An instance segmentation and structured topology reconstruction method for complex geometry buildings

Abstract

The application discloses a kind of instance segmentation and structured topological reconstruction methods of complex geometric building, it is related to three-dimensional model generation technical field, comprising the following steps: S1, building point cloud instance segmentation: using SoftGroup network to the semantic analysis of unmanned aerial vehicle LiDAR point cloud, output contains the independent component instance of each building component semantic label, the fine segmentation of each building component is realized by point cloud instance mask;S2, geometric primitive extraction: according to semantic label, group instance point cloud is carried out, and differential geometric fitting is executed, the application is realized by semantic-geometric coupling modeling The unified expression of irregular structure and regular component, solve the segmentation and parameterization bottleneck in complex building scene in traditional method, also adopt the boundary expression of elliptic curve segment simplification conical top or hemispherical top and support structure, reduce the complexity of calculation, while guarantee the integrity of building skeleton.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional model generation, and in particular to an instance segmentation and structured topological reconstruction method for complex geometric buildings. BACKGROUND

[0002] The instance segmentation and structured topological reconstruction of complex geometric buildings is a technology for extracting semantic building components from unordered point clouds, constructing geometric constraint relationships, and generating regularized three-dimensional models. It has important application value in the fields of historical building digital protection, urban planning, intelligent construction, etc. Through this technology, semantic analysis and parameterized expression of building components can be realized, providing high-precision basic data for structural health monitoring, heritage restoration, and BIM modeling.

[0003] However, existing methods still have significant defects when dealing with scenarios where multiple types of geometric primitives (such as conical roofs, hemispherical roofs) and regular structures (planes, cylinders) coexist:

[0004] (1) Traditional point cloud segmentation methods lack the ability to distinguish between irregular structures (conical roofs, hemispherical roofs) and regular components, resulting in component segmentation adhesion, and traditional geometric primitive fitting strategies are difficult to accurately express the geometric characteristics of complex components;

[0005] (2) The geometric constraint rules in the topological reconstruction process are single, ignoring the multi-level directional combination characteristics of building structures, resulting in poor geometric consistency of the model and frequent topological fractures;

[0006] Therefore, an instance segmentation and structured topological reconstruction method for complex geometric buildings is proposed to solve the above problems. SUMMARY

[0007] The purpose of the present application is to solve the problems in the prior art by providing an instance segmentation and structured topological reconstruction method for complex geometric buildings.

[0008] An instance segmentation and structured topological reconstruction method for complex geometric buildings includes the following steps:

[0009] S1, Building point cloud instance segmentation: using a SoftGroup network to perform semantic analysis on unmanned aerial vehicle LiDAR point clouds, outputting independent component instances containing semantic labels of each building component, and achieving fine segmentation of each building component through point cloud instance masks;

[0010] S2, Geometric primitive extraction: grouping instance point clouds according to semantic labels and performing differential geometric fitting;

[0011] S3, Adjacency matrix construction: generating an adjacency matrix A = [A ij ] based on the spatial proximity of instance masks, determining Ii and I j Two instance adjacency relationship;

[0012] S4, boundary line extraction: for any two adjacent structural instances I i and I j , fracture line modeling is performed according to the structural type combination thereof;

[0013] S5, feature vertex extraction: building key connection points are extracted through multi-plane intersection solution and cylinder end point interception, combined with bounding box corner point extraction auxiliary geometric boundary expression, to form a structured topological connection point set together;

[0014] S6, boundary regularization constraint: a hierarchical constraint model is constructed through geometric grouping, main direction extraction and joint optimization.

[0015] Preferably, in the step S1, a SoftGroup network is used, an instance segmentation threshold τ = 0.8, and a minimum bounding box size of the instance point cloud ≥ 1m 3 , to eliminate noise components smaller than the threshold.

[0016] Preferably, in the step S2, the differential geometric fitting includes:

[0017] For planar semantic components, a RANSAC algorithm is used to fit a plane equation:

[0018] a k x+b k y+c k z+d k = 0

[0019] to obtain a normal vector n k = (a k , b k , c k ), wherein a k , b k , c k , d k are plane parameters;

[0020] For cylindrical semantic components, a cylindrical axial vector and a radius are fitted to construct a cylinder parameter model;

[0021] For conical roof and hemispherical roof semantic components, the highest point coordinates are extracted as the structural center, and the enclosing sphere radius is recorded.

[0022] Preferably, in the step S3, the determination of the adjacency relationship between two instances I i and I j includes calculating the minimum Euclidean distance min||p-q||(p∈I i , q∈I j ) between their point clouds.

[0023]

[0024] where A ij is the adjacency matrix element, indicating whether instances I i and I j are adjacent, and ε is the spatial proximity threshold.

[0025] Preferably, in step S4, for any two adjacent structural instances I i and I j , the fracture line modeling according to the structural type combination thereof comprises:

[0026] For the inter-planar intersection line, the two plane equations π i and π j are solved simultaneously:

[0027] π i : ax + by + cz + d = 0 i i i i

[0028] π j : ax + by + cz + d = 0 j j j j

[0029] The parametric intersection line L ij (t) = P0 + t·(n i × n j ), where t is the position of a point on the line, n i , n j are the normal vectors of the planes π i and π j , and P0 is an arbitrary point on the intersection line of the two planes;

[0030] For the intersection line of a conical or hemispherical roof and a support structure boundary, an elliptic curve segment is used for approximation, and the equation in the local coordinate system is: Connecting the structural vertices V k forms an approximate conical skeleton, the major axis of the elliptic curve segment is taken as the enclosing sphere radius R k , the minor axis is taken as 0.8R k , and the elliptic parameters are optimized by the Levenberg-Marquardt algorithm.

[0031] Preferably, in step S5, for the case of roof three-plane intersection, the plane equations are solved simultaneously to construct a linear equation group, and the Cramer's rule is used to directly solve the three-dimensional space intersection point coordinates V = π1∩π2∩π3, where π i ​​​​​​​​For the planar equation, for the instance with the semantic label of cone top or hemisphere top, the vertex coordinates or the sphere center of the fitting model are directly taken; for the instance with the semantic label of cylinder, the upper and lower endpoints are taken along the cylinder axis direction; for the instance with the semantic label of vertical wall surface, the point cloud is projected to a two-dimensional plane in the direction of the normal vector, an axial bounding box is calculated, and four corner points are extracted as the contour vertices.

[0032] Preferably, in the step S6, the constructing the hierarchical constraint model comprises: firstly, performing geometric grouping on the boundary lines, and dividing them into orthogonal constraint groups G ortho and free-form groups G free ; then, extracting the main direction for the orthogonal constraint groups G ortho , performing K-means clustering (k=2) on the boundary direction angle θ ortho in G i , forcing the orthogonal constraint θ2=θ1+π / 2, and generating a main direction vector set Finally, performing joint optimization on the extracted building boundary lines: defining an objective function E total =E angle +λE conn , wherein: forcing the alignment of the main direction of the boundary of the orthogonal group; maintaining the continuity of the end points through an adjacency matrix; and using the Levenberg-Marquardt algorithm to optimize the boundary end point coordinates, and the free group G free only participates in the calculation of E conn to maintain the topological connectivity, and finally obtaining a building contour structure connected at the beginning and the end and meeting the topological connectivity rule.

[0033] Compared with the existing technology, the present application has the following advantages:

[0034] 1. The present application realizes the unified expression of special-shaped structures and regular components through semantic-geometric coupling modeling, and solves the segmentation and parameterization bottleneck of traditional methods in complex building scenes.

[0035] 2. The present application simplifies the boundary expression of the cone top or hemisphere top and the supporting structure by using an elliptic curve segment, reduces the calculation complexity, and at the same time guarantees the integrity of the building skeleton.

[0036] 3. The present application groups the boundaries according to geometric characteristics, applies main direction constraints to the orthogonal group, and retains geometric degrees of freedom for the free group, and through joint optimization of the coupling direction alignment and topological continuity, overcomes the geometric distortion of inclined surfaces caused by traditional global orthogonal constraints. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 is a principle block diagram of the present application.

[0038] Figure 2 is a general workflow diagram of the present application.

[0039] Figure 3 is an example segmentation flowchart of the present application.

[0040] Figure 4 is a component encoding diagram after example segmentation of the present application.

[0041] Figure 5 is an adjacency matrix diagram of the present application.

[0042] Figure 6 is a plane boundary fitting schematic diagram of the present application.

[0043] Figure 7 is a semisphere top boundary fitting schematic diagram of the present application.

[0044] Figure 8 is a boundary regularization before and after schematic diagram of the present application. DETAILED DESCRIPTION

[0045] In order to make the technical means, creative features, purposes and effects of the present application easy to understand, the present application will be further described below in conjunction with specific embodiments.

[0046] Referring to Figures 1-2 Fig. 1, an example segmentation and structured topological reconstruction method of a complex geometric building includes the following steps:

[0047] S1, building point cloud instance segmentation: a SoftGroup network is used for semantic analysis of unmanned aerial vehicle LiDAR point cloud, and independent component instances containing semantic labels of each building component are output, fine segmentation of each building component is realized through point cloud instance masks, and each building component includes a roof plane, a roof slope, a vertical wall, a decorative column, a semisphere top, a conical top, etc.

[0048] S2, geometric primitive extraction: according to the semantic labels, the instance point cloud is grouped, and differential geometric fitting is performed;

[0049] S3, adjacency matrix construction: an adjacency matrix A = [A ij ] is generated based on the spatial proximity of the instance masks, and the adjacency relationship of two instances I i and I j is determined;

[0050] S4, boundary line extraction: for any two adjacent structure instances I i and I j , fracture line modeling is performed according to the structure type combination;

[0051] S5, feature vertex extraction: extract building key connection points through multi-plane intersection solution and cylinder endpoint interception, combined with bounding box corner extraction to assist geometric boundary expression, together forming a structured topology connection point set;

[0052] S6, boundary regularization constraint: construct a hierarchical constraint model through geometric grouping, principal direction extraction and joint optimization.

[0053] In step S1, the SoftGroup network is used, the instance segmentation threshold τ = 0.8, and the minimum bounding box size of the instance point cloud is ≥1m 3 , eliminating noise components with a size less than the threshold, the SoftGroup network can simultaneously process irregular structures (conical top, hemispherical top) and regular structures (plane, cylinder), and through point cloud instance mask, the fine segmentation of each building component is realized, eliminating the adhesion phenomenon between components.

[0054] In step S2, the differential geometric fitting includes:

[0055] For plane semantic components (roof plane, roof slope, vertical wall), the RANSAC algorithm is used to fit the plane equation:

[0056] a k x+b k y+c k z+d k =0

[0057] The normal vector n k =(a k ,b k ,c k ), where a k , b k , c k , d k are plane parameters;

[0058] For column semantic components (ornamental columns), fit the cylindrical axial vector and radius to construct the cylinder parameter model;

[0059] For conical top and hemispherical top semantic components, extract the highest point coordinates as the structure center and record the enclosing sphere radius.

[0060] In step S3, the determination of the adjacency relationship between I i and I j includes calculating the minimum Euclidean distance between their point clouds min||p-q||(p∈I i , q∈I j ):

[0061]

[0062] Where, Aij For adjacent matrix elements, indicates instance I i and I j are adjacent, and ε is a spatial proximity threshold.

[0063] In the step S4, for any two adjacent structural instances I i and I j , the fracture line modeling according to the structural type combination thereof comprises:

[0064] For the inter-plane intersection line, the two plane equations π i and π j are solved simultaneously:

[0065] π i : a i x+b i y+c i z+d i = 0

[0066] π j : a j x+b j y+c j z+d j = 0

[0067] The parameterized intersection line L ij (t) = P0+t·(n i ×n j ), wherein t is a position indicating a point on the line, n i and n j are normal vectors of the planes π i and π j , and P0 is an arbitrary point on the intersection line of the two planes;

[0068] For the intersection line between the conical or hemispherical roof and the support structure boundary, an elliptic curve segment is adopted for approximation, and the equation in the local coordinate system is: Connecting the structural vertexes V k to form an approximate conical skeleton, the major axis of the elliptic curve segment is taken as the surrounding sphere radius R k , the minor axis is taken as 0.8R k , and the elliptic parameters are optimized through the Levenberg-Marquardt algorithm.

[0069] In the step S5, for the case of roof three-plane intersection, a linear equation group is constructed by solving the plane equations simultaneously, and the Cramer rule is directly used to solve the three-dimensional space intersection point coordinates V = π1∩π2∩π3, wherein π iThe equation is for a plane. For instances with semantic labels of cone or hemispherical top, the vertex coordinates or center of the sphere of the fitted model are directly taken, with the coordinate error not exceeding 5% of the radius of the bounding sphere. For instances with semantic labels of cylinder, the upper and lower endpoints are taken along the cylinder axis. For instance point clouds with semantic labels of vertical wall, the projection is onto a two-dimensional plane in the direction of the normal vector, the axial bounding box is calculated, and the four corner points are extracted as contour vertices, with the corner point spacing error <0.02m.

[0070] In step S6, constructing the hierarchical constraint model includes: first, geometrically grouping the boundary lines into orthogonal constraint groups G. ortho (Including roof plan and vertical wall) and freeform group G free (Including roof slopes, etc.); then, for the orthogonal constraint group G ortho Extract the principal direction for G ortho Middle boundary direction angle θ i Perform K-means clustering (k=2), enforce orthogonality constraint θ2=θ1+π / 2, and generate a set of principal direction vectors. The objective function is defined as a weighted sum of a direction alignment term (calculating the least squares deviation between the orthogonal group boundary direction vector and the principal direction) and an adjacent endpoint continuity term (constraining the endpoint coincidence error of adjacent boundaries to be less than 0.05m) (the weight of the direction term is ≥60%). Finally, the extracted building boundary lines are jointly optimized: the objective function E is defined as follows. total =E angle +λE conn ,in: Force orthogonal group boundaries to align with the main direction; Endpoint continuity is maintained using an adjacency matrix; the Levenberg-Marquardt algorithm is used to optimize the boundary endpoint coordinates. The algorithm terminates when the total energy function decrease ΔE < 1e-4 or the number of iterations exceeds 100, and the free group G... free Only participate in E conn The calculations are performed to maintain topological connectivity, ultimately resulting in a building outline structure that is connected end-to-end and conforms to the rules of topological connectivity.

[0071] Example

[0072] S1, Building Point Cloud Instance Segmentation

[0073] Given a point cloud The network first predicts the semantic labels of each point in parallel. and offset vector toward the instance center Semantic branching is achieved through cross-entropy loss. To optimize classification capabilities, the offset branch uses L1 loss. Constraint geometric consistency:

[0074]

[0075] where N is the total number of points in the point cloud, p i is the coordinate of a point in 3D space, p i = (x i , y i , z i ); s i is the semantic label of the i-th point, is the set of semantic label categories, δ i is the offset vector of point p i to the center of the instance it belongs to, denotes the 3D real space, is the geometric center of the instance that point p i belongs to, y i,c is the semantic ground truth label.

[0076] SoftGroup introduces a probability-aware soft grouping strategy, for each seed point p j , it dynamically aggregates neighboring points p i that satisfy the following conditions:

[0077]

[0078] When the above three points are satisfied, p i , p j belong to the same instance, where d(p i , p j ) is the Euclidean distance between points p i and p j ; r dynamic = a · std({||p i -p j ||2}) is the adaptive radius, where a is the scaling factor, std is the standard deviation of the Euclidean distance between all pairs of points in the point cloud; sim(·) is the semantic similarity function, τ s and τ δ are learnable thresholds.

[0079] For the generated candidate regions, a lightweight U-Net network is used to predict the instance score S m ∈ [0, 1], and the non-maximum suppression (NMS) is used to merge overlapping candidate instances, the network structure is shown in Figure 3 , and the instance mask M m is determined by the following formula:

[0080]

[0081] where, is the seed point pj The soft neighborhood, where γ is the confidence threshold.

[0082] S2, Geometric Element Extraction

[0083] For multiple components of the same semantic type, independent fitting is required based on instance labels, with each instance I... k For a given structural unit, the fitting method is as follows:

[0084] For planar semantic components (including upright walls, roof planes, and roof slopes), RANSAC is used to fit the plane equations:

[0085] a k x+b k y+c k z+d k =0

[0086] Obtain the normal vector n k =(a k ,b k ,c k ), where a k b k c k d k For plane parameters.

[0087] For the semantic component of a column, fit the column axis v. k With radius r k .

[0088] For cone-shaped or hemispherical dome-shaped components, the highest point is extracted as the structural center, and the radius of the enclosing sphere and the information of the highest point are recorded.

[0089] S3, Adjacency Matrix Construction

[0090] In the topological relationship modeling of building components, the adjacency matrix is ​​generated based on spatial proximity constraints, defining two structural instances I. i and I j The adjacency matrix, such as Figure 4 and Figure 5 As shown:

[0091]

[0092] Where A ij The adjacency matrix element represents instance I. i and I j Whether they are adjacent. ε is the spatial proximity threshold.

[0093] S4, Boundary Line Extraction

[0094] After the geometric fitting and adjacency relationship determination of the structural surface patches, the spatial relationship between adjacent structures needs to be further converted into explicit boundary line segments and feature vertices to construct the architectural topology skeleton.

[0095] For any two adjacent structural instances I i and I j , the fracture line modeling is performed according to the combination of their structural types. For planar structures (including vertical wall surfaces, roof planes, and roof slopes), the fitting equations of the two surface patches are directly solved:

[0096] π i : a i x + b i y + c i z + d i = 0

[0097] π j : a j x + b j y + c j z + d j = 0

[0098] The parametric form of the intersection line is obtained:

[0099] L ij (t) = P0 + t · (n i × n j )

[0100] where t represents the position of a point on the line, n i and n j are the normal vectors of the planes π i and π j , and P0 is an arbitrary point on the intersection line of the two planes.

[0101] To simplify the geometric modeling cost of complex curved surface structures while ensuring the integrity and clarity of the architectural skeleton structure, for the combination of cylindrical and conical or hemispherical roof structures in a specific architectural point cloud, the following approximate method is used for structure expression: for the instance of a conical or hemispherical roof, the highest point is extracted as the structure vertex V k , and the adjacent column and plane roof structures are found using the adjacency matrix, which are considered as the support boundaries below the vertex; for the intersection line between the column and the plane, or between the hemispherical roof and the plane, an elliptic curve segment is used to fit the boundary line, representing the approximate boundary of the conical base, which is expressed in the local coordinate system as:

[0102]

[0103] where (x0, y0) are the coordinates of the center of the projected ellipse of the conical base, and a and b are the lengths of the semi-axes of the ellipse.

[0104] The boundary elliptic curve segment is connected with the highest point to construct an approximate conical structure, avoiding reconstruction of a complete free-form surface, while being capable of participating in topology structure optimization, intersection line extraction between planes, and boundary line extraction of a hemispherical top, as shown in Figure 6 and Figure 7 .

[0105] S5. Feature vertex extraction

[0106] To form a structured topology model, corner points need to be extracted from the intersection lines as line segment endpoints or face patch connection points, and the extraction method is as follows:

[0107] For the case of three planes intersecting the roof, the intersection point coordinates V = π1∩π2∩π3 are obtained by solving three plane equations, where π i is a plane equation; for instances with semantic labels of conical top or hemispherical top, the vertex coordinates of the fitting model or the center of the sphere are directly taken; for a column, the upper and lower endpoints are taken along the column axis direction; the vertical plane boundary points are extracted, that is, the axial bounding box of the two-dimensional projection of the face patch point set is calculated, and four corner points are extracted as the contour vertices.

[0108] S6. Boundary regularization constraint

[0109] The boundaries are grouped by geometric primitive types, and an optimization model is established to jointly constrain the boundary direction and connectivity, and the method flow is as follows:

[0110] Let the input boundary set be Each boundary L i is associated with a geometric primitive type t i ; according to the type characteristics, the boundaries are divided into an orthogonal constraint group G ortho (containing types such as planes and facades that require orthogonal / parallel constraints) and a free-form group G free (containing types such as inclined planes that do not require direction constraints), satisfying:

[0111]

[0112] (1) The main direction analysis is performed on the boundaries in G ortho :

[0113] 1) The direction vector of each boundary is calculated , where is the endpoint coordinate;

[0114] 2) The direction angle θ i = arctan(v i ) is K-means clustered (k = 2) to obtain the main direction angles θ1 and θ2, satisfying the orthogonal constraint θ2 = θ1 + π / 2;

[0115] 3) The main direction unit vector set where d k = (cos k , sin k ).

[0116] (2) Define the objective function, including the direction regularization term and the connectivity constraint term:

[0117] 1) Direction regularization term: force G ortho to align with the main direction of the boundary;

[0118]

[0119] 2) Connectivity constraint term: maintain the consistency of the endpoints of adjacent boundaries through the adjacency matrix A;

[0120]

[0121] 3) The total optimization objective is in the form of weighted summation:

[0122] E total = E angle + λE conn

[0123] where λ is the balance weight coefficient, and the optimization variable is the endpoint coordinates of all boundaries

[0124] (3) Convert E total into a nonlinear least squares problem and use the Levenberg-Marquardt algorithm for iterative optimization.

[0125] The boundaries of the free group G free only participate in the calculation of E conn to maintain topological connectivity, and their directions are not explicitly constrained, finally obtaining a building contour structure that is connected at the beginning and end and meets the topological connectivity rules, as shown in Figure 8 .

[0126] From the technical common sense, the present application can be realized by other embodiments without departing from the spirit or essential characteristics thereof. Therefore, the above disclosed embodiments, in all aspects, are only illustrative and not the only. All changes within the scope of the present application or within the scope equivalent to the present application are included in the present application.

Claims

1. A method for instance segmentation and structured topology reconstruction of buildings with complex geometric shapes, characterized in that: Includes the following steps: S1. Building point cloud instance segmentation: The SoftGroup network is used to perform semantic parsing on the UAV LiDAR point cloud, and outputs independent component instances containing semantic labels of each building component. Fine segmentation of each building component is achieved through point cloud instance mask. S2. Geometric primitive extraction: Group instance point clouds according to semantic labels and perform differential geometric fitting; S3. Adjacency Matrix Construction: Generating an adjacency matrix based on the spatial proximity of instance masks. ,determination The adjacency relationship between the two instances; S4. Boundary Line Extraction: For any two adjacent structural instances Fracture lines are modeled based on combinations of structural types. S5. Feature Vertex Extraction: Key connection points of the building are extracted by solving multi-plane intersection lines and intercepting the endpoints of the column. Combined with the extraction of corner points of the bounding box to assist in the geometric boundary expression, a structured set of topological connection points is formed. S6. Boundary regularization constraints: A hierarchical constraint model is constructed through geometric grouping, principal direction extraction, and joint optimization.

2. The method for instance segmentation and structured topology reconstruction of complex geometric buildings according to claim 1, characterized in that: In step S1, a SoftGroup network is used, the instance segmentation threshold τ = 0.8, and the minimum bounding box size of the instance point cloud is ≥ 1m. 3 Eliminate noisy components whose size is smaller than the threshold.

3. The method for instance segmentation and structured topology reconstruction of complex geometric buildings according to claim 1, characterized in that: In step S2, performing differential geometric fitting includes: The RANSAC algorithm is used to fit the plane equation for planar semantic components: Obtain the normal vector ,in For planar parameters; Fit the axial vector and radius of the cylinder to the semantic component of the column, and construct the column parametric model; For semantic components such as conical tops and hemispherical tops, the coordinates of the highest point are extracted as the structural center, and the radius of the enclosing sphere is recorded.

4. The method for instance segmentation and structured topology reconstruction of complex geometric buildings according to claim 1, characterized in that: In step S3, the determination is made. The adjacency relationship between two instances includes calculating the minimum Euclidean distance between their point clouds. : in The elements of the adjacency matrix represent instances. Are they adjacent? This is the spatial proximity threshold.

5. The method for instance segmentation and structured topology reconstruction of complex geometric buildings according to claim 1, characterized in that: In step S4, for any two adjacent structural instances Fracture line modeling based on combinations of structural types includes: For the line of intersection between two planes, solve the equations of the two planes simultaneously. : Find the parameterized intersection line ,in, To represent the position of a point on a straight line, For plane and Normal vector Let be any point on the line of intersection of the two surfaces; For the intersection of the conical or hemispherical dome and the boundary of the supporting structure, an elliptic curve segment approximation is used, and the equation in the local coordinate system is: Connecting the vertices of the structure Forming an approximate conical skeleton, the major axis of the elliptic curve segment is taken as the radius of the enclosing sphere. The minor axis is 0.

8. The ellipse parameters were optimized using the Levenberg-Marquardt algorithm.

6. The method for instance segmentation and structured topology reconstruction of complex geometric buildings according to claim 5, characterized in that: In step S5, for the case where the three planes of the roof intersect, a system of linear equations is constructed by simultaneously solving the plane equations, and the coordinates of the intersection points in three-dimensional space are directly solved using Cramer's rule. ,in The equation is for a plane; for instances with semantic labels of cone or hemispherical top, the vertex coordinates or center of the sphere of the fitted model are directly taken; for instances with semantic labels of cylinder, the upper and lower endpoints are taken along the cylinder axis; for instance point clouds with semantic labels of vertical wall, the projection is onto a two-dimensional plane in the direction of the normal vector, the axial bounding box is calculated, and the four corner points are extracted as contour vertices.

7. The method for instance segmentation and structured topology reconstruction of complex geometric buildings according to claim 1, characterized in that: In step S6, constructing the hierarchical constraint model includes: first, geometrically grouping the boundary lines into orthogonal constraint groups. and free form group Then, for the orthogonal constraint group Extract the main direction, for Middle boundary direction angle Perform K-means clustering (k=2) with forced orthogonality constraints. Generate the main direction vector set Finally, joint optimization is performed on the extracted building boundary lines: the objective function is defined. ,in: Force orthogonal group boundaries to align with the main direction; The endpoint continuity is maintained through an adjacency matrix; the Levenberg-Marquardt algorithm is used to optimize the boundary endpoint coordinates, and the free group... Only participate The calculations are performed to maintain topological connectivity, ultimately resulting in a building outline structure that is connected end-to-end and conforms to the rules of topological connectivity.

Images