Building three-dimensional model intelligent reconstruction method and system based on deep learning

Abstract

The invention relates to the technical field of three-dimensional model intelligent reconstruction, in particular to a building three-dimensional model intelligent reconstruction method and system based on deep learning, and the method comprises the steps: obtaining multi-source data of a target ancient building, and generating a dense point cloud through a motion recovery structure and multi-view stereo matching; identifying a structural component and a non-structural component through semantic segmentation, and reconstructing an initial grid model with semantic label information; analyzing a spatial topological relation of the structural component by adopting a graph neural network, and generating a hierarchical combination mode and a connection constraint rule in combination with a historical construction knowledge base; and according to a connection constraint rule, adaptively adjusting a grid node pose, driving a parameterized component to be accurately assembled, and fusing to obtain a structured three-dimensional model of the target historic building. According to the method, the structural component connection relation of the hidden area of the historic building can be restored under the non-contact acquisition condition, and the structural rationality and geometric accuracy of the three-dimensional model are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional model intelligent reconstruction, in particular to a building three-dimensional model intelligent reconstruction method and system based on deep learning. BACKGROUND

[0002] At present, the three-dimensional modeling of buildings mainly relies on photogrammetry and laser radar scanning technologies, and the external geometric morphology is reconstructed by collecting image or point cloud data, which is widely used in modern buildings. However, the data collection and modeling of ancient buildings face special challenges. Since ancient buildings are usually protected cultural relics, close contact or structural disassembly is usually prohibited, making it difficult to obtain key structural data such as internal beam frames, interlayer dougong gaps and hidden mortise and tenon joints. At the same time, the structure of ancient buildings is complex, and the hidden structure relationship of dougong mortise and tenon is difficult to restore through conventional modeling.

[0003] Therefore, the existing technology can only construct the surface geometric model of the ancient building, and it is difficult to effectively restore the internal structure relationship, component style and combination logic. Especially in the face of weathering erosion, component damage and other situations, the generated model often has low shape accuracy, poor component integrity and unclear structure expression, which not only affects the accuracy of its digital archives, but also interferes with subsequent repair design and structure safety evaluation.

[0004] Therefore, it is urgent to develop a building three-dimensional model intelligent reconstruction method and system based on deep learning, which can deeply analyze the structural intention of ancient buildings and construct an intelligent three-dimensional model with high authenticity and structural interpretability. SUMMARY

[0005] (1) Technical problem to be solved The purpose of the present application is to provide a building three-dimensional model intelligent reconstruction method and system based on deep learning, to solve the technical problem that the three-dimensional model structure logic is broken and the assembly integrity is insufficient due to the missing connection relationship of hidden components or the interference of weathered and damaged components.

[0006] (2) Technical scheme To achieve the above-mentioned purpose, on the one hand, the present application provides a building three-dimensional model intelligent reconstruction method based on deep learning, which comprises: Step S1: obtaining multi-source data of a target ancient building, the multi-source data comprising a visible light image sequence, laser point cloud data and depth image data; obtaining sparse point cloud by a motion recovery structure algorithm from the multi-source data, and obtaining dense point cloud by multi-view stereo matching according to the sparse point cloud; obtaining semantic label information by semantic segmentation operation from the dense point cloud, the semantic label information comprising first information corresponding to structural components and second information corresponding to non-structural components; obtaining an initial mesh model by surface reconstruction from the dense point cloud carrying the semantic label information.

[0007] Step S2: parsing the connection relationship between the structural members through the neural network based on the first information to obtain the hierarchical combination mode and the connection constraint rule of the structural members; and matching the first information and the hierarchical combination mode with the corresponding component parameter template through the preset construction knowledge base.

[0008] Step S3: obtaining the grid nodes corresponding to the structural members in the initial grid model, adjusting the spatial positions of the grid nodes according to the connection constraint rule to obtain a final grid model; performing spatial assembly according to the component parameter template and the hierarchical combination mode to obtain an assembly structure; and fusing the structural member grid regions in the final grid model and the assembly structure according to the connection constraint rule to obtain a structured three-dimensional model of the target ancient building.

[0009] Further, the method for obtaining semantic label information from the dense point cloud through semantic segmentation operation includes: dividing the dense point cloud to obtain a plurality of sub-blocks, extracting the local geometric shape, normal vector distribution, color texture information and spatial adjacency relationship of each sub-block to form a multi-modal feature vector; and obtaining the semantic label information corresponding to each point in the dense point cloud through the pre-trained point cloud semantic segmentation model. The first information includes the category label and confidence score of the column, beam and dougong, and the second information includes the category label and confidence score of the colored drawing, eave and swallow-tail; the category label is corrected when the confidence score is lower than the preset confidence threshold.

[0010] Further, the method for obtaining the hierarchical combination mode and the connection constraint rule of the structural members by parsing the connection relationship between the structural members through the neural network based on the first information includes: clustering the dense point cloud segments with the same category label and spatial adjacency as candidate regions according to the category label of the structural members in the first information and the spatial distribution of the dense point cloud; constructing a topological graph with the candidate regions as nodes, the connection edges of the topological graph being the spatial pose relationship between the nodes; parsing the connection weight of the connection edges between the nodes through the graph neural network; and establishing the hierarchical combination mode and the connection constraint rule of the structural members according to the connection weight.

[0011] Further, the method for establishing the hierarchical combination mode and the connection constraint rule of the structural members according to the connection weight includes: Obtaining the geometric gravity height of the structural member corresponding to the first information in the dense point cloud, sorting the structural member according to the gravity direction according to the geometric gravity height to obtain a sorting result; obtaining an optimized hierarchical combination mode and connection constraint rule according to the sorting result; and performing historical connection verification on the optimized hierarchical combination mode and connection constraint rule.

[0012] Further, the method of performing historical connection verification on the optimized hierarchical combination mode and connection constraint rule comprises: Obtaining a frequency statistical matrix of historical connection instances from the construction knowledge base; calculating the appearance probability of the optimized hierarchical combination mode in the frequency statistical matrix; and replacing the optimized hierarchical combination mode with a historical connection sequence with the highest appearance probability in the frequency statistical matrix when the appearance probability is lower than a preset appearance probability threshold.

[0013] Further, the method of obtaining the grid node corresponding to the structural member in the initial grid model and adjusting the spatial position of the grid node according to the connection constraint rule to obtain the final grid model comprises: Identifying a structural member combination with an interface matching relationship according to the hierarchical combination mode and connection constraint rule of the structural member, obtaining a corresponding set of nodes to be aligned according to the structural member combination, and calculating the spatial pose deviation of the set of nodes to be aligned through the initial grid model.

[0014] Constructing an optimization function with the objective of minimizing the spatial pose deviation and adding a curvature constraint term; updating the spatial coordinates of the set of nodes to be aligned through an iterative optimization algorithm for the optimization function, and obtaining the final grid model when the norm of the spatial coordinate update vector of adjacent iterative steps is less than a preset convergence threshold.

[0015] Further, the method of performing spatial assembly according to the component parameter template and hierarchical combination mode to obtain a structure of an assembly comprises: Retrieving the component parameter template and hierarchical combination mode in a pre-constructed component model library to obtain a parameterized component model; obtaining an interface region in the final grid model and an interface matching mode defined in the connection constraint rule; obtaining a spatial transformation matrix according to the interface region and the interface matching mode; adjusting the interface node position of the parameterized component model according to the spatial transformation matrix and aligning the spatial transformation matrix with the grid node to obtain a spatial positioning model of the structural member; and organizing the spatial positioning model according to the hierarchical combination mode to generate a structural member assembly unit.

[0016] The category label, geometric shape and attachment conditions defined in the preset decorative base surface rules of the non-structural component are calculated by a normal projection method to obtain an attachment model; the structural component assembly unit and the attachment model are spatially combined to obtain an assembly structure.

[0017] Furthermore, the method of adjusting the interface node positions of the parameterized component model according to the spatial transformation matrix and aligning them with the grid nodes to obtain the spatial positioning model of the structural component includes: The boundary contour of the parameterized component is aligned with the adjacent area in the final mesh model, and the joint deviation is obtained through geometric consistency test. When the joint deviation exceeds the preset assembly tolerance threshold, a nonlinear optimization algorithm is used to iteratively adjust the spatial position of the anchor point within the preset degree of freedom constraint range. After each iterative adjustment, the joint deviation is recalculated, and the spatial positioning model of the structural component is obtained when the joint deviation meets the assembly tolerance threshold.

[0018] Furthermore, the method of fusing the structural component grid area in the final grid model with the assembly structure according to connection constraint rules to obtain a structured three-dimensional model of the target ancient building includes: The method comprises obtaining a mesh area of ​​a structural component in a final mesh model, establishing a mapping relationship in combination with the spatial pose of the assembly structure, processing the mesh area according to the mapping relationship and the interface matching method defined in the connection constraint rules to obtain a geometric expression of the structural component, generating an adjacency relationship graph according to the connection constraint rules of the structural component, and combining the geometric expression with the adjacency relationship graph to generate a structured three-dimensional model.

[0019] Based on the same inventive concept, on the other hand, the present invention also provides a deep learning-based intelligent reconstruction system for a three-dimensional building model, the system comprising: The data acquisition and preprocessing module is used to obtain multi-source data of the target ancient building, wherein the multi-source data includes visible light image sequences, laser point cloud data and depth image data; obtain a sparse point cloud from the multi-source data through a motion recovery structure algorithm, and obtain a dense point cloud based on the sparse point cloud through multi-view stereo matching; obtain semantic label information from the dense point cloud through a semantic segmentation operation, wherein the semantic label information includes first information corresponding to structural components and second information corresponding to non-structural components; and obtain an initial mesh model through surface reconstruction of the dense point cloud carrying the semantic label information.

[0020] The structural analysis and template matching module is used to analyze the connection relationship between structural components through a neural network using the first information to obtain the hierarchical combination method and connection constraint rules of the structural components; and match the first information and the hierarchical combination method with the corresponding component parameter template through a preset construction knowledge base.

[0021] The model optimization and assembly output module is used to obtain the grid nodes corresponding to the structural components in the initial grid model, adjust the spatial positions of the grid nodes according to the connection constraint rules to obtain the final grid model; perform spatial assembly according to the component parameter template and hierarchical combination method to obtain the assembly structure; and fuse the structural component grid area in the final grid model with the assembly structure according to the connection constraint rules to obtain a structured three-dimensional model of the target ancient building.

[0022] (3) Beneficial effects Compared with the prior art, the present invention has the following beneficial effects: 1. This invention uses a graph neural network combined with a historical structural knowledge base to analyze and optimize the connection relationship of structural components in hidden areas, and can restore the force transmission path of ancient buildings that conforms to traditional craftsmanship under non-contact collection conditions.

[0023] 2. This invention uses an assembly mechanism that combines parametric template drive with connection constraint rules to achieve high-precision geometric fusion of structural and non-structural components, generate a unified geometric expression, and establish an adjacency relationship diagram, significantly improving the structural rationality, geometric accuracy, and historical authenticity of the three-dimensional model. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flowchart of a method for intelligently reconstructing a three-dimensional building model based on deep learning according to Example 1 of the present invention; Figure 2 This is a schematic diagram of the module composition of the deep learning-based intelligent reconstruction system for building three-dimensional models according to Example 2 of the present invention. DETAILED DESCRIPTION

[0025] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0026] Before giving examples, it is necessary to elaborate on the application scenarios conceived by the present invention. Due to the complex structure of ancient buildings, the variety of components, and the fact that most of them are in a state of cultural relic protection, they cannot be touched at close range or disassembled for measurement, resulting in their three-dimensional modeling relying on manual experience and structural speculation for a long time. There are problems such as high recognition error rate, low modeling efficiency, and fuzzy component connection relationships. Especially in the case of densely distributed multi-layer brackets and mortise and tenon joints with significant interference, traditional methods based on a single data source or pure geometric reconstruction are difficult to effectively restore the structural logic. The present invention is aimed at the above-mentioned application dilemma and proposes a three-dimensional reconstruction method for buildings that integrates multi-source perception, deep learning structural analysis, and parametric model assembly. It is suitable for scenarios such as digital restoration, structural analysis, and BIM integration of cultural relic protection buildings. It has the advantages of being inaccessible, requiring no destruction, and having reasonable and reusable result structures.

[0027] Example 1: Figure 1 As shown, this embodiment provides a method for intelligently reconstructing a three-dimensional building model based on deep learning, the method comprising: Step S1: Acquire multi-source data of the target ancient building, wherein the multi-source data includes a visible light image sequence, laser point cloud data, and depth image data; obtain a sparse point cloud from the multi-source data through a motion recovery structure algorithm, and obtain a dense point cloud based on the sparse point cloud through multi-view stereo matching; obtain semantic label information from the dense point cloud through a semantic segmentation operation, wherein the semantic label information includes first information corresponding to structural components and second information corresponding to non-structural components; obtain an initial mesh model through surface reconstruction of the dense point cloud carrying the semantic label information.

[0028] Step S2: Analyze the connection relationship between structural components using a neural network based on the first information to obtain a hierarchical combination method and connection constraint rules of the structural components; match the first information and the hierarchical combination method with corresponding component parameter templates through a preset construction knowledge base.

[0029] Step S3: Obtain the grid nodes corresponding to the structural components in the initial grid model, and adjust the spatial positions of the grid nodes according to the connection constraint rules to obtain the final grid model; retrieve matching parametric component models from the pre-built component model library according to the component parameter template and the hierarchical combination method; perform spatial assembly according to the component parameter template and the hierarchical combination method to obtain an assembly structure; merge the structural component grid area in the final grid model with the assembly structure according to the connection constraint rules to obtain a structured three-dimensional model of the target ancient building.

[0030] For example, the digital modeling of Ancient Building A was performed using a quadcopter drone equipped with an industrial camera and LiDAR for aerial photography. Depth images penetrate the physical gaps between bracket layers to reconstruct the mortise and tenon joints. The time-of-flight (TOF) depth camera has a minimum penetration capability of 15 cm. Over 800 frames of visible light images were acquired along the flight path, and the laser point cloud achieved an accuracy of 0.5 cm, meeting national standards. A structure-from-motion algorithm was used to process the image sequence, generating a sparse point cloud based on the pose constraints of the laser point cloud. This sparse point cloud was then transformed into a dense point cloud with a density of 12,000 points per square meter through multi-view stereo matching. This dense point cloud was then divided into multiple sub-blocks of fixed size for subsequent feature extraction.

[0031] The dense point cloud with semantically labeled information is converted into an initial mesh model through Poisson surface reconstruction. The spatial poses of the column-beam nodes are analyzed to construct a topological graph. The connecting edges between nodes are defined by their spatial pose relationships. A graph neural network is used to analyze the connection weights of these edges, thereby generating a hierarchical combination relationship from columns to beams to brackets. Parametric templates (such as the mortise and tenon joint dimensions of a hip-and-gable main beam) are matched to the construction knowledge base. The hierarchical rules and interface matching parameters preset in the knowledge base are combined to form connection constraint rules. The nodes of the initial mesh model are adjusted and optimized to complete the assembly of the structural components.

[0032] In particular, the construction of the structural knowledge base includes: converting the bracket and mortise and tenon types and beam-column interface dimensions in the "Engineering Practice Rules" and "Construction Methods" into a parametric template library, for example, the dovetail tenon opening depth is 0.3 times the component width and the slope is 12°, and the depth of the groove at the top of the raised beam column is not less than 1 / 5 of the beam height; statistically calculating the component combination probability matrix based on 27,000 sets of official architectural surveying data, and storing regional parameters according to regional types such as the Jin school, Su school, and Beijing school, for example, the Jin school hump has a height-to-width ratio of 0.8 and the Su school bracket has a curvature radius of 0.5m; at the same time, storing the hierarchical rules and interface matching parameters of structural components, providing a basis for connection relationship analysis and assembly constraints.

[0033] The method of obtaining semantic label information by performing a semantic segmentation operation on the dense point cloud, wherein the semantic label information includes first information corresponding to a structural component and second information corresponding to a non-structural component, includes: Divide the dense point cloud into multiple sub-blocks, extract the local geometric shape, normal vector distribution, color texture information and spatial adjacency of each sub-block to form a multimodal feature vector; pass the multimodal feature vector through a pre-trained point cloud semantic segmentation model to obtain the semantic label information corresponding to each point in the dense point cloud; The first information includes the category labels and confidence scores of columns, beams, and brackets, and the second information includes the category labels and confidence scores of painted walls, ridge tiles, and brackets. When the confidence score is lower than the preset confidence threshold, the category label is modified.

[0034] Exemplarily, the technical features improve the point cloud semantic segmentation accuracy through multi-modal feature fusion, which plays a key role in the complex structure area of ancient buildings. Taking the double-layered dougong area of ancient building A as an example, the point cloud is divided into cubic sub-blocks with a side length of 2.5 meters. This size is derived from the statistical physical size of a typical dougong group and can completely cover the common components. Four types of features are extracted from each sub-block, including local geometry, normal vector distribution, color texture information, and spatial adjacency relationship. Among them, the principal curvature value at the turning point of the dougong is greater than 0.8, the normal variation coefficient of the column base connection area is significantly higher than the average level, and the color saturation distribution of the colored area shows a centralized trend. These features form a multi-modal vector and are input into a residual PointNet++ model for processing.

[0035] When the confidence score output by the model is lower than 0.6, the system starts a two-stage correction mechanism. The confidence threshold of 0.6 is determined by plotting the precision-recall curve on the validation set and taking the inflection point position. The first stage performs neighborhood label voting, and the neighborhood range is set to 15 adjacent points, which is calculated by the average density of dense point cloud. The second stage performs geometric consistency judgment, including normal angle constraint and gravity direction deviation verification. After the correction is completed, the boundary positioning error of the structural component is controlled within 2 centimeters, which meets the requirements of the "Ancient Building Surveying and Mapping Specification" for digital archiving.

[0036] In this embodiment, an existing point cloud semantic segmentation network is used for processing, and a preferred structure such as PointNet++ or KPConv is used. The three-dimensional point cloud coordinates and normal vectors are input, and the component category label of each point is output. The model can be trained through existing ancient building scanning data or small sample artificial annotation data, and has strong structural component recognition ability.

[0037] The method for analyzing the connection relationship between the structural components through the neural network based on the first information to obtain the hierarchical combination mode and the connection constraint rule comprises: According to the category label of the structural component in the first information and the spatial distribution of the dense point cloud, the dense point cloud segments with the same category label and spatial adjacency are clustered into candidate areas; A topological graph is constructed with the candidate areas as nodes, and the connection edges of the topological graph are the spatial pose relationships between the nodes. The connection weight of the connection edge between the nodes is analyzed through the graph neural network. The hierarchical combination mode and the connection constraint rule of the structural component are established according to the connection weight.

[0038] Exemplarily, in the process of analyzing the connection relationship of structural components of ancient buildings, the system first clusters the column, beam and dougong point cloud obtained by semantic segmentation according to the consistency of category labels and spatial proximity to form a candidate region set. The clustering process sets a spatial radius constraint to maintain the integrity of the point cloud segment, for example, the clustering radius of the column and beam in the Wanshou Palace is set to 0.6 meters.

[0039] Subsequently, a topological graph of the spatial pose relationship between nodes is constructed with the candidate regions as nodes. The spatial pose relationship is defined by the component spacing, the interface normal vector angle and the geometric gravity height difference, for example, the measured column spacing is 1.2 meters (fluctuation ± 0.3 meters), the interface normal vector angle cosine value is greater than or equal to 0.98, and the beam column gravity height difference meets the traditional frame reference requirements.

[0040] The constructed topological graph is input into a graph attention network (GAT) to analyze the connection weight of the connection edge between nodes. With 0.85 as the weight threshold, the effective connection edge is screened to obtain the initial component connection sequence, for example, the column-beam connection weight is 0.91, and the beam-dougong connection weight is 0.87. In particular, the weight threshold is set to 0.85, which is determined according to the lower limit of the effective connection edge in the connection weight distribution of a large number of historical component connection samples, which can effectively distinguish between real structure connections and noise connections.

[0041] According to the connection weight obtained by analysis, the hierarchical combination mode and connection constraint rule are generated. When processing the dense area of double-layer dougong, the system identifies the mortise and tenon axial fitting relationship through the normal vector angle constraint, such as the normal vector angle deviation of the dougong mouth and the huagong interface being less than or equal to 3°, and the axis centering error being less than or equal to 1.8 cm, to ensure that the horizontal displacement and rotation limit values in the generated connection constraint rule are clear, providing accurate boundary conditions for subsequent assembly.

[0042] The method for establishing the hierarchical combination mode and connection constraint rule of the structural component according to the connection weight comprises: obtaining the geometric gravity height of the structural component corresponding to the first information in the dense point cloud, sorting the structural component according to the gravity direction to obtain a sorting result; obtaining an optimized hierarchical combination mode and connection constraint rule according to the sorting result; and verifying the optimized hierarchical combination mode and connection constraint rule.

[0043] Exemplarily, there is a clear hierarchical rule in the construction of ancient buildings, which is stored in the construction knowledge base and formed by historical specifications such as “Yingzao Fashi” and “Engineering Practice Regulations” and 27,000 sets of official building surveying data statistics, including typical component connection sequence (such as column-beam-dougong-rafter-roof component) and allowed interface type combination.

[0044] During the connection relationship resolution process, if a connection edge in the topology graph corresponds to a component type combination that is not in the aforementioned hierarchical rules, the edge is marked as a candidate for misconnection. For example, if a column node directly connects to a roof component node, and this combination does not exist in the construction knowledge base, then the edge is a candidate for misconnection. The system removes these candidate edges to eliminate connection paths that do not conform to historical construction logic.

[0045] To optimize the rationality of the remaining connection paths, the system extracts the geometric center of gravity height of each structural component corresponding to the first information from the dense point cloud. The geometric center of gravity height refers to the Z-axis coordinate of the center of gravity of the component mesh model in the three-dimensional coordinate system. The components are sorted according to the direction of gravity (bottom-up), with a height difference threshold set to no less than 0.5 meters. This threshold is derived from the minimum vertical spacing standard between load-bearing components in traditional timber frames.

[0046] Taking the front hall gatehouse of an ancient building as an example, after eliminating incorrectly connected edges where columns directly connect to roof components, the system sorts the components by their geometric center of gravity heights: columns (0 meters), beams (3.2 meters), brackets (4.1 meters), rafters (5.7 meters), and roof components (6.3 meters). This order conforms to the traditional bottom-up gravity transfer logic. The system further combines the sorting results with connection constraint rules to form an optimized hierarchical combination method and connection constraint rules. These rules include the allowed connection types, interface directionality, and relative height difference constraints between components.

[0047] Finally, the system updates the topology based on the optimized rules, generating a complete force transmission path from the load-bearing base to the roof components. For example, the height difference between the bracket and rafter connection is 1.6 meters, with a tolerance of ±0.2 meters. This tolerance is derived from the standard deviation of historical surveying data. Comparisons with the measured structure show that the optimized path is 97.3% consistent with the actual structure.

[0048] The method for performing historical connection verification on the optimized hierarchical combination mode and connection constraint rules includes: Obtain a frequency statistical matrix of historical connection instances from the construction knowledge base; calculate the probability of occurrence of the optimized hierarchical combination method in the frequency statistical matrix; when the occurrence probability is lower than a preset occurrence probability threshold, replace the optimized hierarchical combination method with the historical connection order with the highest occurrence probability in the frequency statistical matrix.

[0049] For example, for the optimized hierarchical combination mode and connection constraint rules, the system performs historical connection verification to ensure the consistency of the path with traditional construction rules.

[0050] The historical connection verification is based on a pre-constructed historical connection instance frequency statistical matrix in the construction knowledge base. This matrix is generated by 27,000 sets of official building surveying data statistics, records the connection sequence between different component types and their frequency of occurrence in historical instances, and is stored separately by regional style (such as Jin, Su, and Beijing) to reflect regional differences.

[0051] The system calculates the probability of the optimized hierarchical combination mode in the frequency statistical matrix. When the probability is lower than the preset threshold of 5%, the system replaces the combination with the highest probability historical connection sequence in the matrix, and retains the interface matching mode and relative height difference constraints defined in the connection constraint rules. The preset threshold of 5% is set according to the statistical confidence level, ensuring that combinations below this value are rare in history and may be incorrect connections.

[0052] For example, in the gable area of ancient buildings, the optimized hierarchical combination mode is "coupling → roof component → decorative component", and the geometric center of gravity height sequence is coupling 4.1 meters → roof component 6.3 meters → decorative component 5.2 meters. According to the search, the probability of this combination in the construction knowledge base is only 0.8%, which is lower than the 5% threshold. The historical frequency matrix shows that the high-frequency path in this area is "coupling → rafter → roof component", with an appearance probability of 41.3%, and the corresponding height sequence is coupling 4.1 meters → rafter 5.7 meters → roof component 6.3 meters.

[0053] Accordingly, the system replaces the original combination with the high-frequency path and adjusts the connection direction in the topology graph to conform to the gravity transmission logic of historical construction. After replacement, the connection weight between the coupling and the rafter is increased from 0.82 to 0.91. According to the actual measurement verification, the spatial consistency of the updated path with the physical structure reaches 98.2%, and the interface direction angle deviation is reduced to 1.5°.

[0054] The method for obtaining the grid nodes corresponding to the structural components in the initial grid model, adjusting the spatial positions of the grid nodes according to the connection constraint rules to obtain the final grid model comprises: According to the hierarchical combination mode of the structural components and the connection constraint rules, identify the structural component combination with interface matching relationship, and obtain the corresponding alignment node set according to the structural component combination; calculate the spatial pose deviation of the alignment node set through the initial grid model.

[0055] An optimization function is constructed to minimize the spatial pose deviation, and a curvature constraint term is added; the optimization function is updated to the spatial coordinates of the alignment node set through an iterative optimization algorithm, and when the spatial coordinate update vector norm of adjacent iteration steps is less than a preset convergence threshold, the final grid model is obtained.

[0056] For example, in the column-beam connection area of ​​the main hall of an ancient building, the initial mesh has a positive Y-axis offset of 3.8 cm at the beam end. The center coordinates of the column top are [2.1, 5.3, 0], and the center of the beam bottom are [2.1, 5.66, 0]. The system identifies the top contact fit type based on the connection diagram and extracts 12 grid points at the column top and 8 grid points at the beam bottom as the alignment node set. The principal component of the spatial pose deviation, ΔY, is calculated to be 3.8 cm.

[0057] Constructing the optimization function: .

[0058] Where N=12 is the number of nodes to be aligned, and the mapping is established by coordinate transformation at the 8 nodes at the bottom of the beam; is the coordinate of the column top node; is the coordinate of the bottom node of the beam; is the average curvature value of the triangle, which is calculated by the difference of the adjacent node normal vectors, where the node normal vector is calculated by weighting the adjacent triangle normal vectors, and the patch normal vector is determined by the cross product of the position vectors of its three vertices; and Triangular patches Average curvature value between initial mesh and optimized mesh; is a rigid transformation, ;in, represents the three-dimensional rotation group, that is, the set of all possible 3D rotation matrices that satisfy and , used to characterize the rotational posture of the component in three-dimensional space; Represents the three-dimensional coordinate vector of a point in a dense point cloud or mesh model (usually expressed as ), in this optimization function, specifically refers to the column top node coordinates Or beam bottom node coordinates ; is the three-dimensional translation vector, and the rotation matrix Together they constitute a rigid transformation , used to adjust the spatial position of the component; the curvature constraint weight λ = 0.3, calculated by inverse calculation based on the ±0.5cm dovetail joint tolerance specified in the "Engineering Practice Rules," ensures that deformation optimization meets the tolerances of traditional woodworking techniques. An iterative optimization algorithm is used to update the node coordinates, terminating when the displacement vector norm is less than 0.001mm. The convergence threshold of 0.001mm is equivalent to the accuracy of traditional dovetail joints. After optimization, the maximum spacing between the bottom surface of the beam and the top surface of the column is 1.5mm, meeting the load-bearing requirements of the raised beam structure and the 2mm assembly tolerance threshold.

[0059] The method for performing spatial assembly according to the component parameter template and the hierarchical combination mode to obtain an assembly structure includes: The component parameter template and the hierarchical combination mode are searched in a pre-built component model library to obtain a parameterized component model; an interface region in the final mesh model and an interface matching mode defined in a connection constraint rule are obtained; a spatial transformation matrix is obtained according to the interface region and the interface matching mode; an interface node position of the parameterized component model is adjusted according to the spatial transformation matrix, and a spatial positioning model of a structural component is obtained by aligning the mesh node; and the spatial positioning model is organized according to the hierarchical combination mode to generate a structural component assembly unit.

[0060] A category label, a geometric shape and an attachment condition defined in a preset decoration base surface rule of the non-structural component are calculated to obtain an attachment model by a normal projection method; and the structural component assembly unit and the attachment model are spatially combined to obtain an assembly structure.

[0061] For example, in a hall roof construction in an ancient building, a main beam with a span of 4.8 meters in a hip-and-gable roof type needs to be assembled above a column-doung-arch combination structure according to the category label information of the structural component. The interface region (i.e., the beam end node set and the column top node set) corresponding to the main beam in the final mesh model is taken as an alignment object, and a spatial transformation matrix for interface alignment is calculated according to the interface matching mode defined in the connection constraint rule. The matrix includes a rotation compensation parameter (measured from the normal direction of the column top node set) of 5.2° on the Y axis and a translation vector (calculated from the difference between the column top center coordinates (2.1, 5.3, 0) and the beam theoretical gravity center (2.1, 5.66, 0)) of 3.6 cm, which are used to correct the interface node position of the parameterized main beam model to make it spatially aligned with the mesh node, thereby generating a spatial positioning model of the structural component.

[0062] Subsequently, the spatial positioning model is organized according to the hierarchical combination mode, and assembly constraints between the beam and the column-doung-arch are applied according to a mortise-and-tenon fitting angle of 12° and a dovetail depth fitting tolerance of ±0.5 mm defined in the connection constraint rule, thereby generating a stable structural component assembly unit.

[0063] For the non-structural component, an attachment model of the non-structural component is obtained by calculating an attachment position of the non-structural component according to the category label information and the geometric shape of the non-structural component and the attachment condition defined in the decoration base surface rule by using the normal projection method, and the non-structural component and the structural component assembly unit are spatially combined to generate a complete assembly structure.

[0064] The actual mortise-and-tenon fitting gap meets the tolerance requirement of the Engineering Practice Rules and Regulations, and the normal angle deviation between the attached non-structural component and the decoration base surface is not more than 1.5°.

[0065] The method of adjusting the interface node positions of the parameterized component model according to the spatial transformation matrix and aligning them with the grid nodes to obtain a spatial positioning model of the structural component includes: The boundary contour of the parameterized component is aligned with the adjacent area in the final mesh model, and the joint deviation is obtained through geometric consistency test. When the joint deviation exceeds the preset assembly tolerance threshold, a nonlinear optimization algorithm is used to iteratively adjust the spatial position of the anchor point within the preset degree of freedom constraint range. After each iterative adjustment, the joint deviation is recalculated, and the spatial positioning model of the structural component is obtained when the joint deviation meets the assembly tolerance threshold.

[0066] For example, the spatial registration mechanism provided by the present invention ensures accuracy through a dual tolerance system: the structural component assembly tolerance threshold of 2 mm is derived from the upper limit of the mortise and tenon fit tolerance in the "Engineering Practice Rules", and the non-structural component pattern continuity threshold of 0.5 pixels corresponds to the 4K texture mapping standard; the component's degree of freedom constraint range is ±10 mm for translation and ±3° for rotation. This range is calculated based on the thermal expansion and contraction tolerance of 0.8% annual deformation rate of wood, ensuring that the model maintains structural stability and historical appearance consistency in an environment of -20°C to 40°C.

[0067] For example, at the interface between the ridge beam and the ridge tile of ancient building A, geometric consistency testing detected a 3.5mm misalignment in the X-axis joint, exceeding the 2mm assembly tolerance threshold. The system employed anchor point degree-of-freedom constraints within the parametric beam model to construct an objective function that minimized the joint profile projection error. Iterative optimization was then performed using the Levenberg–Marquardt algorithm. By the sixth iteration, a 1.2° Z-axis rotation compensated for the misalignment, reducing it to 2.5mm. By the 12th iteration, the displacement vector norm converged to 0.0003mm, corresponding to a 1.2mm misalignment. The corrected ridge beam met the assembly tolerance requirements, and finite element analysis verified that the uniformity of the contact pressure distribution with the tile was improved by 82% compared to the pre-correction level.

[0068] The method of fusing the structural component grid area in the final grid model with the assembly structure according to connection constraint rules to obtain a structured three-dimensional model of the target ancient building includes: The method comprises obtaining a mesh area of ​​a structural component in a final mesh model, establishing a mapping relationship in combination with the spatial pose of the assembly structure, processing the mesh area according to the mapping relationship and the interface matching method defined in the connection constraint rules to obtain a geometric expression of the structural component, generating an adjacency relationship graph according to the connection constraint rules of the structural component, and combining the geometric expression with the adjacency relationship graph to generate a structured three-dimensional model.

[0069] Exemplarily, after completing node refinement and component assembly, the system establishes a projection mapping relationship from vertex to surface for the structural component mesh area in the final mesh model and the parametric component model that has been positioned, based on the spatial posture of the assembly structure. The judgment threshold of the mapping residual is 2mm, which is derived from the upper limit of the mortise and tenon fit tolerance stipulated in the "Engineering Practice Rules". Combined with the interface fit method defined in the connection constraint rules, geometric replacement or fusion is performed within the interface neighborhood: when the bidirectional Hausdorff distance does not exceed 1.5mm and the normal angle does not exceed 5°, the original mesh triangle is replaced by the parametric component surface; otherwise, within a neighborhood radius of 40mm, the vertex coordinates are weightedly fused according to the mapping correspondence to obtain the geometric expression of the structural component. The neighborhood radius of 40mm comes from the statistics of typical mortise and tenon lengths and shoulder widths in the construction knowledge base.

[0070] In this example, six 0.5m columns with a diameter error of no more than 0.02m and three main beams with a span of 4.8m in the front porch area of ​​the main hall were replaced / fused to form a continuous, seamless geometric representation. Subsequently, an adjacency graph was constructed based on the connection constraints between components. Valid connection edges with a weight threshold of 0.85 or above were retained, and the connection type and force transmission direction were recorded in the edge attributes. The weight threshold of 0.85 was derived from the valid lower limit of the historical sample distribution. The force transmission direction was determined from low to high based on the geometric center of gravity height, for example, column → beam → bracket → rafter → roof component. In this area, the system generated a total of 18 edges, including typical connections such as "column-beam (mortise and tenon joint, upward)" and "beam-bracket (mortise and tenon joint, upward)." Finally, the geometric representation was associated with the adjacency graph to obtain a structured 3D model that is both renderable and contains component-level connection type and force transmission direction information for subsequent assembly verification and structural analysis.

[0071] Example 2: Based on the same inventive concept, Figure 2 As shown, this embodiment also provides a deep learning-based intelligent reconstruction system for a three-dimensional building model, the system comprising: The data acquisition and preprocessing module is used to obtain multi-source data of the target ancient building, wherein the multi-source data includes visible light image sequences, laser point cloud data and depth image data; obtain a sparse point cloud from the multi-source data through a motion recovery structure algorithm, and obtain a dense point cloud based on the sparse point cloud through multi-view stereo matching; obtain semantic label information from the dense point cloud through a semantic segmentation operation, wherein the semantic label information includes first information corresponding to structural components and second information corresponding to non-structural components; and obtain an initial mesh model through surface reconstruction of the dense point cloud carrying the semantic label information.

[0072] The structural analysis and template matching module is configured to analyze the connection relationship between the structural components by a neural network based on the first information, to obtain a hierarchical combination mode and a connection constraint rule of the structural components, and to match the first information and the hierarchical combination mode to a corresponding component parameter template based on a preset construction knowledge base.

[0073] The model optimization and assembly output module is configured to obtain a grid node corresponding to a structural component in an initial grid model, to adjust the spatial position of the grid node based on the connection constraint rule to obtain a final grid model, to perform spatial assembly based on the component parameter template and the hierarchical combination mode to obtain an assembly structure, and to fuse the structural component grid region in the final grid model and the assembly structure based on the connection constraint rule to obtain a structured three-dimensional model of the target ancient building.

[0074] It should be noted that, as for the system in the above embodiments, the specific manner in which each module operates has been described in detail in the embodiments related to the method, and will not be described in detail here.

[0075] Finally, it should be noted that, although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some of the technical features, and any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for intelligent reconstruction of building three-dimensional models based on deep learning, characterized in that: The method comprises: Acquire multi-source data of the target ancient building, the multi-source data including visible light image sequences, laser point cloud data, and depth image data; apply a structure-from-motion algorithm to the multi-source data to obtain a sparse point cloud, and apply multi-view stereo matching to obtain a dense point cloud based on the sparse point cloud; apply a semantic segmentation operation to the dense point cloud to obtain semantic label information, the semantic label information including first information corresponding to structural components and second information corresponding to non-structural components; and apply surface reconstruction to the dense point cloud carrying the semantic label information to obtain an initial mesh model; The first information is used to analyze the connection relationship between structural components through a neural network to obtain a hierarchical combination method and connection constraint rules of the structural components; the first information and the hierarchical combination method are matched with a corresponding component parameter template through a preset construction knowledge base; The grid nodes corresponding to the structural components in the initial grid model are obtained, and the spatial positions of the grid nodes are adjusted according to the connection constraint rules to obtain the final grid model; the assembly structure is obtained by spatial assembly according to the component parameter template and the hierarchical combination method; the structural component grid area in the final grid model is merged with the assembly structure according to the connection constraint rules to obtain a structured three-dimensional model of the target ancient building.

2. The method for intelligent reconstruction of a building three-dimensional model based on deep learning according to claim 1, characterized in that: The method of obtaining semantic label information by performing a semantic segmentation operation on the dense point cloud, wherein the semantic label information includes first information corresponding to a structural component and second information corresponding to a non-structural component, includes: Divide the dense point cloud into multiple sub-blocks, extract the local geometric shape, normal vector distribution, color texture information and spatial adjacency of each sub-block to form a multimodal feature vector; pass the multimodal feature vector through a pre-trained point cloud semantic segmentation model to obtain the semantic label information corresponding to each point in the dense point cloud; The first information includes the category labels and confidence scores of columns, beams, and brackets, and the second information includes the category labels and confidence scores of painted walls, ridge tiles, and brackets. When the confidence score is lower than the preset confidence threshold, the category label is modified.

3. The method for intelligent reconstruction of a building three-dimensional model based on deep learning according to claim 2, characterized in that: The method of analyzing the connection relationship between structural components using a neural network to obtain the hierarchical combination mode and connection constraint rules of the structural components includes: Clustering dense point cloud segments having the same category label and spatially adjacent to each other into candidate regions according to the category label of the structural component in the first information and the spatial distribution of the dense point cloud; A topological graph is constructed with the candidate areas as nodes, wherein the connecting edges of the topological graph represent the spatial posture relationships between the nodes; the connection weights of the connecting edges between the nodes in the topological graph are analyzed through a graph neural network; and a hierarchical combination method and connection constraint rules of structural components are established based on the connection weights.

4. The method for intelligent reconstruction of a building three-dimensional model based on deep learning according to claim 3, characterized in that: The method for establishing a hierarchical combination mode and connection constraint rules of structural components according to the connection weights includes: Obtaining the geometric center of gravity height of the structural component corresponding to the first information in the dense point cloud, sorting the structural components according to the geometric center of gravity height to obtain a sorting result; obtaining an optimized hierarchical combination method and connection constraint rules based on the sorting result; and performing historical connection verification on the optimized hierarchical combination method and connection constraint rules.

5. The method for intelligent reconstruction of a building three-dimensional model based on deep learning according to claim 4, characterized in that: The method for performing historical connection verification on the optimized hierarchical combination mode and connection constraint rules includes: Obtain a frequency statistical matrix of historical connection instances from the construction knowledge base; calculate the probability of occurrence of the optimized hierarchical combination method in the frequency statistical matrix; when the occurrence probability is lower than a preset occurrence probability threshold, replace the optimized hierarchical combination method with the historical connection order with the highest occurrence probability in the frequency statistical matrix.

6. The method for intelligently reconstructing a building three-dimensional model based on deep learning according to claim 3, characterized in that: The method of obtaining grid nodes corresponding to structural components in the initial grid model and adjusting the spatial positions of the grid nodes according to connection constraint rules to obtain the final grid model includes: Identify a combination of structural components with an interface matching relationship based on a hierarchical combination mode of the structural components and connection constraint rules, and obtain a corresponding set of nodes to be aligned based on the combination of structural components; calculate the spatial pose deviation of the set of nodes to be aligned using an initial mesh model; An optimization function is constructed with the goal of minimizing the spatial pose deviation, and a curvature constraint term is added. The optimization function is used to update the spatial coordinates of the set of nodes to be aligned through an iterative optimization algorithm, and the final mesh model is obtained when the norm of the spatial coordinate update vectors of adjacent iterative steps is less than a preset convergence threshold.

7. The method for intelligently reconstructing a building three-dimensional model based on deep learning according to claim 6, wherein: The method for obtaining an assembly structure by performing spatial assembly according to the component parameter template and the hierarchical combination method includes: The component parameter template and hierarchical combination method are retrieved from a pre-built component model library to obtain a parametric component model; the interface area in the final grid model and the interface matching method defined in the connection constraint rule are obtained; a spatial transformation matrix is ​​obtained based on the interface area and the interface matching method; the interface node positions of the parametric component model are adjusted according to the spatial transformation matrix and aligned with the grid nodes to obtain a spatial positioning model of the structural component; the spatial positioning model is organized according to the hierarchical combination method to generate a structural component assembly unit; The category label, geometric shape and attachment conditions defined in the preset decorative base surface rules of the non-structural component are calculated by a normal projection method to obtain an attachment model; the structural component assembly unit and the attachment model are spatially combined to obtain an assembly structure.

8. The method for intelligently reconstructing a three-dimensional building model based on deep learning according to claim 7, characterized in that: The method of adjusting the interface node positions of the parameterized component model according to the spatial transformation matrix and aligning them with the grid nodes to obtain a spatial positioning model of the structural component includes: The boundary contour of the parameterized component is aligned with the adjacent area in the final mesh model, and the joint deviation is obtained through geometric consistency test. When the joint deviation exceeds the preset assembly tolerance threshold, a nonlinear optimization algorithm is used to iteratively adjust the spatial position of the anchor point within the preset degree of freedom constraint range. After each iterative adjustment, the joint deviation is recalculated, and the spatial positioning model of the structural component is obtained when the joint deviation meets the assembly tolerance threshold.

9. The method for intelligently reconstructing a three-dimensional building model based on deep learning according to claim 7, wherein: The method of fusing the structural component grid area in the final grid model with the assembly structure according to connection constraint rules to obtain a structured three-dimensional model of the target ancient building includes: The method comprises obtaining a mesh area of ​​a structural component in a final mesh model, establishing a mapping relationship in combination with the spatial pose of the assembly structure, processing the mesh area according to the mapping relationship and the interface matching method defined in the connection constraint rules to obtain a geometric expression of the structural component, generating an adjacency relationship graph according to the connection constraint rules of the structural component, and combining the geometric expression with the adjacency relationship graph to generate a structured three-dimensional model.

10. A deep learning-based intelligent building 3D model reconstruction system, characterized by: The system comprises: A data acquisition and preprocessing module is used to obtain multi-source data of the target ancient building, the multi-source data including visible light image sequences, laser point cloud data, and depth image data; obtain a sparse point cloud from the multi-source data through a structure-from-motion algorithm, and obtain a dense point cloud based on the sparse point cloud through multi-view stereo matching; obtain semantic label information from the dense point cloud through a semantic segmentation operation, the semantic label information including first information corresponding to structural components and second information corresponding to non-structural components; and obtain an initial mesh model through surface reconstruction of the dense point cloud carrying the semantic label information; a structure analysis and template matching module, configured to analyze the connection relationship between structural components using a neural network to obtain a hierarchical combination mode and connection constraint rules of the structural components; and match the first information and the hierarchical combination mode with a corresponding component parameter template using a preset structural knowledge base; The model optimization and assembly output module is used to obtain the grid nodes corresponding to the structural components in the initial grid model, adjust the spatial positions of the grid nodes according to the connection constraint rules to obtain the final grid model; perform spatial assembly according to the component parameter template and hierarchical combination method to obtain the assembly structure; and fuse the structural component grid area in the final grid model with the assembly structure according to the connection constraint rules to obtain a structured three-dimensional model of the target ancient building.

Images