Urban surface feature three-dimensional model simplification processing method, system and device and storage medium

Abstract

The embodiment of the invention discloses an urban surface feature three-dimensional model simplification processing method, system and device and a storage medium, and realizes efficient simplification of urban surface features and accurate reservation of structural information. Aiming at the problems of key feature loss, low semantic segmentation precision, easy model structure deformation and the like caused by a unified simplification strategy in a traditional method, a semantic segmentation technology based on deep learning is proposed, and a city scene is segmented into categories of buildings, vegetation and the like in combination with multi-feature fusion (geometry, color and topology) and a graph convolutional network, so that the segmentation accuracy is remarkably improved; a personalized simplification strategy is adopted, the edge collapse sequence of the building is restrained through an improved QEM algorithm, the simplification rate reaches 90% or above, a facade structure is reserved, the vegetation simplification rate exceeds 76%, and the overall form is maintained; and in combination with grid filtering and region growth optimization, noise is effectively eliminated, and the boundary is smoothed.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of urban surveying and mapping technology, specifically to a method, system, device, and storage medium for simplifying three-dimensional models of urban features. Background Art

[0002] In the field of 3D model processing, traditional methods usually adopt a uniform simplification strategy to deal with different types of terrain features, without fully considering the structural and feature differences of various terrain features such as buildings and trees in large urban scenes.

[0003] In semantic segmentation, many machine learning-based algorithms currently focus on 3D point cloud data processing. For data types like urban grid models, which have continuous surfaces and rich topological information, there is relatively little research on deep learning, resulting in suboptimal segmentation quality and efficiency. A common approach is to directly apply machine learning algorithms for point cloud data to surface grid models, lacking targeted optimization for the unique properties of grid models and applying deep learning algorithms to small-scale applications, thus lacking applications for large-scale semantic segmentation of urban 3D models.

[0004] In the model simplification stage, the classic QEM algorithm, when processing urban grid models, simply uses the sum of the distances from a point to its adjacent face as the error metric, without considering the significant differences in the structure of different urban models. This results in severe deformation of some model structures during the simplification process. In particular, when processing artificial features such as buildings, it is unable to effectively preserve their key structures, and when processing natural features such as vegetation, it is unable to effectively simplify the structure, leading to redundancy in the model structure.

[0005] The existing technology has the following problems:

[0006] 1) The unified simplification strategy cannot meet the diverse needs of urban land features. It is easy to oversimplify or lose key information of complex land features such as buildings, or undersimplify relatively simple objects such as trees, resulting in redundant model data and difficulty in accurately preserving key features.

[0007] 2) When the semantic segmentation algorithm for point cloud data is ported to the urban grid model, the topological information and continuous surface characteristics of the grid model are not fully utilized, resulting in inaccurate segmentation results and blurred boundary recognition. Subsequent simplification operations based on inaccurate semantic information are also difficult to achieve the desired effect, resulting in low model processing efficiency.

[0008] 3) When the classic QEM algorithm is applied to simplify urban models, it is not optimized in combination with the characteristics of the urban models themselves. As the simplification rate increases, the edges of the ground features in the model become unclear, the corners are missing, the structure of artificial features such as buildings is damaged, and the visual quality drops significantly. This makes it impossible to meet the dual requirements of model accuracy and efficiency for urban digital applications. Summary of the Invention

[0009] To address this, embodiments of the present invention provide a method, system, device, and storage medium for simplifying three-dimensional models of urban features, thereby solving the technical problems of existing technologies that are unable to retain key information and structural features of features to the greatest extent while reducing the amount of model data, resulting in low simplification rates, poor visual quality, weak practicality, and easy structural deformation. These technologies are also unable to meet the complexity and special requirements of urban models and are not suitable for the topological structure and continuous surface characteristics of urban grid models.

[0010] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0011] According to a first aspect of the present invention, a method for simplifying a three-dimensional model of urban features is provided, the method being applied to an urban real-scene model, comprising:

[0012] S1. The urban real scene model is over-segmented to generate uniform regions with geometric and photometric features. A graph structure is constructed using the uniform regions. Each segmented region is used as a graph node, and the relationship between regions is used as a graph edge. The segmented regions are classified using GCN and ECC, and semantic labels for terrain, vegetation, buildings, and water are output.

[0013] S2. Based on semantic tags, set differentiated simplification parameters for different land cover types. Buildings are smoothed in planar areas using a grid filtering algorithm while retaining sharp features, and vegetation is simplified at a high ratio. The current segmented area to be simplified is grown to divide the grid into planar and non-planar areas. Based on the area to which the vertex belongs, it is divided into vertices, edges, and corners and assigned corresponding weights. The QEM algorithm is used for simplification, and the edge collapse order is constrained by the corresponding weights to generate a simplified 3D model of urban land cover.

[0014] Furthermore, the urban reality model is over-segmented to generate uniform regions with geometric and photometric features, including:

[0015] The grid triangular facets of the urban real-scene model are trained using a random forest classifier based on the features of feature values, elevation, scale, density, and color to learn the probability that each facet is non-planar.

[0016] The trained random forest classifier is used to classify the triangular faces of the grid into planar and non-planar faces;

[0017] The probability expression for a patch being non-planar is:

[0018]

[0019] Among them, F iLet τ be an feature vector formed by concatenating the features of eigenvalues, elevation, scale, density, and color. Let τ be a decision tree, and let P be the prediction probability of decision tree t. t ∈[0,1], the latent label of face i is represented as L={0,1}, li=0 represents a plane, and li=1 represents a non-plane.

[0020] Furthermore, the urban reality model is over-segmented to generate uniform regions with geometric and photometric features, and the process further includes:

[0021] The non-planar probabilities of each facet are sorted from high to low to form a non-planar probability queue.

[0022] Intercalate the patches in the non-planar probability queue to region r in turn, and let A be the set of adjacent patches of r;

[0023] Initialize the label value of r and use MRF to calculate the label value of the current face in set A;

[0024] Determine if the initial label value of r is equal to the label value of the current face. If they are equal, insert the current face into region r and gradually aggregate the mesh faces into a group of locally homogeneous regions.

[0025] Gradually aggregate the mesh patches into a set of locally homogeneous regions, including:

[0026] The α-β exchange graph cutting algorithm is used to minimize the energy function U(X) to accumulate and form the currently segmented patch regions;

[0027] If no more faces can be accumulated, the segmentation growth stops and the region r is cleared;

[0028] The growth process restarts from the remaining facet with the highest planar probability until all mesh faces have been processed. This growth process is repeated.

[0029] The energy function U(X) is composed of the univariate term ψ i (x i ) and binary terms composition:

[0030]

[0031] Where A represents the neighboring face of the current growth region, x i and x r Let λ represent the binary labels that patch i and region r will receive, respectively. d and λ m To balance the weights of unary and binary terms.

[0032] Furthermore, each segmented region is treated as a graph node, and the relationships between regions are treated as graph edges. GCN and ECC are used to classify the segmented regions, outputting semantic labels for terrain, vegetation, buildings, and water bodies, including:

[0033] Obtain the node features and edge features of each segmented region, concatenate the node features and edge features and input them into the MLP. The MLP outputs a 64-dimensional feature vector as the hidden state of the GRU.

[0034] ReLU activation and normalization are applied to each hidden layer of all MLPs, and the calculated edge features are used as input to FGN, outputting edge weights;

[0035] The hidden state of the GRU is updated using the output edge weights, and the GRU is optimized using ECC. The segmentation region labels are output and the segmentation region labels L(s) are assigned to the corresponding labels. k ) is converted into surface tags L(i), which are semantic tags for terrain, vegetation, buildings, and water.

[0036] Furthermore, the building utilizes a mesh filtering algorithm to smooth planar regions while preserving sharp features, including:

[0037] The normals of the triangle faces are smoothed using a bilateral filtering algorithm, and then the mesh vertices are updated using the new normals. The process iterates between normal filtering and vertex updating, specifically including:

[0038] Obtain a mesh M and its corresponding vertex and edge set; given a triangle face, calculate the initial normal of the triangle face.

[0039] The expression for calculating the initial normal is:

[0040]

[0041] Among them, f i Let f represent the triangular faces, where v1, v2, and v3 are faces. i The three vertices;

[0042] After initializing the normals of each face, the direction of the normals is optimized using a bilateral filtering algorithm and weights based on spatial distance and normal proximity to generate new normals.

[0043] The vertex position is updated by calculating the weighted average of the surface normals. The updated vertex is then used as the basis for generating new surface normals in the next iteration. The process of initializing surface normals, filtering normals, and updating vertexes is repeated iteratively until convergence is achieved.

[0044] Furthermore, the current segmented region to be simplified is subjected to region growing, dividing the grid into planar and non-planar regions, including:

[0045] Obtain the region type of the segmented region to be simplified. If the region type is an unsegmented triangular face or a small planar region, calculate the average value of the normal direction of its neighboring regions and assign the average value to the region with the smallest included angle to form a new region. If the new region is still smaller than a preset threshold, reassign each face in the region to its neighboring region.

[0046] If the region type is a non-smooth planar boundary, the boundary faces are filtered out and the number of faces in the neighboring regions is calculated. Each boundary face will be assigned to a region with more faces.

[0047] Furthermore, the area to which a vertex belongs is divided into vertices, edges, and corners, and corresponding weights are assigned. The QEM algorithm is used for simplification, and the edge collapse order is constrained by the corresponding weights to generate a simplified 3D model of urban features, including:

[0048] A vertex is a face if it belongs to only one partition region, an edge if it belongs to two partition regions, and a corner if it belongs to three partition regions. The weights of faces, edges, and corners are set to 1, 10, and 100, respectively.

[0049] According to a second aspect of the present invention, a system for simplifying three-dimensional models of urban features is provided, the system comprising:

[0050] The urban model semantic segmentation module is used to oversegment the urban real scene model, generate uniform regions with geometric and photometric features, and construct a graph structure using the uniform regions. Each segmented region is used as a graph node, and the relationship between regions is used as a graph edge. The segmented regions are classified through GCN and ECC, and semantic labels of terrain, vegetation, buildings, and water are output.

[0051] The model simplification module is used to set differentiated simplification parameters for different land cover types based on semantic labels. Buildings are smoothed in planar areas using a grid filtering algorithm while retaining sharp features, and vegetation is simplified at a high ratio. The module performs region growing on the segmented area to be simplified, dividing the grid into planar and non-planar regions. Based on the region to which the vertex belongs, it is divided into vertices, edges, and corners and assigned corresponding weights. The QEM algorithm is used for simplification, and the edge collapse order is constrained by the corresponding weights to generate a simplified 3D model of urban land cover.

[0052] According to a third aspect of the present invention, a device for simplifying and processing three-dimensional models of urban features is provided, the device comprising: a processor and a memory;

[0053] The memory is used to store one or more program instructions;

[0054] The processor is configured to run one or more program instructions to perform the steps of a method for simplifying a three-dimensional model of urban features as described in any of the preceding claims.

[0055] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, wherein when executed by a processor, the computer program implements the steps of the method for simplifying a three-dimensional model of urban features as described in any of the preceding claims.

[0056] The embodiments of the present invention have the following advantages:

[0057] This invention achieves efficient simplification of urban features while accurately preserving structural information. Addressing the issues of key feature loss, low semantic segmentation accuracy, and easily deformable model structures caused by uniform simplification strategies in traditional methods, this invention proposes a deep learning-based semantic segmentation technique. This technique combines multi-feature fusion (geometric, color, topological) with graph convolutional networks to segment urban scenes into categories such as buildings and vegetation, significantly improving segmentation accuracy. A personalized simplification strategy is employed, using an improved QEM algorithm to constrain the edge collapse order for buildings, achieving a simplification rate of over 90% while preserving facade structure. Vegetation simplification exceeds 76% while maintaining its overall shape. Furthermore, grid filtering and region growing optimization effectively eliminate noise and smooth boundaries. Attached Figure Description

[0058] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0059] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0060] Figure 1 This is a schematic diagram of the logical structure of a simplified processing system for three-dimensional urban feature models provided in an embodiment of the present invention;

[0061] Figure 2 A flowchart illustrating a method for simplifying a 3D model of urban features, provided in an embodiment of the present invention;

[0062] Figure 3This is a schematic diagram of the urban model semantic segmentation technology route in a method for simplifying urban 3D feature models provided in an embodiment of the present invention;

[0063] Figure 4 This is a schematic diagram of the oversegmentation algorithm in a method for simplifying a 3D model of urban features provided in an embodiment of the present invention;

[0064] Figure 5 This is a schematic diagram of the segmented region classification map network structure in a method for simplifying three-dimensional urban feature models provided in an embodiment of the present invention;

[0065] Figure 6 This is a schematic diagram of the urban model simplification technology route in an urban feature 3D model simplification processing method provided in an embodiment of the present invention;

[0066] Figure 7 This is a schematic diagram of the regional growth optimization calculation process in a method for simplifying a 3D model of urban features provided in an embodiment of the present invention;

[0067] Figure 8 This is a schematic diagram showing the comparison of buildings before and after simplification in a method for simplifying a 3D model of urban features according to an embodiment of the present invention;

[0068] Figure 9 This is a schematic diagram showing the comparison of facade structure before and after simplification in a method for simplifying three-dimensional models of urban features, provided in another embodiment of the present invention.

[0069] Figure 10 This is a schematic diagram comparing data before and after simplification in a method for simplifying a 3D model of urban features, provided in another embodiment of the present invention.

[0070] Figure 11 This is a schematic diagram showing the comparison of data before and after simplification in a method for simplifying a three-dimensional model of urban features, provided in another embodiment of the present invention. Detailed Implementation

[0071] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0072] 1) AI: Artificial Intelligence refers to the ability of a computer system to simulate human intelligent activities, and is used in this invention for key steps such as land feature classification and semantic segmentation.

[0073] 2) QEM algorithm: Quadric Error Metrics is a classic 3D model simplification algorithm that simplifies the network by minimizing the quadratic error metric.

[0074] 3) GCN: Graph Convolutional Network, a deep learning model specifically designed for processing graph-structured data. In this invention, it is used to learn segmentation region features and achieve classification in the semantic segmentation of urban models.

[0075] 4) RF: Random Forest classifier is an ensemble learning algorithm based on decision trees. It consists of multiple decision trees that are used to determine the planar and non-planar properties of grid patches based on various features, providing basic information for subsequent model processing.

[0076] 5) ExMAT: External MedialAxis Transform, which establishes the connection relationship between segments by calculating the radius of the outer shrinking sphere of the model.

[0077] To address the aforementioned technical problems of failing to retain key information and structural features of ground features to the greatest extent while reducing model data volume, resulting in low simplification rate, poor visual quality, weak practicality, and easy structural deformation, which makes it difficult to meet the complexity and special requirements of urban models and is not suitable for the topological structure and continuous surface characteristics of urban grid models.

[0078] refer to Figure 1 This invention discloses a system for simplifying three-dimensional urban feature models, which includes an urban model semantic segmentation module 1 and a model simplification module 2.

[0079] Corresponding to the urban feature 3D model simplification processing system disclosed above, this invention also discloses an urban feature 3D model simplification processing method. The following details the urban feature 3D model simplification processing method disclosed in this invention, in conjunction with the urban feature 3D model simplification processing system described above.

[0080] refer to Figure 2 This invention discloses a method for simplifying 3D models of urban features, which is applied to urban real-scene models and includes:

[0081] S1. The urban real scene model is over-segmented to generate uniform regions with geometric and photometric features. A graph structure is constructed using the uniform regions. Each segmented region is used as a graph node, and the relationship between regions is used as a graph edge. The segmented regions are classified using GCN and ECC, and semantic labels for terrain, vegetation, buildings, and water are output.

[0082] S2. Based on semantic tags, set differentiated simplification parameters for different land cover types. Buildings are smoothed in planar areas using a grid filtering algorithm while retaining sharp features, and vegetation is simplified at a high ratio. The current segmented area to be simplified is grown to divide the grid into planar and non-planar areas. Based on the area to which the vertex belongs, it is divided into vertices, edges, and corners and assigned corresponding weights. The QEM algorithm is used for simplification, and the edge collapse order is constrained by the corresponding weights to generate a simplified 3D model of urban land cover.

[0083] First, the urban scene's mesh model is oversegmented, dividing the urban mesh into a set of uniform regions with geometric and photometric features. Then, a graph structure is constructed, where nodes represent segmented regions and edges represent interactions between them. Feature embedding is performed on this graph, using point cloud and handcrafted features as node and edge features, respectively. This is processed by a Multilayer Perceptron (MLP) and a Gated Recurrent Unit (GRU) to learn features representing each segmented region. Finally, Conditional Edge Convolution (ECC) is used to classify the segmented regions, converting the final fragment labels into polygon labels, thus achieving semantic segmentation of the 3D urban scene, classifying the urban model into terrain, vegetation, buildings, water, vehicles, and boats. The technical approach is as follows: Figure 3 .

[0084] Further, the urban real-scene model is over-segmented to generate uniform regions with geometric and photometric features, including: training a random forest classifier on the grid triangular faces of the urban real-scene model using features of feature values, elevation, scale, density, and color, learning the probability that each facet is non-planar; using the trained random forest classifier to classify the grid triangular faces into planar and non-planar faces; the probability expression for a facet being non-planar is:

[0085]

[0086] Among them, F i Let τ be an feature vector formed by concatenating the features of eigenvalues, elevation, scale, density, and color. Let τ be a decision tree, and let P be the prediction probability of decision tree t. t ∈[0,1], the latent label of face i is represented as L={0,1}, li=0 represents a plane, and li=1 represents a non-plane.

[0087] (1) Oversegmentation technique based on planar properties

[0088] Oversegmentation based on planar characteristics is the first step in semantic segmentation of urban models, aiming to decompose the urban grid into a set of homogeneous segmented regions based on geometric and photometric properties. This step includes two main stages: planar and non-planar classification and incremental segmentation.

[0089] Planar and Non-planar Classification: In this stage, the triangular faces of the mesh are classified into two categories: planar and non-planar. A set of features is designed to describe these faces, including eigenvalue-based features (linearity, flatness, sphericity, curvature, and perpendicularity), elevation-based features (absolute elevation, relative elevation, and multi-scale elevation), scale-based features (InMAT radius: the radius of an inwardly reduced sphere under 3D median transformation), density-based features (number of vertices and triangle face density), and color-based features (greenness and HSV histogram). These features are concatenated into a feature vector F. i For each facet i, a random forest (RF) classifier is trained to learn the probability that each facet is non-planar. Thus, a probability map is learned through the classifier instead of performing binary classification, which is used for subsequent segmentation aggregation.

[0090] Furthermore, the urban reality model is over-segmented to generate uniform regions with geometric and photometric features, and the process further includes:

[0091] The non-planar probabilities of each facet are sorted from high to low to form a non-planar probability queue.

[0092] Intercalate the patches in the non-planar probability queue to region r in turn, and let A be the set of adjacent patches of r;

[0093] Initialize the label value of r and use MRF to calculate the label value of the current face in set A;

[0094] Determine if the initial label value of r is equal to the label value of the current face. If they are equal, insert the current face into region r and gradually aggregate the mesh faces into a group of locally homogeneous regions.

[0095] Gradually aggregate the mesh patches into a set of locally homogeneous regions, including:

[0096] The α-β exchange graph cutting algorithm is used to minimize the energy function U(X) to accumulate and form the currently segmented patch regions;

[0097] If no more faces can be accumulated, the segmentation growth stops and the region r is cleared;

[0098] The growth process restarts from the remaining facet with the highest planar probability until all mesh faces have been processed. This growth process is repeated.

[0099] The energy function U(X) is composed of the univariate term ψ i (x i ) and binary terms composition:

[0100]

[0101] Where A represents the neighboring face of the current growth region, x i and x r Let λ represent the binary labels that patch i and region r will receive, respectively. d and λ m To balance the weights of unary and binary terms.

[0102] Incremental Segmentation: A graph cut method is used to segment all triangles, gradually aggregating mesh patches into a set of locally homogeneous regions using the planar and non-planar probability graphs obtained in the previous step. Region growth is achieved by solving a binary labeling problem. Starting with the patch with the highest planar probability (i.e., the current region initially has only one initial patch), if a neighboring patch i receives the same label, the patch is grown into region r. The label value of the patch in each growth iteration is computed using a Markov random field (MRF). The energy function U(X) consists of a univariate term ψ. i (x i ) and binary terms composition:

[0103]

[0104] Here, x represents the neighboring facets of the current growth region r (i.e., the facets directly connected to r). i and x r Let represent the binary labels that face i and region r will receive, respectively. A label is added to the current region only if a neighboring face receives the same label as the current region. Before minimizing the energy function, we fix the label of the current region to 0 (i.e., x). r ≡0). After optimization, when x i When λ = 0, patch i is added to r. d ≥0 and λ m A value ≥ 0 represents the weight that balances unary and binary terms. A larger λ... d This can lead to too many segments with smaller undersegmentation errors. Conversely, a larger λ... m It may result in fewer segments, but may introduce larger undersegmentation errors.

[0105] The α-β exchange graph cutting algorithm minimizes the energy function U(X) to accumulate the currently segmented patch region. If no more patches can be accumulated, the segmentation growth stops, and growth restarts from the patch with the highest planar probability among the remaining patches. The growth process is repeated until all mesh patches have been processed. The algorithm flow is as follows: Figure 4 As shown.

[0106] (2) Region segmentation and classification techniques based on graph convolutional networks

[0107] By oversegmenting, a set of segmented regions can be obtained. Each segmented region is treated as a graph node, and the relationships between regions are treated as graph edges. A graph convolutional network (GCN) is used to achieve semantic segmentation of the grid.

[0108] Node features: In the graph structure, each node represents a segmentation region and contains two types of features, which are generated based on the vertices and centroids of the segmentation region: (a) F l (s k ) 256 It is obtained using PointNet, whose input features are point clouds randomly sampled from grid vertices and face centroids, with a feature vector size of 128×6, containing XYZ and RGB information; (b)F h (s k ) 48 It consists of handcrafted features, including the same type of features used for planar and non-planar classification, and four additional shape features that capture local geometric differences, namely:

[0109]

[0110] Edge Features: Compared to graphs based on 3D points or triangular facets, graphs based on segmented regions can better capture global contextual relationships and highlight the differences in segmented regions that are often present in urban scenes. To this end, global features need to meet two conditions. First, global features should be generalizable, meaning they can be captured across different scenes. Second, global features should be established between graph nodes with significant feature differences. To fully utilize planar and non-planar segmented segments and establish meaningful relationships between segmented regions, a graph containing the following four types of edges is proposed:

[0111] (a) Parallel edges: Edges that connect parallel planar segments. Two planar segments are considered parallel if the angle between their supporting planes is less than a threshold. These edges primarily connect planar segments belonging to man-made objects.

[0112] (b) Ground Edges: Edges connecting segments and their local ground planes. A local ground plane is defined as the lowest and largest planar segment within the cylindrical neighborhood around the segment's boundary vertex. These edges primarily capture the relationships between the ground and all non-ground objects.

[0113] (c) ExMAT Edges: First, an ExMAT (i.e., the radius of the outer shrinking sphere of the 3D medium axis transformation) is constructed on the fragment, and then graph edges connecting the fragments connected by the outer shrinking sphere are introduced. Since the outer skeleton typically corresponds to the connection points between objects, ExMAT edges allow connecting adjacent fragments that belong to different objects.

[0114] (d) Spatial Nearest Neighbor Edges: First, a 3D Delaunay triangulation is constructed using the input mesh vertices and face centers. If any pair of points between two segments is connected by a Delaunay edge, then they are connected by an edge. This type of edge can encode contextual information at different scales. In particular, these edges help capture relationships between urban objects from near distance (for objects close to the ground) to far distance (for objects far from the ground).

[0115] In the edges of the graph above, we define the edge feature F. h (e k，k+1 ) = log(F h (s k ) / F h (s k+1 )), where s k and s k+1 It is by the edge e k，k+1 The two segmented regions are connected. Two additional edge features are also introduced, defined as the mean and standard deviation of the vertex offsets of the segmented region boundaries, where the offsets are defined using the nearest point pairs between the two segmented regions.

[0116] Segmentation and Classification. Based on graphs and features, we utilize Graph Convolutional Networks (GCNs) to classify segments. Node features and handcrafted features learned from PointNet are concatenated and fed into a Multilayer Perceptron (MLP), outputting a 64-dimensional feature vector as the hidden state of a Gated Recurrent Unit (GRU). ReLU activation and batch normalization are applied to each hidden layer of all MLPs. The computed edge features are used as input to a Filter Generator Network (FGN), a series of MLPs with widths of 32, 128, and 64, outputting a 64-dimensional edge feature vector. The output edge weights are then used to update the hidden state, and the GRU is optimized using Conditional Edge Convolution (ECC). The final classification result is obtained through Edge Conditional Convolution (ECC), which takes both node and edge features as input. The output segmentation region label is L(s). k Then, it is converted into face labels L(i), and the network structure is as follows: Figure 5 .

[0117] Furthermore, each segmented region is treated as a graph node, and the relationships between regions are treated as graph edges. GCN and ECC are used to classify the segmented regions, outputting semantic labels for terrain, vegetation, buildings, and water bodies, including:

[0118] Obtain the node features and edge features of each segmented region, concatenate the node features and edge features and input them into the MLP. The MLP outputs a 64-dimensional feature vector as the hidden state of the GRU.

[0119] ReLU activation and normalization are applied to each hidden layer of all MLPs, and the calculated edge features are used as input to FGN, outputting edge weights;

[0120] The hidden state of the GRU is updated using the output edge weights, and the GRU is optimized using ECC. The segmentation region labels are output and the segmentation region labels L(s) are assigned to the corresponding labels. k ) is converted into surface tags L(i), which are semantic tags for terrain, vegetation, buildings, and water.

[0121] This invention employs different simplification strategies for different land cover types. For buildings, important structural elements need to be preserved, while flat facades should be simplified as much as possible while maintaining their planar structure. To address these issues, a grid filtering algorithm is used to smooth planar areas while retaining sharp features.

[0122] For vegetation, the numerous small structures (such as leaves) result in a large number of triangular facets. These small structures don't need to be depicted in detail in urban scenes; only their basic shapes are required. Therefore, vegetation models need to be simplified significantly. For other terrain features, general simplification procedures are applied.

[0123] First, the building model is filtered, using a mesh filtering algorithm to smooth planar regions while preserving sharp features. Then, simplification calculations are performed on all types of models. First, region growing is performed, dividing the mesh into planar and non-planar regions. Based on the region to which a vertex belongs, it is classified into three categories: vertex, edge, and corner, and assigned different weights. Finally, the QEM algorithm is used for simplification. The order of edge collapse is constrained by the planar regions detected in the previous step to prevent the mesh structure from collapsing as the simplification rate increases. The specific technical route is as follows: Figure 6 As shown.

[0124] (1) Model mesh filtering technology

[0125] In the mesh filtering step, face normals and face vertex positions are optimized separately. Specifically, a bilateral filtering algorithm is used to smooth the normals of the triangular faces, and then the new normals are used to update the mesh vertices, iterating between normal filtering and vertex updating. This step modifies the geometry of the mesh model while preserving the topological relationships.

[0126] Surface normal filtering: Given a mesh M, its vertex set is represented as V(M) = {v i The edge set is represented as E(M) = {e_i = 1, ..., n}. i，j =v j -v i ;v i v j ∈V}. Where e i，j This indicates connecting two vertices v. i and v j The edges. Since the basic operation object is a triangular face, the initial normal of each face is defined as a unit vector perpendicular to the face plane.

[0127] Furthermore, the building utilizes a mesh filtering algorithm to smooth planar regions while preserving sharp features, including: smoothing the normals of triangular faces using a bilateral filtering algorithm, and then updating the mesh vertices using the new normals, iterating between normal filtering and vertex updating, specifically including: obtaining a mesh M and its corresponding vertex set and edge set, given a triangular face and calculating the initial normals of the triangular face;

[0128] The expression for calculating the initial normal is:

[0129]

[0130] Among them, f i Let f represent the triangular faces, where v1, v2, and v3 are faces. i The three vertices;

[0131] After initializing the normals of each face, the direction of the normals is optimized using a bilateral filtering algorithm and weights based on spatial distance and normal proximity to generate new normals.

[0132] After initializing the normals for each face, a bilateral filtering algorithm is used to optimize the direction of the normals. This algorithm involves two types of weights: a) spatial distance-based weights. b) Weights based on normal proximity Two faces f i and f j Spatial distance is through their centers of mass and The distance between them was calculated using the Euclidean distance. and n fj These represent the normals of the two faces. Both weight functions are defined in a Gaussian-like form, as shown below:

[0133]

[0134] Where, σ dist It is the variance parameter of proximity; θ is the user-specified angle threshold, and a smaller value produces a better discrimination effect. and Both are non-negative functions. As the distance between adjacent faces increases, decreases rapidly; when two normals become less similar, The influence of a smaller weight is smaller, and vice versa. Therefore, a smaller weight value will suppress the mutual influence between two adjacent surfaces. In general, σ dist The value is between 20° and 30°.

[0135] When extracting the neighborhood of a face, faces that cross fold regions should be avoided. An adaptive topological neighborhood query scheme is used. Each neighborhood set contains multiple faces with similar normal directions. Let... f i The neighborhood set of a query surface f is included in which each element is included if the following two conditions are met: a) it is related to the query surface f. i (a) Share at least one vertex; (b) The normal to the adjacent face is equal to f. i The dot product of the normals is greater than cosθ. Furthermore, the query surface f... i It is also included in itself In the middle. Calculate surface f i Filtering normal as follows:

[0136]

[0137] Where w i It is a normalization term that ensures the result is a unit vector. It is a face f i The area. To put it simply, It is the neighborhood The weighted average of the inner normals. For meshes with non-uniform area sizes, topological neighborhood queries can provide better results than range-based neighborhood queries because they only include neighboring faces with similar orientations. As can be seen from the above formula, the new surface normals... It is the weighted average of its neighbors. Since all weights are positive, the above formula can be intuitively interpreted as a smoothing operation on the face normal, thus inherently preserving the basic structure of the model.

[0138] Vertex position update: Calculated using a weighted average based on face normals. For a non-boundary vertex, a ring neighborhood containing at least three faces is called a face neighborhood. These faces provide reliable guidance for vertex updates. In this embodiment of the invention, the face center c fk and vertex v i connecting lines between Used to replace projection h i，k e i，k Using the exponential kernel function on h i，k There is no significant difference when projecting, so the projected value is normalized by dividing by the number of adjacent faces. Therefore, the vertex displacement value Δv i Calculated using the following formula:

[0139]

[0140] Where, |N vi (f)|0 represents the number of adjacent faces; Indicates v i adjacent face f k one of the, f k The filtered normal.

[0141] An iteration is finally completed through vertex updates: v i ←v i +Δv i The updated vertices can then generate new face normals for the next iteration. This process iteratively executes face normal initialization, normal filtering, and vertex update until convergence is achieved. Theoretically, convergence is achieved when the displacement value for each vertex falls within a specified threshold range.

[0142] The vertex position is updated by calculating the weighted average of the surface normals. The updated vertex is then used as the basis for generating new surface normals in the next iteration. The process of initializing surface normals, filtering normals, and updating vertexes is repeated iteratively until convergence is achieved.

[0143] (2) Model grid region growth optimization technique

[0144] Traditional planar region growing methods first select a seed face as the starting point, choosing a known planar region. Then, starting from the seed face, the planar region is gradually grown, adding adjacent faces whose normal vectors are close to the seed face. The addition condition is determined by calculating the angle or distance between the face and the seed plane's normal vector. The planar region growing process stops when no more faces meet the addition condition, and all faces are divided into several distinct planar regions. Although traditional planar region growing methods perform well for simple planar region segmentation, in complex scenes or with noisy mesh data, the planar segmentation results suffer from three common problems: a) unsegmented triangular faces; b) excessively small planar regions; and c) unsmooth planar boundaries.

[0145] For cases a) and b), calculate the average of the normal directions of their neighboring regions and assign them to the region with the smallest included angle. If the new region is still smaller than the threshold, each face in that region must be reassigned to one of its neighboring regions. A constraint is set to determine whether processing is necessary: ​​the minimum number of faces in each region must be no less than 10. For case c), filter out boundary faces, then calculate the number of faces in the neighboring regions, and each boundary face will be assigned to a region with more faces. The region growth optimization calculation process is as follows: Figure 7 As shown.

[0146] Furthermore, the current segmented region to be simplified is subjected to region growing, dividing the grid into planar and non-planar regions, including:

[0147] Obtain the region type of the segmented region to be simplified. If the region type is an unsegmented triangular face or a small planar region, calculate the average value of the normal direction of its neighboring regions and assign the average value to the region with the smallest included angle to form a new region. If the new region is still smaller than a preset threshold, reassign each face in the region to its neighboring region.

[0148] If the region type is a non-smooth planar boundary, the boundary faces are filtered out and the number of faces in the neighboring regions is calculated. Each boundary face will be assigned to a region with more faces.

[0149] (3) Model mesh simplification technique

[0150] The basic principle of the Quadrature Edge Collapse (QEM) algorithm is to merge the two vertices of an edge into one vertex and delete the edge and its two adjacent triangular faces. The algorithm progressively reduces the input mesh according to a cost function until the desired simplification ratio is achieved. For the cost function, the QEM algorithm is a classic method for calculating the collapse cost, using the sum of the distances from a point to its adjacent faces as an error metric. This algorithm is widely used due to its efficiency and simplification quality.

[0151] In the original QEM algorithm, each vertex is associated with a quadratic matrix Q. v Related. Given a face p = [a, b, c, d], the square of the distance from vertex v = [x, y, z, 1] to that face is:

[0152] d(v) 2 =(p T v) 2 =v T (pp T v = v T (K p )v

[0153] The sum of the squares of the distances from vertex v to all its adjacent faces is:

[0154]

[0155] For an edge l(v) i v j Its quadratic matrix is ​​obtained by adding the quadratic matrices of its two vertices, i.e. Define edge l(v) i v j The cost function is as follows:

[0156]

[0157] Where v is a new vertex.

[0158] The QEM algorithm can effectively simplify general 3D models. However, buildings have many planar areas with sharp features. If the original QEM algorithm is used for mesh simplification, these planar areas may be deformed, and the overall shape may be destroyed when the simplification rate is increased. Therefore, embodiments of the present invention detect the typical shape of the building, simplify the constraints based on these shapes, and preserve sharp features such as edges and corners. It is worth noting that a balance needs to be struck between feature preservation and simplification efficiency.

[0159] In the preceding steps, a set of segmented regions were predefined, and corresponding constraints needed to be designed for edges and corners. Vertices were classified into three types and assigned different weights based on the regions they belonged to: a vertex belonging to only one segmented region was a face; belonging to two segmented regions was an edge; and belonging to three segmented regions was a corner. The weights for faces, edges, and corners were set to 1, 10, and 100, respectively.

[0160] Furthermore, the area to which the vertex belongs is divided into vertices, edges, and corners, and corresponding weights are assigned. The QEM algorithm is used for simplification, and the edge collapse order is constrained by the corresponding weights to generate a simplified 3D model of urban features. This model includes: if a vertex belongs to only one segmented area, it is a face; if it belongs to two segmented areas, it is an edge; and if it belongs to three segmented areas, it is a corner. The weights of faces, edges, and corners are set to 1, 10, and 100, respectively.

[0161] Large-scale 3D urban model data typically exhibits high detail, large scale, and complexity in practical applications, making the processing and rendering of these models extremely time-consuming and computationally intensive. Therefore, it is necessary to simplify the 3D models to reduce data volume and improve rendering efficiency while maintaining sufficient visual quality.

[0162] 1) Figure 8 This is a set of 3D models of buildings, containing a large amount of detail and complex geometry, with a total data volume of 117,747 triangles. A categorical simplification method is used for simplification.

[0163] The simplified building model has 11,633 triangles, a simplification rate of 90.12%, significantly reducing the data volume of the building model. A visual comparison of the models before and after simplification shows that the simplification method effectively removed redundant details while preserving the main shape and features of the building. The simplified model is visually closer to the original model in quality, allowing users to better understand the building's structure.

[0164] like Figure 9 The image shows a visual comparison before and after the simplification of the facade structure. Although the number of triangular facets in the simplified model is greatly reduced, the facade structure and its detailed information remain intact, and the visualization effect is almost unaffected.

[0165] 2) Analysis of simplified vegetation results

[0166] 3D vegetation model data typically contains a large number of fine details and complex geometric structures. These details are often not critical for some application scenarios, but they can lead to massive data volumes and low rendering efficiency. (See Table 1 and...) Figure 10 , Figure 11As can be seen, the simplification method significantly reduces the amount of data in the model, effectively removing a large number of small details while maintaining the overall shape and structure of the vegetation. This demonstrates the effectiveness of the simplification method.

[0167] Table 1 Comparison of data volume before and after vegetation model simplification.

[0168] Data Number Simplify the number of slices in front Simplify the number of subsequent pieces Simplification rate Data 1 190803 45573 76.1％ Data 2 237712 22900 90.4％

[0169] The embodiments of the present invention have the following advantages:

[0170] 1) Achieve accurate identification and classification of different types of land features in large urban scenes, and formulate personalized simplification strategies for the characteristics of various land features to reduce the amount of model data while preserving key information and structural features of land features to the greatest extent.

[0171] 2) Optimize the semantic segmentation technique for urban grid models, design deep learning algorithms that adapt to their topological structure and continuous surface characteristics, improve the quality and efficiency of semantic segmentation, and provide a reliable semantic foundation for subsequent accurate simplification.

[0172] 3) Improve traditional model simplification algorithms, especially for the complexity and special characteristics of urban models. Classify and design simplification steps, optimize the simplification process, avoid structural deformation, and ensure that the model still has good visual quality and practicality under high simplification rate.

[0173] The core of this invention's embodiments is:

[0174] 1) AI-based land cover classification and personalized simplification algorithm: Utilize deep learning to accurately classify land cover, and formulate exclusive simplification strategies for different land cover such as buildings and trees to ensure that key features are retained while reducing the amount of data;

[0175] 2) Semantic segmentation technology for urban models based on deep learning: By designing a method that combines multi-feature fusion, oversegmentation and graph convolutional networks to adapt to the characteristics of urban grid models, high-precision and high-efficiency semantic segmentation is achieved, providing a reliable basis for subsequent simplification and solving the dilemma of traditional algorithms in urban grid model processing.

[0176] 3) Structural preservation urban model simplification techniques: including grid filtering, regional growth optimization and improved QEM algorithm, to ensure the structural stability of the model during the simplification process from multiple aspects, especially the protection of the key structure of the building model, and to overcome the drawbacks of classic simplification algorithms in urban model applications.

[0177] In addition, embodiments of the present invention also provide an urban feature 3D model simplification processing device, the device comprising: a processor and a memory; the memory for storing one or more program instructions; the processor for running one or more program instructions to perform the steps of an urban feature 3D model simplification processing method as described in any of the preceding embodiments.

[0178] In addition, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method for simplifying a three-dimensional model of urban features as described in any of the preceding claims.

[0179] In this embodiment of the invention, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0180] The various methods, steps, and logic diagrams disclosed in the embodiments of this invention can be implemented or executed. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor reads information from the storage medium and, in conjunction with its hardware, completes the steps of the above methods.

[0181] The storage medium can be memory, such as volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0182] Among them, non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory.

[0183] Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (Synchlink DRAM, SLDRAM), and direct memory bus RAM (DRRAM).

[0184] The storage media described in the embodiments of the present invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0185] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using a combination of hardware and software. When applied as software, the corresponding functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of computer programs from one place to another. Storage media can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0186] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for simplifying three-dimensional models of urban features, characterized in that, The method is applied to urban reality models and includes: S1. The urban real scene model is over-segmented to generate uniform regions with geometric and photometric features. A graph structure is constructed using the uniform regions. Each segmented region is used as a graph node, and the relationship between regions is used as a graph edge. The segmented regions are classified using GCN and ECC, and semantic labels for terrain, vegetation, buildings, and water are output. S2. Based on semantic tags, set differentiated simplification parameters for different land cover types. Buildings are smoothed in planar areas using a grid filtering algorithm while retaining sharp features, and vegetation is simplified at a high ratio. The current segmented area to be simplified is grown to divide the grid into planar and non-planar areas. Based on the area to which the vertex belongs, it is divided into vertices, edges, and corners and assigned corresponding weights. The QEM algorithm is used for simplification, and the edge collapse order is constrained by the corresponding weights to generate a simplified 3D model of urban land cover.

2. The method for simplifying three-dimensional models of urban features as described in claim 1, characterized in that, The urban reality model is over-segmented to generate uniform regions with geometric and photometric features, including: The grid triangular facets of the urban real-scene model are trained using a random forest classifier based on the features of feature values, elevation, scale, density, and color to learn the probability that each facet is non-planar. The trained random forest classifier is used to classify the triangular faces of the grid into planar and non-planar faces; The probability expression for a patch being non-planar is: Among them, F i Let τ be an feature vector formed by concatenating the features of eigenvalues, elevation, scale, density, and color. Let τ be a decision tree, and let P be the prediction probability of decision tree t. t ∈[0,1], the latent label of face i is represented as L={0,1}, li=0 represents a plane, and li=1 represents a non-plane.

3. The method for simplifying three-dimensional models of urban features as described in claim 2, characterized in that, The process of over-segmenting the urban reality model to generate uniform regions with geometric and photometric features also includes: The non-planar probabilities of each facet are sorted from high to low to form a non-planar probability queue. Intercalate the patches in the non-planar probability queue to region r in turn, and let A be the set of adjacent patches of r; Initialize the label value of r and use MRF to calculate the label value of the current face in set A; Determine if the initial label value of r is equal to the label value of the current face. If they are equal, insert the current face into region r and gradually aggregate the mesh faces into a group of locally homogeneous regions. Gradually aggregate the mesh patches into a set of locally homogeneous regions, including: The α-β exchange graph cutting algorithm is used to minimize the energy function U(X) to accumulate and form the currently segmented patch regions; If no more faces can be accumulated, the segmentation growth stops and the region r is cleared; The growth process restarts from the remaining facet with the highest planar probability until all mesh faces have been processed. This growth process is repeated. The energy function U(X) is composed of the univariate term ψ i (xi) and binary terms composition: Where A represents the neighboring face of the current growth region, x i and x r Let λ represent the binary labels that patch i and region r will receive, respectively. d and λ m To balance the weights of unary and binary terms.

4. The method for simplifying three-dimensional models of urban features as described in claim 3, characterized in that, Each segmented region is treated as a graph node, and the relationships between regions are treated as graph edges. GCN and ECC are used to classify the segmented regions, outputting semantic labels for terrain, vegetation, buildings, and water bodies, including: Obtain the node features and edge features of each segmented region, concatenate the node features and edge features and input them into the MLP. The MLP outputs a 64-dimensional feature vector as the hidden state of the GRU. ReLU activation and normalization are applied to each hidden layer of all MLPs, and the calculated edge features are used as input to FGN, outputting edge weights. The hidden state of the GRU is updated using the output edge weights, and the GRU is optimized using ECC. The segmentation region labels are output and the segmentation region labels L(s) are assigned to the corresponding labels. k ) is converted into surface tags L(i), which are semantic tags for terrain, vegetation, buildings, and water.

5. The method for simplifying three-dimensional models of urban features as described in claim 1, characterized in that, The building utilizes a grid filtering algorithm to smooth planar regions while preserving sharp features, including: The normals of the triangle faces are smoothed using a bilateral filtering algorithm, and then the mesh vertices are updated using the new normals. The process iterates between normal filtering and vertex updating, specifically including: Obtain a mesh M and its corresponding vertex and edge set; given a triangle face, calculate the initial normal of the triangle face. The expression for calculating the initial normal is: Among them, f i Let f represent the triangular faces, where v1, v2, and v3 are faces. i The three vertices; After initializing the normals of each face, the direction of the normals is optimized using a bilateral filtering algorithm and weights based on spatial distance and normal proximity to generate new normals. The vertex position is updated by calculating the weighted average of the surface normals. The updated vertex is then used as the basis for generating new surface normals in the next iteration. The process of initializing surface normals, filtering normals, and updating vertexes is repeated iteratively until convergence is achieved.

6. The method for simplifying three-dimensional models of urban features as described in claim 5, characterized in that, Perform region growing on the current segmented region to be simplified, dividing the mesh into planar and non-planar regions, including: Obtain the region type of the segmented region to be simplified. If the region type is an unsegmented triangular face or a small planar region, calculate the average value of the normal direction of its neighboring regions and assign the average value to the region with the smallest included angle to form a new region. If the new region is still smaller than a preset threshold, reassign each face in the region to its neighboring region. If the region type is a non-smooth planar boundary, the boundary faces are filtered out and the number of faces in the neighboring regions is calculated. Each boundary face will be assigned to a region with more faces.

7. The method for simplifying three-dimensional models of urban features as described in claim 6, characterized in that, The city is divided into vertices, edges, and corners based on their respective regions and assigned corresponding weights. The QEM algorithm is used for simplification, and the edge collapse order is constrained by the corresponding weights to generate a simplified 3D model of urban features, including: A vertex is a face if it belongs to only one partition region, an edge if it belongs to two partition regions, and a corner if it belongs to three partition regions. The weights of faces, edges, and corners are set to 1, 10, and 100, respectively.

8. A system for simplifying and processing three-dimensional models of urban features, characterized in that, The system includes: The urban model semantic segmentation module is used to oversegment the urban real scene model, generate uniform regions with geometric and photometric features, and construct a graph structure using the uniform regions. Each segmented region is used as a graph node, and the relationship between regions is used as a graph edge. The segmented regions are classified through GCN and ECC, and semantic labels of terrain, vegetation, buildings, and water are output. The model simplification module is used to set differentiated simplification parameters for different land cover types based on semantic labels. Buildings are smoothed in planar areas using a grid filtering algorithm while retaining sharp features, and vegetation is simplified at a high ratio. The module performs region growing on the segmented area to be simplified, dividing the grid into planar and non-planar regions. Based on the region to which the vertex belongs, it is divided into vertices, edges, and corners and assigned corresponding weights. The QEM algorithm is used for simplification, and the edge collapse order is constrained by the corresponding weights to generate a simplified 3D model of urban land cover.

9. A device for simplifying and processing three-dimensional models of urban features, characterized in that, The device includes: a processor and a memory; The memory is used to store one or more program instructions; The processor is configured to run one or more program instructions to perform the steps of a method for simplifying a three-dimensional model of urban features as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method for simplifying a three-dimensional model of urban features as described in any one of claims 1 to 7.

Images