Method and system for quickly constructing three-dimensional space model of multi-level building

Abstract

The invention discloses a rapid construction method and system for a three-dimensional space model of a multi-level building, and relates to the field of building model construction, and the method comprises the following operation steps: S1, constructing a dynamic demand-oriented detail level model; s2, automatic component modeling based on image segmentation; s3, performing multi-version difference judgment and model updating; and S4, cloud collaboration and load optimization. According to the method and the system for quickly constructing the three-dimensional space model of the multilevel building, the region is dynamically divided through the semantic weight, so that the rendering efficiency during model construction can be effectively improved, the memory can be effectively reduced during rendering, the visual fidelity can be optimized, key curvature characteristics can be forcibly reserved for a high semantic weight region, and the construction efficiency is improved. Based on this, deformation distortion caused by simplification can be avoided, a visible area in the model can be automatically identified, an invisible area can be degraded, only the contour line frame of the building is reserved, and the overall memory occupation can be greatly reduced.

Description

Technical Field

[0001] The invention relates to the field of building model construction, and in particular to a method and system for quickly constructing a three-dimensional space model of a multi-level building. Background Art

[0002] The architectural model is between the plane drawing and the actual three-dimensional space. It organically links the two together and is a three-dimensional model. The architectural model helps to consider the design creation, can intuitively reflect the design intention, and make up for the limitations of the drawings in expression. It is both a part of the designer's design process and a form of expression of design. It is widely used in urban construction, real estate development and other aspects. The architectural model is an architectural term that expresses the spatial effect of the design plan with its unique imagery.

[0003] When constructing a three-dimensional spatial model of a building, an LOD model is used. However, it cannot dynamically adjust details according to the application scenario, which will lead to a waste of resources during assembly, operation and viewing. In addition, when constructing the model, the image of the building needs to be segmented, but traditional image segmentation technology has a low recognition rate for complex materials in building components, and manual labeling is required, resulting in a long overall time consumption. When updating, most of the time, the entire model needs to be rebuilt, which is time-consuming and will result in additional costs.

[0004] Therefore, it is necessary to propose a method and system for quickly constructing a three-dimensional spatial model of a multi-level building to solve the above problems. Summary of the invention

[0005] The main purpose of the present invention is to provide a method and system for quickly constructing a three-dimensional space model of a multi-level building, which can effectively solve the problems in the background technology.

[0006] To achieve the above object, the technical solution adopted by the present invention is: A method for quickly constructing a three-dimensional space model of a multi-level building includes the following steps: S1: Build a dynamic demand-oriented level of detail model. Based on the traditional LOD model, introduce a dynamic semantic weight mechanism to adjust the level of detail of different areas in real time according to the application scenario of the building; S2: Automated component modeling based on image segmentation, using a multimodal image segmentation network to intelligently identify and segment building components, and generating a parametric component library in combination with point cloud data; S3: Multi-version difference identification and model update, using geometric hash coding and semantic difference trees to identify the changed areas of the model and remodel only the updated parts; S4: Cloud collaboration and load optimization, using edge computing and GPU clusters to achieve distributed rendering of buildings, combined with LOD dynamic loading strategy to reduce the load on terminal devices.

[0007] Preferably, the S1 specifically includes the following steps: S101: Data preparation, collecting geometric data, semantic data, and scene data of buildings, where geometric data includes point cloud data, BIM models, and GIS data of buildings; semantic data includes functional attributes and spatial topological relationships of buildings; scene data includes application scenario definitions, key areas, and background areas of buildings. Preprocess the collected data, clean the geometric data, and align the semantic data; S102: Generate a dynamic LOD model, assign semantic weights to different building areas according to the application scenario requirements of the building, and divide the LOD levels based on this; S103: semantic fusion and topological mapping, binding high semantic weight areas and low semantic weight areas with associated LOD models to ensure that semantic attributes are dynamically adjusted with the model. Based on the IFC standard, the spatial topological relationship of the building is constructed. According to the user's perspective and application requirements, the high-precision model of the high semantic weight area and the simplified model of the low semantic weight area are dynamically loaded; S104: The dynamic LOD model is stored in blocks and a spatial index is established to support fast retrieval and loading. GPU accelerated rendering technology is used in combination with the LOD dynamic loading strategy to perform real-time rendering and optimization of the LOD model. The LOD dynamic loading strategy includes on-demand loading: dynamically loading high-precision model blocks according to the user's perspective; ray tracing optimization: ray tracing technology is used for high semantic weight areas to improve rendering quality.

[0008] Preferably, the S102 specifically includes the following steps: S10201: Assignment of semantic weights: According to the application scenario requirements, the semantic weights of different areas are assigned into high semantic weight areas and low semantic weight areas. High semantic weight areas are areas with important functionality and greater impact on safety and overall structure in the building; low semantic weight areas are areas that can be partially damaged or modified without causing fatal impact on the core functions and safety of the building; S10202: LOD level division, according to the semantic weight, the LOD level is dynamically divided: LOD0: only retain the building outline with the lowest progress; LOD1: simple building geometry; LOD2: building geometry with details; LOD3: high-precision building geometry; S10203: Perform geometric simplification on low semantic weight areas. Use the QEM algorithm to simplify the mesh in low semantic weight areas, reduce the number of vertices and faces, and reduce the amount of geometric data by 40%-60% based on the lack of obvious visual differences.

[0009] Preferably, the step S2 specifically includes the following steps: S201: Collection and preprocessing of multimodal data, input RGB images: high-resolution building facades and roof photos; point cloud data: building geometry information obtained by LiDAR and structured light scanning; thermal infrared images: data used to identify building material properties; semantic labels: based on the component properties in the BIM model, preprocess the above data: affine transformation and ICP algorithm are used to align RGB images and point cloud data to eliminate perspective differences; for problems such as moiré and uneven illumination, non-local mean filtering and generative adversarial networks are used to repair image defects; S202: Multi-task image segmentation, improves Mask R-CNN by building a dual-branch structure, including an RGB branch and a thermal infrared branch, introduces the SENet module, splices the feature maps extracted by the RGB branch and the thermal infrared branch in the channel dimension to which they belong, obtains the fused feature map, performs global average pooling on the fused feature map, compresses the feature map of each channel into a scalar, obtains the global feature information of the channel, performs a nonlinear transformation on the global feature information through a fully connected layer, and then obtains the attention weight of each channel through a Sigmoid function, multiplies the attention weight with the fused feature map channel by channel, weights the feature map, highlights important channel information, and is used to suppress noise interference. Based on this, a feature map that combines RGB and thermal infrared features is obtained; S203: Component-level semantic labeling: segment the mask of the building component instance using the improved Mask R-CNN, and associate the segmented mask with the detailed information of the building component in the BIM model. For each segmented instance, match it with the corresponding component in the BIM model, find the most likely corresponding component in the BIM model according to the location and shape characteristics of the instance, and assign a corresponding BIM semantic label to each segmented instance to achieve preliminary classification of the component. After determining the BIM semantic labels of the segmented instances, these instances are bound to the IFC standard attributes, and the IFC standard attribute information of each component is extracted from the BIM model. The attribute information is associated with the corresponding segmented instances and stored in the database for subsequent query and analysis; S204: Generation of topological relationships: treating the segmented building components as nodes in a graph, constructing a graph structure according to the spatial position relationship between the components, defining a feature vector for each node component, including the type and attribute information of the component, using a graph neural network to learn the graph structure, and updating the feature vector of the node based on a message passing mechanism to learn the topological relationship between the components, and finally obtaining a graph structure containing the topological relationship between the building components; S205: Parametric modeling of building components. Extract the size and curvature parameters of building components from the point cloud, generate editable CAD templates, call the preset template library according to the type of building, automatically adjust the parameters, splice the components into a complete building model based on the Poisson reconstruction algorithm, detect logical conflicts between components and repair them, and adjust the LOD level of the components according to the application scenarios of the components.

[0010] Preferably, in S202, the RGB branch specifically includes: inputting the RGB image into the RGB branch, the RGB image is used to identify the appearance features of the building, and using ResNet to extract features from the RGB image, thereby obtaining a series of RGB feature maps of different scales for subsequent feature fusion and target segmentation; The thermal infrared branch specifically includes: inputting thermal infrared images into the thermal infrared branch, the thermal infrared images are used to reflect the thermal radiation characteristics of the building, and are used to provide information about the material and temperature distribution of the building. Feature extraction is also performed on them, based on which a series of thermal infrared feature maps of different scales are obtained.

[0011] Preferably, the step S3 specifically includes the following steps: S301: Collect different versions of BIM model data. The different versions of BIM model data may come from different design stages, construction stages, or new versions generated due to design changes. Preprocess the collected BIM model data, including data format unification and model cleaning. S302: extracting features from geometric elements of building components in the BIM model, performing hash coding on the extracted features, and generating a unique hash value for converting complex features into digital codes of fixed length; Analyze the semantic information of the elements of the building components in the BIM model, including the type of elements, attributes and the relationship between them, and build a semantic difference tree based on the semantic information. The nodes of the tree represent the elements in the BIM model, and the edges between the nodes represent the relationship between the elements. By comparing different versions of the semantic difference tree, it is used to identify the changes at the semantic level in the model, including the addition, deletion and modification of elements and attributes. By comparing the geometric hash codes and semantic difference trees of different versions of the BIM model, the changed area of ​​the model is determined; S303: Remodel the changed area. According to the changed area determined in S302, extract the element information that needs to be updated, prepare the tools and environment required for modeling, and regenerate the 3D model of the elements whose geometric shapes have changed according to the new geometric features; for the newly added elements, create a new 3D model according to their semantic information and design requirements; for the elements whose attributes have changed, update their attribute information, merge the updated part after remodeling with the unchanged part, and generate a new BIM model version based on this.

[0012] Preferably, the S4 specifically comprises the following steps: S401: Build a cloud data center and deploy multiple GPU clusters with powerful graphics processing capabilities to undertake the main rendering computing tasks. At the same time, edge computing devices are deployed at the edge nodes of the network. The edge computing devices are close to the end users to reduce data transmission delays, and then the 3D model data to be rendered is uploaded to the cloud server; S402: Task allocation. After receiving the rendering task, the cloud server decomposes the rendering task into multiple subtasks according to the idle state and computing power of the GPU cluster, and allocates them to different GPU clusters for parallel processing. The edge computing device is responsible for collecting information from the terminal device, including the user's network status and device performance. At the same time, it performs preliminary processing on some of the received data, including simple geometric transformation and data compression, to reduce the burden on the cloud server, and transmits the processed data to the cloud server again for further rendering calculation. During the rendering process, data interaction and synchronization are performed between the GPU clusters to ensure the consistency of the rendering results. S403: Combined with the LOD dynamic loading strategy, the load on the terminal device is reduced. When the terminal device starts the rendering application, the performance information of the terminal device is collected, including the processing power of the CPU and GPU, the memory size, and the user's network status. According to the performance information of the terminal device and the user's network status, as well as the user's current viewing angle and observation distance, the LOD level model in step S10202 is dynamically loaded. According to the dynamic loading strategy, the LOD level model data of the corresponding detail level is loaded from the cloud server to the terminal device, and the memory of the terminal device is saved. When a certain LOD level model is no longer within the user's observation range, it is unloaded from the memory of the terminal device.

[0013] A multi-level building three-dimensional space model rapid construction system, including a dynamic LOD management module, a multimodal component modeling module, a version difference processing module, and a cloud collaborative rendering module. The dynamic LOD management module is used to realize dynamic adjustment and optimization of the building detail level; The multimodal component modeling module is used to automatically identify, segment and parameterize building components; The version difference processing module is used to identify the model change area and update it incrementally; The cloud collaborative rendering module is used to perform distributed rendering and terminal load optimization on the building model; The dynamic LOD management module includes a data preprocessing unit, a semantic weight allocation module, a LOD hierarchical division module, and a dynamic loading strategy module. The data preprocessing unit is used to collect point cloud data, BIM data, and GIS data related to the building and integrate them. At the same time, the purpose of the building components is extracted from the BIM data and an adjacency matrix is ​​constructed to describe the connectivity of the building components. The key areas of the building are defined, and the building components within 5m from the viewpoint are regarded as high-detail areas. The above data are aligned and cleaned; the semantic weight allocation module is used to divide the area of ​​the building into high semantic weight areas and low semantic weight areas, and at the same time, the structural weight of the building is quantified, including based on the mechanics of the components. Parameter normalization calculation, manual labeling and grading of building functions; the LOD level division module is used to divide LOD into LOD0-3, and perform edge folding operations on components in low semantic weight areas, retaining feature edges with curvature changes >15° to prevent structural distortion, set the maximum allowable error to 0.05m, and reduce the number of triangles by 40%-60%; the dynamic loading strategy module is used to build a quadtree spatial index, calculate the blocks within the frustum in real time, and set high details to load the near-field area of ​​the building radius, load LOD3 at 0-10m, load LOD2 at 10-50m, enable ray tracing reflection and shadow calculation for high semantic weight areas, and use baked light maps for low semantic weight areas.

[0014] Preferably, the multimodal component modeling module includes a multimodal data fusion network, a building component semantic annotation module, a topological relationship generation module, and a parametric modeling module, wherein the multimodal data fusion network includes a dual-branch improved MaskR-CNN, the dual branches include an RGB branch and a thermal infrared branch, the RGB branch extracts the appearance features of the building based on ResNet, and the thermal infrared branch is used to extract the material attribute features of the building, and at the same time, the SENet channel attention mechanism is introduced to fuse the features of the above branches and suppress noise; The building component semantic annotation module calculates the minimum bounding box AABB and centroid coordinates of the segmented components, searches for components with spatial position overlap > 80% in the BIM model, calculates the Hausdorff distance between the segmented component point cloud and the candidate component in the BIM model, and sets the threshold to 0.1m. The matching IFC entity attributes are attached to the components. Specifically, based on the segmented component point cloud P and the candidate component point cloud Q in the BIM model, the following is calculated: ; H is the Hausdorff distance, if Threshold, the two are considered to be geometrically matched and can be given the same semantic label; The topological relationship generation module is used to design the building graph neural network. Its node features include the type, size parameters and location coordinates of building components, and the edge connection rules are the physical collision and functional association between building components. The parametric modeling module includes extracting the parameters of relevant buildings from the point cloud to generate CAD templates and detecting logical conflicts, including using the GJK algorithm to detect component penetration and verifying the rules between building components according to building specifications.

[0015] Preferably, the version difference processing module includes extracting the features of the building based on the geometric hash encoder, including calculating the volume, surface curvature, and center of mass coordinates of each building component, and inputting the above-calculated features into the SHA-256 algorithm after splicing, and constructing a tree structure based on the IFC attributes to identify the addition and deletion of elements and attribute changes, and based on this, perform difference detection on the building, including geometric changes: inconsistent hash values; attribute changes: comparing the IFC attribute dictionary, and setting a local reconstruction strategy, including column radius adjustment: retaining the unchanged part of the point cloud, and only performing Poisson reconstruction on the new radius area, adding windows: digging holes in the wall point cloud, and data merging: using the KD tree spatial index to quickly locate the insertion position.

[0016] Compared with the prior art, the present invention provides a method and system for quickly constructing a three-dimensional space model of a multi-level building, which has the following beneficial effects: This method and system for quickly constructing a three-dimensional spatial model of a multi-level building can effectively improve the rendering efficiency when building the model, effectively reduce memory usage when rendering, and optimize visual fidelity. It can force the retention of key curvature features in high semantic weight areas, thereby avoiding deformation and distortion caused by simplification. It can automatically identify visible areas in the model, downgrade invisible areas, and retain only the outline of the building, which can significantly reduce overall memory usage.

[0017] This method and system for rapidly constructing a multi-level three-dimensional spatial model of a building can improve the accuracy of material recognition of the building based on thermal infrared data through multi-modal feature fusion, and can also associate segmented components with BIM attributes, effectively improving the efficiency of semantic annotation, based on which the components of the building can be automatically annotated.

[0018] This method and system for quickly building a three-dimensional spatial model of a multi-level building can reduce the need for full model reconstruction and improve iteration efficiency through difference discrimination technology. It is especially suitable for scenes that are frequently modified during the construction phase. When viewing the model, high-precision model blocks can be dynamically loaded according to the user's perspective, combined with ray tracing optimization technology to improve rendering quality and speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0020] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods. Example

[0021] like Figure 1 As shown, a method for quickly constructing a three-dimensional space model of a multi-level building includes the following steps: S1: Build a dynamic demand-oriented level of detail model. Based on the traditional LOD model, introduce a dynamic semantic weight mechanism to adjust the level of detail of different areas in real time according to the application scenario of the building. The specific steps include the following: S101: Data preparation, collecting geometric data, semantic data, and scene data of buildings, where geometric data includes point cloud data, BIM models, and GIS data of buildings; semantic data includes functional attributes and spatial topological relationships of buildings; scene data includes application scenario definitions, key areas, and background areas of buildings. Preprocess the collected data, clean the geometric data, and align the semantic data; S102: Generate a dynamic LOD model, assign semantic weights to different building areas according to the application scenario requirements of the building, and divide the LOD levels based on this, which specifically includes the following steps: S10201: Assignment of semantic weights: According to the application scenario requirements, the semantic weights of different areas are assigned into high semantic weight areas and low semantic weight areas. High semantic weight areas are areas with important functionality and greater impact on safety and overall structure in the building; low semantic weight areas are areas that can be partially damaged or modified without causing fatal impact on the core functions and safety of the building; S10202: LOD level division, according to the semantic weight, the LOD level is dynamically divided: LOD0: only retain the building outline with the lowest progress; LOD1: simple building geometry; LOD2: building geometry with details; LOD3: high-precision building geometry; S10203: Simplify the geometry of low semantic weight areas. Use the QEM algorithm to simplify the mesh of low semantic weight areas, reduce the number of vertices and faces, and reduce the amount of geometric data by 40%-60% based on the lack of obvious visual differences. S103: semantic fusion and topological mapping, binding high semantic weight areas and low semantic weight areas with associated LOD models to ensure that semantic attributes are dynamically adjusted with the model. Based on the IFC standard, the spatial topological relationship of the building is constructed. According to the user's perspective and application requirements, the high-precision model of the high semantic weight area and the simplified model of the low semantic weight area are dynamically loaded; S104: The dynamic LOD model is stored in blocks and a spatial index is established to support fast retrieval and loading. GPU accelerated rendering technology is used in combination with the LOD dynamic loading strategy to perform real-time rendering and optimization of the LOD model. The LOD dynamic loading strategy includes on-demand loading: dynamically loading high-precision model blocks according to the user's perspective; ray tracing optimization: ray tracing technology is used for high semantic weight areas to improve rendering quality.

[0022] S2: Automated component modeling based on image segmentation uses a multimodal image segmentation network to intelligently identify and segment building components, and generates a parametric component library based on point cloud data. The specific steps include: S201: Collection and preprocessing of multimodal data, input RGB images: high-resolution building facades and roof photos; point cloud data: building geometry information obtained by LiDAR and structured light scanning; thermal infrared images: data used to identify building material properties; semantic labels: based on the component properties in the BIM model, preprocess the above data: align RGB images and point cloud data using affine transformation and ICP algorithm to eliminate perspective differences; use non-local mean filtering and generative adversarial networks to repair image defects for problems such as moiré and uneven illumination; S202: Multi-task image segmentation, improve Mask R-CNN by building a dual-branch structure, including an RGB branch and a thermal infrared branch, introduce the SENet module, splice the feature maps extracted by the RGB branch and the thermal infrared branch in the channel dimension to which they belong, obtain the fused feature map, perform global average pooling on the fused feature map, compress the feature map of each channel into a scalar, obtain the global feature information of the channel, perform nonlinear transformation on the global feature information through a fully connected layer, and then obtain the attention weight of each channel through a Sigmoid function, multiply the attention weight by the fused feature map channel by channel, weight the feature map, highlight important channel information, and suppress noise interference, based on which a feature map that combines RGB and thermal infrared features is obtained, wherein the RGB branch specifically includes: inputting an RGB image into the RGB branch, the RGB image is used to identify the appearance features of the building, and ResNet is used to extract features from the RGB image, based on which a series of RGB feature maps of different scales are obtained for subsequent feature fusion and target segmentation; The thermal infrared branch specifically includes: inputting the thermal infrared image into the thermal infrared branch, the thermal infrared image is used to reflect the thermal radiation characteristics of the building, and is used to provide information about the material and temperature distribution of the building. Feature extraction is also performed on the thermal infrared image, based on which a series of thermal infrared feature maps of different scales are obtained; S203: Component-level semantic labeling: segment the mask of the building component instance using the improved Mask R-CNN, and associate the segmented mask with the detailed information of the building component in the BIM model. For each segmented instance, match it with the corresponding component in the BIM model, find the most likely corresponding component in the BIM model according to the location and shape characteristics of the instance, and assign a corresponding BIM semantic label to each segmented instance to achieve preliminary classification of the component. After determining the BIM semantic labels of the segmented instances, these instances are bound to the IFC standard attributes, and the IFC standard attribute information of each component is extracted from the BIM model. The attribute information is associated with the corresponding segmented instances and stored in the database for subsequent query and analysis; S204: Generation of topological relationships: treating the segmented building components as nodes in a graph, constructing a graph structure according to the spatial position relationship between the components, defining a feature vector for each node component, including the type and attribute information of the component, using a graph neural network to learn the graph structure, and updating the feature vector of the node based on a message passing mechanism to learn the topological relationship between the components, and finally obtaining a graph structure containing the topological relationship between the building components; S205: Parametric modeling of building components. Extract the size and curvature parameters of building components from the point cloud, generate editable CAD templates, call the preset template library according to the type of building, automatically adjust the parameters, splice the components into a complete building model based on the Poisson reconstruction algorithm, detect logical conflicts between components and repair them, and adjust the LOD level of the components according to the application scenarios of the components.

[0023] S3: Multi-version difference identification and model update, using geometric hash coding and semantic difference trees to identify the changed areas of the model, and to remodel only the updated parts, specifically including the following steps: S301: Collect different versions of BIM model data. The different versions of BIM model data may come from different design stages, construction stages, or new versions generated due to design changes. Preprocess the collected BIM model data, including data format unification and model cleaning. S302: extracting features from geometric elements of building components in the BIM model, performing hash coding on the extracted features, and generating a unique hash value for converting complex features into digital codes of fixed length for quick comparison and matching. If the hash values ​​of geometric elements in different versions of the BIM model are the same, it is considered that these elements have not changed geometrically. Analyze the semantic information of the elements of the building components in the BIM model, including the type and attributes of the elements and the relationship between them. Construct a semantic difference tree based on the semantic information. The nodes of the tree represent the elements in the BIM model, and the edges between the nodes represent the relationship between the elements. By comparing different versions of the semantic difference tree, it is used to identify the changes at the semantic level in the model, including the addition, deletion, and modification of the attributes of the elements. By comparing the geometric hash codes and semantic difference trees of different versions of the BIM model, the changed area of ​​the model is determined. The changed area can be the elements with changed geometry, the elements that are added or deleted, and the parts with changed element attributes. S303: Remodel the changed area. According to the changed area determined in S302, extract the element information that needs to be updated, prepare the tools and environment required for modeling, and regenerate the 3D model of the elements whose geometric shapes have changed according to the new geometric features. If the radius of the column in the building component has changed, use the new radius value to rebuild the 3D model of the column; for the newly added elements, create a new 3D model according to their semantic information and design requirements. If a new window is added to the building component, create a 3D model of the window according to the type and size information of the window; for the elements whose attributes have changed, update their attribute information. If the material of the wall is changed from concrete to masonry, update the material attributes of the wall, merge the updated part after remodeling with the unchanged part, and generate a new BIM model version based on this.

[0024] S4: Cloud collaboration and load optimization, using edge computing and GPU clusters to achieve distributed rendering of buildings, combined with LOD dynamic loading strategy to reduce the load on terminal devices, specifically including the following steps S401: Build a cloud data center and deploy multiple GPU clusters with powerful graphics processing capabilities to undertake the main rendering computing tasks. At the same time, edge computing devices are deployed at the edge nodes of the network. The edge computing devices are close to the end users to reduce data transmission delays, and then the 3D model data to be rendered is uploaded to the cloud server; S402: Task allocation. After receiving the rendering task, the cloud server decomposes the rendering task into multiple subtasks according to the idle state and computing power of the GPU cluster, and allocates them to different GPU clusters for parallel processing. For a large-scale architectural scene rendering task, the scene can be divided into multiple sub-areas according to the area, and the rendering task of each sub-area is allocated to one or more GPUs for processing; the edge computing device is responsible for collecting information from the terminal device, including the user's network status and device performance. At the same time, it performs preliminary processing on some of the received data, including simple geometric transformation and data compression, to reduce the burden on the cloud server, and transmits the processed data to the cloud server again for further rendering calculation; during the rendering process, data interaction and synchronization are performed between the various GPU clusters to ensure the consistency of the rendering results. When a GPU completes the rendering of a sub-area, it needs to feed back the rendering results to other related GPUs for overall synthesis and optimization; S403: Combined with the LOD dynamic loading strategy, the load on the terminal device is reduced. When the terminal device starts the rendering application, the performance information of the terminal device is collected, including the processing power of the CPU and GPU, the memory size, and the user's network status. According to the performance information of the terminal device and the user's network status, as well as the user's current viewing angle and observation distance, the LOD level model in step S10202 is dynamically loaded. According to the dynamic loading strategy, the LOD level model data of the corresponding detail level is loaded from the cloud server to the terminal device, and the memory of the terminal device is saved. When a certain LOD level model is no longer within the user's observation range, it is unloaded from the memory of the terminal device. Example

[0025] A system for rapidly constructing a three-dimensional spatial model of a multi-level building, comprising a dynamic LOD management module, a multimodal component modeling module, a version difference processing module, and a cloud collaborative rendering module. The dynamic LOD management module is used to realize dynamic adjustment and optimization of the building detail level. The multimodal component modeling module is used to automatically identify, segment and parameterize building components; The version difference processing module is used to identify the model change areas and update them incrementally; The cloud collaborative rendering module is used for distributed rendering of building models and terminal load optimization; The dynamic LOD management module includes a data preprocessing unit, a semantic weight allocation module, a LOD hierarchical division module, and a dynamic loading strategy module. The data preprocessing unit is used to collect and integrate point cloud data, BIM data, and GIS data related to the building. At the same time, it extracts the purpose of building components from BIM data and constructs an adjacency matrix to describe the connectivity of building components, defines the key areas of the building, and regards the building components within 5m from the viewpoint as high-detail areas. The above data are aligned and cleaned; the semantic weight allocation module is used to divide the area of ​​the building into high semantic weight areas and low semantic weight areas, and quantify the structural weight of the building, including normalization calculation based on component mechanical parameters, manual calculation of the building structure, and so on. The functions of objects are labeled and graded, such as 0-1, fire-fighting facilities = 1.0, ordinary walls = 0.3; the LOD level division module is used to divide LOD into LOD0-3, and perform edge folding operations on components in low semantic weight areas, retaining feature edges with curvature changes > 15° to prevent structural distortion, set the maximum allowable error to 0.05m, and reduce the number of triangles by 40%-60%; the dynamic loading strategy module is used to build a quadtree spatial index, calculate the blocks within the frustum in real time, and set high details to load the near-field area of ​​the building radius, load LOD3 for 0-10m, load LOD2 for 10-50m, enable ray tracing reflection and shadow calculation for high semantic weight areas, and use baked light maps for low semantic weight areas.

[0026] The multimodal component modeling module includes a multimodal data fusion network, a building component semantic annotation module, a topological relationship generation module, and a parametric modeling module. The multimodal data fusion network includes a dual-branch improved MaskR-CNN. The dual branches include an RGB branch and a thermal infrared branch. The RGB branch extracts the appearance features of the building based on ResNet, and the thermal infrared branch is used to extract the material attribute features of the building. At the same time, the SENet channel attention mechanism is introduced to fuse the features of the above branches and suppress noise. The building component semantic annotation module calculates the minimum bounding box AABB and centroid coordinates of the segmented components, searches for components with spatial position overlap > 80% in the BIM model, calculates the Hausdorff distance between the segmented component point cloud and the candidate component in the BIM model, and sets the threshold to 0.1m. The matching IFC entity attributes are attached to the components. Specifically, based on the segmented component point cloud P and the candidate component point cloud Q in the BIM model, the following is calculated: H is the Hausdorff distance, if Threshold, the two are considered to be geometrically matched and can be given the same semantic label; The topological relationship generation module is used to design the building graph neural network. Its node features include the type, size parameters and location coordinates of building components, and the edge connection rules are the physical collision and functional association between building components. The parametric modeling module includes extracting the parameters of relevant buildings from the point cloud to generate CAD templates and detecting logical conflicts, including using the GJK algorithm to detect component penetration and verifying the rules between building components according to building specifications.

[0027] The version difference processing module includes extracting the features of the building based on the geometric hash encoder, including calculating the volume, surface curvature, and center of mass coordinates of each building component. At the same time, the features of the above calculations are spliced ​​and input into the SHA-256 algorithm. At the same time, a tree structure is built based on IFC attributes to identify the addition and deletion of elements and changes in attributes. Based on this, the difference detection of buildings is performed, including geometric changes: inconsistent hash values; attribute changes: comparing the IFC attribute dictionary and setting local reconstruction strategies, including column radius adjustment: retaining the unchanged part of the point cloud, and only performing Poisson reconstruction on the new radius area, adding windows: digging holes in the wall point cloud, and data merging: using the KD tree spatial index to quickly locate the insertion position.

[0028] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for rapidly constructing a three-dimensional spatial model of a multi-level building, characterized in that: The steps include: S1: Build a dynamic demand-oriented level of detail model. Based on the traditional LOD model, introduce a dynamic semantic weight mechanism to adjust the level of detail of different areas in real time according to the application scenario of the building; S2: Automated component modeling based on image segmentation, using a multimodal image segmentation network to intelligently identify and segment building components, and generating a parametric component library in combination with point cloud data; S3: Multi-version difference identification and model update, using geometric hash coding and semantic difference trees to identify the changed areas of the model and remodel only the updated parts; S4: Cloud collaboration and load optimization, using edge computing and GPU clusters to achieve distributed rendering of buildings, combined with LOD dynamic loading strategy to reduce the load on terminal devices.

2. A method for rapidly constructing a three-dimensional space model of a multi-level building according to claim 1, characterized in that: The S1 specifically includes the following steps: S101: Data preparation, collecting geometric data, semantic data, and scene data of buildings, where geometric data includes point cloud data, BIM models, and GIS data of buildings; Semantic data includes functional attributes and spatial topological relationships of buildings; scene data includes application scene definitions, key areas and background areas of buildings. The collected data is preprocessed, the geometric data is cleaned, and the semantic data is aligned. S102: Generate a dynamic LOD model, assign semantic weights to different building areas according to the application scenario requirements of the building, and divide the LOD levels based on this; S103: semantic fusion and topological mapping, binding high semantic weight areas and low semantic weight areas with associated LOD models to ensure that semantic attributes are dynamically adjusted with the model. Based on the IFC standard, the spatial topological relationship of the building is constructed. According to the user's perspective and application requirements, the high-precision model of the high semantic weight area and the simplified model of the low semantic weight area are dynamically loaded; S104: The dynamic LOD model is stored in blocks and a spatial index is established to support fast retrieval and loading. GPU accelerated rendering technology is used in combination with the LOD dynamic loading strategy to perform real-time rendering and optimization of the LOD model. The LOD dynamic loading strategy includes on-demand loading: dynamically loading high-precision model blocks according to the user's perspective; ray tracing optimization: ray tracing technology is used for high semantic weight areas to improve rendering quality.

3. A method for rapidly constructing a three-dimensional space model of a multi-level building according to claim 2, characterized in that: The S102 specifically includes the following steps: S10201: Assignment of semantic weights: According to the application scenario requirements, the semantic weights of different areas are assigned into high semantic weight areas and low semantic weight areas. High semantic weight areas are areas with important functionality and greater impact on safety and overall structure in the building; low semantic weight areas are areas that can be partially damaged or modified without causing fatal impact on the core functions and safety of the building; S10202: LOD level division, according to the semantic weight, the LOD level is dynamically divided: LOD0: only retain the building outline with the lowest progress; LOD1: simple building geometry; LOD2: building geometry with details; LOD3: high-precision building geometry; S10203: Perform geometric simplification on low semantic weight areas. Use the QEM algorithm to simplify the mesh in low semantic weight areas, reduce the number of vertices and faces, and reduce the amount of geometric data by 40%-60% based on the lack of obvious visual differences.

4. The method for rapidly constructing a three-dimensional space model of a multi-level building according to claim 1, characterized in that: The S2 specifically includes the following steps: S201: Collection and preprocessing of multimodal data, input RGB images: high-resolution building facades and roof photos; point cloud data: building geometry information obtained by LiDAR and structured light scanning; thermal infrared images: data used to identify building material properties; semantic labels: based on the component properties in the BIM model, preprocess the above data: affine transformation and ICP algorithm are used to align RGB images and point cloud data to eliminate perspective differences; for problems such as moiré and uneven illumination, non-local mean filtering and generative adversarial networks are used to repair image defects; S202: Multi-task image segmentation, improves Mask R-CNN by building a dual-branch structure, including an RGB branch and a thermal infrared branch, introduces the SENet module, splices the feature maps extracted by the RGB branch and the thermal infrared branch in the channel dimension to which they belong, obtains the fused feature map, performs global average pooling on the fused feature map, compresses the feature map of each channel into a scalar, obtains the global feature information of the channel, performs a nonlinear transformation on the global feature information through a fully connected layer, and then obtains the attention weight of each channel through a Sigmoid function, multiplies the attention weight with the fused feature map channel by channel, weights the feature map, highlights important channel information, and is used to suppress noise interference. Based on this, a feature map that combines RGB and thermal infrared features is obtained; S203: Component-level semantic labeling: segment the mask of the building component instance using the improved Mask R-CNN, and associate the segmented mask with the detailed information of the building component in the BIM model. For each segmented instance, match it with the corresponding component in the BIM model, find the most likely corresponding component in the BIM model according to the location and shape characteristics of the instance, and assign a corresponding BIM semantic label to each segmented instance to achieve preliminary classification of the component. After determining the BIM semantic labels of the segmented instances, these instances are bound to the IFC standard attributes, and the IFC standard attribute information of each component is extracted from the BIM model. The attribute information is associated with the corresponding segmented instances and stored in the database for subsequent query and analysis; S204: Generation of topological relationships: treating the segmented building components as nodes in a graph, constructing a graph structure according to the spatial position relationship between the components, defining a feature vector for each node component, including the type and attribute information of the component, using a graph neural network to learn the graph structure, and updating the feature vector of the node based on a message passing mechanism to learn the topological relationship between the components, and finally obtaining a graph structure containing the topological relationship between the building components; S205: Parametric modeling of building components. Extract the size and curvature parameters of building components from the point cloud, generate editable CAD templates, call the preset template library according to the type of building, automatically adjust the parameters, splice the components into a complete building model based on the Poisson reconstruction algorithm, detect logical conflicts between components and repair them, and adjust the LOD level of the components according to the application scenarios of the components.

5. A method for rapidly constructing a three-dimensional space model of a multi-level building according to claim 4, characterized in that: In S202, the RGB branch specifically includes: inputting the RGB image into the RGB branch, the RGB image is used to identify the appearance features of the building, and using ResNet to extract features from the RGB image, based on which a series of RGB feature maps of different scales are obtained for subsequent feature fusion and target segmentation; The thermal infrared branch specifically includes: inputting thermal infrared images into the thermal infrared branch, the thermal infrared images are used to reflect the thermal radiation characteristics of the building, and are used to provide information about the material and temperature distribution of the building. Feature extraction is also performed on them, based on which a series of thermal infrared feature maps of different scales are obtained.

6. The method for rapidly constructing a three-dimensional space model of a multi-level building according to claim 1, characterized in that: The S3 specifically includes the following steps: S301: Collect different versions of BIM model data. The different versions of BIM model data may come from different design stages, construction stages, or new versions generated due to design changes. Preprocess the collected BIM model data, including data format unification and model cleaning. S302: extracting features from geometric elements of building components in the BIM model, performing hash coding on the extracted features, and generating a unique hash value for converting complex features into digital codes of fixed length; Analyze the semantic information of the elements of the building components in the BIM model, including the type of elements, attributes and the relationship between them, and build a semantic difference tree based on the semantic information. The nodes of the tree represent the elements in the BIM model, and the edges between the nodes represent the relationship between the elements. By comparing different versions of the semantic difference tree, it is used to identify the changes at the semantic level in the model, including the addition, deletion and modification of elements and attributes. By comparing the geometric hash codes and semantic difference trees of different versions of the BIM model, the changed area of ​​the model is determined; S303: Remodel the changed area. According to the changed area determined in S302, extract the element information that needs to be updated, prepare the tools and environment required for modeling, and regenerate the 3D model of the elements whose geometric shapes have changed according to the new geometric features; for the newly added elements, create a new 3D model according to their semantic information and design requirements; for the elements whose attributes have changed, update their attribute information, merge the updated part after remodeling with the unchanged part, and generate a new BIM model version based on this.

7. The method for rapidly constructing a three-dimensional space model of a multi-level building according to claim 1, characterized in that: The S4 specifically includes the following steps: S401: Build a cloud data center and deploy multiple GPU clusters with powerful graphics processing capabilities to undertake the main rendering computing tasks. At the same time, edge computing devices are deployed at the edge nodes of the network. The edge computing devices are close to the end users to reduce data transmission delays, and then the 3D model data to be rendered is uploaded to the cloud server; S402: Task allocation: After receiving the rendering task, the cloud server decomposes the rendering task into multiple subtasks according to the idle state and computing power of the GPU cluster, and allocates them to different GPU clusters for parallel processing; Edge computing devices are responsible for collecting information about terminal devices, including the user's network status and device performance. At the same time, they perform preliminary processing on some of the received data, including simple geometric transformation and data compression, to reduce the burden on the cloud server, and transmit the processed data to the cloud server for further rendering calculations. During the rendering process, data is exchanged and synchronized between GPU clusters to ensure the consistency of rendering results. S403: Combined with the LOD dynamic loading strategy, the load on the terminal device is reduced. When the terminal device starts the rendering application, the performance information of the terminal device is collected, including the processing power of the CPU and GPU, the memory size, and the user's network status. According to the performance information of the terminal device and the user's network status, as well as the user's current viewing angle and observation distance, the LOD level model in step S10202 is dynamically loaded. According to the dynamic loading strategy, the LOD level model data of the corresponding detail level is loaded from the cloud server to the terminal device, and the memory of the terminal device is saved. When a certain LOD level model is no longer within the user's observation range, it is unloaded from the memory of the terminal device.

8. A system for rapidly constructing a three-dimensional space model of a multi-level building, using a method for rapidly constructing a three-dimensional space model of a multi-level building as described in any one of claims 1 to 7, comprising a dynamic LOD management module, a multimodal component modeling module, a version difference processing module, and a cloud collaborative rendering module, characterized in that: The dynamic LOD management module is used to realize dynamic adjustment and optimization of the building detail level; The multimodal component modeling module is used to automatically identify, segment and parameterize building components; The version difference processing module is used to identify the model change area and update it incrementally; The cloud collaborative rendering module is used to perform distributed rendering and terminal load optimization on the building model; The dynamic LOD management module includes a data preprocessing unit, a semantic weight allocation module, an LOD level division module, and a dynamic loading strategy module. The data preprocessing unit is used to collect point cloud data, BIM data, and GIS data related to the building and integrate them, and at the same time extract the purpose of building components from the BIM data and construct an adjacency matrix to describe the connectivity of building components, define the key areas of the building, regard the building components within 5m from the viewpoint as high-detail areas, and align and clean the above data; The semantic weight allocation module is used to divide the area of ​​the building into a high semantic weight area and a low semantic weight area, and quantify the structural weight of the building, including normalizing and calculating the mechanical parameters of the components and manually labeling and grading the functions of the building; The LOD level division module is used to divide LOD into LOD0-3, and perform edge folding operations on components in low semantic weight areas, retaining feature edges with curvature changes >15° to prevent structural distortion, set the maximum allowable error to 0.05m, and reduce the number of triangles by 40%-60%; the dynamic loading strategy module is used to build a quadtree spatial index, calculate blocks within the frustum in real time, and set high details to load the near-field area of ​​the building radius, load LOD3 at 0-10m, load LOD2 at 10-50m, enable ray tracing reflection and shadow calculation for high semantic weight areas, and use baked light maps for low semantic weight areas.

9. A system for rapidly constructing a three-dimensional spatial model of a multi-level building according to claim 8, characterized in that: The multimodal component modeling module includes a multimodal data fusion network, a building component semantic annotation module, a topological relationship generation module, and a parametric modeling module, wherein the multimodal data fusion network includes a dual-branch improved MaskR-CNN, the dual branches include an RGB branch and a thermal infrared branch, the RGB branch extracts the appearance features of the building based on ResNet, and the thermal infrared branch is used to extract the material attribute features of the building, and at the same time introduces the SENet channel attention mechanism to fuse the features of the above branches and suppress noise; The building component semantic annotation module calculates the minimum bounding box AABB and centroid coordinates of the segmented components, searches for components with spatial position overlap>80% in the BIM model, calculates the Hausdorff distance between the segmented component point cloud and the candidate component in the BIM model, and sets the threshold to 0.1m. The matching IFC entity attributes are attached to the component. Specifically, based on the segmented component point cloud P and the candidate component point cloud Q in the BIM model, the following is calculated: ; H is the Hausdorff distance, if Threshold, the two are considered to be geometrically matched and can be given the same semantic label; The topological relationship generation module is used to design the building graph neural network. Its node features include the type, size parameters and location coordinates of building components, and the edge connection rules are the physical collision and functional association between building components. The parametric modeling module includes extracting the parameters of relevant buildings from the point cloud to generate CAD templates and detecting logical conflicts, including using the GJK algorithm to detect component penetration and verifying the rules between building components according to building specifications.

10. A system for rapidly constructing a three-dimensional space model of a multi-level building according to claim 8, characterized in that: The version difference processing module includes extracting the features of the building based on the geometric hash encoder, including calculating the volume, surface curvature, and centroid coordinates of each building component, and inputting the above calculated features into the SHA-256 algorithm after splicing, and building a tree structure based on the IFC attributes to identify the addition and deletion of elements and attribute changes, and based on this, perform difference detection of the building, including geometric changes: inconsistent hash values; Attribute changes: compare IFC attribute dictionaries and set local reconstruction strategies, including column radius adjustment: retain the unchanged part of the point cloud and perform Poisson reconstruction only on the new radius area, new windows: dig holes in the wall point cloud, data merging: use KD tree spatial index to quickly locate the insertion position.

Images