A method and system for building interior layout inference and BIM digital twin model generation based on external point cloud

Abstract

A method and system for inferring building interior layout and generating BIM digital twin models based on external point clouds are disclosed, relating to the fields of building information modeling and point cloud processing technology. Addressing the problem that it is difficult to reconstruct the internal layout from external building observation information alone, and even more difficult to further form a complete BIM model usable for application analysis, the method specifically includes: acquiring and preprocessing the building's external point cloud, attitude correction, and facade clustering; generating semantically enhanced point clouds through projection, rasterization, and semantic parsing, and extracting point cloud slices to obtain a regularized external plan layout; constructing a network input containing external structural tensors and potential grid priors, and inferring a complete plan containing walls, doors, windows, and rooms through a multi-branch generative adversarial network; generating a complete BIM digital twin model through single-layer 3D extrusion, floor information recovery, and multi-layer BIM instantiation. This invention enables automated processing from building external point clouds to internal layout reconstruction and multi-layer BIM model generation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of building information modeling and point cloud processing, and particularly relates to a building internal layout inference and BIM digital twin model generation method and system based on external point cloud. BACKGROUND

[0002] The existing building information model (BIM) construction usually relies on indoor scanning, manual modeling or indoor drawing data. For existing buildings, especially buildings lacking indoor data, the above methods have the problems of high acquisition cost, long implementation period, and obvious limitations of site and authority.

[0003] The existing Scan-to-BIM technology generally obtains indoor point cloud through indoor laser scanning, RGB-D acquisition or mobile mapping, and then performs geometric reconstruction and semantic modeling. Such methods have high requirements for acquisition equipment, operating conditions and human participation; in the presence of occlusion, noise, complex decoration or prohibited entry, indoor data is often difficult to obtain completely.

[0004] In contrast, building external point cloud can be easily obtained through multi-view photogrammetry, unmanned aerial vehicle surveying or laser scanning. However, external point cloud mainly reflects the building shell geometry and facade information, and usually cannot directly give the indoor wall, door, room partition and floor organization, so it is difficult to directly form a complete BIM model that can be used for simulation and operation.

[0005] Some existing solutions attempt to infer the internal layout according to the building contour, window distribution or empirical rules, or use generative models to complete the layout. However, the former has strong rule dependence and limited adaptability, and the latter often takes known floor plan as input, which is difficult to apply to scenarios with only building external observation information; at the same time, the existing evaluation method mainly focuses on geometric similarity, and cannot fully reflect the usability of the generated model in applications such as traffic and evacuation.

[0006] Therefore, there is still a need for a technical solution that can restore the internal layout of a building and generate a multi-layer BIM digital twin model based only on building external point cloud without indoor scanning. SUMMARY

[0007] In view of the problem in the prior art that it is difficult to restore the internal layout only by building external observation information and difficult to further form a complete BIM digital twin model that can be used for application analysis, the present application provides a building internal layout inference and BIM digital twin model generation method based on external point cloud, comprising:

[0008] S1. Obtain building external three-dimensional point cloud data and perform preprocessing;

[0009] S2. The pre-processed point cloud is subjected to pose correction, facade division and semantic analysis to generate an architectural facade point cloud with semantic labels, and based on this, a regularized architectural exterior plan layout is generated;

[0010] S3. A network input tensor is constructed, which at least includes exterior structural information extracted from the exterior plan layout, and latent axis network prior information generated based on architectural geometric features;

[0011] S4. The network input tensor is input into a trained interior layout inference model to infer a complete architectural plan layout containing walls, doors, windows and room areas;

[0012] S5. The complete architectural plan layout is subjected to three-dimensional geometric stretching and semantic instantiation, and combined with floor information recovered from the facade point cloud, a multi-story building BIM digital twin model is generated.

[0013] Further, in S2, the generation of the regularized architectural exterior plan layout specifically includes:

[0014] S21. The pose-corrected point cloud is projected onto a two-dimensional plane to extract the building contour, and the point cloud is divided into multiple facade point cloud clusters according to the contour;

[0015] S22. Each of the facade point cloud clusters is rasterized into an image, and wall, window and entrance door areas are identified, and semantic labels are mapped back to the three-dimensional point cloud;

[0016] S23. Extract the horizontal point cloud slice of the target floor and project it to generate a pixel-level semantic plan, and perform vectorization and geometric regularization processing on the pixel-level semantic plan to obtain the regularized architectural exterior plan layout.

[0017] Further, in S3, the construction of the network input tensor specifically includes:

[0018] S31. Extract the exterior walls, exterior windows and exterior doorways from the exterior plan layout to generate an exterior structure tensor;

[0019] S32. Generate candidate axis lines according to the architectural exterior contour corner points and the midpoints of the inter-window wall segments, and filter them according to the architectural interior area mask, and rasterize the filtered axis lines into a latent axis network tensor;

[0020] S33. Concatenate the exterior structure tensor and the latent axis network tensor in the channel dimension to form the network input tensor.

[0021] Further, S4 specifically includes:

[0022] S4.1. Construct a multi-branch generative adversarial network as the internal layout inference model. The network includes: a shared feature extraction backbone, a structural branch for outputting binary masks of components or non-components, a segmentation branch for outputting semantic categories of walls, doors, windows, and rooms, and an RGB visualization branch for outputting a visual image of the building plan.

[0023] S4.2. Input the network input tensor into the multi-branch generative adversarial network, and use the shared features to extract the multi-scale spatial features between the building outline, window distribution and potential grid lines;

[0024] S4.3. Through the structural branch, the semantic category segmentation branch, and the RGB visualization branch, the binary structural results, semantic segmentation results, and visualization reconstruction results of the internal building components are output simultaneously to form the complete building plan layout.

[0025] Furthermore, in S5, the generation of a multi-story building BIM digital twin model specifically includes:

[0026] S51. Perform component analysis and single-layer 3D extrusion on the complete building floor plan to generate single-layer BIM semantic objects and establish topological relationships between components;

[0027] S52. Analyze the vertical distribution pattern of windows and floor slabs in the point cloud of the building facade, and automatically restore the building floor height and total number of floors;

[0028] S53. Based on the recovered floor information, the single-story BIM model is copied and adjusted along the vertical direction to generate the multi-story building BIM digital twin model, and the vertical spaces such as staircases and elevator shafts are modeled in a continuous manner.

[0029] It also provides a system for inferring the interior layout of buildings and generating BIM digital twin models based on external point clouds, including:

[0030] The external point cloud acquisition module is used to acquire and preprocess the 3D point cloud data of the building exterior.

[0031] The external layout parsing module is used to perform attitude correction, facade division and semantic parsing on the preprocessed point cloud, generate a building facade point cloud with semantic tags, and generate a regularized building external plan layout based on this.

[0032] A data construction module is used to construct a network input tensor, wherein the tensor includes at least external structural information extracted from the external planar layout, and prior information of the potential grid generated based on the building geometric features;

[0033] The internal layout inference module is used to input the network input tensor into the trained internal layout inference model and infer and generate a complete building plan layout including interior walls, doors, windows and room areas.

[0034] The BIM digital twin modeling module is used to perform three-dimensional geometric stretching and semantic instantiation of the complete building floor plan, and combine the floor information recovered from the facade point cloud to generate a multi-story building BIM digital twin model.

[0035] A computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the building interior layout inference and BIM digital twin model generation method based on any of the preceding claims.

[0036] A computer-readable storage medium is also provided, which stores a computer program that, when executed by a processor, implements the building interior layout inference and BIM digital twin model generation method based on any of the preceding claims.

[0037] The beneficial effects of this invention are:

[0038] 1. Reduce reliance on data acquisition for building interior modeling:

[0039] This invention uses the external 3D point cloud of a building as the basic input, eliminating the need for interior scanning data of the building as a prerequisite. This helps to reduce the difficulty of data acquisition for modeling the interior of existing buildings and expands the applicable scenarios.

[0040] 2. Improve the automation level of internal layout restoration:

[0041] This invention automatically recovers the internal layout information of walls, doors, windows and rooms by performing attitude correction, facade semantic parsing, external plan generation and internal layout inference on the external point cloud of a building, reducing manual intervention.

[0042] 3. It helps improve the structural consistency of the layout restoration results:

[0043] By introducing latent grid priors, attention mechanisms, edge constraints, and adaptive symmetry constraints, this invention helps to enhance the consistency of generated layouts in terms of boundary continuity, structural regularity, and spatial organization.

[0044] 4. Supports integrated modeling of building interior and exterior structures:

[0045] This invention combines the facade information obtained from external point cloud analysis with the inferred internal layout to generate a multi-story building BIM digital twin model that includes the organizational relationships of walls, doors and windows, rooms and floors.

[0046] 5. Supports functional availability verification for digital twin applications:

[0047] After BIM instantiation is completed, this invention can further perform fire simulation, evacuation simulation or traffic analysis based on spatial connectivity to verify the functional consistency of the generated digital twin model. Attached Figure Description

[0048] Figure 1 This is a flowchart of the overall method of the present invention;

[0049] Figure 2 It is a flowchart for generating a standardized building exterior floor plan;

[0050] Figure 3 This is a comparative schematic diagram of point cloud coordinate correction before and after the present invention;

[0051] Figure 4 This is a schematic diagram illustrating the facade point cloud clustering process and results based on the two-dimensional projection analysis method of this invention. Figure 4 (a) is a point cloud model of the building. Figure 4 (b) Density map of the point cloud projected onto a two-dimensional plane. Figure 4 (c) shows the results of the elevation point cloud clustering;

[0052] Figure 5 This is a schematic diagram of the intensity of the window and wall in different channels and the results of their segmentation, according to the present invention. Figure 5 (a), (b), and (c) show the results for the red, green, and blue channels, respectively.

[0053] Figure 6 This is a schematic diagram of the facade semantic parsing process of the present invention. Figure 6 (a) Clouds on the facade. Figure 6 (b) shows the result of RGB average projection. Figure 6 (c) shows the binary classification results for windows and walls. Figure 6 (d) shows the regularized door and window results. Figure 6 (e) shows the result of back-projection back to the point cloud;

[0054] Figure 7 This is a schematic diagram of point cloud slice generation according to the present invention;

[0055] Figure 8 This is a schematic diagram of the processing flow of this invention from point cloud slicing to standardized external layout, wherein... Figure 8 (a) is a pixel-level semantic plane graph. Figure 8 (b) is the result of geometric primitive abstraction. Figure 8 (c) shows the results of the primitive sorting. Figure 8 (d) shows the regularized outer contour result after window insertion.Figure 8 (e) shows the standardized external layout result;

[0056] Figure 9 This is a schematic diagram of the construction process of the training dataset driven by the building rules of this invention;

[0057] Figure 10 This is a schematic diagram of the multi-branch generative adversarial network structure of the present invention;

[0058] Figure 11 This is a schematic diagram of the process for constructing a multi-layer BIM digital twin model and assessing the consistency of fire and evacuation functions according to the present invention.

[0059] Figure 12 This is a schematic diagram comparing the topological connectivity of the generated layouts of two test samples and the corresponding 3D BIM digital twin models of the present invention. Different colors represent different navigable areas detected by Pathfinder, and red rectangles represent topological error locations.

[0060] Figure 13 This is a comparative schematic diagram of the present invention and the comparative method in the key stage of fire-evacuation coupling simulation, with red circles indicating exit locations;

[0061] Figure 14 This is a comparison graph of the cumulative number of evacuees for the present invention and the comparative method. Detailed Implementation

[0062] The technical solution of the present invention will be further described below with reference to embodiments, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention. In the following embodiments, process equipment or devices not specifically specified are all conventional equipment or devices in the art. Unless specifically specified, the technical means used in the embodiments of the present invention are all conventional means well known to those skilled in the art.

[0063] Example 1, combined with Figure 1 This embodiment illustrates that the present invention provides a method for inferring the interior layout of a building and generating a BIM digital twin model based on external point clouds, comprising:

[0064] S1. Acquire and preprocess 3D point cloud data of the building exterior;

[0065] Specifically, let the point cloud set outside the building be represented as:

[0066]

[0067] Where, p i This represents the i-th point in the point cloud. For its three-dimensional spatial coordinates, ai This represents the attribute value of a point, which is either the reflectance intensity of the LiDAR point cloud or the RGB color information of the point cloud obtained based on multi-view photogrammetry, where N is the total number of points in the cloud.

[0068] The preprocessing includes the following steps:

[0069] S1.1: Acquire multi-view images or laser scan data of the building exterior:

[0070] Acquire raw external observation data that characterizes the geometric shape and openings of the building facade. The data source is either multi-view images of the building exterior or laser scan data of the building exterior. When using multi-view images, it is preferable to ensure overlapping areas between adjacent images to meet the requirements of subsequent 3D reconstruction. Let the acquired image set be:

[0071]

[0072] Among them, I k Let m represent the k-th image, and m be the total number of images acquired.

[0073] S1.2: Reconstruct the point cloud of the building's exterior:

[0074] After acquiring multi-view images, the camera pose is first recovered using the SfM algorithm to generate a sparse point cloud, and then a dense point cloud of the building exterior is generated using the MVS algorithm. When the original data is laser scan data, the point cloud of the building exterior can be read directly. This process yields the point cloud set P of the building exterior.

[0075] If the original data source is a LiDAR scan, the image reconstruction process is skipped, and the process proceeds directly to the subsequent point cloud preprocessing stage. Regardless of the data source, this step yields an external point cloud set P that describes the geometry of the building's outer shell.

[0076] S1.3: Point cloud preprocessing:

[0077] The point cloud obtained from S1.2 is preprocessed to reduce the impact of noise and unify the benchmark for subsequent analysis.

[0078] First, outliers are removed using statistical filtering or radius filtering; second, voxel meshing is used for downsampling; finally, the point cloud is unified to the same reference coordinate system. After preprocessing, a suitable external point cloud for attitude analysis and facade resolution is obtained.

[0079]

[0080] Among them, s vLet P' be the voxel side length, and P′ be the set of downsampled point clouds. Finally, the point clouds are unified into the same reference coordinate system to avoid introducing additional errors due to inconsistent coordinate references in subsequent processing. After processing in step 1, a building exterior point cloud with less noise, clear structure, and suitable for geometric analysis is obtained, providing basic data for subsequent point cloud attitude correction and facade analysis.

[0081] S2. Perform attitude correction, facade division and semantic parsing on the preprocessed point cloud to generate a building facade point cloud with semantic tags, and generate a regularized building exterior plan layout based on this.

[0082] S3. Construct a network input tensor, wherein the tensor includes at least external structural information extracted from the external planar layout, and prior information of the potential grid generated based on the building geometric features;

[0083] S4. Input the network input tensor into the trained internal layout inference model to infer and generate a complete building plan layout including walls, doors, windows and room areas;

[0084] S5. Perform three-dimensional geometric stretching and semantic instantiation on the complete building floor plan, and combine it with the floor information recovered from the facade point cloud to generate a multi-story building BIM digital twin model.

[0085] Furthermore, in S2, the generation of a regularized building exterior plan layout specifically includes:

[0086] S21. Project the pose-corrected point cloud onto a two-dimensional plane to extract the building outline, and divide the point cloud into multiple facade point cloud clusters based on the outline;

[0087] S22. Rasterize each of the facade point cloud clusters into an image, identify the wall, window and entrance door areas, and map the semantic labels back to the three-dimensional point cloud;

[0088] S23. Extract horizontal point cloud slices of the target floor and project them to generate a pixel-level semantic planar map. Perform vectorization and geometric regularization processing on the pixel-level semantic planar map to obtain the regularized building exterior planar layout.

[0089] Specifically, S2 is used to convert the building's external point cloud into a regularized external layout representation, and the specific process is as follows: Figure 2 As shown. Specifically, it includes posture correction, facade division, facade semantic parsing, semantic back projection, horizontal slice generation, and planar standardization processing.

[0090] In S21, point cloud pose correction is performed: the acquired external point clouds of buildings often have arbitrary poses and inconsistent orientations of the main facade. To ensure the stability of subsequent slicing, projection, and contour extraction, pose correction needs to be performed first to align the building's vertical orientation with the Z-axis of the coordinate system and to ensure that the orientation of the main facade is consistent with the coordinate axis direction. The effects before and after correction are as follows: Figure 3 As shown.

[0091] First, the RANSAC algorithm is used to extract candidate ground planes from the point cloud. The planar model in 3D space can be represented as:

[0092]

[0093] Where (a, b, c) are the plane normal vectors, and d is the offset term. For any point p in the point cloud... i =(x i , y i ,z i The distance from the plane to the given plane can be expressed as:

[0094]

[0095] When D i Less than the preset threshold ϵ p When the corresponding point is considered as an interior point of the plane, multiple candidate planes are obtained through random sampling and interior point evaluation. The candidate plane with the largest convex hull area is then selected as the ground estimation plane.

[0096] After identifying the ground plane, calculate its normal vector n. g Then rotate the normal vector to coincide with the +Z axis; then translate the point cloud according to the ground elevation so that the ground is located in the Z = 0 plane.

[0097] After completing the vertical correction, a large vertical plane is identified as a candidate plane for the main facade. The normal vector of the main facade is projected onto the XY plane and the angle θ between it and the +X axis is calculated. Then, the point cloud is rotated around the Z axis to align the direction of the main facade with the coordinate axis.

[0098] In S21, after attitude correction, the external point cloud of the building is divided into several facade point cloud clusters for separate analysis of doors, windows, and walls:

[0099] The corrected point cloud is projected onto the XY plane to extract the closed outline polygon B of the building. Let the edge set of the polygon be {e}. j For any three-dimensional point p i If its planar projection point is on a certain side e j The distance is less than the threshold δ dis Then, the point is assigned to the elevation point cloud cluster corresponding to that edge, resulting in:

[0100]

[0101] Among them, P j Let m represent the cloud cluster of the j-th facade point, and m be the number of sides of the building's outer contour, which is also the actual number of facades.

[0102] Figure 4 The process and results of facade point cloud clustering based on two-dimensional projection analysis are presented. Figure 4 (a) is a point cloud model of the building. Figure 4 (b) is the density map formed by projecting the point cloud onto a two-dimensional plane. Figure 4 (c) is the result of facade clustering based on the building outline boundary.

[0103] In S22, the point cloud clusters of each facade are rasterized into images in order to perform facade semantic parsing:

[0104] For each facade point cloud cluster, it is orthogonally projected onto a two-dimensional plane with the principal normal vector of that facade as the normal. Let the coordinates of the projection plane be (u, v), where u typically corresponds to the horizontal coordinate and v to the vertical coordinate. To convert continuous coordinates into discrete pixel indices, an adaptive resolution rasterization method is used. Taking a facade facing the X direction as an example, from point coordinates to pixel index (i... u i v The mapping of ) can be defined as:

[0105]

[0106]

[0107] Among them, u min , u max and v min , v max These represent the boundary range of the current facade point cloud in the projected coordinate system, and W and H represent the width and height of the output facade image, respectively.

[0108] During rasterization, the attributes of multiple points projected onto the same pixel are aggregated to obtain the corresponding pixel value. This attribute can be either reflection intensity or an RGB color value.

[0109] Regarding semantic segmentation and regularization: Identifying windows, entrance doors, and wall areas based on facade images. Figure 5 The intensity distribution and segmentation results of windows and walls in different color channels are shown. The ability to distinguish between window areas and wall areas can be improved by utilizing the response differences of different channels.

[0110] A threshold τ is set on the rasterized facade image to initially extract candidate opening regions; then, morphological processing and connected component analysis are performed on the candidate regions to remove discrete noise and connect broken regions.

[0111] Calculate geometric descriptors such as area, color ratio, density, range, and aspect ratio for candidate connected regions, and retain regions that meet the conditions according to preset discrimination rules.

[0112] The reserved area is regularized into a rectangle, and the window area and the entrance door area are distinguished by the area's elevation location, scale and distribution characteristics.

[0113] Figure 6 (a) to Figure 6 (e) illustrates the entire process of facade semantic analysis and back projection, in which Figure 6 (e) shows the result after remapping the two-dimensional semantic labels back to the three-dimensional point cloud. Based on the established point cloud-pixel mapping relationship, the semantic labels in the two-dimensional facade are reversed and assigned to the corresponding three-dimensional points to obtain a set of facade point clouds with semantic labels.

[0114] After semantic backprojection, the point cloud simultaneously possesses geometric coordinates and semantic category information.

[0115] To obtain a plan representing the external structure of a standard floor, at the target height Z... slice Extracting horizontal point cloud slices of thickness δ from the vicinity, the specific process is as follows: Figure 7 As shown:

[0116]

[0117] in, For the extracted horizontal point cloud slices, Let i be the i-th three-dimensional point in set P. Slice the target area. The thickness of the slice. for The vertical coordinates;

[0118] The point cloud slices are projected onto the XY plane to obtain a pixel-level semantic plane map. When multiple points are projected onto the same pixel position, conflict resolution is performed according to a preset semantic priority, with the preferred priority being window > wall > background.

[0119] This step also requires vectorization and normalization of the pixel-level semantic plane graph to obtain a regularized external layout suitable for network input and subsequent modeling.

[0120] First, the connected components of the wall area are extracted and abstracted into geometric primitives such as line segments or polyline segments; then, the geometric primitives are sorted in combination with the building's outer contour.

[0121] Apply orthogonal constraints to the endpoints of geometric primitives and connect adjacent endpoints to form a closed outer contour; window openings are inserted into the corresponding wall segments in the form of line segments.

[0122] Then, the regularized vector layout is standardized by uniform scaling: its bounding box is calculated, centered on the preset canvas, and scaled according to a uniform ratio; the walls are drawn using a preset pixel width, resulting in a standardized external layout diagram.

[0123]

[0124] Among them, Γ wall Γ represents the set of regularized exterior wall segments. window This represents the set of window opening line segments embedded in the exterior wall. It also outputs a standardized physical scale `s` for subsequent reconstruction of the 3D geometry from the 2D layout. The final standardized external layout result is shown below. Figure 8 As shown in (e), its step-by-step processing from pixel-level semantic plane graph to regularized layout is as follows: Figure 8 (a) to Figure 8 As shown in (d).

[0125] Furthermore, in S3, the construction of the network input tensor specifically includes:

[0126] S31. Extract the exterior walls, exterior windows, and exterior door passages from the external plan layout to generate an external structure tensor;

[0127] S32. Generate candidate axes based on the corner points of the building's outer contour and the midpoints of the wall segments between windows, and filter them according to the building's internal area mask, then rasterize the filtered axes into potential axis tensors.

[0128] S33. The external structure tensor and the potential axis tensor are concatenated in the channel dimension to form the network input tensor.

[0129] Specifically, S3 is used to construct the multichannel tensor representation required for training samples or inference inputs, and uses externally observable structural information and the axiom prior derived from external geometry as conditions for internal layout inference. The specific process is as follows: Figure 9 As shown, since it is difficult to reliably determine the position of the internal walls based solely on the external contour and window information, this embodiment introduces a potential grid prior into the input tensor to enhance the structural constraints for inferring the internal layout.

[0130] Complete architectural plan semantic encoding:

[0131] To enable the network to learn the spatial relationships between different components, the complete building plan is encoded as a multi-channel semantic representation. The complete building plan can be derived from public datasets, existing drawings, or manually annotated results.

[0132] Let the complete set of architectural plan semantics be:

[0133]

[0134] Wherein: S wall This refers to the wall area, including exterior and interior walls; S door Indicates the doorway area; S window Indicates the window area; S room Indicates the accessible space area inside the building; S bg Indicates the background area.

[0135] A multi-channel independent encoding method is used to encode the walls, doors, windows, rooms, and background as independent channels to construct a semantic tensor:

[0136]

[0137] Each channel T c All images are two-dimensional binary images with the same size as the network input. A pixel value of 1 indicates that the location belongs to the semantic category, and a pixel value of 0 indicates that it does not belong to the category.

[0138] By making each semantic category an independent channel, it becomes easier to apply semantic supervision separately during the training phase.

[0139] Construction of external input samples:

[0140] Since this invention is an inference task that involves extrapolation from the outside, the network input only retains structural information that can be obtained from external observations of the building.

[0141] From the complete semantic tensor T gt Extract the exterior wall, exterior window, and exterior door components, and remove the interior wall, interior door, and room semantics to obtain the external structure tensor:

[0142]

[0143] Among them, T ext Represents the external structure tensor. Indicates an external wall passageway. Indicates an external access point. Indicates an external window passage.

[0144] Through the above processing, the training input is consistent with the actual deployment scenario.

[0145] Potential grid generation:

[0146] To further constrain the internal layout, a potential grid is automatically generated from the external structure and used as an explicit prior input network. The potential grid is preferably generated using the following two types of rules:

[0147] The first type of rule is based on the corner points of the outer contour. Let the set of vertices of the standardized outer contour of the building be:

[0148]

[0149] For each corner point v i Generate a horizontal axis and a vertical axis passing through the point, respectively, to obtain a set of candidate axes for the corner point:

[0150]

[0151] in, This is the set of candidate axes generated from the corner points of the building's outer contour. A horizontal candidate axis is generated through the corner point. To generate a vertical candidate axis through the corner point, This is the set of corner points of the building's outer contour.

[0152] Candidate axes generated from the outer contour corner points are used to indicate potential structural transitions and bay boundaries.

[0153] Another type of candidate axis is generated from the midpoint of the wall segment between adjacent windows:

[0154]

[0155] in, , This represents the center coordinates of two adjacent windows identified on the same building facade. To indicate and Midpoint between;

[0156] By generating candidate axes perpendicular or parallel to the facade through the midpoint of each window wall segment, a set of window axes can be obtained:

[0157]

[0158] in, Let be the set of all candidate axes generated from the midpoint of the wall between windows. It is a single axis in this set;

[0159] Finally, the two types of candidate axes are merged to obtain the initial set of axes:

[0160]

[0161] Where G is the initial set of axes.

[0162] The initial grid set is obtained by merging the two types of axes. This grid is automatically generated by external geometry and window information and does not depend on the original design drawings.

[0163] Internal area filtering and grid gridding: The candidate axes may contain redundant axes that extend to the outside of the building, so filtering is required based on the internal area.

[0164] Let the interior area of ​​the building be R. int Then only candidate axes located within the internal region will be retained:

[0165]

[0166] Among them, G int This is the set of filtered candidate axes.

[0167] After filtering by internal regions, only candidate axes located inside the building are retained.

[0168] The filtered candidate axis set is rasterized into a binary axis lattice channel T of the same size as the input image. axis .

[0169] Training input tensor construction:

[0170] external structure tensor T ext With axis channel T axis The data is then concatenated along the channel dimension to form the final network input.

[0171]

[0172] The final input includes both external structural information and prior information about the grid lines.

[0173] S3 also includes data augmentation: during the training phase, data augmentation is performed on the input-output sample pairs (X, T). gt Simultaneously perform enhancement operations such as rotation, mirroring, and scaling to improve the model's adaptability to changes in pose and scale.

[0174] Enhanced operations maintain the spatial correspondence between inputs and labels.

[0175] Furthermore, S4 specifically includes:

[0176] S4.1. Construct a multi-branch generative adversarial network as the internal layout inference model. The network includes: a shared feature extraction backbone, a structural branch for outputting binary masks of components or non-components, a segmentation branch for outputting semantic categories of walls, doors, windows, and rooms, and an RGB visualization branch for outputting architectural plan visualization images.

[0177] S4.2. Input the network input tensor into the multi-branch generative adversarial network, and use the shared features to extract the multi-scale spatial features between the building outline, window distribution and potential grid lines;

[0178] S4.3. Through the structural branch, the semantic segmentation branch, and the RGB visualization branch, the binary structural results, semantic segmentation results, and visualization reconstruction results of the internal building components are output simultaneously to form the complete building plan layout.

[0179] Specifically, in S4, the shared feature backbone network: Figure 10 The multi-branch generative adversarial network structure used in this embodiment is shown. Let the generator be denoted as G(⋅), its input being the multi-channel tensor X constructed in S33, and its output being the layout prediction result. As shown below:

[0180]

[0181] in, The output of the binary structure branch, The output of the semantic segmentation branch, This is the output of the regression branch.

[0182] The generator preferably adopts an encoder-decoder structure. In the encoding stage, the spatial relationship between the outer contour, window distribution and potential grid is extracted, and in the decoding stage, the positions of the internal walls, room partitions and door openings are restored.

[0183] A shared backbone network is used to extract global spatial features that can be used by all task branches.

[0184] The features obtained from the shared backbone are denoted as:

[0185]

[0186] Here, Enc(⋅) represents the encoder part.

[0187] Based on shared features, multiple output branches are set up:

[0188] The first branch is a binary structure branch, used to output the binary results of the building component region and the non-component region:

[0189]

[0190] The second branch is the semantic segmentation branch, which outputs category results such as walls, doors, windows, rooms, and background:

[0191]

[0192] After the output is activated by Softmax, pixel-level probability distributions for each category can be obtained.

[0193] The third branch is the RGB visualization branch, used to output layout images with preset color encoding rules:

[0194]

[0195] The RGB visualization branch is used to provide additional visual supervision.

[0196] The three branches together constitute the overall output of the generator:

[0197]

[0198] The multi-branch structure allows the network to be constrained by the structural layer, semantic layer, and visual layer simultaneously.

[0199] To enhance the response to key regional features, spatial attention or channel attention mechanisms are introduced into the network.

[0200] Let the shared feature map be F. An attention weight map A is obtained through convolutional mapping and activation functions. Then, this weight map is used to element-wise weight the original features to obtain the enhanced features.

[0201]

[0202] in, This represents element-wise multiplication. Attention weighting is used to increase the feature weights of potential axis regions, outer contour boundary regions, and door / window neighborhoods.

[0203] For buildings with an overall tendency towards symmetry, symmetry constraints can be further introduced:

[0204] To leverage this architectural prior, this implementation performs horizontal and vertical symmetry analysis on the building's outer contour and incorporates the results into network training. Specifically, the degree of symmetry between the building's outer contour and window distribution is detected in both the horizontal and vertical directions, and candidate axes of symmetry are determined accordingly. This is then used to generate the layout. Calculate the differences between the original layout and its horizontal and vertical mirror images. Let S be the layout of the original layout. x (⋅) and S y (⋅) represent horizontal and vertical mirror transformations respectively, then the symmetry loss is defined as:

[0205]

[0206] The symmetry constraint, in the form of a loss term, guides the generation of results to maintain consistency with the overall symmetry trend reflected in the building's external form.

[0207] To improve the clarity of wall boundaries and door / window opening boundaries, edge constraints are introduced during training; for the RGB branch, color consistency constraints can also be introduced.

[0208] Let Δ denote the Laplacian edge operator, and let Y denote the true semantic layout. sem Then the edge loss can be expressed as:

[0209]

[0210] Edge constraints are used to enhance the consistency between the generated layout boundary and the actual layout boundary.

[0211] Color consistency constraints are used to maintain consistency between the generated visual layout and the target color code.

[0212] To ensure the authenticity of the generated results, a multi-scale discriminator D(⋅) is introduced. The discriminator receives the combination of input conditions X and the generated output to determine whether the generated layout matches the input external structural conditions.

[0213] The multi-scale discriminator evaluates the authenticity of layout organization and local connectivity from both global and local scales.

[0214] A composite loss function is used to jointly optimize the generator network. The total loss function is defined as follows:

[0215]

[0216] The total loss includes adversarial loss, feature matching loss, binary structure loss, semantic segmentation loss, RGB branching loss, edge loss, and symmetry loss; λ1, λ2…λ6 are the weight coefficients of each loss.

[0217] The adversarial loss can be written as:

[0218]

[0219] Feature matching loss can be expressed as:

[0220]

[0221] Among them, D t (⋅) represents the feature map of the t-th layer of the discriminator, N t This indicates the number of features in this layer.

[0222] The preferred semantic branching loss method is weighted cross-entropy loss.

[0223]

[0224] Among them, w c This represents the category weight of class c, used to alleviate the problem of unbalanced pixel counts among categories such as walls, doors, windows, and backgrounds.

[0225] Through joint optimization, an internal layout result that satisfies both structural and semantic constraints is obtained.

[0226] After semantic decoding, a complete architectural floor plan is obtained, including exterior walls, interior walls, doors, windows, room areas, and background areas.

[0227] Furthermore, in S5, the generation of a multi-story building BIM digital twin model specifically includes:

[0228] S51. Perform component analysis and single-layer 3D extrusion on the complete building floor plan to generate single-layer BIM semantic objects and establish topological relationships between components;

[0229] S52. Analyze the vertical distribution pattern of windows and floor slabs in the point cloud of the building facade, and automatically restore the building floor height and total number of floors;

[0230] S53. Based on the recovered floor information, the single-story BIM model is copied and adjusted along the vertical direction to generate the multi-story building BIM digital twin model, and the vertical spaces such as staircases and elevator shafts are modeled in a continuous manner.

[0231] Specifically, S5 includes single-layer semantic 3D stretching, single-layer BIM instantiation, floor information recovery, multi-layer BIM extension, and functional consistency assessment. The specific process is as follows: Figure 11 As shown.

[0232] Single-layer semantic 3D stretching:

[0233] First, based on the complete planar layout output by S4, the walls, doors, windows, and room areas are geometrically vectorized and their components are analyzed. Let the set of wall centerline segments be:

[0234]

[0235] Among them, l i This represents the centerline of the i-th wall. Given the wall thickness t... w and floor height H f Each wall segment can be offset in two dimensions and stretched along the vertical direction to generate a three-dimensional wall solid.

[0236]

[0237] Among them, W i For a 3D wall solid, Extrude(⋅) represents a 3D extrusion operation.

[0238] The three-dimensional opening range of door and window components is determined based on their position and width in the two-dimensional plane and the height information obtained from the semantic parsing of the facade, thereby forming single-layer wall, door, window and room space objects.

[0239] Single-layer semantic BIM instantiation:

[0240] After completing the single-layer geometric extrusion, the geometric objects need to be converted into BIM semantic objects. Let the set of single-layer room areas be:

[0241]

[0242] Where, r i Each room area is an independent room region object. Each closed room region can be instantiated as a space object and a topological relationship can be established with the enclosing walls. Door and window objects establish an opening relationship with their host walls.

[0243] After instantiation, a single-layer model contains both geometric entities and object semantic relationships.

[0244] Multi-story building floor number identification and floor height restoration:

[0245] To recover the building's floor height and number of floors, the vertical repetition patterns of windows and floor slabs in the facade semantic point cloud can be analyzed to estimate the floor height H. f and the total number of floors N f .

[0246] Preferably, peak clustering can be performed on the elevation distribution of windows on the facade to obtain floor height zones, and the floor height can be estimated based on the spacing between adjacent floor height zones; if necessary, the total building height H can be combined. tot Calculate the number of floors:

[0247]

[0248] This yields information about the building's floor organization.

[0249] Multi-layer BIM generation:

[0250] This allows a single-layer semantic BIM model to be vertically expanded into a multi-layer BIM digital twin model. For the k-th layer, its overall elevation can be represented as:

[0251]

[0252] Then, translate the single-layer walls, doors, windows, and space objects along the Z-axis to the corresponding floor level to generate the wall collection W for each floor. k Gate set D k Windows collection k and room set R k .

[0253] When there are differences in the outer contour or window distribution of different floors, external layout inputs can be generated for the corresponding floors and internal layout inference and BIM instantiation can be performed separately.

[0254] Vertical structural members between floors and overall BIM organization:

[0255] A continuous model is created for vertical core areas such as stairwells, elevator shafts, vertical pipe shafts, and public transportation spaces to establish vertical organizational relationships between floors.

[0256] Each floor object is assigned a floor number, elevation, component category, and spatial attribute information to form the overall hierarchical organization of the building.

[0257] BIM format writing and digital twin model generation:

[0258] Write all semantic objects into IFC or other BIM data format files to generate a complete multi-story building BIM digital twin model.

[0259] After being written out, the two-dimensional layout results are converted into a three-dimensional semantic model that can be used for building operation and maintenance, path analysis, and digital twin simulation.

[0260] During application verification, the building space connectivity network can be extracted based on multi-layer BIM, and evacuation path analysis, fire simulation, or traffic analysis can be performed. Figure 12 The generated layouts of two test samples and the corresponding 3D BIM models are shown, with topological connectivity comparisons. Different colors represent different navigable areas, and red rectangles indicate topological errors. Combined with... Figure 12 It can be seen that the topological connectivity of the layout generated by this invention and its corresponding 3D BIM model can reflect the correctness of the internal spatial organization restoration.

[0261] For example, rooms, corridors, stairwells, and entrances can be used as nodes, and doorways and traffic connections can be used as edges to construct a spatial graph structure. Based on this, indicators such as total evacuation time, critical paths, or bottleneck locations can be calculated. Figure 13 The comparison results of the method of this invention and the comparative method in the key stage of fire-evacuation coupling simulation are shown, where the red circles indicate the exit locations. Combined with... Figure 13 It can be seen that the model generated by this invention is closer to the reference model in terms of evacuation organization and exit accessibility.

[0262] Figure 14 The cumulative evacuation figures curves of the method of this invention and the comparative method are shown. These curves can be used to characterize the functional consistency of the generated model during the evacuation process. When the generated model and the reference model meet the preset consistency conditions in terms of spatial connectivity, traffic logic, or application indicators, the generated model can be considered to meet the corresponding application requirements.

[0263] Let M be a certain functional index of the real model. gt The corresponding functional index of the generated model is M. pred Then its relative error can be expressed as:

[0264]

[0265] When multiple core functional indicators meet the preset tolerance range, the generated multi-layer BIM model can be considered to have good application consistency.

[0266] It also provides a system for inferring the interior layout of buildings and generating BIM digital twin models based on external point clouds, including:

[0267] The external point cloud acquisition module is used to acquire and preprocess the 3D point cloud data of the building exterior.

[0268] The external layout parsing module is used to perform attitude correction, facade division and semantic parsing on the preprocessed point cloud, generate a building facade point cloud with semantic tags, and generate a regularized building external plan layout based on this.

[0269] A data construction module is used to construct a network input tensor, wherein the tensor includes at least external structural information extracted from the external planar layout, and prior information of the potential grid generated based on the building geometric features;

[0270] The internal layout inference module is used to input the network input tensor into the trained internal layout inference model and infer and generate a complete building plan layout including interior walls, doors, windows and room areas.

[0271] The BIM digital twin modeling module is used to perform three-dimensional geometric stretching and semantic instantiation of the complete building floor plan, and combine the floor information recovered from the facade point cloud to generate a multi-story building BIM digital twin model.

[0272] Specifically, the system includes the following functional modules:

[0273] 1. External Point Cloud Acquisition Module: This module acquires and preprocesses 3D point cloud data of the building exterior. It can acquire multi-view image data or laser scan data of the building exterior. If it is image data, it reconstructs the 3D point cloud of the building exterior based on the structured beam adjustment algorithm and the multi-view stereo matching algorithm. If it is laser scan data, it directly reads the 3D point cloud. At the same time, it performs preprocessing operations on the acquired 3D point cloud of the building exterior, removes outliers by statistical filtering or radius filtering, performs voxel downsampling using the voxel mesh method, and unifies all point cloud data to the same coordinate reference system. The final output is a set of building exterior point clouds with low noise, clear structure, and suitable for subsequent geometric analysis.

[0274] 2. External Layout Analysis Module: This module first performs attitude correction on the preprocessed point cloud. It then uses the RANSAC algorithm to detect and select the candidate plane with the largest convex hull area as the ground plane. The point cloud is rotated and translated to align the building's vertical orientation with the Z-axis of the coordinate system, with the ground located on the Z=0 plane. Next, the main facade is identified and rotated around the Z-axis to align it with the coordinate axis direction. Subsequently, the corrected point cloud is projected onto the XY plane to extract the building's closed contour. Based on the contour boundaries and the distance threshold from points to edges, it is divided into multiple facade point cloud clusters. After rasterizing each facade point cloud cluster, walls, windows, and entrance door areas are identified. Semantic labels are then mapped back to the 3D point cloud to obtain a building facade point cloud with semantic labels. Finally, horizontal point cloud slices of the target floor are extracted to generate a pixel-level semantic plan view. Through semantic priority conflict resolution, vectorization, geometric regularization, and standardization, a regularized building external plan layout is generated.

[0275] 3. Data Construction Module: This module first encodes the complete building floor plan into a multi-channel semantic tensor containing walls, doors, windows, rooms, and background. From this semantic tensor, only external walls, windows, and doorways that can be observed externally are extracted to generate an external structure tensor. Then, horizontal and vertical candidate axes are generated based on the corner points of the building's outer contour. Combined with the midpoints of the wall segments between windows, candidate axes perpendicular / parallel to the facade are generated. After merging, an initial set of candidate axes is formed. Redundant axes are filtered based on the building's internal area mask, and the filtered axes are rasterized into a potential axis grid tensor. Finally, the external structure tensor and the potential axis grid tensor are concatenated in the channel dimension to form a network input tensor that simultaneously contains external structure information and potential axis grid prior information. Data augmentation operations such as rotation, mirroring, and scaling can also be performed on the input samples and supervision labels simultaneously to improve the adaptability of subsequent model inference.

[0276] 4. Internal Layout Inference Module: This module incorporates a trained multi-branch generative adversarial network as the internal layout inference model. This model includes a shared feature extraction backbone, a structural branch that outputs binary masks of components / non-components, and a segmentation branch that outputs semantic categories of walls / doors / windows / rooms. The module inputs the network input tensor output by the data construction module into this model. Through the shared feature extraction backbone, it extracts multi-scale spatial features between the building's outer contour, window distribution, and potential grid lines. Then, through the structural branch and semantic segmentation branch, it simultaneously outputs the binary structural results and semantic segmentation results of the building's internal components. After mask filtering, connected component analysis, and region merging, invalid information is removed and independent room areas are identified. Finally, a complete building floor plan containing interior walls, doors, windows, and room areas is generated.

[0277] 5. BIM Digital Twin Modeling Module: This module first analyzes the complete building floor plan, extracting wall centerlines, door and window positions and dimensions. Given wall thickness and floor height, it performs 2D offset on the walls and 3D stretching along the vertical direction to generate single-layer wall, door, window and room space objects. These are then transformed into BIM semantic objects containing component categories, dimensions, and other information, and semantic topological relationships between doors and windows and host walls, and between rooms and enclosing walls are established. Next, the vertical distribution information of windows and floor slabs is extracted from the point cloud of the building facade with semantic tags. The building floor height and total number of floors are recovered through peak clustering and other methods, and standard and non-standard floors are identified. Finally, based on the floor information, the single-layer BIM model is extended vertically. Standard floors are directly copied and generated, while non-standard floors undergo separate internal layout inference and instantiation. Vertical core areas such as stairwells and elevator shafts are modeled in a continuous manner. Floor numbers, elevations, component categories and other attribute information are added to each floor object. After integration, all semantic objects are written into IFC or other standardized BIM data formats to generate a complete multi-story building BIM digital twin model.

[0278] The system can also be supplemented with a functional consistency verification module, as an extension of the five core modules mentioned above. This module is used to verify the usability of the generated multi-layer BIM digital twin model at the application level. Based on the multi-layer BIM model, this module constructs a building space connectivity network, using rooms, corridors, stairwells, and entrances as nodes, and doorways and traffic connections as edges to build the spatial graph structure. Through fire simulation, evacuation simulation, or traffic analysis, it calculates functional indicators such as total evacuation time, cumulative evacuees, and critical paths, and compares the relative errors of the generated model with those of the real model.

[0279]

[0280] Among them, E RSET This represents relative error. When multiple core functional indicators meet the preset tolerance requirements, the generated multi-layer BIM digital twin model is deemed to have a high degree of consistency with the real building in terms of functionality.

[0281] The above modules are logical functional divisions based on method steps. In practical applications, each module can be implemented by one or more processing units in hardware, software, or a combination of hardware and software. The functions between modules can be reasonably combined, split, or integrated according to actual needs without affecting the overall implementation effect of the technical solution of this invention.

Claims

1. A method for inferring the interior layout of a building and generating a BIM digital twin model based on external point clouds, characterized in that, include: S1. Acquire and preprocess 3D point cloud data of the building exterior; S2. Perform attitude correction, facade division and semantic parsing on the preprocessed point cloud to generate a building facade point cloud with semantic tags, and generate a regularized building exterior plan layout based on this. S3. Construct a network input tensor, wherein the tensor includes at least external structural information extracted from the external planar layout, and prior information of the potential grid generated based on the building geometric features; S4. Input the network input tensor into the trained internal layout inference model to infer and generate a complete building plan layout including walls, doors, windows and room areas; S5. Perform three-dimensional geometric stretching and semantic instantiation on the complete building floor plan, and combine it with the floor information recovered from the facade point cloud to generate a multi-story building BIM digital twin model.

2. The method for inferring the building's interior layout and generating a BIM digital twin model based on external point clouds according to claim 1, characterized in that, In S2, the generation of a regularized building exterior plan layout specifically includes: S21. Project the pose-corrected point cloud onto a two-dimensional plane to extract the building outline, and divide the point cloud into multiple facade point cloud clusters based on the outline; S22. Rasterize each of the facade point cloud clusters into an image, identify the wall, window and entrance door areas, and map the semantic labels back to the three-dimensional point cloud; S23. Extract horizontal point cloud slices of the target floor and project them to generate a pixel-level semantic planar map. Perform vectorization and geometric regularization processing on the pixel-level semantic planar map to obtain the regularized building exterior planar layout.

3. The method for inferring the building's interior layout and generating a BIM digital twin model based on external point clouds according to claim 2, characterized in that, In S3, the construction of the network input tensor specifically includes: S31. Extract the exterior walls, exterior windows, and exterior door passages from the external plan layout to generate an external structure tensor; S32. Generate candidate axes based on the corner points of the building's outer contour and the midpoints of the wall segments between windows, and filter them according to the building's internal area mask, then rasterize the filtered axes into potential axis tensors. S33. The external structure tensor and the potential axis tensor are concatenated in the channel dimension to form the network input tensor.

4. The method for inferring the building's interior layout and generating a BIM digital twin model based on external point clouds according to claim 3, characterized in that, S4 specifically includes: S4.

1. Construct a multi-branch generative adversarial network as the internal layout inference model. The network includes: a shared feature extraction backbone, a structural branch for outputting binary masks of components or non-components, a segmentation branch for outputting semantic categories of walls, doors, windows, and rooms, and an RGB visualization branch for outputting a visual image of the building plan. S4.

2. Input the network input tensor into the multi-branch generative adversarial network, and use the shared features to extract the multi-scale spatial features between the building outline, window distribution and potential grid lines; S4.

3. Through the structural branch, the semantic segmentation branch, and the RGB visualization branch respectively, the binary structural results, semantic segmentation results, and visualization reconstruction results of the internal building components are output simultaneously to form the complete building plan layout.

5. The method for inferring the building's interior layout and generating a BIM digital twin model based on external point clouds according to claim 4, characterized in that, In S5, the generation of a multi-story building BIM digital twin model specifically includes: S51. Perform component analysis and single-layer 3D extrusion on the complete building floor plan to generate single-layer BIM semantic objects and establish topological relationships between components; S52. Analyze the vertical distribution pattern of windows and floor slabs in the point cloud of the building facade, and automatically restore the building floor height and total number of floors; S53. Based on the recovered floor information, the single-story BIM model is copied and adjusted along the vertical direction to generate the multi-story building BIM digital twin model, and the vertical spaces such as staircases and elevator shafts are modeled in a continuous manner.

6. A system for inferring the interior layout of buildings and generating BIM digital twin models based on external point clouds, characterized in that, include: The external point cloud acquisition module is used to acquire and preprocess the 3D point cloud data of the building exterior. The external layout parsing module is used to perform attitude correction, facade division and semantic parsing on the preprocessed point cloud, generate a building facade point cloud with semantic tags, and generate a regularized building external plan layout based on this. A data construction module is used to construct a network input tensor, wherein the tensor includes at least external structural information extracted from the external planar layout, and prior information of the potential grid generated based on the building geometric features; The internal layout inference module is used to input the network input tensor into the trained internal layout inference model and infer and generate a complete building plan layout including interior walls, doors, windows and room areas. The BIM digital twin modeling module is used to perform three-dimensional geometric stretching and semantic instantiation of the complete building floor plan, and combine the floor information recovered from the facade point cloud to generate a multi-story building BIM digital twin model.

7. A computer device, including a memory and a processor, characterized in that, The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the method for inferring the building interior layout and generating a BIM digital twin model based on external point clouds as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method for inferring the building interior layout and generating a BIM digital twin model based on external point clouds as described in any one of claims 1-5.

Images