Method and device for reconstructing gain of coupling between information collected by multi-source in large scene and light-weight structure, equipment and medium

Abstract

The application discloses a large-scene multi-source acquisition information and light structured coupling gain reconstruction method and device, equipment and medium, the method comprises the following steps: generating an initial parameterized building model by using point cloud data; correcting the initial parameterized building model by using a relationship perception assembly network; comprehensively checking the corrected parameterized building model according to multi-source acquisition information; and performing multi-source information coupling gain according to the checking result to obtain a building information model. The application changes the processing mode of the scene observation data acquisition and model reconstruction as a relatively fixed one-way processing process. After the initial parameterized building model is reconstructed by using multi-source acquisition information, the multi-source acquisition information is used as scene evidence to roughly correct the initial parameterized building model and perform multi-source information coupling gain, so that the complementary advantages of the multi-source acquisition information are fully exerted, and the precision, completeness and structural consistency of the building information model are improved.

Description

Technical Field

[0001] This application relates to the field of digital building technology, and in particular to a method, apparatus, equipment and medium for reconstructing large-scale multi-source information acquisition and lightweight structured coupling gain. Background Technology

[0002] With the development of digitalization and intelligence in the construction industry, how to transform large-scale real-world building scenarios into building information models to provide a reliable digital foundation for applications such as digital archiving of existing buildings, acquisition of current status models, renovation and expansion, and project verification has become an important issue that urgently needs to be addressed in the field of building digitalization.

[0003] Current lightweight reconstruction methods generally treat scene observation data acquisition and model reconstruction as a relatively fixed, one-way process. They utilize acquired scene observation data (such as point cloud data) to perform geometric reconstruction to transform the architectural scene into a Building Information Model (BIM). While scene observation data can obtain spatial geometric information of real architectural scenes, it is susceptible to issues such as local occlusion, missing observations, and noise interference, especially in large-scale real-world architectural scenes. These problems can lead to various defects in the reconstructed BIM, including geometric information deviations, incorrect component category identification, and disordered spatial relationships between components, ultimately affecting the accuracy and completeness of the final BIM model.

[0004] Therefore, the existing technology still needs to be improved and enhanced. Summary of the Invention

[0005] The technical problem to be solved by this application is to provide a method, apparatus, device and medium for reconstructing large-scale multi-source information acquisition and lightweight structured coupling gain, in order to address the shortcomings of the existing technology.

[0006] To address the aforementioned technical problems, the first aspect of this application provides a method for reconstructing large-scene multi-source acquisition information and lightweight structured coupling gain, wherein the method specifically includes: Acquire multi-source acquisition information of a large-scale real building scene to be reconstructed, wherein the multi-source acquisition information includes point cloud data, multi-view images, depth maps, and normal maps; An initial parametric building model is generated based on the point cloud data, wherein the initial parametric building model includes the geometric information of building components and the structural relationships between components; The initial parametric building model is corrected using a relation-aware assembly network, wherein the relation-aware assembly network is configured with attention within building components, attention between building components, and cross attention between building components and point cloud data; The corrected parametric building model is comprehensively checked based on the point cloud data and the multi-view images to obtain a comprehensive check result. The comprehensive check includes one or more of geometric consistency check, semantic check and layout structure check. Based on the multi-source collected information and the comprehensive inspection results, a component-level modification strategy is determined through a large model. Then, based on the component-level modification strategy, multi-source information coupling gain is applied to the modified parametric building model to obtain a building information model.

[0007] The reconstruction method based on large-scale scene multi-source acquisition information and lightweight structured coupling gain, wherein acquiring the multi-source acquisition information of the large-scale real-world architectural scene to be reconstructed specifically includes: Collect multi-source scene observation information of a large-scale real building scene to be reconstructed, wherein the multi-source scene observation information includes point cloud data, multi-view images, depth information, surface orientation information, and acquisition posture; Based on the time information and the acquisition posture, the observation data in the multi-source scene observation information are transformed into the scene coordinate system, and the spatial correspondence between the observation data is established. A depth map is generated based on the depth information; A normal map is generated based on the surface orientation information, and the surface orientation information is mapped to the point cloud data based on the spatial correspondence, so that the three-dimensional points in the point cloud data are attached with point cloud normal attributes.

[0008] The reconstruction method for large-scale scene multi-source acquisition information and lightweight structured coupling gain, wherein generating an initial parametric building model based on the point cloud data specifically includes: Identify building components in the point cloud data; Each building component is described using a parametric volume form to obtain its geometric information, which includes not only location and direction of extension, but also thickness, height, and one or more component categories. The structural relationships between components are determined based on the geometric information of all building components; An initial parametric building model is constructed based on geometric information and structural relationships between components.

[0009] The reconstruction method for large-scale multi-source information acquisition and lightweight structured coupling gain, wherein the step of using a relation-aware assembly network to correct the initial parametric building model specifically includes: For each building component in the initial parametric building model, a building component feature containing modification tags is generated based on the geometric information of the building component and the structural relationships between the components; Point cloud data is encoded to extract point cloud features; Multi-attention fusion is performed on the building component features and the point cloud features of all building components to obtain the fused features of each building component. The multi-attention fusion includes attention within building components, attention between building components, and cross attention between building components and point cloud data. The geometric information correction amount for each building component is determined based on the fusion characteristics of all building components, and the structural relationships between components are predicted. The initial parametric building model is corrected based on the geometric information correction amount and the predicted structural relationships between components.

[0010] The reconstruction method for large-scene multi-source acquisition information and lightweight structured coupling gain, wherein the step of correcting the initial parametric building model based on the geometric information correction amount and the predicted structural relationship between components specifically includes: The geometric information correction amount is used to correct the geometric information in the initial parametric building model, and the structural relationships between components in the initial parametric building model are updated by predicting the structural relationships between components. The initial parametric building model is then modified using a relation-aware assembly network until the predicted geometric information correction amount meets the preset requirements.

[0011] The reconstruction method for large-scale scene multi-source acquisition information and lightweight structured coupling gain, wherein the component-level modification strategy includes: When the inspection result indicates that a building component is missing, the component-level modification strategy is to construct a coarse component. When the comprehensive inspection results indicate the presence of redundant or erroneous building components, the component-level modification strategy is to delete the building components. When the comprehensive inspection results indicate that the geometric information of a building component is abnormal, the component-level modification strategy is to adjust the geometric information of the building component. When the comprehensive inspection result indicates that the component category or the structural relationship between components is incorrect, the component-level modification strategy is to correct the component category or the structural relationship between components.

[0012] The reconstruction method for large-scale scene multi-source acquisition information and lightweight structured coupling gain, wherein, after applying multi-source information coupling gain to the modified parametric building model according to the component-level modification strategy to obtain the building information model, the method further includes: The building information model is standardized according to preset specifications, which include one or more of industry building codes, design preferences, and specific project rules.

[0013] The second aspect of this application provides a reconstruction device for large-scene multi-source acquisition information and lightweight structured coupling gain, wherein the reconstruction device for large-scene multi-source acquisition information and lightweight structured coupling gain specifically includes: The multi-source acquisition information module is used to acquire multi-source acquisition information of the large-scale real building scene to be reconstructed, wherein the multi-source acquisition information includes point cloud data, multi-view images, depth maps and normal maps; The generation module is used to generate an initial parametric building model based on the point cloud data; A coarse correction module is used to correct the initial parametric building model using a relation-aware assembly network, wherein the relation-aware assembly network is configured with attention within building components, attention between building components, and cross attention between building components and point cloud data. The multi-source information coupling gain module is used to perform a comprehensive check on the corrected parametric building model based on the point cloud data and the multi-view image to obtain a comprehensive check result, determine the component-level modification strategy through the large model based on the multi-source acquisition information and the comprehensive check result, and perform multi-source information coupling gain on the corrected parametric building model based on the component-level modification strategy to obtain a building information model. The comprehensive check includes one or more of geometric consistency check, semantic check, and layout structure check.

[0014] A third aspect of this application provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the reconstruction method of large-scene multi-source acquisition information and lightweight structured coupling gain as described above.

[0015] A fourth aspect of this application provides a terminal device, which includes: a processor and a memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the reconstruction method of large-scene multi-source acquisition information and lightweight structured coupling gain as described above.

[0016] Beneficial effects: 1. This application changes the processing mode of scene observation data acquisition and model reconstruction as a relatively fixed one-way process. After reconstructing the parametric building model using multi-source acquisition information, the multi-source acquisition information is used as scene evidence to roughly correct the parametric building model and to enhance the coupling gain of multi-source information. This fully leverages the complementary advantages between multi-source acquisition information and improves the accuracy, integrity and structural consistency of the building information model.

[0017] 2. The lightweight parametric building model in this application represents each building component with the geometric information of the building components and the structural relationship between the components. This allows large-scale real-world building scenes to retain the geometric, semantic and structural information of the building components while maintaining low model complexity and high processing efficiency, providing a stable intermediate carrier for subsequent model analysis, editing and gain.

[0018] 3. After performing multi-source information coupling gain on the parametric building model, this application further standardizes the parametric building model with multi-source information coupling gain using preset specifications. This integrates building specifications, engineering rules, and actual modeling requirements into the processing of the reconstruction results, enabling the reconstructed building information model to gradually evolve from geometric scene recovery into a standardized, unified, and exchangeable building information model. This enhances the connection between the reconstructed building information model and existing building software and engineering application processes. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating the reconstruction method of large-scene multi-source acquisition information and lightweight structured coupling gain provided in the embodiments of this application.

[0021] Figure 2 A network framework diagram for constructing recognition and relationship-aware assembly.

[0022] Figure 3 This is a flowchart illustrating an example of a method for reconstructing large-scene multi-source acquisition information and lightweight structured coupling gain provided in an embodiment of this application.

[0023] Figure 4 This is a comparison diagram of the reconstruction effect of the partial building structure in this application and the plan view.

[0024] Figure 5 The results are parametric volume reconstructions of beam and column building components.

[0025] Figure 6 This is a comparison diagram of the initial parametric building model and the parametric building model corrected using a relation-aware assembly network.

[0026] Figure 7 A schematic diagram illustrating the standardized processing of instantiation results.

[0027] Figure 8This is a schematic diagram of the reconstruction device for large-scene multi-source acquisition information and lightweight structured coupling gain provided in the embodiments of this application.

[0028] Figure 9 A schematic block diagram of the terminal device provided in the embodiments of this application. Detailed Implementation

[0029] This application provides a method, apparatus, device, and medium for reconstructing large-scene multi-source acquired information with lightweight structured coupling gain. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application.

[0030] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0031] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0032] It should be understood that the sequence number and size of each step in this embodiment do not imply the order of execution. The execution order of each process is determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application embodiment.

[0033] Research has revealed that current lightweight reconstruction methods generally treat scene observation data acquisition and model reconstruction as a relatively fixed, one-way process. They utilize acquired scene observation data (such as point cloud data) to perform geometric reconstruction, transforming the architectural scene into a Building Information Model (BIM). While scene observation data can obtain spatial geometric information of real architectural scenes, it is susceptible to issues such as local occlusion, missing observations, and noise interference, especially in large-scale real-world architectural scenes. These problems lead to various defects in the reconstructed BIM, including geometric information deviations, incorrect component category identification, and disordered spatial relationships among components, ultimately affecting the accuracy and completeness of the final BIM model.

[0034] For example, while automatic reconstruction methods based on two-dimensional layouts can describe the spatial organization of large scenes with a relatively compact data structure, exhibiting good automation and lightweight characteristics, they only collect scene observation data for building reconstruction. Furthermore, their output mainly describes volumetric attributes such as two-dimensional spatial boundaries, wall thickness, vertical range, and door and window heights, without directly modeling the objects.

[0035] While structured scene reconstruction methods based on 3D planes can directly predict 3D structures such as walls and windows from scene observations, preserving more 3D geometric information compared to 2D layouts while maintaining a relatively compact representation, they only collect scene observation data for building reconstruction. Furthermore, the plane is essentially surface geometry. To create building components with actual thickness and independent instance identities, further plane matching, instance association, and determination of the hierarchical relationships between doors / windows and walls are required. In particular, when large scenes have occlusions, missing elements, or complex adjacent structures, the same component may correspond to multiple incomplete planes, introducing uncertainty into the matching and association processes and affecting subsequent volumetric modeling and component attribute expansion.

[0036] While geometric reconstruction methods for volumetric building components can obtain building information models with thickness and volume, they only collect scene observation data for building reconstruction and require setting various geometric thresholds, matching rules, and scene assumptions. They are also susceptible to the cumulative effects of intermediate step errors under conditions of complex structures, occlusion, noise, and incomplete scanning.

[0037] Based on this, this application provides a reconstruction method for large-scale scene multi-source acquisition information and lightweight structured coupling gain. The method includes acquiring multi-source acquisition information of a large-scale real-world building scene to be reconstructed; generating an initial parametric building model based on the point cloud data; correcting the initial parametric building model using a relation-aware assembly network; performing a comprehensive check on the corrected parametric building model based on the point cloud data and the multi-view images to obtain a comprehensive check result; determining a component-level modification strategy based on the multi-source acquisition information and the comprehensive check result through a large model; and applying multi-source information coupling gain to the corrected parametric building model based on the component-level modification strategy to obtain a building information model. This application changes the processing mode of scene observation data acquisition and model reconstruction as a relatively fixed unidirectional process. After reconstructing the parametric building model using multi-source acquisition information, the multi-source acquisition information is used as scene evidence to perform coarse correction and multi-source information coupling gain on the parametric building model, giving full play to the complementary advantages between multi-source acquisition information and improving the accuracy, completeness, and structural consistency of the building information model.

[0038] It should be noted that the execution entity of each embodiment of this application can be a computing service system with data processing, network communication, and program execution functions, such as an electronic system capable of realizing the above functions, a reconstruction system for large-scene multi-source information acquisition and lightweight structured coupling gain, etc. The following uses a reconstruction system for large-scene multi-source information acquisition and lightweight structured coupling gain (hereinafter referred to as "the system") as an example to describe the following embodiments.

[0039] The application content will be further explained below with reference to the accompanying drawings and the description of the embodiments.

[0040] This embodiment provides a method for reconstructing large-scale scene multi-source acquisition information and lightweight structured coupling gain, such as... Figure 1 , Figure 2 and Figure 3 As shown, the reconstruction method of large-scene multi-source acquisition information and lightweight structured coupling gain specifically includes: S10. Acquire multi-source information of the large-scale real building scene to be reconstructed.

[0041] Specifically, the large-scale real-world building scene to be reconstructed refers to a large-scale real-world building scene that requires conversion of its building information model, such as existing public buildings or large commercial buildings with a floor area exceeding a certain square meter that need to be digitally modeled. The system can acquire multi-source acquisition information corresponding to the large-scale real-world building scene through various acquisition methods, such as UAV aerial photography, ground laser scanning, handheld mobile scanners, and multi-view camera shooting. This multi-source acquisition information can include point cloud data, multi-view images, depth maps, and normal maps. For example, acquisition personnel can continuously walk inside the large-scale real-world building scene with handheld or backpack 3D scanning equipment to sample corridors, rooms, and other areas to be reconstructed, acquiring laser point clouds, multi-view images, and multi-source scene observation information such as depth, normal, and acquisition pose, and forming multi-source acquisition information based on this multi-source scene observation information.

[0042] In one embodiment, acquiring multi-source information of the large-scale real-world building scene to be reconstructed specifically includes: Collect multi-source scene observation information of large-scale real building scenes to be reconstructed; Based on the time information and the acquisition posture, the observation data in the multi-source scene observation information are transformed into the scene coordinate system, and the spatial correspondence between the observation data is established. A depth map is generated based on the depth information; A normal map is generated based on the surface orientation information, and the surface orientation information is mapped to the point cloud data based on the spatial correspondence, so that the three-dimensional points in the point cloud data are attached with point cloud normal attributes.

[0043] Specifically, the multi-source scene observation information is obtained by scanning a large-scale real-world building scene using a 3D scanning device. This multi-source scene observation information includes point cloud data, multi-view images, depth information, surface orientation information, and acquisition posture. During continuous scanning, the 3D scanning device uses lidar, inertial measurement unit, and imaging sensors for synchronous positioning and mapping to obtain point cloud data, multi-view images, depth information, and surface orientation information. Then, based on the geometric matching between continuous observations and the device's motion information, the attitude information (including position and orientation) of the scanning device over time is estimated. Loop closure detection or global optimization are then used to reduce accumulated positioning errors and obtain the acquisition posture at each moment.

[0044] In addition, for large scenes that require multiple independent scans, common control points or markers that can be observed together are set or retained between different scan areas, and adjacent scan ranges have a certain degree of spatial overlap. After obtaining the scan data of independent scans, it is convenient to register different scan data based on the common control points or markers to obtain multi-source scene observation information.

[0045] Next, as Figure 3 As shown, after acquiring multi-source scene observation information, different types of observation data can be paired. Specifically, using the time synchronization information and acquisition pose provided by the scanning device, combined with the pre-calibrated relative positions and attitude relationships between different sensors, point cloud data, multi-view images, depth information, and surface orientation information can be transformed into a unified scene coordinate system, and spatial correspondences between different types of observation data can be established. For multi-source scene observation information obtained from multiple independent scans, the initial spatial transformation relationship between each scan dataset is first solved using common control points or markers observed in different scan areas. Then, automatic fine registration is performed using overlapping point clouds between adjacent scans to correct the initial spatial transformation relationship. Finally, the corrected initial spatial transformation relationship is used to unify the multi-source scene observation information obtained from multiple independent scans into the same scene coordinate system.

[0046] Then, the paired multi-source scene observation information is preprocessed to form multi-source acquisition information. This preprocessing includes outlier removal and voxel downsampling of the registered point cloud data; converting the registered multi-view images into image formats suitable for visual understanding while preserving the correspondence between the multi-view images and the 3D scene regions, such as perspective views and panoramic views; organizing depth information into depth maps corresponding to the image pixels of the multi-view images; and organizing surface orientation information into normal maps corresponding to the multi-view images and point cloud normal attributes that can be attached to each 3D point of the point cloud, thus obtaining multi-source acquisition information for the large-scale realistic building scene to be reconstructed.

[0047] S20. Generate a parametric building model based on the point cloud data, wherein the parametric building model includes the geometric information of building components and the structural relationships between components.

[0048] Specifically, the parametric building model is a structured building representation extracted from point cloud data. It retains only the geometric information of building components and the structural relationships between building components in the large-scale real building scene to be reconstructed, while removing a large amount of redundant discrete point data from the point cloud data, thereby achieving a lightweight representation of the model.

[0049] In one embodiment, generating a parametric building model based on the point cloud data specifically includes: S21. Identify building components in the point cloud data; S22. Describe each building component using parametric volume form to obtain the geometric information of each building component; S23. Determine the structural relationships between components based on the geometric information of all building components; S24. Construct an initial parametric building model based on geometric information and structural relationships between components.

[0050] In step S21, the building components are architectural units with independent functions and physical attributes, such as walls, doors, windows, beams, and columns, from the large-scale real-world architectural scene to be reconstructed. For example, Figure 2 As shown, the process of identifying building components can be carried out in the following steps: First, the point cloud data is input into the point cloud feature encoder (such as the point cloud feature encoder in Sonata) to extract point cloud features that can characterize the building's spatial structure and local geometric shape; then, the point cloud features are converted into feature representations suitable for component prediction through the feature mapping module (such as the feature mapping module in Sonata); finally, the feature representations are input into the autoregressive component proposal network (such as the Qwen2.5 0.5B model), which predicts building components such as walls, doors, windows, beams, and columns one by one until all component proposals for the large-scale real building scene to be reconstructed are completed, and finally, all building components in the point cloud data are identified.

[0051] In step S22 above, after identifying the building components in the point cloud data, each building component is described using a parametric volumetric form, i.e., represented by a cube. Specifically, the central axis position and extension direction of the building component are used, and information such as thickness, height, and component category are further recorded. The geometric information of the building component is determined using its position (i.e., the position information of the two endpoints of the central axis), extension direction, thickness, height, and component category, and this geometric information is used to represent the building component to obtain a parametric representation of the building component. The geometric information includes position and extension direction, and may also include one or more of thickness, height, and component category. In a typical implementation, the geometric information includes position, thickness, height, and component category.

[0052] The embodiments of this application directly reconstruct using parametric volumetric components with location, extension direction, thickness, height, and component type, which can directly correspond to actual building components such as walls, doors, windows, beams, and columns, reducing additional geometric inference and manual processing in the process of converting planar structures into building components.

[0053] In step S23 above, the structural relationships between components refer to spatial connections, subordination, and adjacency relationships between different building components. For example, a rectangular door component may belong to a wall component, or there may be a perpendicular intersecting connection between two walls. These structural relationships are determined based on geometric information and point cloud data. The determination process can involve using a learnable vector as the structural relationship between each building component, marking each component with a modification flag indicating that modification is needed, and then using a relationship-aware assembly network to predict the structural relationships between components to initialize the structural relationships for each building component. Of course, in practical applications, other methods can also be used to determine the structural relationships between components, such as based on preset building structure design rules.

[0054] In step S24 above, after obtaining geometric information and structural relationships between components, the geometric information can be used as a parameter representation of the building framework, and spatial assembly can be performed according to the structural relationships between components to obtain a complete initial parametric building model.

[0055] This application uses parametric volumetric components such as walls, doors, windows, beams, and columns as building components, and explicitly describes the structural relationships between building components. This enables large-scale discrete point cloud data to be converted into a limited number of lightweight models with clear parameters and architectural semantics. While reducing model complexity and subsequent processing costs, it retains the geometric, volumetric, and relational information required by building components, facilitating subsequent local editing and modification.

[0056] S30. The initial parametric building model is corrected using a relation-aware assembly network.

[0057] Specifically, the relation-aware assembly network is used to coarsely refine the initial parametric building model to improve the accuracy of geometric information and structural relationships between components. This relation-aware assembly network is configured with internal attention within building components, inter-building component attention, and cross-attention between building components and point cloud data. Internal attention within building components is used to capture the correlation between different parameter dimensions of the same component; inter-building component attention is used to capture the spatial positional correlation between different building components; and cross-attention between building components and point cloud data is used to perform cross-attention between building components and point cloud data to align and fuse the real spatial information carried by the building components and point cloud data.

[0058] In one embodiment, such as Figure 2As shown, the relationship-aware assembly network can include a multi-attention fusion module and a prediction module. The multi-attention fusion module can include self-attention units within building components, self-attention units between building components, and cross-attention units between building components and point clouds, connected sequentially. The prediction module includes a geometric prediction head and a relationship prediction head. Specifically, the self-attention units within building components perform self-attention within the building component to capture the association between different parameter dimensions of the same component; the self-attention units between building components perform self-attention between building components to capture the spatial positional association between different building components; and the cross-attention units between building components and point cloud data perform cross-attention between building components and point cloud data to align and fuse the real spatial information carried by the building components and point cloud data, obtaining fused features containing spatial context information for each building component. The geometric prediction head is used to predict the geometric information correction amount of the building components based on the fused features, and the relationship prediction head is used to predict the structural relationships between building components based on the fused features. The structural relationships between components can include the adjacency relationship between adjacent walls and the subordinate relationship between doors / windows and the host wall, etc.

[0059] In one embodiment, the modification of the initial parametric building model using a relation-aware assembly network specifically includes: S31. For each building component in the initial parametric building model, generate building component features containing modification markers based on the geometric information of the building component and the structural relationship between the components; S32. Perform point cloud encoding on the point cloud data to extract point cloud features; S33. Perform multi-attention fusion on the building component features and the point cloud features of all building components to obtain the fused features of each building component; S34. Determine the geometric information correction amount for each building component and predict the structural relationship between components based on the fusion characteristics of all building components. S35. Correct the initial parametric building model based on the geometric information correction amount and the predicted structural relationship between components.

[0060] In step S31, the building component features may include parameter tokens, relationship tokens, and modification markers. The parameter tokens are determined based on the geometric information of the building component and carry parameters such as the building component's central axis, thickness, height, and component category. The relationship tokens are used to record the structural relationships between the building component and other components. The modification markers indicate whether the component needs gain adjustment; unmarked building components remain unchanged. Specifically, such as... Figure 2 As shown, building components are represented by cubes. The characteristics of a building component include the position information of the two endpoints of the central axis, thickness, height, component category, structural relationships between components, and modification markers. It can be represented as follows: , , in, Indicates the first The architectural features of each building component. Indicates the first The cube representation of each building component Indicates modification flag, and This indicates the position information of the two endpoints of the central axis. Indicates thickness, Indicates altitude, Indicates the first The component categories of each building component. Indicates the first The structural relationships between building components.

[0061] In step S32, the point cloud features are obtained by encoding the point cloud data. For example, the point cloud data is encoded using a point cloud feature encoder to extract the point cloud features corresponding to the point cloud data.

[0062] In step S33, multi-attention fusion refers to sequentially performing attention within building components, attention between building components, and cross-attention between building components and point cloud data. Specifically, firstly, all building component features and point cloud features are input into the multi-attention fusion module. Intra-building component self-attention is performed through the building component self-attention unit, modeling the relationships within a single building component and clarifying the dependencies of its own parameter dimensions. Then, inter-building component self-attention is performed through the building component self-attention unit, calculating attention weights within all building components to capture the spatial relationships between different building components. Finally, cross-attention calculation between building components and point cloud data is performed through the building component and point cloud cross-attention unit. This constructs query vectors based on each building component and key and value vectors based on point cloud features for attention calculation, injecting the real spatial information carried by the point cloud data into the building components and correcting deviations caused by relying solely on parameter information. Furthermore, in practical applications, the relation-aware assembly network can adopt a stacked form of multiple attention fusion modules. That is, after completing a multi-attention fusion through a layer of attention fusion modules, the next layer of attention fusion modules in the cascade of multi-attention fusion can be used for iterative fusion. After multiple layers of iterative fusion, the fusion features of each building component are output.

[0063] In step S34, after obtaining the fusion features, the fusion features are input into the geometric prediction head and the relationship prediction head respectively. The geometric prediction head predicts the geometric information correction amount of each building component, and the relationship prediction head generates the inter-component structural relationship between each building component and other building components.

[0064] In step S35, after obtaining the geometric information correction amount and the predicted structural relationship between components, the geometric information correction amount can be used to correct the geometric information of the building components in the initial parametric building model, and the predicted structural relationship between components can be used to replace the structural relationship between the building components in the initial parametric building model to obtain the corrected parametric building model.

[0065] Furthermore, in practical applications, in order to further improve the accuracy of the corrected parametric building model, it can be determined whether the corrected parametric building model needs to be roughly corrected again based on the amount of geometric information correction. If it needs to be roughly corrected again, the current corrected parametric building model can be used as the initial parametric building model, and the relation-aware assembly network can be used to correct the initial parametric building model again, so as to achieve iterative correction of geometric information and structural relationships between components.

[0066] Based on this, in one embodiment, correcting the initial parametric building model based on the geometric information correction amount and the predicted structural relationship between components specifically includes: The geometric information correction amount is used to correct the geometric information in the initial parametric building model, and the structural relationships between components in the initial parametric building model are updated by predicting the structural relationships between components. The initial parametric building model is then modified using a relation-aware assembly network until the predicted geometric information correction amount meets the preset requirements.

[0067] Specifically, the geometric information correction amount is the adjustment offset of the geometric information of a single building component. Each building component that needs to be corrected corresponds to a set of geometric information correction amounts. The geometric information correction amount includes the adjustment range of each physical information item in the geometric information. For example, if the geometric information includes position, extension direction, thickness, height and component type, then the geometric information correction amount includes the adjustment range of position, extension direction, thickness and height.

[0068] Furthermore, when correcting the geometric information in the initial parametric building model using the geometric information correction amount, the building components that need correction can be determined first based on the geometric information correction amount. For example, each correction item in the geometric information correction amount can be compared with a preset threshold corresponding to each correction item. If there is a correction item greater than the preset threshold, the building component is determined to need correction. Alternatively, the mean or weighted value of each correction item in the geometric information correction amount can be obtained, and the mean or weighted value can be compared with a comprehensive threshold. If it is greater than the comprehensive threshold, the building component is determined to need correction, etc. In addition, after correcting the geometric information in the initial parametric building model using the geometric information correction amount, the correction flag corresponding to the building component is configured to indicate that it needs to be modified. For example, 1 indicates that it needs to be modified, and 0 indicates that it does not need to be modified, etc. Then, the geometric information correction amount of the building component that needs to be corrected is superimposed on the geometric information of the building component to obtain the corrected geometric information, and the structural relationship between components in the initial parametric building model is updated using the predicted structural relationship between components to obtain the corrected initial parametric building model.

[0069] Next, based on the geometric information correction amount, it is determined whether the modified initial parametric building model needs further modification. For example, it checks whether the number of components requiring modification in the current modification exceeds a preset threshold, or whether there is at least one geometric information correction amount, and whether the adjustment range of at least one physical information item in that geometric information correction amount meets the preset requirements corresponding to that physical information item. If the modified initial parametric building model needs further modification, the relation-aware assembly network is used again to modify the initial parametric building model until the predicted geometric information correction amount meets the preset requirements. Meeting the preset requirements for the predicted geometric information correction amount can mean that all building components do not need modification, or that the number of iterations reaches a preset number.

[0070] This application embodiment gradually corrects the position, orientation, length, thickness, and height of components through an iterative correction process, thereby gradually eliminating the errors caused by initial identification and parameter estimation. This allows the geometric information and structural relationships between components to continuously approach the actual situation of the large-scale real building scene to be reconstructed, resulting in a corrected parametric building model with more accurate geometry and explicit structural relationships.

[0071] It is understood that after obtaining the initial parametric building model, this application uses a relation-aware assembly network to learn the information interaction between building components, between building components, and between building components and point cloud data. This enables building components to simultaneously perceive their own geometric state, the structural relationships between adjacent components, and the geometric evidence carried by the point cloud data. Then, using paired relationships such as wall adjacency and the subordination of doors and windows to the host wall as relation supervision, the position, extension direction, thickness, and height of building components are gradually updated through a continuous correction process from coarse geometric information to accurate geometric information. At the same time, the structural relationships between components are predicted and corrected, enabling building components to be further precisely aligned with the large-scale real building scene to be reconstructed and to form a structurally consistent parametric building model.

[0072] S40. Based on the point cloud data and the multi-view images, a comprehensive inspection is performed on the corrected parametric building model to obtain the comprehensive inspection results.

[0073] Specifically, after obtaining the corrected parametric building model, a comprehensive check is performed on the model by combining scene evidence from different sources (i.e., information collected from different sources in multi-source acquisition information) to automatically determine whether there are missing components, redundant components, category errors, or obvious geometric deviations in the corrected parametric building model. The comprehensive check includes a geometric consistency check based on the first distance from point cloud data to the surface of the corrected parametric building model and the second distance from the local point cloud data corresponding to the building component to the surface of the corresponding parametric building component; an image semantic check based on the consistency between the building components in the corrected parametric building model and the building components appearing in the corresponding target image regions in multi-view images; and a layout structure check based on the structural relationships between components.

[0074] The geometric consistency check includes component missingness checks based on a first distance and / or component error checks based on a second distance. The component missingness check based on the first distance can be performed as follows: First, calculate the nearest distance from each 3D point in the point cloud data to the surface of all building components in the corrected parametric building model to obtain the first distance; then, spatially cluster all abnormal 3D points whose first distance exceeds a first preset distance threshold. When abnormal 3D points form a continuous cluster region, and the number and spatial range of abnormal 3D points in this continuous cluster region both exceed their respective preset thresholds, a large-area abnormal point cloud region is considered to exist; finally, combining the corresponding multi-view images, multimodal intelligent analysis is used to determine whether the large-area abnormal point cloud region belongs to building structures such as walls, doors, windows, beams, and columns. If the large-area abnormal point cloud region belongs to building structures such as walls, doors, windows, beams, and columns, then the corrected parametric building model is considered to have missing components.

[0075] Furthermore, the inspection process based on the second distance for component deviation can be as follows: For each building component, a local point cloud is extracted by extending a preset distance outside the spatial range of the building component, and the nearest distance from each local 3D point in the local point cloud to the surface of the building component in the corrected parametric building model is calculated to obtain the second distance corresponding to each local 3D point; then, the local 3D points that are too close to each other on the component surface are deduplicated. If the second distance corresponding to more than a preset number of local 3D points in the deduplicated local point cloud exceeds the second preset distance threshold, and interference such as furniture is excluded by multimodal analysis, then the building component is considered to have obvious errors. In addition, when more than a preset number of abnormal local 3D points are within the threshold range, the position, thickness or height deviation is judged based on the average deviation of the point cloud data in the thickness and height directions. If a deviation exists, the position, thickness or height deviation is recorded simultaneously.

[0076] Image semantic checking is based on the consistency between architectural components in the corrected parametric architectural model and their corresponding target image regions in multi-view images. Specifically, for architectural components in the corrected parametric architectural model, the local 3D scene region corresponding to the architectural component and its surrounding area is taken as the target region. Then, based on the spatial correspondence established between multi-view images, acquisition posture, and 3D scene, candidate images that can observe the target region are selected from the multi-view images. From the selected candidate images, several target view images with relatively complete display of the target component, less occlusion, and different viewing directions are selected. Then, multimodal intelligent analysis is used to determine whether the type, quantity, and relative position of the architectural components actually appearing in the several target view images are consistent with the type, quantity, and relative position of the architectural components in the corrected parametric architectural model. If they are consistent, the image semantic check passes; if they are inconsistent, the image semantic check fails.

[0077] The layout structure check is based on the structural relationships between components. Specifically, using the structural relationships between components in the modified parametric building model, structural anomalies such as wall breaks, abnormal overlaps, missing host walls for doors and windows, and incorrect component relationships are checked. If a structural anomaly is detected, the layout structure check is deemed to have failed, and the abnormal building component is identified. If no structural anomalies are detected, the layout structure check is deemed to have passed.

[0078] S50. Based on the multi-source collected information and the comprehensive inspection results, a component-level modification strategy is determined through the large model, and the modified parametric building model is subjected to multi-source information coupling gain based on the component-level modification strategy to obtain a building information model.

[0079] Specifically, the comprehensive inspection results can include missing building components, redundant or erroneous building components, abnormal geometric information of building components, and errors in component categories or structural relationships between components. The component-level modification strategy can be proposed by the large language model agent using point cloud data, multi-view images, depth maps, and normal maps as scene evidence. This can include: when the inspection result indicates a missing building component, the component-level modification strategy is to construct a coarse component; when the comprehensive inspection result indicates a redundant or erroneous building component, the component-level modification strategy is to delete the building component; when the comprehensive inspection result indicates abnormal geometric information of the building component, the component-level modification strategy is to adjust the geometric information of the building component (e.g., perform movement, scaling, or size adjustment on the building component); when the comprehensive inspection result indicates errors in component categories or structural relationships between components, the component-level modification strategy is to correct the component categories or structural relationships between components.

[0080] Furthermore, in practical applications, after applying multi-source information coupling gain to the modified parametric building model according to the component-level modification strategy, the parametric building model after multi-source information coupling gain can be used as the initial parametric building model. The relation-aware assembly network is then used to modify the initial parametric building model, forming a process of coarse correction using the relation-aware assembly network and iterative correction using multi-source information coupling gain. This allows the modified components to re-align geometrically with the real scene through multi-layer information interaction within the components, between components, and between the components and the point cloud, and to synchronously restore or update their structural relationships. Unmodified components remain relatively stable, improving the accuracy of the obtained building information model of a large-scale real building scene.

[0081] In one embodiment, after performing multi-source information coupling gain on the modified parametric building model according to the component-level modification strategy to obtain a building information model, the method further includes: The building information model is standardized according to preset specifications.

[0082] Specifically, the preset specifications include one or more of the following: industry building codes, design preferences, and specific project rules. Industry building codes are current architectural design and construction standards that clearly define general rules such as the dimensional modules of building components, component spacing, connection methods, and door and window opening dimensions. Design preferences are personalized modeling requirements proposed by the client or the modeling project, such as simplified treatment of component fillets or rules for retaining non-load-bearing partitions. Specific project rules correspond to the specific requirements of this modeling project, such as the requirement to retain all original structural components of the existing building for the XX renovation project.

[0083] For example, when standardizing the building information model according to preset specifications, the adjacent and intersection relationships between walls can be determined as different structural forms such as corner connections, T-shaped connections, and cross connections according to industry building codes. The connection methods of components can be adjusted according to specific project rules, and the orientation of doors and windows can be adjusted according to design preferences.

[0084] After obtaining the building information model, the embodiments of this application standardize the building information model, which enhances the connection between the automatic reconstruction results and actual building engineering applications. It can also reduce the manual conversion and repetitive modeling work between collecting data from real scenes and editing building information models, and improve the usability of reconstruction results in practical applications such as digital archiving of existing buildings, acquisition of current status models, renovation and expansion and engineering verification.

[0085] Furthermore, to verify the effectiveness of the reconstruction method using large-scene multi-source acquisition information and lightweight structured coupling gain provided in this application embodiment, this application verifies the effectiveness of constructing an initial parametric building model using geometric information and structural relationships between components. As shown in Table 1, this application can directly recover building components such as walls, doors, and windows with clear spatial range and volumetric attributes from scene observations. The component matching indices for exterior walls and interior walls reach 93.78% and 92.57%, respectively, with an average thickness error of 2.80 cm for interior walls, and 79.50% of the matched interior wall thickness errors are controlled within 3 cm. The component matching indices for doors and windows reach 91.14% and 88.97%, respectively, with average height errors of 2.86 cm and 9.23 cm, respectively. In particular, regarding the recovery of interior wall thickness, the accuracy within 3 cm is significantly higher than existing two-dimensional layout and planar representation methods, indicating that the central axis parametric volumetric components can effectively recover key architectural attributes such as wall thickness, opening height, and component spatial range.

[0086] Table 1. Comparison of Component-Level Volume Reconstruction Results for Different Architectural Scene Representation Methods

[0087] Notes: Match-F1 is the F1 score for matching, CL-Offset is the centerline offset error, T-MAE is the mean absolute error of topology, T-Acc@3 indicates the top 3 in topology accuracy, T-Acc@5 indicates the top 5 in topology accuracy, H-MAE is the mean absolute error of geometric height, marked with " "This application represents the ablation experimental version."

[0088] Meanwhile, Table 2 shows that this application maintains high quality in both 2D layout and 3D planar reconstruction while directly expressing the volumetric properties of components. Specifically, the F1 score for corner points reaches 93.1%, and the F1 score for angles reaches 91.1%. In the 3D planar evaluation, the F1 score for window components reaches 91.4% under strict intersection-union ratio conditions, higher than SpatialLM's 87.5%. This indicates that using parametric volumetric components does not sacrifice layout or surface geometry recovery capabilities, but rather further obtains volumetric information at the building component level while maintaining the quality of scene structure reconstruction.

[0089] Table 2. Comparison of 2D layout and 3D planar reconstruction results for different architectural scene representation methods.

[0090] Notes: Floor Plan represents a floor plan, Plane represents a plane, Cuboid represents a cube, Room represents a room, Corner represents a corner point, Angle represents an angle, 2D Layout represents a two-dimensional layout, 3D Plan represents a three-dimensional plane, Prec. (Precision) represents precision, Rec. (Recall) represents recall, F1 represents F1 score, IoU2D@0.25 represents the 2D intersection-union ratio with a threshold of 0.25, IoU2D@02 represents the 2D intersection-union ratio with a threshold of 0.5, Wall represents a wall, Door represents a door, and Window represents a window.

[0091] like Figure 4 As shown, compared with the planar representation method, this application can more completely maintain the wall thickness, component extension range and the consistent relationship between doors and windows and walls in the same local building structure, further demonstrating the advantages of parametric volume representation in the restoration of complex building structures. Figure 5 The reconstruction results of beam and column components are further demonstrated, showing that the parametric representation is not only applicable to major building components such as walls, doors, and windows, but can also be extended to other components with clear volume and architectural semantics such as beams and columns, providing further verification for the formation of a unified lightweight parametric building component representation.

[0092] Next, to verify the ability of the relation-aware continuous refinement mechanism to improve the initial reconstruction results, joint refinement of parametric volumetric components was first performed within the entire scene. As shown in Table 3, after relation-aware refinement, the component matching indices for exterior walls, interior walls, doors, and windows reached 94.42%, 92.73%, 92.20%, and 89.76%, respectively. Among them, the accuracy of the interior wall with a thickness of 3 cm further improved from 79.50% in the initial component proposal stage to 81.27%. In contrast, when using planar representation under the same refinement framework, the component matching indices for exterior walls, interior walls, doors, and windows were 87.94%, 87.31%, 88.02%, and 76.44%, respectively, with the accuracy of the interior wall with a thickness of 3 cm being only 20.45%. The results show that using parametric volumetric components with clear component range and thickness information as intermediate states, and combining component relationships with point cloud observations for continuous refinement, can further improve the geometric accuracy and component consistency of the initial model.

[0093] Table 3. Full-scene relationship-aware refinement results under different intermediate representations

[0094] Note: Method represents the method, AssemblyNet (Plane) represents a parametric set of building components represented by a plane, AssemblyNet (Cuboid) represents a parametric set of building components represented by a cube, Exterior Wall represents the exterior wall, and Interior Wall represents the interior wall.

[0095] Furthermore, Figure 6 Qualitative verification was conducted on common coarse-grained component deviations encountered during actual iterations. Figure 6 It can be seen that, given only a rough component state, the refinement network can further correct the position and scale deviations of the components, restore the alignment relationships between local components, and align and eliminate redundant components. This result indicates that in the multi-source information coupling gain process, the large language model agent does not need to directly complete precise geometric fitting, but can primarily be responsible for identifying model problems and proposing coarse-grained modifications such as adding, deleting, moving, scaling, and adjusting relationships. The relationship-aware component refinement network then completes the subsequent fine geometric alignment and structural relationship restoration, thus supporting the iterative gain process of "multi-source inspection, coarse-grained modification, fine-grained correction, and re-inspection."

[0096] Finally, it was verified that the building information model can directly support building code adaptation and model standardization, enabling further conversion of AI reconstruction results into actual building models. Figure 7As shown, after reconstructing the geometric and structural relationships of the components, this application can directly instantiate the building model using the wall adjacency relationships and the subordinate relationships between doors / windows and the host wall. For the same set of parametric components and their relationships, different wall connection forms such as corner connections, cross connections, and overall wall connections can be formed according to different building connection rules, without the need to reconstruct the scene geometry; for door / window components, the corresponding wall can be determined and the opening association can be completed based on their host relationship.

[0097] The results show that the lightweight parametric building model obtained in this application not only contains geometric information, but also retains explicit structural relationships that can be used for subsequent building processing. This allows the reconstruction results to adjust the component connection methods according to industry building codes, corporate design preferences and specific project rules while keeping the basic component representation unchanged, and further convert them into standardized building entities, providing a basis for interfacing with building software such as Revit.

[0098] Based on the above-mentioned reconstruction method using multi-source acquisition information and lightweight structured coupling gain for large-scale scenes, this embodiment provides a reconstruction device using multi-source acquisition information and lightweight structured coupling gain for large-scale scenes, such as... Figure 8 As shown, the reconstruction device for large-scene multi-source acquisition information and lightweight structured coupling gain specifically includes: The multi-source acquisition information module 100 is used to acquire multi-source acquisition information of the large-scale real building scene to be reconstructed, wherein the multi-source acquisition information includes point cloud data, multi-view images, depth maps and normal maps; Generation module 200 is used to generate an initial parametric building model based on the point cloud data; The coarse correction module 300 is used to correct the initial parametric building model using a relation-aware assembly network, wherein the relation-aware assembly network is configured with attention within building components, attention between building components, and cross attention between building components and point cloud data. The multi-source information coupling gain module 400 is used to perform a comprehensive check on the corrected parametric building model based on the point cloud data and the multi-view image to obtain a comprehensive check result, determine a component-level modification strategy through a large model based on the multi-source acquisition information and the comprehensive check result, and perform multi-source information coupling gain on the corrected parametric building model based on the component-level modification strategy to obtain a building information model. The comprehensive check includes one or more of geometric consistency check, semantic check, and layout structure check.

[0099] Based on the above-described reconstruction method of large-scene multi-source acquisition information and lightweight structured coupling gain, this embodiment provides a computer-readable storage medium storing one or more programs, which can be executed by one or more processors to implement the steps in the reconstruction method of large-scene multi-source acquisition information and lightweight structured coupling gain as described in the above embodiment.

[0100] Based on the reconstruction method using multi-source information acquisition and lightweight structured coupling gain in large-scale scenarios, this application also provides a terminal device, such as... Figure 9 As shown, it includes at least one processor 20; a display screen 21; and a memory 22, and may also include a communications interface 23 and a bus 24. The processor 20, display screen 21, memory 22, and communications interface 23 can communicate with each other via the bus 24. The display screen 21 is configured to display a preset user guide interface in the initial setup mode. The communications interface 23 can transmit information. The processor 20 can invoke logical instructions in the memory 22 to execute the methods described in the above embodiments.

[0101] Furthermore, the logical instructions in the aforementioned memory 22 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0102] The memory 22, as a computer-readable storage medium, can be configured to store software programs, computer-executable programs, such as program instructions or modules corresponding to the methods in the embodiments of this disclosure. The processor 20 executes functional applications and data processing by running the software programs, instructions, or modules stored in the memory 22, thereby implementing the methods in the above embodiments.

[0103] The memory 22 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 22 may include high-speed random access memory (RAM) and non-volatile memory. Examples include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, as well as transient storage media.

[0104] Furthermore, the specific process of loading and executing multiple instruction processors in the aforementioned storage medium and terminal device has been described in detail in the above method, and will not be repeated here.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A reconstruction method for large-scale scene multi-source information acquisition and lightweight structured coupling gain, characterized in that, The method for reconstructing large-scale scene multi-source acquisition information and lightweight structured coupling gain specifically includes: Acquire multi-source acquisition information of a large-scale real building scene to be reconstructed, wherein the multi-source acquisition information includes point cloud data, multi-view images, depth maps, and normal maps; An initial parametric building model is generated based on the point cloud data, wherein the initial parametric building model includes the geometric information of building components and the structural relationships between components; The initial parametric building model is corrected using a relation-aware assembly network, wherein the relation-aware assembly network is configured with attention within building components, attention between building components, and cross attention between building components and point cloud data; The corrected parametric building model is comprehensively checked based on the point cloud data and the multi-view images to obtain a comprehensive check result. The comprehensive check includes one or more of geometric consistency check, semantic check and layout structure check. Based on the multi-source collected information and the comprehensive inspection results, a component-level modification strategy is determined through a large model. Then, based on the component-level modification strategy, multi-source information coupling gain is applied to the modified parametric building model to obtain a building information model.

2. The reconstruction method for large-scene multi-source acquisition information and lightweight structured coupling gain according to claim 1, characterized in that, The acquisition of multi-source information of the large-scale real-world architectural scene to be reconstructed specifically includes: Collect multi-source scene observation information of a large-scale real building scene to be reconstructed, wherein the multi-source scene observation information includes point cloud data, multi-view images, depth information, surface orientation information, and acquisition posture; Based on the time information and the acquisition posture, the observation data in the multi-source scene observation information are transformed into the scene coordinate system, and the spatial correspondence between the observation data is established. A depth map is generated based on the depth information; A normal map is generated based on the surface orientation information, and the surface orientation information is mapped to the point cloud data based on the spatial correspondence, so that the three-dimensional points in the point cloud data are attached with point cloud normal attributes.

3. The reconstruction method for large-scene multi-source acquisition information and lightweight structured coupling gain according to claim 1, characterized in that, The generation of the initial parametric building model based on the point cloud data specifically includes: Identify building components in the point cloud data; Each building component is described using a parametric volume form to obtain its geometric information, which includes not only location and direction of extension, but also thickness, height, and one or more component categories. The structural relationships between components are determined based on the geometric information of all building components; An initial parametric building model is constructed based on geometric information and structural relationships between components.

4. The reconstruction method for large-scene multi-source acquisition information and lightweight structured coupling gain according to claim 1, characterized in that, The specific steps of using a relation-aware assembly network to correct the initial parametric building model include: For each building component in the initial parametric building model, a building component feature containing modification tags is generated based on the geometric information of the building component and the structural relationships between the components; Point cloud data is encoded to extract point cloud features; Multi-attention fusion is performed on the building component features and the point cloud features of all building components to obtain the fused features of each building component. The multi-attention fusion includes attention within building components, attention between building components, and cross attention between building components and point cloud data. The geometric information correction amount for each building component is determined based on the fusion characteristics of all building components, and the structural relationships between components are predicted. The initial parametric building model is corrected based on the geometric information correction amount and the predicted structural relationships between components.

5. The reconstruction method for large-scene multi-source acquisition information and lightweight structured coupling gain according to claim 4, characterized in that, The step of correcting the initial parametric building model based on the geometric information correction amount and the predicted structural relationship between components specifically includes: The geometric information correction amount is used to correct the geometric information in the initial parametric building model, and the structural relationships between components in the initial parametric building model are updated by predicting the structural relationships between components. The initial parametric building model is then modified using a relation-aware assembly network until the predicted geometric information correction amount meets the preset requirements.

6. The reconstruction method for large-scene multi-source acquisition information and lightweight structured coupling gain according to claim 1, characterized in that, The component-level modification strategy includes: When the inspection result indicates that a building component is missing, the component-level modification strategy is to construct a coarse component. When the comprehensive inspection results indicate the presence of redundant or erroneous building components, the component-level modification strategy is to delete the building components. When the comprehensive inspection results indicate that the geometric information of a building component is abnormal, the component-level modification strategy is to adjust the geometric information of the building component. When the comprehensive inspection result indicates that the component category or the structural relationship between components is incorrect, the component-level modification strategy is to correct the component category or the structural relationship between components.

7. The reconstruction method for large-scene multi-source acquisition information and lightweight structured coupling gain according to claim 1, characterized in that, After performing multi-source information coupling gain on the modified parametric building model according to the component-level modification strategy to obtain the building information model, the method further includes: The building information model is standardized according to preset specifications, which include one or more of industry building codes, design preferences, and specific project rules.

8. A reconstruction device for large-scale multi-source information acquisition and lightweight structured coupling gain, characterized in that, The reconstruction device for large-scale multi-source information acquisition and lightweight structured coupling gain specifically includes: The multi-source acquisition information module is used to acquire multi-source acquisition information of the large-scale real building scene to be reconstructed, wherein the multi-source acquisition information includes point cloud data, multi-view images, depth maps and normal maps; The generation module is used to generate an initial parametric building model based on the point cloud data; A coarse correction module is used to correct the initial parametric building model using a relation-aware assembly network, wherein the relation-aware assembly network is configured with attention within building components, attention between building components, and cross attention between building components and point cloud data. The multi-source information coupling gain module is used to perform a comprehensive check on the corrected parametric building model based on the point cloud data and the multi-view image to obtain a comprehensive check result, determine the component-level modification strategy through the large model based on the multi-source acquisition information and the comprehensive check result, and perform multi-source information coupling gain on the corrected parametric building model based on the component-level modification strategy to obtain a building information model. The comprehensive check includes one or more of geometric consistency check, semantic check, and layout structure check.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which can be executed by one or more processors to implement the steps in the reconstruction method of large-scene multi-source acquisition information and lightweight structured coupling gain as described in any one of claims 1-7.

10. A terminal device, characterized in that, include: Processor and memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the reconstruction method of large-scene multi-source acquisition information and lightweight structured coupling gain as described in any one of claims 1-7.

Images