method for calculating the comprehensive visual importance score of each semantic region includes the following steps:

[0100] Four feature parameters were extracted for each semantic region: relative area ratio of components, average observation frequency, structural recognition information content, and contour saliency. The feature parameters were then normalized so that their values ​​were uniformly mapped to the [0,1] interval.

[0101] Obtain the weight coefficients corresponding to each feature parameter. The weight coefficients are determined by regression analysis based on user visual attention data in urban scenarios.

[0102] Based on the normalized feature parameters and corresponding weight coefficients, a comprehensive visual importance scoring function is constructed. The comprehensive visual importance score S is calculated for each semantic region. The calculation method is: S=w1·x1+w2·x2+w3·x3+w4·x4, where x1, x2, x3, and x4 represent the normalized relative area ratio of the component, the average observation frequency, the amount of structural recognition information, and the contour saliency, respectively. w1, w2, w3, and w4 represent the corresponding weight coefficients, and satisfy w1+w2+w3+w4=1.

[0103] In some embodiments, the comprehensive visual importance score can also be calculated using a nonlinear model, including mapping each feature parameter using a weighted nonlinear function or a machine learning model to obtain a score result that better reflects the characteristics of visual perception.

[0104] Furthermore, the method for dividing semantic regions into different simplified constraint levels based on comprehensive visual importance scores includes the following steps:

[0105] Obtain the comprehensive visual importance score S corresponding to all semantic regions, and normalize the score so that its value range is [0,1].

[0106] Based on preset grading rules, each semantic region is classified into different levels: when S≥τ1, the corresponding semantic region is classified as high-fidelity; when τ2≤S<τ1, the corresponding semantic region is classified as standard simplification; when S<τ2, the corresponding semantic region is classified as radical simplification.

[0107] Wherein, τ1 and τ2 are preset grading thresholds, and satisfy 0 < τ2 < τ1 < 1.

[0108] In some embodiments, τ1 and τ2 are adaptively determined using statistical methods, including:

[0109] The comprehensive visual importance score S of all semantic regions is sorted, and the percentile threshold is determined according to the score distribution. Semantic regions with scores in the top P1 percentile are classified as high-fidelity, semantic regions with scores between P2 and P1 percentiles are classified as standard simplification, and semantic regions with scores in the bottom P2 percentile are classified as radical simplification. P1 and P2 are preset ratio parameters that satisfy 0 < P2 < P1 < 1. In some embodiments, P1 is between 0.6 and 0.8, and P2 is between 0.2 and 0.4.

[0110] Furthermore, the method for establishing a semantically relevant simplified control model based on simplified constraint levels and their simplified constraints includes the following steps:

[0111] Assign a corresponding simplification constraint level to each semantic region, and determine the set of control parameters for that semantic region based on the simplification constraint level. The set of control parameters includes at least the target simplification rate r, the error tolerance ε, and the feature preservation weight γ.

[0112] Based on the set of control parameters, constraint functions are constructed for each semantic region to limit vertex position offset, edge folding priority and geometric feature preservation strength during mesh simplification. The constraint functions include vertex displacement constraint functions and feature preservation constraint functions.

[0113] A unified mesh simplification cost function is constructed by weighting and combining geometric error terms, semantic constraint terms, and feature preservation terms to obtain a comprehensive cost function C used to guide edge folding operations. Its expression is: C = w g ·E g +w s ·E s +w f ·E f , of which E g E is the geometric error term. s E is a semantic constraint penalty term. f For the feature preservation constraint term, w g w s w f The corresponding weight parameters are assigned values ​​based on the simplification constraint level of the semantic region;

[0114] The control parameter sets and corresponding cost function parameters of each semantic region are associated and stored, and the mapping relationship between semantic regions and simplified control parameters is constructed to form a semantically related simplified control model;

[0115] During the mesh simplification process, based on the semantic region to which the currently processed edge or vertex belongs, the corresponding control parameters and cost function weights are read from the semantically related simplification control model to dynamically adjust the simplification strategy.

[0116] It should be noted that the semantically related simplification control model is used to establish a mapping relationship between semantic region categories and control parameters and constraint rules in the mesh simplification process, thereby achieving differentiated simplification control for different semantic regions.

[0117] The semantically relevant simplification control model construction method described in this application extracts feature parameters from each semantic region in the semantically labeled grid from dimensions such as the relative area ratio of components, average observation frequency, structural recognition information content, and contour saliency. Weights are determined by combining these parameters with user visual attention data from urban scenes, and a comprehensive visual importance score is calculated. Based on this score, the semantic regions are divided into high-fidelity, standard simplification, and radical simplification levels, and differentiated simplification constraint strategies are set for each level. This ensures the structural integrity and visual fidelity of key components while improving the simplification efficiency of secondary components. On this basis, a semantically relevant hierarchical simplification control model is established to constrain and regulate the simplification process of the three-dimensional grid, achieving a synergistic unity between the fidelity of key components and the overall model performance optimization.

[0118] In some embodiments, based on a simplified control model, a semantic constraint-based mesh simplification process is performed on the three-dimensional mesh model of urban buildings to obtain a simplified mesh model, including the following steps:

[0119] Candidate folded edges in the 3D mesh model of urban buildings are identified under the guidance of a simplified control model, and the geometric quadratic error cost of each folded edge is calculated as the basic cost component.

[0120] The semantic regions to which the two vertices of each candidate folded edge belong are checked based on the simplified control model. If they belong to different regions, a semantic cross-domain penalty coefficient is added to the base cost. The penalty intensity is determined by the difference in simplification levels between the two regions.

[0121] For candidate folded edges within high-fidelity semantic regions, a sparse penalty term related to the density of the remaining vertices in the region is added based on a simplified control model, so that the high-fidelity region adaptively reduces the folding rate during the simplification process.

[0122] After each folding operation is completed, the cost of all candidate folding edges in the neighborhood of the affected vertex is updated according to the simplified control model, and their order is adjusted in the priority queue to ensure that the next folding operation proceeds according to the current optimal strategy.

[0123] Driven by a simplified control model and a priority queue, the folding operation is executed cyclically to simplify the 3D mesh model of urban buildings while ensuring both geometric accuracy and semantic structure protection, until the preset simplification goal is achieved and a simplified mesh model is obtained.

[0124] The semantically constrained mesh simplification method described in this application identifies candidate folding edges and calculates the geometric quadratic error as the basic cost under the guidance of a simplification control model. At the same time, it superimposes cross-domain penalty coefficients based on the semantic regions to which the vertices at both ends of the folding edge belong, and adds a sparse penalty term related to the density of the remaining vertices to high-fidelity regions to differentiate and control the folding priority. After each fold, the cost of candidate edges in the affected neighborhood is updated in real time and rearranged in the priority queue to drive subsequent folds to proceed according to the optimal strategy. Under this mechanism, the folding operation is executed cyclically, so that the three-dimensional mesh of urban buildings is gradually simplified while taking into account both geometric accuracy and semantic structure protection, until the preset goal is achieved, resulting in a simplified mesh model with stable structure and consistent semantics.

[0125] In some embodiments, the method for performing geometric feature preservation processing on a simplified mesh model is as follows:

[0126] Extract building outline feature lines from the 3D mesh model of urban buildings, including roof edge lines, floor division lines, building volume transition lines, and window-wall boundary lines;

[0127] The feature lines of the building's outer contour are segmented and fitted with polyline to obtain polyline segments, and the endpoints of the polyline segments are used as anchor points to construct a three-dimensional feature constraint framework.

[0128] During the mesh simplification process, displacement restrictions are imposed on vertices falling within the influence range of the 3D feature constraint framework, so that vertices do not deviate from the nearest polyline segment by more than the preset tolerance when moving, and the building volume perception and outline continuity are adaptively maintained under different LOD levels.

[0129] When a region containing a polyline segment loses vertex support due to simplification, a retained vertex is inserted on the polyline segment to ensure that the contour features in the simplified mesh continue to exist, while maintaining visual and structural continuity at different LOD levels.

[0130] The geometric feature preservation method described in this application extracts outer contour feature lines such as roof edge lines, floor division lines, volume transition lines, and window-wall boundary lines from the three-dimensional mesh model of urban buildings, performs segmented polyline fitting, and constructs a three-dimensional feature constraint framework using endpoints. During the mesh simplification process, displacement restrictions are applied to vertices within the constraint range to ensure that their offset does not exceed a preset tolerance, thereby maintaining the perception of building volume and contour continuity at different LOD levels. When a local area loses feature support due to simplification, retained vertices are inserted on the corresponding polyline segments to maintain the contour structure, achieving a coordinated guarantee of visual stability and structural continuity during the simplification process.

[0131] In some embodiments, the method for error evaluation of simplified mesh models at each LOD level based on a 3D mesh model of urban buildings is as follows:

[0132] Several sampling points are uniformly determined on the surface of the simplified mesh model at each LOD level;

[0133] Calculate the shortest undirected point-to-surface distance from each sampling point to the surface of the 3D mesh model of the city buildings, and express the geometric distance error as the root mean square of the absolute values ​​of the distances to all sampling points.

[0134] Compare the unit normal vectors of the simplified mesh model and the 3D mesh model of the city building at the same sampling point, and take the average deviation of the angle to represent the normal deflection error;

[0135] The closed volume of each semantic region in the three-dimensional mesh model and the simplified model of urban buildings is calculated separately, and the volume change rate is calculated by weighting according to the simplification constraint level of each semantic region to obtain the semantic volume error;

[0136] The geometric distance error, normal deflection error, and semantic volume error are normalized and weighted to obtain a comprehensive error score.

[0137] The error assessment method described in this application involves uniformly distributing sampling points on the surface of the simplified mesh model at each LOD level, calculating the shortest point-to-surface distance from the sampling points to the original 3D urban building mesh model, and representing the geometric distance error in the form of root mean square. Simultaneously, the normal deflection error is obtained by comparing the angle between the unit normal vectors at corresponding positions. Furthermore, the closed volume of each semantic region in the original and simplified models is calculated, and the semantic volume error is obtained by weighting the calculations based on the simplification constraint levels. On this basis, the geometric distance error, normal deflection error, and semantic volume error are normalized and weighted to form a comprehensive error score, thereby achieving a quantitative evaluation of the accuracy and semantic preservation of models at different LOD levels.

[0138] In some embodiments, the simplified mesh model at each LOD level is evaluated based on the urban building 3D mesh model, and local areas where the error exceeds a preset threshold are corrected and controlled according to the evaluation results, including the following steps:

[0139] Local areas where the error exceeds a preset threshold are identified, all non-compliant faces are marked in units of triangular faces, and a preset buffer range is expanded based on the spatial bounding box of the non-compliant faces to determine the correction domain.

[0140] Within the correction domain, a Loop subdivision is performed on each non-compliant triangular facet by one level, so that the midpoint of each edge is uniformly subdivided into four sub-triangular faces, thereby increasing vertex density and improving the degree of freedom of local geometric adjustment.

[0141] Using the surface of a local area corresponding to the 3D mesh model of urban buildings as the optimization target, the Laplacian surface deformation method is used to iteratively optimize the vertex positions within the correction domain, so that the vertex projection positions onto the surface of the 3D mesh model of urban buildings gradually move closer.

[0142] When the local area error still exceeds the preset threshold, repeat the process of illegal area localization, adaptive mesh refinement and local surface optimization until the local area error converges to below the preset threshold.

[0143] The error-driven local correction control method described in this application assesses the error of the simplified mesh model at each LOD level, identifies local regions where the error exceeds a preset threshold, and marks the non-compliant faces as units of triangular faces. The correction domain is determined by combining spatial bounding box expansion. Within this domain, a first-level loop subdivision is performed on the non-compliant faces to increase vertex density and local adjustment degrees of freedom. The surface of the corresponding region of the original urban building 3D mesh model is used as the optimization target, and the Laplacian surface deformation method is employed to iteratively optimize the vertices, gradually approximating the original surface. When the local error still does not meet the standard, the region location, refinement, and optimization steps are repeated until the error converges below the preset threshold, thereby achieving adaptive closed-loop correction and stable control of the simplified model's accuracy.

[0144] In some embodiments, when the LOD level reaches a preset simplification level, determining whether there are semantic components in the current simplified mesh model that cannot retain geometric structure includes the following steps:

[0145] Semantic component patch extraction steps: Under the current LOD level, extract the patch set corresponding to each semantic component, and count the total number of patches in the patch set;

[0146] Candidate Unretainable Component Marking Step: When the number of faces of a semantic component is lower than the preset minimum number of faces threshold, the semantic component is marked as a candidate unretainable component;

[0147] Shape integrity verification step: Calculate the ratio of the shortest axis to the longest axis of the bounding box for the candidate unretainable component. When the ratio is lower than the preset shape fidelity threshold, it is determined that the semantic component has lost its geometric recognizability.

[0148] Geometric interleaving test steps: Project the set of facets of the candidate component that cannot be retained and the set of facets of its adjacent semantic regions onto the maximum projection plane of the candidate component's bounding box, and calculate the ratio of the overlapping area of ​​the projection to the total projection area of ​​the candidate component. When the ratio exceeds 30%, it is determined that the geometric structure of the semantic component cannot be retained independently.

[0149] Final determination step: Semantic components that meet either the shape integrity test or the geometric interpenetration test are identified as semantic components that cannot retain their geometric structure.

[0150] The method for determining semantic components that cannot be preserved as described in this application first extracts the set of faces corresponding to each semantic component and counts the number of faces after the LOD level reaches a preset simplification level. When the number of faces is lower than a preset minimum threshold, it is marked as a candidate component that cannot be preserved. Then, the candidate component is subjected to shape integrity test. The geometric recognizability is judged by calculating the ratio of the shortest axis to the longest axis of its bounding box, and a geometric interpenetration test is performed. The candidate component and its adjacent facets are projected onto the maximum projection plane, and the overlap area ratio is calculated to evaluate its independent preservation ability. When either the shape integrity test or the geometric interpenetration test condition is met, it is determined that the semantic component cannot preserve its geometric structure, thereby providing a basis for subsequent alternative expression processing.

[0151] In some embodiments, if there are semantic components whose geometric structure cannot be preserved, then the semantic components are represented using alternative expressions, and the simplified mesh models of each LOD level are combined to finally obtain multi-level LOD model data, including the following steps:

[0152] Based on the semantic type, alternative representation methods are selected for the components that cannot be retained, including: for planar hole components such as windows and doorways, a bulletin board texture replacement method is used, embedding an orthogonal projection texture image with a transparent channel into the wall position where the original semantic component is located. The orthogonal projection texture image is generated by the front view rendering image of the semantic component corresponding to the urban building 3D mesh model; for cantilever decorative components such as balconies and eaves, a normal map overlay method is used to bake the geometric details of the semantic component onto the normal map of the supporting wall, and the concavity and convexity are visually restored through lighting calculations; for linear components such as columns and decorative moldings, a skeleton line plus procedural instance rendering method is used, recording the component direction with a single polyline, and instantiating a preset low-face-count section template along the skeleton line during rendering; for point-like decorative components with extremely small volume, a particle system positioning point is used to record, and a bulletin board sprite is used to replace it during rendering;

[0153] The world coordinates, spatial orientation, and semantic category labels of all semantic components that use alternative expressions are fully recorded. At the same time, the additional rendering resources required for each alternative expression, including texture images, normal maps, section templates, and sprite atlases, are packaged and stored together with the multi-level LOD model data.

[0154] The semantic components after the replacement process are combined with the simplified mesh models of each LOD level to generate the multi-level LOD model data.

[0155] The method for generating multi-level LOD models by replacing semantic components described in this application selects the corresponding replacement expression method according to the semantic type after detecting semantic components that cannot retain their geometric structure: for planar hole components such as windows and doorways, a bulletin board texture is used as a replacement, and an orthogonal projection texture with a transparent channel generated by the original model's frontal rendering is embedded in the original wall position; for cantilever decorative components such as balconies and eaves, a normal map overlay method is used to bake the geometric details onto the normal map of the supporting wall to restore the visual concavity and convexity effect; for linear components such as columns and decorative moldings, a skeleton line plus procedural instance rendering method is used, with polylines recording the direction and low-face-count section templates instantiated along the line; for point-like decorative components with extremely small volume, particle positioning points are used to record them, and bulletin board sprites are used as replacements during rendering; at the same time, the world coordinates, spatial orientation, and semantic tags of each replacement component are fully recorded, and the required texture images, normal maps, section templates, sprite atlases, and other resources are packaged together with the simplified mesh models of each LOD level to generate multi-level LOD model data, thereby significantly reducing geometric complexity while maintaining the overall visual expression effect.

[0156] This application provides a static LOD processing device for urban digital twin scenes, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the static LOD processing method for urban digital twin scenes in the first embodiment described above.

[0157] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the static LOD processing method for urban digital twin scenes in the above embodiments.

[0158] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for static LOD processing of urban digital twin scenes, characterized in that, Includes the following steps: Semantic recognition and region division are performed on building components in a 3D mesh model of urban buildings to obtain a semantically labeled mesh containing multiple semantic regions. Based on the semantic annotation grid, a semantically related simplified control model is established, so that each semantic region corresponds to a corresponding simplified constraint condition. Based on the simplified control model, semantic constraint mesh simplification processing is performed on the three-dimensional mesh model of urban buildings to obtain a simplified mesh model; Perform geometric feature preservation processing on the simplified mesh model to stabilize the building outline features in the simplified mesh model; Based on the urban building 3D mesh model, the simplified mesh model of each LOD level is evaluated for error, and the local area where the error exceeds the preset threshold is corrected and controlled according to the evaluation results. Based on the simplified control model, semantic constraint mesh simplification processing is performed on the 3D mesh model of urban buildings to obtain a simplified mesh model, including the following steps: Candidate folded edges in the 3D mesh model of urban buildings are identified under the guidance of a simplified control model, and the geometric quadratic error cost of each folded edge is calculated as the basic cost component. Based on the simplified control model, the semantic regions to which the two vertices of each candidate folded edge belong are checked. If they belong to different regions, a semantic cross-region penalty coefficient is added to the basic cost. For candidate folded edges within high-fidelity semantic regions, a sparse penalty term related to the density of the remaining vertices in the region is added based on a simplified control model, so that the high-fidelity region adaptively reduces the folding rate during the simplification process. After each folding operation is completed, the cost of all candidate folding edges in the neighborhood of the affected vertex is updated according to the simplified control model, and their order is adjusted in the priority queue to ensure that the next folding operation proceeds according to the current optimal strategy. Driven by a simplified control model and a priority queue, the folding operation is executed cyclically to simplify the 3D mesh model of urban buildings while ensuring both geometric accuracy and semantic structure protection, until the preset simplification target is achieved and a simplified mesh model is obtained. The method for performing geometric feature preservation processing on simplified mesh models is as follows: Extracting building outline feature lines from a 3D mesh model of urban buildings; The feature lines of the building's outer contour are segmented and fitted with polyline to obtain polyline segments, and the endpoints of the polyline segments are used as anchor points to construct a three-dimensional feature constraint framework. During the mesh simplification process, displacement constraints are imposed on vertices falling within the influence range of the 3D feature constraint framework, and the building volume perception and outline continuity are adaptively maintained at different LOD levels. When a region containing a polyline segment loses vertex support due to simplification, a retained vertex is inserted on the polyline segment to ensure that the contour features in the simplified mesh continue to exist, while maintaining visual and structural continuity at different LOD levels.

2. The method for static LOD processing of urban digital twin scenes according to claim 1, characterized in that, The semantic recognition and region division of building components in the 3D mesh model of the city buildings are performed to obtain a semantically labeled mesh containing multiple semantic regions, including the following steps: Using triangular facets as basic units, local geometric descriptors including face normals, Gaussian curvature, mean curvature, and facet area are extracted. Construct an adjacency graph of adjacent faces, using faces as nodes and shared edges as edges to express local geometric topological relationships; A multi-layer graph convolutional network is used to perform message passing and feature aggregation on the face adjacency graph, and each face is assigned a building component category label to obtain a set of face semantic category labels. Perform connectivity analysis on the set of semantic category labels for facets, merge spatially adjacent and class-consistent facets to form preliminary semantic regions, and obtain a preliminary semantic annotation grid. For the boundary transition region, boundary segments are extracted along the shared edges of adjacent semantic regions. Hard and soft boundaries are determined based on the angle between the normal vectors of the patches and the curvature gradient. Sharp cutting and boundary alignment are performed on the hard boundaries, and gradual transition assignment is performed on the soft boundaries. Patches are assigned to adjacent regions, and finally a semantic annotation mesh with clear spatial ownership and continuous boundaries is obtained.

3. The method for static LOD processing of urban digital twin scenes according to claim 2, characterized in that, Based on the semantic annotation grid, a semantically related simplified control model is established, so that each semantic region corresponds to a corresponding simplified constraint condition, including the following steps: For each semantic region and its corresponding building component in the semantic annotation grid, feature parameters are extracted from four dimensions: relative area ratio of components, average observation frequency, structural recognition information content, and contour saliency. The weights of each dimension are determined by combining large-scale urban scene user visual attention data, and then the comprehensive visual importance score of each semantic region is calculated. Based on the comprehensive visual importance score, the semantic region is divided into three simplification constraint levels: high fidelity, standard simplification, and radical simplification. Set corresponding simplified constraints for the semantic regions of each simplified constraint level; Based on the simplified constraint levels and their simplified constraints, a semantically relevant simplified control model is established.

4. The method for static LOD processing of urban digital twin scenes according to claim 1, characterized in that, The method for error evaluation of simplified mesh models at each LOD level based on the 3D mesh model of urban buildings is as follows: Several sampling points are uniformly determined on the surface of the simplified mesh model at each LOD level; Calculate the shortest undirected point-to-surface distance from each sampling point to the surface of the 3D mesh model of the city buildings, and express the geometric distance error as the root mean square of the absolute values ​​of the distances to all sampling points. Compare the unit normal vectors of the simplified mesh model and the 3D mesh model of the city building at the same sampling point, and take the average deviation of the angle to represent the normal deflection error; The closed volume of each semantic region in the three-dimensional mesh model and the simplified model of urban buildings is calculated separately, and the volume change rate is calculated by weighting according to the simplification constraint level of each semantic region to obtain the semantic volume error; The geometric distance error, normal deflection error, and semantic volume error are normalized and weighted to obtain a comprehensive error score.

5. A static LOD processing method for urban digital twin scenes according to claim 1, characterized in that, Based on the aforementioned 3D mesh model of urban buildings, the simplified mesh model at each LOD level is evaluated for error. Based on the evaluation results, local areas where the error exceeds a preset threshold are corrected and controlled, including the following steps: Local areas where the error exceeds a preset threshold are identified, all non-compliant faces are marked in units of triangular faces, and a preset buffer range is expanded based on the spatial bounding box of the non-compliant faces to determine the correction domain. Within the correction domain, a Loop subdivision is performed on each non-compliant triangular facet by one level, so that the midpoint of each edge is uniformly subdivided into four sub-triangular faces, thereby increasing vertex density and improving the degree of freedom of local geometric adjustment. Using the surface of a local area corresponding to the 3D mesh model of urban buildings as the optimization target, the Laplacian surface deformation method is used to iteratively optimize the vertex positions within the correction domain, so that the vertex projection positions onto the surface of the 3D mesh model of urban buildings gradually move closer. When the local area error still exceeds the preset threshold, repeat the process of illegal area location, adaptive mesh refinement and local surface optimization until the local area error converges to below the preset threshold.

6. The method for static LOD processing of urban digital twin scenes according to claim 1, characterized in that, The static LOD processing method for urban digital twin scenes also includes the following steps: When the LOD level reaches the preset simplification level, determine whether there are semantic components in the current simplified mesh model that cannot retain the geometric structure; If there are semantic components that cannot preserve the geometric structure, then the semantic components are represented by alternative expressions, and the simplified mesh models of each LOD level are combined to finally obtain multi-level LOD model data; If there are no semantic components that cannot preserve the geometric structure, the simplified mesh models of each LOD level are directly integrated to obtain multi-level LOD model data.

7. A static LOD processing device for urban digital twin scenes, characterized in that, include: The memory, the processor, and the computer program stored on the memory and executable on the processor, the computer program being configured to implement the steps of the static LOD processing method for urban digital twin scenes as claimed in any one of claims 1 to 6.

8. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the static LOD processing method for urban digital twin scenes as described in any one of claims 1 to 6.

Images