An automatic contour extraction method for the triangular network model of buildings based on segmentation optimization

Abstract

The present invention discloses a method for automatically extracting the contour of a building triangular mesh model based on segmentation optimization. The method includes: Step 1, preparing input data, that is, the building triangular mesh model; Step 2, using the progressive region growth algorithm and considering planar connectivity to extract the main plane of the building model, and only projecting the main elevation onto the two-dimensional space to obtain a line segmentation vector map; Step 3, supplementing the candidate set in two ways, namely, adaptive intersection and topological regularization for line segmentation, and analyzing the element relationships in the line segmentation; Step 4, constructing a graph structure based on the candidate set, including nodes and node features, edges and edge features; Step 5, constructing a graph neural network framework based on message passing to classify the nodes in the graph; Step 6, mapping the extracted structural features back to the geometric space to obtain a vector map of the building contour. The present invention solves the problems of incomplete boundary contour structure and contour vectorization caused by occlusion and boundary blur in the three-dimensional space of buildings.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision three-dimensional reconstruction, and particularly relates to a boundary vectorization extraction technology for a building triangular mesh model based on segmentation optimization. Background Art

[0002] The building boundary footprint information is the most important data basis in many geographical applications, and is widely used in fields such as map drawing, three-dimensional building reconstruction, urban planning, and emergency response. Due to occlusion and boundary blur problems in complex scenarios, it poses challenges to the accurate positioning and extraction of the boundaries of individual buildings. How to recover the occluded contours and extract the complete vectorized contours separated from the surrounding environment from the data with missing or unclear building boundary information is the key to the automatic extraction of building boundary footprints.

[0003] It is an intuitive extraction idea to extract the straight lines fitting the boundary from building images or three-dimensional data according to the boundary characteristics of the building, and classify and analyze the extracted straight lines fitting the boundary. The existing technologies mainly adopt three techniques: the edge detection-based method uses an edge detection algorithm to extract edge features from the original data for straight line fitting; the segmentation-based method divides the building image or three-dimensional data into several regions and performs straight line fitting according to the feature information in the regions; the deep learning-based method automatically extracts boundary information by learning the feature information in the original data and performs straight line fitting and classification. Currently, traditional extraction methods rely on manually designed features and rules, and the algorithms are complex and not accurate and robust enough. Most of the rapidly developing deep learning methods in recent years are processed based on image data and have not widely involved the extraction of building boundaries from three-dimensional data. Especially in the real world, there are phenomena of boundary adhesion and ambiguity in building data of complex scenes, resulting in errors or even mistakes in the extracted boundaries compared with the real boundaries and thus unable to be put into subsequent applications. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to avoid the difficulty of existing methods in extracting the building vector polygon boundaries required for geographical and cartographic applications from occluded or boundary-unclear three-dimensional data, and to provide a method for automatically extracting the contours of a building triangular mesh model based on segmentation optimization. The strategies of line segmentation adaptive intersection and generating a candidate set according to simple topological rules adopted in the present invention can recover the missing or unclear boundaries as much as possible, and a deep learning method based on a graph neural network is used to select the final boundary vector map from the candidate set. Compared with the existing methods, the present invention can extract complete high-precision vectorized polygons from three-dimensional data occluded or with unclear boundaries in complex scenes. Therefore, this method has important application value and broad application prospects.

[0005] The technical solution adopted by the present invention to solve its technical problems is:

[0006] The present invention provides a method for automatically extracting the contour of a building triangular mesh model based on segmentation optimization, and the method includes the following steps:

[0007] Step 1, preparation of the building triangular mesh model;

[0008] Step 2, use the progressive region growth method to extract connected main plane primitives from the triangular patches. The elevation primitives are projected onto the two-dimensional space to obtain a set of line segmentation vectors, and the line segmentation set is initialized as a graph g=(v,e) as the initial line segmentation, where v is the midpoint of the line segmentation and e is the line segment between the nodes;

[0009] Step 3, first generate partial candidate line segmentations according to topological rules for the initial line segmentation obtained in Step 2, and then further expand the candidate line segmentations by adaptively intersecting to obtain a vector graph of the final candidate line segmentations;

[0010] Step 4, initialize the candidate line segmentation vector graph as a graph structure G=(L,E), each candidate line segmentation is a node of the graph, and define node features, construct an adjacency matrix to represent the connection relationship between the line segmentations, where L is the node of the graph and E is the edge of the graph;

[0011] Step 5, construct a graph neural network framework based on message passing for line segmentation classification, prepare training data and formulate a training strategy to obtain the classification result of the line segmentation;

[0012] Step 6, map the classification result of the line segmentation back to the geometric space to obtain the final building contour vector graph.

[0013] Further, the specific implementation manner of Step 2 is as follows;

[0014] Step 2.1, after encrypting the vertices of the building single triangular mesh, use the QTPS algorithm to extract the initial plane primitives based on the point cloud; then find the triangular patches whose vertices fall on the same initial plane primitive to form a plane support domain, and calculate the plane equation of each plane support domain according to the opposition inference theory; use a progressive region growth method to classify the patches that have not been assigned to the existing plane primitives. When all patches are assigned, they are the main plane primitives,

[0015] Step 2.2, project the elevations in the building main plane primitives onto the two-dimensional space, obtain a set of line segmentations of the roughly approximated contour through an edge detection algorithm, and initialize the line segmentation set as a graph g=(v,e) as the initial line segmentation, where v is the midpoint of the line segmentation and e is the line segment between the nodes.

[0016] Further, the patch classification rule satisfies the following conditions:

[0017] (1) The current patch and the adjacent unlabeled patch The normal vector angle of is less than the angle threshold, and the range is between the minimum angle threshold and π;

[0018] (2) Adjacent unlabeled patches and the current patch The normal vector angle of the corresponding plane of the supervoxel is less than the angle threshold.

[0019] Furthermore, the specific implementation method of step 3 is as follows;

[0020] Step 3.1, define the zero degree of the nodes in the graph g=(v,e), that is, the zero degree (n1, n2) of the line segment division, where v is the midpoint of the line segment division, e is the line segment between the nodes, and n1 and n2 respectively represent the number of times the two endpoints intersect with other line segment divisions. When n1 = 0 or n2 = 0 in the zero degree of the line segment division l i Then the zero degree (n1, n2) of the line segment division = 0;

[0021] Step 3.2, first generate some candidate line segment divisions using topological rules, that is, when certain conditions are met, the line segment division is rigidly transformed to the position of the current line segment division as a candidate for the current line segment division to expand the candidate set, that is, the conditions are as follows:

[0022] Query the topological relationship set of the initial line segment division in the set of line segment divisions with zero degree. When two line segment divisions l1 and l2 have a parallel relationship and the intersection l1∩:l2≠0, that is, there is a line segment division l3 between one endpoint of l1 and one endpoint of l2, and the line segment division l3 is rigidly transformed to the other endpoint of l1(l2) as the candidate set of the line segment division. The determination of the parallel relationship is as follows:

[0023]

[0024] Step 3.3, then generate a candidate set through adaptive intersection, that is, extend the set of line segment divisions {l i |i = 0,1,2…n} obtained by expanding the candidate set in step 2.2 uniformly outward along the two endpoints of the line segment until it collides with other line segment divisions. Here, n represents the number of line segment divisions. When the number of collisions of the line segment division reaches the preset threshold, the updated vector graph g'=(v',e') of the candidate line segment division is obtained.

[0025] Furthermore, the specific implementation method of step 4 is as follows;

[0026] Step 4.1, initialize the vector graph g'=(v',e') of the candidate line segment division as the graph structure G=(L,E), where v' is the midpoint of the line segment division, e' is the line segment between the nodes, L={l1,l2,…,l m} is the node of the graph, m represents the number of nodes, and use the line segment division li It is represented that \(E = \{e\) ij |i, j\in[1, n]\} are the edges of the graph, representing the relationship between line segments \(l\) i ; \(n\) is the number of line segments

[0027] Step 4.2, define the node feature \(f\) v \(\in\mathbb{R}\) D : including the position information, geometric information and context information of the line segment; construct the adjacency matrix \(E\) n×n \(=(E_1, E_2, \cdots, E\) n ) to represent the connection relationship of nodes. The vector \(E\) n×n in the adjacency matrix \(E\) i \(=(e\) i1 , e\) i2 , \cdots, e\) in )^T\), \(e\) ij , \(j\in[1, n]\), \(^T\) represents the transpose; there are common points between line segments and they are not collinear, that is, it is considered that the line segments have a connection relationship and satisfy the condition

[0028]

[0029] That is, (1) there exists a point \(m\), \(m\in l_1\) and \(m\in l_2\); (2) a pair of line segments satisfies where \(\theta\) is the angle threshold

[0030] Therefore, \(e\) ij is defined as follows

[0031]

[0032] Furthermore, the position information of the line segment includes the starting point coordinates and the ending point coordinates; the geometric information includes the length, direction vector, normal vector, and the height of the plane element to which the line segment belongs; the context information includes the number of line segments connected to the neighborhood of the line segment and the positions of the line segments connected to the endpoints of the line segment

[0033] Furthermore, in step 5, the specific method for constructing a graph neural network framework based on message passing for line segment classification is as follows

[0034] Step 5.1, define a message passing operator based on the graph neural convolutional network Take the graph \(G=(L, E)\) as the input and output the transformed graph \(G'\), that is After transformation The features and even the topology of the subsequent graph can be learned and updated; the process of message passing can be divided into two steps: message aggregation and feature update. Message aggregation refers to aggregating the features of the neighbor nodes of the current node to generate an aggregated feature vector; after obtaining the aggregated feature vector, feature update updates the features of the current node by aggregating the features of the neighborhood nodes with the features of the current node, which is achieved through a linear or non-linear update function; the message passing function is expressed as:

[0035]

[0036] where update and aggregate represent update and aggregation respectively, represents the node feature of node v at the k-th layer, represents the node feature of node v at the (k + 1)-th layer, is the set of neighbor nodes of node v, represents the node feature of node u at the k-th layer, represents that node u is a neighbor node of node v, is the aggregation function of neighbor nodes.

[0037] Step 5.2, after updating all node features, the output layer is responsible for mapping the learned node representations to the output space of the line segmentation classification task. The output layer contains a fully connected layer, which is used to map the node representations to class probabilities and output confidence scores, indicating whether the corresponding line segment is selected as the building contour;

[0038] Step 5.3, in the boundary classification task, the cross-entropy loss is used to measure the difference between the predicted class probability distribution of the model and the true class probability distribution;

[0039] Step 5.4, use the training data to train the graph neural network model, set the initial learning rate, set the batch size, and the trained graph neural network model is used for the classification task of candidate line segments.

[0040] Furthermore, in Step 5.1, the aggregation method is average aggregation or sum aggregation or maximum aggregation or attention aggregation.

[0041] Furthermore, in Step 5.3, the weighted binary cross-entropy loss is used as the loss function:

[0042]

[0043] H and are the true value and the prediction of the building edge confidence respectively, and λ is a constant.

[0044] Furthermore, the specific implementation method of Step 6 is as follows;

[0045] Step 6.1: Select the line segmentations with the category of the final building outline according to the category probability scores based on the classification results of line segmentations.

[0046] Step 6.2: Map the selected line segmentations back to the geometric space to obtain the vector graph of the final building outline.

[0047] The beneficial effects produced by the present invention are as follows:

[0048] (1) The present invention uses topological rules and a prior-based adaptive method to generate a candidate set of line segmentations, which can make up for the data defects of occluded or unclear building boundaries in complex scenes, so as to generate a complete and closed polygon outline close to the real world.

[0049] (2) Initialize the candidate set of line segmentations as the graph structure of a graph neural network, and use the learning ability of the graph neural network for features to select the final building boundary line segmentations from the candidate set, avoiding manually designing complex energy functions and improving the accuracy of the boundaries;

[0050] (3) Extract the boundary contour vector polygon based on the building triangulation data, which can make full use of the three-dimensional information of the original data and obtain the final vectorized graph, facilitating the use in applications such as geography, cartography, and 3D reconstruction. Description of the Drawings

[0051] The present invention will be further described below in conjunction with the drawings and embodiments. In the drawings:

[0052] Figure 1 is the flowchart of the embodiment of the present invention;

[0053] Figure 2 is the schematic diagram of the generation of the candidate set of line segmentations in the embodiment of the present invention; where (a) is the original triangulation model; (b) is the generation of the initial line segmentations by the two-dimensional projection of the elevation primitives; (c) is the generation of part of the candidate set of line segmentations by topological rules; (d) is the further expansion of the candidate set of line segmentations by adaptive intersection;

[0054] Figure 3 is the schematic diagram of the graph neural network in the embodiment of the present invention for classifying line segmentations to generate the vector graph of the final building boundary contour; where (a) is the initialization of the graph structure of the candidate set of line segmentations; (b) is the vector graph of the final building contour. Detailed Embodiment

[0055] The following further describes the specific technical solutions of the present invention according to the drawings and embodiments.

[0056] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0057] To solve the problem that the building boundary is occluded or unclear in complex scenarios, resulting in incomplete and inaccurate outlines, and finally the final boundary does not match the true outline, the technical solution of the present invention will be specifically described below in conjunction with the accompanying drawings and embodiments.

[0058] As Figure 1 shown, an automatic contour extraction method for a building triangular network model based on segmentation optimization according to an embodiment of the present invention includes the following steps:

[0059] Step 1, data preparation. This method requires a building triangular network model as the processing data;

[0060] Step 2, use the progressive region growth method to extract connected main plane primitives from the building triangular patches, and project the elevation primitives among them into the two-dimensional space to obtain a line segmentation vector map. The specific method is as follows:

[0061] Step 2.1, after encrypting the vertices of the building single triangular network, use the QTPS algorithm to extract the initial plane primitives based on the point cloud; then find the triangular patches whose vertices fall on the same initial plane primitive to form a plane support domain, and calculate the plane equation of each plane support domain according to the contrapositive reasoning theory; use the progressive region growth method to classify the patches that have not been assigned to the existing plane primitives into the plane support domain. When all patches are assigned, they are the main plane primitives of the building. Among them, the QTPS algorithm and the progressive region growth method are from Zhu et al. (Zhu, X., Liu, X., Zhang, Y., et al. Robust 3-D plane segmentation from airborne point clouds based on quasi-a-contrario theory[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2021, 14: 7133 - 7147).

[0062] The classification rules simultaneously satisfy the following conditions:

[0063] (1) The current patch and the adjacent unlabeled patch have a normal vector angle less than the angle threshold, and the range is between the minimum angle threshold and π;

[0064] (2) The normal vector angle between the plane of the supervoxel corresponding to the adjacent unlabeled patch and the current patch is less than the angle threshold.

[0065] Step 2.2: Project the elevation in the building main body plane elements into a two-dimensional space, obtain a line segmentation set of a rough approximate contour through an edge detection algorithm, and initialize the line segmentation set as a graph g=(v, e) as the initial line segmentation, where v is the midpoint of the line segmentation and e is the line segment between nodes, as Figure 2 shown in (b).

[0066] Step 3: First generate some candidate line segmentations from the initial line segmentation obtained in Step 2.2 according to topological rules, and then further expand the candidate line segmentations by adaptively intersecting; the generation method of the candidate line segmentations is as Figure 2 shown, and the specific method is as follows:

[0067] Step 3.1: Define the zero degree of the nodes in the graph g=(v, e), that is, the zero degree (n1, n2) of the line segmentation. n1 and n2 respectively represent the number of times the two endpoints intersect with other line segmentations. When the degree of the line segmentation l i satisfies n1 = 0 or n2 = 0, then the zero degree (n1, n2) of the line segmentation = 0;

[0068] Step 3.2: First generate some candidate line segmentations using topological rules, that is, when certain conditions are met, the line segmentation undergoes a rigid body transformation to the position of the current line segmentation as a candidate for the current line segmentation to expand the candidate set, as Figure 2 shown in (c), that is, the conditions are as follows:

[0069] Query the topological relationship set of the initial line segmentation in the set of line segmentations with zero degree. When two line segmentations l1 and l2 have a parallel relationship and the intersection l1∩l2≠0, that is, there is a line segmentation l3 between one endpoint of l1 and one endpoint of l2, and the line segmentation l3 undergoes a rigid body transformation to the other endpoint of l1 (l2) as the candidate set of the line segmentation. The determination of the parallel relationship is as follows:

[0070]

[0071] Step 3.3: Then generate a candidate set through adaptive intersection, that is, extend the line segmentation set {l i |i = 0, 1, 2…n} at both endpoints of the line segment uniformly outward until it collides with other line segmentations. Here, n represents the number of line segmentations. When the number of collisions of the line segmentation reaches a preset threshold (generally taken as 2), obtain the updated vector graph g'=(v, e) of the candidate line segmentation, as Figure 2 shown in (d).

[0072] Step 4: Initialize the candidate line segmentation vector graph as a graph structure. Each candidate line segmentation is a node of the graph, and define the node features and construct an adjacency matrix to represent the connection relationship between the line segmentations, asFigure 3 As shown in (a), the specific method is as follows:

[0073] Step 4.1: Initialize the vector graph g′=(v, e) segmented by the updated candidate lines as the graph structure G=(L, E), where L={l1, l2, …, l m} are the nodes of the graph, m represents the number of nodes, and the line segment l i indicates that there is an edge e between two midpoints v1 and v2 in (g′=(v, e)), and the edge e is called a line segment l i (the two endpoints of the line segment are v1 and v2); E={e ij |i, j ∈ [1, n]} is the edge of the graph, representing the relationship between the line segments l i . n is the number of line segments.

[0074] Step 4.2: Define the node feature f v ∈ R D : including the position information of the line segment (starting point coordinates, ending point coordinates), geometric information (such as length, direction vector, normal vector, height of the plane primitive to which the line segment belongs), and context information (the number of line segments connected to the neighborhood of the line segment and the position of the line segments connected to the endpoints of the line segment); construct the adjacency matrix E n×n =(E1, E2, …, E n ) to represent the connection relationship of the nodes. The vector E n×n in the adjacency matrix E i =(e i1 , e i2 , …, e in )′, e ij (j ∈ [1, m]), ′ represents the transpose. If there is a common point between the line segments and they are not collinear, it is considered that the line segments have a connection relationship, and the condition is satisfied as follows:

[0075]

[0076] That is, (1) there exists a point m, m ∈ l1 and m ∈ l2; (2) a pair of line segments satisfies where θ is the angle threshold (generally, 15° can be taken);

[0077] Therefore, e ij is defined as follows:

[0078]

[0079] Step 5: Construct a graph neural network framework based on message passing for line segment classification, prepare training data and formulate a training strategy to obtain the classification result of the line segments. The specific method is as follows:

[0080] Step 5.1: Define a message passing operator based on the graph neural convolutional network Take the graph G=(L, E) as the input and output the transformed graph G′, that is After transformation the features and even the topology of the graph can be learned and updated. The process of message passing can be divided into two steps: message aggregation and feature update. Message aggregation refers to aggregating the features of the neighbor nodes of the current node to generate an aggregated feature vector. The aggregation method can be average aggregation, sum aggregation, maximum aggregation, or attention aggregation, etc. These information will be used to update the representation of the target node. After obtaining the aggregated feature vector, feature update updates the feature of the current node by aggregating the features of the neighborhood nodes and the current node, which is realized through a linear or non-linear update function. The message passing function can be expressed as:

[0081]

[0082] where update and aggregate represent update and aggregation respectively, represents the node feature of node v at the k-th layer, represents the node feature of node v at the k+1-th layer, is the set of neighbor nodes of node v, represents the node feature of node u at the k-th layer, represents that node u is a neighbor node of node v, is the aggregation function of neighbor nodes.

[0083] usually contains multiple layers of message aggregation and node update layers to capture the high-order neighborhood information of the nodes in the graph and learn more complex graph structure features.

[0084] Step 5.2: After updating all node features, the output layer is responsible for mapping the learned node representation to the output space of the line segmentation classification task. The output layer usually contains a fully connected layer, which is used to map the node representation to the class probability and output the confidence score, indicating whether the corresponding line segmentation is selected as the building contour;

[0085] Step 5.3: In the boundary classification task, the cross-entropy loss is used to measure the difference between the predicted class probability distribution of the model and the true class probability distribution. The loss function uses the weighted binary cross-entropy loss:

[0086]

[0087] H and are the true value and the prediction of the building edge confidence respectively. Generally, the weight parameter λ is used to balance the problem of sample class imbalance, and generally it can be taken as 3.

[0088] Step 5.4: Use the training data to train the graph neural network model. During the training process, the parameters of the model are adjusted through steps such as forward propagation, loss calculation, backpropagation, and parameter update. Set the initial learning rate, set the batch size, and use the individual buildings in the complex environment for network training and testing according to a certain ratio. The ground truth is obtained by manually selecting the corner points of the boundary contour of each building and outlining the boundary segments. The trained graph neural network model is used for the classification task of candidate line segmentation.

[0089] Step 6: Map the classification result of line segmentation back to the geometric space to obtain the final building contour vector map as Figure 3 (b) shown. The specific generation method of the final building contour vector map is as follows:

[0090] Step 6.1: Select the line segmentations with the category of the final building contour according to the category probability scores based on the classification result of line segmentation;

[0091] Step 6.2: Map the selected line segmentations back to the geometric space to obtain the final building contour vector map.

[0092] In specific implementation, the method proposed by the technical solution of the present invention can be automatically run by those skilled in the art using computer software technology. The system device for implementing the method, such as a computer-readable storage medium storing the corresponding computer program of the technical solution of the present invention and a computer device including running the corresponding computer program, should also be within the protection scope of the present invention.

[0093] It should be understood that those skilled in the art can make improvements or transformations according to the above description, and all such improvements and transformations should fall within the protection scope of the appended claims of the present invention.

Claims

1. An automatic contour extraction method for a triangular network model of a building based on segmentation optimization, characterized in that, It includes the following steps: Step 1, prepare the building triangular mesh model; Step 2, use the progressive region growth method to extract connected main plane primitives from the triangular patches. The elevation primitives are projected into the two-dimensional space to obtain a set of line segmentation vectors, and the line segmentation set is initialized as a graph g=(v,e) as the initial line segmentation, where v is the midpoint of the line segmentation and e is the line segment between the nodes; Step 3, first generate partial candidate line segmentations from the initial line segmentation obtained in Step 2 according to topological rules, and then further expand the candidate line segmentations by adaptively intersecting them to obtain the vector graph of the final candidate line segmentations; Step 4, initialize the candidate line segmentation vector graph as a graph structure G=(L,E), each candidate line segmentation is a node of the graph, and define node features, construct an adjacency matrix to represent the connection relationship between line segmentations, where L is the node of the graph and E is the edge of the graph; Step 5, construct a graph neural network framework based on message passing for line segmentation classification, prepare training data and formulate a training strategy to obtain the classification result of the line segmentation; Step 6, map the classification result of the line segmentation back to the geometric space to obtain the final building contour vector graph.

2. The automatic contour extraction method for the triangular network model of a building based on segmentation optimization according to claim 1, characterized in that: The specific implementation method of Step 2 is as follows; Step 2.1, after encrypting the vertices of the building single triangular mesh, use the QTPS algorithm to extract the initial plane primitives based on the point cloud; then find the triangular patches whose vertices fall on the same initial plane primitive to form a plane support domain, and calculate the plane equation of each plane support domain according to the opposite reasoning theory; use the progressive region growth method to classify the patches that have not been assigned to the existing plane primitives. When all patches are assigned, they are the main plane primitives. Step 2.2, project the elevations in the building main plane primitives into the two-dimensional space, obtain a set of line segmentations with a roughly approximated contour through the edge detection algorithm, and initialize the line segmentation set as a graph g=(v,e) as the initial line segmentation, where v is the midpoint of the line segmentation and e is the line segment between the nodes.

3. The automatic contour extraction method for the triangular network model of a building based on segmentation optimization according to claim 2, characterized in that: The patch classification rules satisfy the following conditions: (1) Current patch and the normal vector angle of the adjacent unlabeled patch is less than the angle threshold, ranging from the minimum angle threshold to π; (2) Adjacent unlabeled patches and the current patch The normal vector angle of the plane of the corresponding supervoxel is less than the angle threshold.

4. The automatic contour extraction method for the triangular network model of a building based on segmentation optimization according to claim 1, characterized in that: The specific implementation method of Step 3 is as follows; Step 3.1, define the zero degree of the nodes in the graph g=(v,e), that is, the zero degree (n1, n2) of the line segment division, where v is the midpoint of the line segment division, e is the line segment between the nodes, and n1 and n2 respectively represent the number of times the two endpoints intersect with other line segment divisions. When in the zero degree of the line segment division l i n1 = 0 or n2 = 0 in the zero degree, then the zero degree (n1, n2) of the line segment division = 0; Step 3.2, first generate partial candidate line segmentations using topological rules, that is, when certain conditions are met, the line segmentation undergoes a rigid body transformation to the position of the current line segmentation as a candidate for the current line segmentation to expand the candidate set, that is, the conditions are as follows: Query the topological relationship set of the initial line segmentation in the set of line segmentations at zero degrees. When two line segmentations l1 and l2 have a parallel relationship and the intersection l1∩l2≠0, that is, there is a line segmentation l3 between one end point of l1 and one end point of l2, and the line segmentation l3 undergoes a rigid body transformation to the other end point of l1(l2) as the candidate set of the line segmentation. The determination of the parallel relationship is as follows: Step 3.3, then generate a candidate set through adaptive intersection, that is, extend the line segment set {l i |i = 0, 1, 2…n} obtained by expanding the candidate set in Step 2.2 uniformly outward along the two endpoints of the line segment until it collides with other line segments. Here, n represents the number of line segments. When the number of collisions of the line segments reaches a preset threshold, the vector graph g'=(v', e') of the updated candidate line segments is obtained.

5. The automatic contour extraction method for the triangular network model of a building based on segmentation optimization according to claim 1, characterized in that: The specific implementation method of Step 4 is as follows; Step 4.1, initialize the vector graph g′=(v′,e′) of candidate line segmentation as the graph structure G=(L,E), where v′ is the midpoint of line segmentation, e′ is the line segment between nodes, L={l1,l2,…,l m} are the nodes of the graph, m represents the number of nodes, and the line segmentation l i is represented by, E={e ij |i,j∈[1,n]} are the edges of the graph, representing the relationship between line segmentations l i ; n is the number of line segmentations; Step 4.2, define the node feature f v ∈R D : including the position information, geometric information, and context information of the line segmentation; construct the adjacency matrix E n×n =(E1, E n , …, E n ) to represent the connection relationship between nodes. The vector E n×n in the adjacency matrix E i =(e i1 , e i2 , …, e in )′, e ij , j ∈ [1, n], ′ represents transpose; if there are common points and the line segments are not collinear between the line segmentations, it is considered that there is a connection relationship between the line segmentations, satisfying the condition That is, (1) there exists a point m, m ∈ l1 and m ∈ l2; (2) a pair of line segmentations satisfies where θ is an angle threshold; Therefore, e ij is defined as follows:

6. The automatic contour extraction method for the triangular network model of a building based on segmentation optimization according to claim 5, characterized in that: The position information of the line segmentation includes the starting point coordinates and the ending point coordinates; the geometric information includes the length, direction vector, normal vector, and the height of the plane primitive to which the line segmentation belongs; the context information includes the number of line segmentations connected to the neighborhood of the line segmentation and the position of the line segmentations connected to the endpoints of the line segmentation.

7. The automatic contour extraction method for the triangular network model of a building based on segmentation optimization according to claim 1, characterized in that: In Step 5, the specific method for constructing a graph neural network framework based on message passing for line segmentation classification is as follows: Step 5.1, define a message passing operator based on a graph neural convolutional network Take the graph G = (L, E) as the input and output the transformed graph G', that is After the transformation the features and even the topology of the graph can be learned and updated; the process of message passing can be divided into two steps: message aggregation and feature update. Message aggregation refers to aggregating the features of the neighbor nodes of the current node to generate an aggregated feature vector; after obtaining the aggregated feature vector, feature update updates the features of the current node by aggregating the features of the neighborhood nodes and the features of the current node, which is realized through a linear or non-linear update function; the message passing function is expressed as: Among them, "update" and "aggregate" represent update and aggregation respectively, represents the node feature of node v at the k-th layer, represents the node feature of node v at the (k + 1)-th layer, is the set of neighbor nodes of node v, represents the node feature of node u at the k-th layer, indicates that node u is a neighbor node of node v, is the aggregation function of neighbor nodes; In step 5.2, after updating all node features, the output layer is responsible for mapping the learned node representations to the output space of the line segmentation classification task. The output layer contains a fully connected layer for mapping the node representations to class probabilities and outputting confidence scores, indicating whether the corresponding line segment is selected as the building outline. In step 5.3, the cross-entropy loss is used in the boundary classification task to measure the difference between the class probability distribution predicted by the model and the true class probability distribution. In step 5.4, the graph neural network model is trained using the training data. Set the initial learning rate and batch size. The trained graph neural network model is used for the classification task of candidate line segments.

8. The automatic contour extraction method of the triangular network model of a building based on segmentation optimization according to claim 7, characterized in that: In step 5.1, the aggregation method is average aggregation or sum aggregation or maximum aggregation or attention aggregation.

9. The automatic contour extraction method for the triangular network model of a building based on segmentation optimization according to claim 7, characterized in that: In step 5.3, the weighted binary cross-entropy loss is used as the loss function: H and are the predictions of the true value and the building edge confidence respectively, and λ is a constant.

10. The automatic contour extraction method for the triangular network model of a building based on segmentation optimization according to claim 1, characterized in that: The specific implementation of step 6 is as follows; In step 6.1, select the line segments with the class of the final building outline according to the classification results of the line segments by the class probability scores. In step 6.2, map the selected line segments back to the geometric space to obtain the final building outline vector map.

Images