Existing building current situation rapid modeling method based on point cloud and BIM fusion

Abstract

The invention discloses an existing building current situation rapid modeling method based on point cloud and BIM fusion, and belongs to the technical field of building information models.The method comprises the steps that multi-source heterogeneous point cloud data acquisition and preprocessing are executed, and three-dimensional laser scanning and unmanned aerial vehicle oblique photography data are acquired and fused; building contour and structure feature automatic identification is executed, and a segmentation method combining local geometric features and a deep neural network is adopted; performing floor separation and door and window hole level collaborative detection based on a height histogram and geometric semantics combined screening method; parameterized model matching and conversion based on the BIM component template library are executed; executing local correction and attribute supplementation driven by human-computer interaction; according to the method, efficient and automatic conversion from the point cloud data to the BIM model is realized, and the current situation investigation period of the existing building is remarkably shortened.

Description

Technical Field

[0001] This invention relates to the field of building information modeling technology, specifically to a method for rapid modeling of existing buildings based on the fusion of point cloud and BIM. Background Technology

[0002] With the rapid advancement of urban renewal and old city renovation projects, the need for accurate acquisition and digital modeling of existing building status information is becoming increasingly urgent. Building Information Modeling (BIM), as the core data carrier for the entire building lifecycle management, can effectively support the renovation design, construction management, and operation and maintenance decisions of existing buildings. However, many existing buildings were constructed a long time ago and lack complete design drawings and as-built data, making it extremely difficult to obtain their status information. Traditional manual surveying methods are not only time-consuming and labor-intensive, but also unable to fully cover the complex structural details of buildings, and the measurement accuracy is insufficient to meet the requirements of subsequent planning and analysis.

[0003] Chinese invention CN120374886A discloses a point cloud modeling method and system based on feature extraction. The method first acquires 3D point cloud data by laser scanning of the target building component area and preprocesses it. Then, it constructs a multi-scale local neighborhood centered on each point, calculates the covariance matrix, and decomposes the eigenvalues. Normalized curvature indices are extracted and normalized using the Sigmoid function to generate an edge probability map for screening candidate edge feature points. An improved PointNet++ network is used for semantic segmentation to output the corresponding component category, and a set of key feature points is selected based on the component category. A three-level B-spline surface model is constructed using the feature point cloud in the key feature point set as control points to fit the geometric shape of the target component. The spatial distribution density is dynamically adjusted using a density adjustment function for adaptive reconstruction. Based on the adaptively reconstructed 3D model, a weighted complete graph is formed using the key feature point set. An improved Kruskal algorithm is used to generate a minimum spanning tree to determine the initial connectivity. Simulated annealing is introduced to iteratively optimize the connectivity. Kuratowski's theorem is used for expandability verification to identify and process non-planar regions. Finally, the QEM algorithm is applied for lightweight processing to complete the 3D modeling.

[0004] However, the technical solutions of the above inventions have the following shortcomings: First, they mainly target the detailed modeling of single building components, using B-spline surface models for geometric fitting. This makes it difficult to adapt to the needs of overall modeling of existing buildings for the collaborative expression of multiple types of components and standardized parameter output, and cannot directly generate parametric component models that conform to BIM specifications. Second, they use a single laser scanning data source, lacking the ability to fuse and process multi-source heterogeneous data such as 3D laser scanning and UAV oblique photography. When facing large existing buildings or complex surrounding environments, the integrity and efficiency of data acquisition are difficult to guarantee. Third, the feature extraction method focuses on local curvature features and fails to effectively utilize the unique structural regularity of existing buildings for floor separation detection and door and window opening identification. It lacks the ability to acquire the most critical floor height information and opening location information in the survey of the current status of existing buildings. Fourth, the fully automated processing flow lacks a human-computer interaction mechanism for operators to make local corrections and supplement attribute information, making it difficult to meet the requirements of planning analysis for the integrity of building attributes. Fifth, the quality verification only uses a one-way inspection method and does not establish a closed-loop calibration mechanism between the model and the original data. When there are deviations in the automatic modeling results, it is difficult to perform targeted parameter adjustments and iterative optimization.

[0005] In recent years, scholars both domestically and internationally have conducted extensive research on point cloud semantic segmentation and automated Scan-to-BIM. Significant progress has been made in deep learning-based point cloud segmentation methods. Network models such as AttTransNet have improved local feature extraction capabilities through attention mechanisms, and transfer learning frameworks have enhanced the generalization performance of models on new datasets. In building facade segmentation, RandLA-Net, combined with an intensity correction model, has achieved automatic identification of facade components, while DLA-Net has improved the semantic segmentation accuracy of large-scale building facade point clouds through dual local attention feature learning. Regarding automatic BIM generation, parametric modeling methods have achieved automatic model creation by matching extracted geometric features with a BIM component library, and deep learning-based Scan-to-BIM frameworks can handle spatial elements and furniture objects in complex interior scenes. However, existing technologies still face challenges in handling overall modeling tasks of existing buildings, including low data fusion efficiency, inaccurate floor structure identification, insufficient automation in BIM conversion, and a lack of human-machine collaboration mechanisms. These limitations make it difficult to meet the dual requirements of urban renewal projects for both the time required for current status surveys and the accuracy of modeling.

[0006] To address the aforementioned technical challenges, there is an urgent need to propose a point cloud and BIM fusion method for rapid modeling of existing buildings. This method should be able to efficiently process multi-source heterogeneous point cloud data, automatically identify building outlines and structural features, accurately detect floor divisions and door / window openings, generate standard BIM component models based on a parametric template library, achieve local corrections and attribute supplementation through human-computer interaction, and establish a closed-loop quality verification mechanism to ensure modeling accuracy. This would significantly shorten the survey cycle for existing buildings and improve modeling quality. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a rapid modeling method for existing buildings based on the integration of point cloud and BIM. This method solves the technical problems of difficulty in obtaining information on the current status of existing buildings and low efficiency of manual modeling, and realizes efficient and automated conversion from point cloud data to BIM model, significantly shortening the current status survey cycle and improving modeling accuracy.

[0008] To achieve the above objectives, this invention provides a method for rapid modeling of existing buildings based on the fusion of point cloud and BIM, comprising the following steps:

[0009] The process involves acquiring and preprocessing multi-source heterogeneous point cloud data. Raw point cloud data of existing buildings and their surrounding environment are acquired using 3D laser scanning equipment or UAV oblique photography equipment. Coordinate transformation and multi-view registration are performed on the raw point cloud data to obtain a global point cloud dataset. Noise removal processing based on a combination of statistical filtering and radius filtering is performed on the global point cloud dataset, and adaptive downsampling processing based on voxel mesh is performed to obtain a preprocessed point cloud dataset.

[0010] The automatic identification steps for building outline and structural features are performed. For each point in the preprocessed point cloud dataset, a local geometric feature descriptor is calculated. The local geometric feature descriptor includes normal vector, curvature and linearity. A segmentation method based on region growing and deep neural network is used to perform semantic segmentation on the preprocessed point cloud dataset to identify and extract the building outline point cloud and the building structure boundary feature points.

[0011] The process involves collaborative detection of floor division and door / window openings. A height distribution histogram is constructed vertically based on the point cloud of the building's main outline. The floor division locations are identified through peak detection and valley analysis. Planar detection and edge extraction are performed on the point cloud of each floor area to identify candidate areas for door / window openings. A joint screening method using geometric constraints and semantic confidence is used to determine the precise location and size parameters of the door / window openings, and topological relationships between adjacent floors are established.

[0012] Perform parametric model matching and conversion steps based on the BIM component template library. Retrieve candidate component templates that match the geometric features of the building structure boundary feature points from the pre-built parametric template library. Calculate the matching degree score between the candidate component templates and the extracted geometric features. Select the component template with the highest matching degree score as the target template. Perform parametric instantiation processing on the target template according to the extracted geometric parameters to generate the BIM component model.

[0013] The system performs local correction and attribute supplementation steps driven by human-computer interaction. It overlays the generated BIM component model with the original point cloud data through a visual interactive interface. It receives user correction instructions for component position, size and shape and performs corresponding model adjustment operations. It receives component attribute information input by the user and writes it into the corresponding BIM component model, generating a 3D information model of the existing building containing geometric and attribute information.

[0014] The closed-loop quality verification and model consistency calibration steps are executed. The geometric deviation distribution between the existing building 3D information model and the preprocessed point cloud dataset is calculated. When the geometric deviation exceeds the preset threshold, a deviation calibration instruction is generated and fed back to the parameterized model matching and conversion steps to perform parameter correction. The process is iterated until the geometric deviation meets the preset accuracy requirements, and the final existing building 3D information model is output.

[0015] The present invention has the following beneficial effects:

[0016] First, this invention achieves effective fusion of 3D laser scanning and UAV oblique photography data through multi-source heterogeneous point cloud data acquisition and preprocessing steps. It adopts a noise removal method combining statistical filtering and radius filtering and an adaptive downsampling method based on voxel grids, which can adapt to the characteristics of point cloud data under different acquisition devices and scene conditions. Compared with a single data source solution, it improves the integrity of data acquisition and processing efficiency, laying a high-quality data foundation for subsequent feature extraction and model generation.

[0017] Secondly, the present invention employs a segmentation method combining local geometric feature descriptors and deep neural networks in the automatic identification step of building outline and structural features. This fully utilizes the structural regularity features of existing buildings to achieve accurate identification of the main outline and structural boundaries of buildings. The identification accuracy is significantly improved compared with traditional curvature feature methods.

[0018] Third, this invention innovatively utilizes height distribution histograms for floor separation detection through a collaborative detection step involving floor separation and door / window openings. It also achieves precise positioning of door / window openings through joint screening using geometric constraints and semantic confidence, establishing topological relationships between floors and meeting the core needs of existing building status surveys for structural hierarchy information.

[0019] Fourth, this invention achieves automatic conversion from point cloud features to standard BIM component models through parametric model matching and conversion steps based on the BIM component template library. The accuracy of component template selection is ensured through a matching degree scoring mechanism. The generated BIM model meets industry standard requirements and can be directly used for subsequent planning analysis and design applications.

[0020] Fifth, the present invention provides a visual interactive interface through human-computer interaction-driven local correction and attribute supplementation steps, enabling operators to make targeted corrections to the automatic modeling results and supplement attribute information, thus balancing modeling efficiency and model integrity and improving the practical value of the modeling results.

[0021] Sixth, this invention establishes a two-way verification mechanism between the model and the point cloud through closed-loop quality verification and model consistency calibration steps. When the detected geometric deviation exceeds the threshold, the parameter correction process is automatically triggered for iterative optimization to ensure that the geometric accuracy of the final output model meets the preset requirements, thus achieving controllable assurance of modeling quality. Attached Figure Description

[0022] Figure 1 This is the overall flowchart of the method for rapid modeling of existing buildings based on the integration of point cloud and BIM, which is the basis of this invention.

[0023] Figure 2 This is a detailed flowchart of the multi-source heterogeneous point cloud data acquisition and preprocessing steps of the present invention.

[0024] Figure 3 This is a detailed flowchart of the automatic identification steps for building outlines and structural features in this invention.

[0025] Figure 4 This is a detailed flowchart of the collaborative detection steps for floor division and door / window opening levels in this invention.

[0026] Figure 5 This is a detailed flowchart of the parametric model matching and conversion steps based on the BIM component template library of this invention.

[0027] Figure 6 This is a detailed flowchart of the local correction and attribute supplementation steps driven by human-computer interaction in this invention.

[0028] Figure 7 This is a detailed flowchart of the closed-loop quality verification and model consistency calibration steps of this invention. Detailed Implementation

[0029] Reference Appendix Figures 1-7 The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0030] like Figure 1As shown, the rapid modeling method for existing buildings based on point cloud and BIM fusion provided by this invention includes six core steps: multi-source heterogeneous point cloud data acquisition and preprocessing step S1, automatic identification of building outline and structural features step S2, collaborative detection of floor division and door / window opening hierarchy step S3, parametric model matching and conversion based on BIM component template library step S4, human-computer interaction driven local correction and attribute supplementation step S5, and closed-loop quality verification and model consistency calibration step S6. These six steps form a deeply coupled closed-loop collaborative process, with the output of the previous step serving as the key input for the next step. The verification result of step S6 can drive parameter optimization in step S4, achieving continuous improvement in modeling accuracy.

[0031] Step S1 is the acquisition and preprocessing step of multi-source heterogeneous point cloud data, such as... Figure 2 As shown.

[0032] In this step, raw point cloud data of the existing building and its surrounding environment is first acquired using either a 3D laser scanning device or a drone oblique photography device. The 3D laser scanning device employs a terrestrial 3D laser scanner with a scanning accuracy better than 3mm at 50m, a scanning field of view covering 360 degrees horizontally and at least 270 degrees vertically, and a single-point cloud density of at least 1000 points per square meter. The drone oblique photography device uses a multi-rotor drone equipped with a five-lens oblique camera, with a camera resolution of at least 42 megapixels. It uses the Structure from Motion algorithm to perform 3D reconstruction and generate a dense point cloud with a density of at least 500 points per square meter. For large existing buildings or complex scenes, both devices can be used simultaneously for data acquisition to ensure complete coverage. The terrestrial laser scanning primarily acquires detailed data of the building facade and ground floor, while the drone oblique photography primarily acquires data of the building roof and upper-floor facades.

[0033] During the data acquisition planning phase, a scanning scheme needs to be developed based on the size and complexity of the existing buildings. For small buildings with a floor area of ​​less than 2,000 square meters, 4 to 6 scanning stations are sufficient for complete coverage; for medium-sized buildings with a floor area of ​​2,000 to 5,000 square meters, 8 to 12 scanning stations are required; and for large buildings with a floor area of ​​more than 5,000 square meters, more than 15 scanning stations are required, in conjunction with oblique photography using drones. The deployment of scanning stations should adhere to the principle that the overlap rate of the field of view between adjacent stations is no less than 30% to ensure the reliability of subsequent registration processing.

[0034] After acquiring the raw point cloud data, coordinate unification transformation and multi-view registration are performed. Coordinate unification transformation converts the point cloud data from various data sources to a unified engineering coordinate system. A seven-parameter coordinate transformation is performed using known control points to ensure a coordinate accuracy better than 0.01m after transformation. The seven-parameter coordinate transformation includes three translation parameters, three rotation parameters, and one scale parameter. The transformation formula uses the Bursa-Wolf model. Control points should be evenly distributed around the building, with no fewer than four control points, and measurements should be taken using a high-precision total station or GNSS-RTK equipment.

[0035] Multi-view registration processing employs a fine registration method based on the Iterative Closest Point (ICP) algorithm, with registration accuracy controlled within 0.005m. When multiple scanning stations or multiple flight data exist, a coarse registration method based on feature point matching is used for initial alignment, followed by ICP fine registration to obtain a global point cloud dataset. Coarse registration uses FPFH feature descriptors for point cloud feature extraction, and the RANSAC algorithm is used for feature point pair matching to estimate the initial transformation matrix. The root mean square error calculation formula during the registration process is as follows:

[0036] ,

[0037] in, To register the root mean square error, To register the number of points, The first point in the source cloud The coordinate vectors of the registration points In the target point cloud The coordinate vector of the nearest point, This represents the Euclidean distance. When... Registration is considered converged when the error difference is less than 0.005m, and the registered global point cloud dataset is output. The ICP algorithm uses a point-to-plane variant to improve the convergence speed, with a maximum number of iterations set to 100 and a convergence threshold set to an error difference of less than 0.0001m between two consecutive iterations.

[0038] Noise removal processing based on a combination of statistical filtering and radius filtering is performed on the global point cloud dataset. Statistical filtering first calculates the average distance between each point and its K nearest neighbors. Then calculate the mean of the average distances of all points. and standard deviation The average distance is greater than Points that are identified as outliers are removed, among which... The value range is 20 to 50 points. The value range is from 1.0 to 2.0. The outlier determination criterion for statistical filtering is:

[0039] ,

[0040] in, For the first The point and its The average distance between the nearest neighbors, The mean of the average distances between all points. The standard deviation of the average distance of all points. This is the threshold value, representing a multiple of the standard deviation. Statistical filtering is primarily used to remove outlier noise points caused by factors such as measurement random errors and atmospheric refraction. These noise points exhibit significant isolation characteristics in their spatial distribution, showing a statistically significant difference from the clustered distribution of points on the actual building surface.

[0041] Radius filtering constructs a radius based on each point as the center. Given a spherical neighborhood, count the number of points within the spherical neighborhood, and then select the points with a number less than [a certain value]. Points that are considered sparse points are removed, among which... The value ranges from 0.05m to 0.15m. The value range is 5 to 10 points. Radius filtering is mainly used to remove sparse noise points caused by laser scanning occlusion and edge effects. These noise points are located at the boundaries of the scanning blind zone or the edges of objects, and the point density in their neighborhood is significantly lower than that in the normal area. The combined use of the two filtering methods can effectively eliminate noise points from different sources while preserving the true geometric details of the building surface.

[0042] Statistical filtering is primarily used to remove outlier noise points caused by random errors in scanning equipment, atmospheric refraction, and multipath effects. These points are typically located far from the actual building surface. Radius filtering is mainly used to remove sparse and isolated points caused by occlusion, edge effects, and equipment blind spots. These points usually lack sufficient neighborhood support and therefore do not have reliable geometric meaning. The combined use of these two filtering methods can effectively remove outliers and sparse points caused by equipment noise, environmental interference, and occlusion, while preserving a valid point cloud representing the true geometry of the building. Based on practical experience, combined filtering can typically remove 3% to 8% of the noise points in the original point cloud.

[0043] The denoised point cloud is subjected to adaptive downsampling based on voxel grids. Adaptive downsampling dynamically sets the voxel grid size according to the original point cloud density, ensuring a reduction in data volume while preserving architectural structural features. The voxel grid size setting rule is: when the original point cloud density... When the number of points is greater than 10,000 per cubic meter, the voxel mesh size Set to 0.02m to 0.05m; when the original point cloud density When the number of points is 1000 to 10000 per cubic meter, the voxel grid size is... Set to 0.05m to 0.10m; when the original point cloud density Downsampling is not performed when the number of points is less than 1000 per cubic meter. Downsampling uses the in-voxel centroid method, taking the average coordinates of all points within each voxel as the representative point coordinates for that voxel. The formula for calculating the voxel centroid coordinates is:

[0044] ,

[0045] in, Let the body mass center coordinate vector be... The number of points within a voxel. For voxels The coordinate vectors of each point. Adaptive downsampling has better feature preservation capabilities than fixed-size downsampling, retaining more detailed information in areas of high point cloud density and avoiding information loss due to over-downsampling in areas of low point cloud density. The output after downsampling is a preprocessed point cloud dataset. This serves as input data for subsequent processing steps. In typical applications, the amount of downsampled point cloud data is usually 10% to 20% of the original data, significantly reducing the computational burden of subsequent processing while maintaining the integrity of the building's structural features.

[0046] Step S2 is the automatic identification step for building outline and structural features, such as... Figure 3 As shown.

[0047] In this step, a local geometric feature descriptor is first calculated for each point in the preprocessed point cloud dataset. The local geometric feature descriptor includes three dimensions: normal vector, curvature, and linearity, which can effectively characterize the local geometric properties of the point cloud. Taking the target point... Construct a center with a radius of spherical neighborhood , The value ranges from 0.3m to 0.5m. The covariance matrix is ​​calculated for a point set within a spherical neighborhood. :

[0048] ,

[0049] in, For point The set of points within the neighborhood covariance matrix The number of points in the neighborhood. For the neighboring region The coordinate vector of each point The centroid coordinate vector of all points in the neighborhood, with superscript... This represents the vector transpose. For the covariance matrix... Perform eigenvalue decomposition to obtain three eigenvalues and its corresponding eigenvectors Normal vector Take the corresponding smallest eigenvalue eigenvectors This indicates the direction of the normal to the surface at that point. Curvature Calculated based on the ratio of the smallest eigenvalue to the sum of the three eigenvalues:

[0050] ,

[0051] in, For point The curvature value ranges from 0 to 1 / 3. A smaller curvature value indicates a flatter surface at that point, while a larger curvature value indicates a more pronounced curvature or that the point is an edge feature. Linearity Calculated based on the ratio of the difference between the largest and second-largest eigenvalues ​​to the largest eigenvalue:

[0052] ,

[0053] in, For point The linearity value ranges from 0 to 1. A larger linearity value indicates that the point cloud distribution at that point is closer to linear features such as edges or ridges, while a smaller linearity value indicates that the point cloud distribution at that point is closer to planar features.

[0054] After calculating the local geometric feature descriptors, a segmentation method based on a combination of region growing and deep neural networks is used to perform semantic segmentation on the preprocessed point cloud dataset. The region growing algorithm performs initial segmentation based on normal vector consistency constraints, starting from the seed point and segmenting points where the angle between normal vectors is less than a threshold. And the curvature difference is less than the threshold. Adjacent points are grouped into the same region. The value ranges from 15 degrees to 30 degrees. The value ranges from 0.05 to 0.1. The region growth terminates when the angle or curvature difference between the normal vectors of the points to be expanded exceeds a threshold, resulting in several candidate regions. The formula for calculating the angle between normal vectors is:

[0055] ,

[0056] in, For point With point The angle between the normal vectors, and The normal vectors of the two points are respectively. This indicates that the absolute value is taken to eliminate the ambiguity of the normal vector direction.

[0057] Point clouds of each candidate region are input into a pre-trained point cloud semantic segmentation neural network for classification and recognition. The point cloud semantic segmentation neural network employs an encoder-decoder architecture based on an attention mechanism. The encoder extracts local and global features of the point cloud level by level through multi-layer ensemble abstraction modules, while the decoder recovers point-level semantic predictions through a feature propagation module. The network input is a 6-dimensional feature vector consisting of the 3D coordinates and local geometric feature descriptors of each point within the candidate region. The output is the probability distribution of each point belonging to each semantic category. The semantic categories include five main categories: walls, ground, roof, columns, and beams, covering the main structural components of existing buildings. The formula for calculating the semantic prediction confidence is:

[0058] ,

[0059] in, For point The semantic prediction confidence, For a set of semantic categories, For point Category The posterior probability, For point The input feature vector is used. Points with a semantic prediction confidence score higher than 0.8 are retained as high-confidence segmentation results, while low-confidence points are corrected for category using a neighborhood voting method.

[0060] After semantic segmentation, the building's main outline point cloud and structural boundary feature points are extracted. The building's main outline point cloud is composed of point clouds with semantic categories of walls, ground, and roof, representing the building's external envelope. Structural boundary feature points are those with linearity values ​​greater than a threshold. point, Values ​​ranging from 0.6 to 0.8 are used for these points located at the edges, ridges, and junctions of building components, serving as crucial data for subsequent geometric parameter extraction. Boundary feature point extraction also requires satisfying curvature constraints. , The value is set to 0.1 to 0.15 to filter out false boundary points caused by noise in flat areas.

[0061] The output of step S2 includes two types of data: building outline point cloud and building structure boundary feature points. The number of points in the building outline point cloud is typically 60% to 80% of the preprocessed point cloud dataset, representing the spatial distribution of the building's external envelope. The number of points in the building structure boundary feature points is typically 5% to 15% of the preprocessed point cloud dataset, concentrated at the geometric boundaries of building components. These two types of data serve as input to step S3, used for subsequent floor segmentation detection and door / window opening recognition. Step S2 and step S1 form a data coupling relationship; the quality of the preprocessed point cloud dataset output by step S1 directly affects the feature extraction accuracy of step S2. Step S2 and step S3 form a feature transfer relationship; the completeness of the building outline point cloud determines the reliability of floor segmentation detection, and the accuracy of the building structure boundary feature points determines the accuracy of door / window opening recognition.

[0062] Step S3 is the collaborative detection step for floor division and door / window opening hierarchy, such as... Figure 4 As shown.

[0063] In this step, a height distribution histogram is first constructed based on the point cloud of the building's main outline along the vertical direction. The point cloud of the building's main outline is then divided into several height intervals along the vertical direction (Z-axis), with intervals between the height intervals. The height range is set to 0.1m to 0.3m, and the number of points within each height interval is counted to form the original histogram distribution. ,in This represents the total number of height intervals. For the first The number of points within each height interval. Gaussian smoothing is applied to the original histogram distribution to eliminate local fluctuations; the smoothing kernel size is set to 3 to 5 intervals. The Gaussian smoothing calculation formula is:

[0064] ,

[0065] in, For the smoothed first The number of points in each height interval. To smooth the kernel half-width, The standard deviation is Gaussian kernel function value, The value ranges from 1.0 to 1.5.

[0066] Peak detection and valley analysis are performed on the smoothed histogram distribution to identify floor division locations. Peak locations correspond to the height of the floor slab, where the point cloud density is significantly increased due to the presence of the floor slab's horizontal plane; valley locations correspond to the spaces between floors, where the point cloud density is significantly reduced due to the lack of horizontal structure. Peak detection uses a local maximum search method, and valley detection uses a local minimum search method, with the detection window size set to 5 to 9 intervals. The criteria for determining floor division locations are as follows:

[0067] ,

[0068] in, This represents the number of points at the peak. This represents the number of points at adjacent valley values. The significance threshold for separation is set, ranging from 0.3 to 0.5. The location of the valley value that meets the judgment criteria is determined as the floor separation height. , This refers to the number of floors.

[0069] After determining the floor division locations, the point cloud of the building's main outline is divided into several floor areas according to the division height. Planar detection and edge extraction are performed on the point cloud of each floor area to identify candidate areas for door and window openings. Planar detection uses the RANSAC algorithm to extract the main wall surface of the floor facade. The number of RANSAC iterations is set to 1000 to 5000, and the distance threshold is set to 0.02m to 0.05m. After the main wall surface is extracted, the point density within each grid cell on the wall surface is calculated. Grid cells with significantly lower point densities than the surrounding areas are marked as candidate opening areas. The point density calculation formula is:

[0070] ,

[0071] in, grid coordinates Point density at that location, This represents the number of points within the grid cell. Let be the area of ​​the grid cell. The criteria for determining candidate areas for openings are:

[0072] ,

[0073] in, The average point density of the main wall surface. The density ratio threshold ranges from 0.1 to 0.3.

[0074] Edge detection is performed on candidate opening regions to obtain the opening boundary contours. Edge detection uses the Canny edge detection algorithm to extract edges from the point density image, and then Hough transform is used to detect straight line segments within the edges. The rectangular region enclosed by these straight line segments is taken as the opening boundary contour. Door and window openings are distinguished based on the geometric features of the boundary contours: the lower boundary of a door opening is flush with the floor level, and its height is typically greater than 1.8m; the lower boundary of a window opening is at the height of a windowsill, typically 0.8m to 1.2m above the floor level. The criteria for determining the type of door and window opening are as follows:

[0075] ,

[0076] in, The height of the lower boundary of the opening. The floor height The tolerance for the doorway opening is set at 0.1m. The lower limit for the windowsill height is set at 0.6m.

[0077] A joint screening method combining geometric constraints and semantic confidence was used to determine the precise location and dimensional parameters of door and window openings. Geometric constraints included: opening width ranging from 0.6m to 3.0m, opening height ranging from 0.8m to 3.5m, and opening aspect ratio ranging from 0.3 to 2.5. Semantic confidence was obtained by calculating the semantic segmentation consistency of the opening boundary points; the semantic consistency score was the proportion of boundary points belonging to the wall category. The joint screening criteria were as follows:

[0078] ,

[0079] in, The overall score for the tunnel entrance, Scoring the geometric constraint satisfaction. Score semantic consistency. and The geometric and semantic weights are 0.6 and 0.4 respectively. Candidate regions with a comprehensive score higher than the threshold of 0.7 are identified as valid door and window openings, and their location coordinates and size parameters (width) are extracted. ,high Height above the ground ).

[0080] Establish topological relationships between adjacent floors. These relationships include: vertical connections between floors (adjacent vertically), alignment of door and window openings on the facade (vertical and horizontal alignment), and dependency relationships between components (doors and windows belong to walls). Topological relationships are represented by a graph structure, where nodes represent each floor and its contained components, and edges represent the relationships between nodes. The topological relationship graph is used for establishing relationships between subsequent BIM component models and for transferring attribute information.

[0081] Step S4 is the parametric model matching and conversion step based on the BIM component template library, such as... Figure 5 As shown.

[0082] In this step, candidate component templates that match the geometric features of the building structure boundary feature points are first retrieved from a pre-built parametric template library for building components. This library contains parametric model definitions for common building component types, including walls, floors, roofs, columns, beams, doors, and windows. Each component template definition includes: a component type identifier, a list of geometric parameters (length, width, height, thickness, etc.), geometric parameter constraints (parameter value range, constraint relationships between parameters), parametric geometric representation (geometric shape generation rules based on parameters), and attribute field definitions (material, purpose, grade, etc.). Template retrieval uses a combination of component type matching and geometric parameter range filtering. The input is the component type and geometric parameters extracted in steps S2 and S3, and the output is a set of candidate component templates that meet the type matching and whose parameters are within the constraint range. .

[0083] The parametric template library for building components adopts a hierarchical classification system. The first layer comprises major component categories: enclosing structural components, load-bearing structural components, and ancillary facility components. The second layer categorizes components by type: enclosing structural components include exterior walls, interior walls, and curtain walls; load-bearing structural components include columns, beams, slabs, and foundations; and ancillary facility components include doors, windows, and stairs. The third layer categorizes components by subtype, further subdivided according to material, form, and function. For example, the wall type includes brick walls, concrete walls, and block walls; the door type includes hinged doors, sliding doors, and folding doors. The template library contains no fewer than 200 pre-installed component templates, covering common major structural and enclosing component forms found in existing buildings.

[0084] The parametric geometric representation of component templates employs a combination of constraint solving and solid modeling. Constraint solving establishes geometric constraints between parameters, enabling parameter linkage and automatic updates. For example, in wall components, there are constraints between wall length, wall width, and the positions of door and window openings; when the wall length parameter changes, the opening position parameter automatically adjusts proportionally. Solid modeling uses a solid geometry construction method to generate the 3D geometric shape of the component. Basic operations include extrusion, rotation, and Boolean operations. Taking a window component as an example, its parametric geometric representation rules are as follows: based on the window frame width, window frame height, window frame thickness, and the number of window sashes, an extruded outer frame solid is first generated; then, a separator strip solid is generated based on the number of window sashes; finally, a Boolean union operation is performed to obtain the complete window frame geometry.

[0085] Calculate the matching score between the candidate component template and the extracted geometric features. The matching score consists of two parts: a geometric distance score and a semantic consistency score. The geometric distance score is calculated by determining the normalized Euclidean distance between the geometric parameter vector of the candidate component template and the extracted geometric parameter vector.

[0086] ,

[0087] in, Scoring for geometric distance, Let be the geometric parameter vector of the candidate component template. For the extracted geometric parameter vector, This represents the Euclidean norm of a vector. The semantic consistency score calculates the consistency between the semantic category of the candidate component template and the identified semantic category.

[0088] ,

[0089] in, Score semantic consistency. The semantic category of the candidate component template. The semantic category to be identified. The matching score is calculated using the following formula:

[0090] ,

[0091] in, Score the match rate. and The geometric and semantic weights are 0.7 and 0.3 respectively. The component template with the highest matching score is selected as the target template. .

[0092] Based on the extracted geometric parameters, parametric instantiation is performed on the target template to generate a BIM component model. Parametric instantiation assigns the extracted geometric parameter values ​​to the corresponding parameter variables of the target template, generating specific geometric shapes through parametric geometric expression rules. Taking a wall component as an example, the parametric geometric expression rule is: based on the wall's start-point coordinates, end-point coordinates, height, and thickness parameters, a cuboid geometric shape is generated, extending half a thickness to each side and upwards by a height value, with the line connecting the start and end points as the central axis. The parametric instantiation process outputs the BIM component model. Each BIM component model contains geometry, spatial location, component type, and parameter values.

[0093] After the BIM component model is generated, the relationships between components are established. Based on the topological relationships established in step S3, corresponding spatial and subordinate relationships are established between the BIM component models. Spatial relationships include adjacency relationships (shared boundaries), containment relationships (doors and windows are contained within walls), and support relationships (walls support floor slabs), etc. Relationship establishment is achieved by calculating the spatial topology between component geometries, and the judgment methods include bounding box intersection detection, boundary surface coplanarity detection, and distance threshold judgment. Establishing relationships between components is a key step in achieving the integrity of the BIM model, supporting subsequent applications such as structural analysis, clash detection, and quantity surveying.

[0094] A parameter transfer relationship is established between steps S4 and S3, with the door and window opening position and size parameters extracted in step S3 serving as input for door and window component template matching. A closed-loop feedback relationship is established between steps S4 and S6, where the geometric deviation detected in step S6 can drive step S4 to re-execute the parameter matching and instantiation process, achieving iterative optimization of modeling accuracy. This closed-loop mechanism is one of the core innovations of this invention, distinguishing it from existing technologies by improving modeling efficiency through parameter correction rather than model reconstruction.

[0095] Step S5 is a local correction and attribute supplementation step driven by human-computer interaction, such as... Figure 6 As shown.

[0096] In this step, the generated BIM component model is first overlaid with the original point cloud data through a visual interactive interface. The visual interactive interface is implemented using a 3D graphics rendering engine, supporting view rotation, zoom, and panning operations, as well as layered display and transparency adjustment of the BIM component model and point cloud data. The interface provides two working modes: model browsing mode and editing mode. In model browsing mode, users can view the overlay effect of the BIM component model and point cloud and identify areas that need correction; in editing mode, users can select specific components and perform correction operations.

[0097] The system receives user commands to correct the position, size, and shape of components and executes corresponding model adjustments. Correction operation types include: position translation (moving components along the X, Y, and Z axes), size adjustment (modifying the length, width, height, or thickness parameters of components), shape modification (adjusting the geometric topology of components), component deletion (removing incorrectly identified components), and component addition (manually adding missing components). Users input correction commands through the interactive interface controls, and the system updates the BIM component model in real time and re-renders it. Correction operations employ an incremental update strategy, recalculating only the affected components and their related components to avoid the performance overhead of global reconstruction.

[0098] The system receives component attribute information input by the user and writes it into the corresponding BIM component model. Attribute information includes: component name, material type, function, construction year, maintenance status, and ownership information. Attribute input is achieved through the attribute editing panel, which displays a list of attribute fields for the selected component. Users can edit the attribute values ​​of each field. Attribute values ​​support data types such as text, numeric, date, and enumeration. After the attribute information is written, it is stored together with the component's geometric information to form complete BIM component data.

[0099] Generate a 3D information model of the existing building containing geometric and attribute information. The existing building 3D information model is stored and exchanged in IFC format, the international standard data format for Building Information Modeling, which supports data interoperability between different BIM software. During model export, the BIM component model is converted into corresponding IFC entities according to the IFC data model, the relationships between entities are defined, and a model file conforming to IFC specifications is generated.

[0100] The IFC entity mapping rules are as follows: wall components are mapped to IfcWall entities, floor slab components to IfcSlab entities, roof components to IfcRoof entities, column components to IfcColumn entities, beam components to IfcBeam entities, door components to IfcDoor entities, and window components to IfcWindow entities. Each IFC entity contains three main parts: geometric representation (IfcProductDefinitionShape), spatial location (IfcLocalPlacement), and attribute set (IfcPropertySet). The geometric representation uses boundary representation (B-Rep) or sweptSolid to describe the three-dimensional shape of the component. Spatial location uses a local coordinate system to describe the position and orientation of the component relative to the building origin. The attribute set contains the component's non-geometric attribute information, such as material, purpose, and construction date.

[0101] A model receiving relationship is established between step S5 and step S4, with the BIM component model output by step S4 serving as the object for visualization and editing in step S5. A model providing relationship is established between step S5 and step S6, with the existing building 3D information model output by step S5 serving as the object for quality verification in step S6. Human-computer interaction correction can quickly correct identification errors and parameter deviations in the automatic modeling process, and is a key step in ensuring the usability of the final model.

[0102] Step S6 is the closed-loop quality verification and model consistency calibration step, such as... Figure 7 As shown.

[0103] In this step, the geometric deviation distribution between the existing building's 3D information model and the preprocessed point cloud dataset is first calculated. The surface of the existing building's 3D information model is then uniformly sampled to obtain a model sampling point set. The sampling density was set to 100 to 500 points per square meter. For each point in the model sampling point set... Calculate the distance to the nearest point in the preprocessed point cloud dataset as the point-level deviation. :

[0104] ,

[0105] in, For model sampling points Point-level deviation, To preprocess the point cloud dataset, For points in a point cloud, This represents the Euclidean distance. Point-level deviation calculation uses a KD-tree data structure for nearest neighbor lookup to improve computational efficiency.

[0106] The mean and standard deviation of all point-level deviations are used as measures of the geometric deviation distribution.

[0107] ,

[0108] ,

[0109] in, The mean of point-level deviations. The standard deviation is the point-level deviation. This represents the number of sampling points for the model. The geometric deviation distribution can also be visualized using deviation histograms and deviation heatmaps to help users locate areas with significant deviations.

[0110] When the geometric deviation exceeds a preset threshold, a deviation calibration command is generated and fed back to the parametric model matching and transformation steps to perform parameter correction. The preset threshold is set as: the mean of point-level deviations. Not exceeding 0.02m, point-level deviation standard deviation Not exceeding 0.03m. The criteria for determining if the deviation exceeds the limit are:

[0111] ,

[0112] in, The mean threshold is set to 0.02m. The standard deviation threshold is 0.03m. This represents a logical OR operation. When the deviation exceeds the limit condition, the deviation distribution is analyzed to identify components with large deviations, and parameter correction instructions are generated for those components. The parameter correction instructions include: target component identifier, correction parameter name, correction direction, and correction amount. The correction amount is calculated based on the deviation distribution, and the least squares optimization method is used to solve for the parameter adjustment value that minimizes the deviation. The objective function for parameter correction optimization is:

[0113] ,

[0114] in, This is the optimal parameter correction vector. For parameter correction vector, For the model sampling points after parameter correction Point-level deviation.

[0115] After the deviation calibration command is fed back to step S4, parameter correction is performed on the BIM model of the target component, the component geometry is updated, and the geometric deviation calculation in step S6 is re-executed. This process is iterative until the geometric deviation meets the preset accuracy requirements or the maximum number of iterations is reached. The preset accuracy requirements are set as follows: the overall fitting error between the model and the point cloud does not exceed 0.05m, and the error in the dimensions of door and window openings does not exceed 0.02m. The maximum number of iterations is set to 5 to 10. After the iterations converge, the final 3D information model of the existing building is output.

[0116] The closed-loop quality verification mechanism is one of the core innovations of this invention, enabling controllable assurance of modeling accuracy. Unlike the unidirectional verification method used in existing technologies, this invention establishes a bidirectional information flow between the model and the point cloud: the forward flow extracts geometric features from the point cloud data and generates a BIM model, while the reverse flow extracts parameter correction information from model deviation analysis and feeds it back to the model generation stage. This closed-loop feedback mechanism can automatically identify and locate components with large deviations, and make targeted parameter adjustments, avoiding the computational overhead of global reconstruction and improving the efficiency and accuracy of the modeling process. The closed-loop mechanism also supports multi-round iterative optimization, continuously improving the geometric accuracy of the model through gradual approximation, ultimately meeting the preset accuracy requirements.

[0117] In one specific embodiment, the method of this invention was used to rapidly model the current state of a six-story existing residential building in an urban renewal project. The building has a total height of approximately 20 meters, a floor area of ​​approximately 3,000 square meters, was built in the 1990s, and uses a brick-concrete structure. Due to the lack of original design drawings, a BIM model needs to be reconstructed using point cloud data for subsequent renovation design.

[0118] During the data acquisition phase, a terrestrial 3D laser scanner (Leica RTC360 model) was used to scan eight sites, covering approximately 200 million points. Simultaneously, four UAV oblique photogrammetry flights were conducted at an altitude of 80 meters, reconstructing a point cloud of approximately 80 million points. The total data acquisition time was approximately six hours, including equipment setup, scanning execution, and data transmission.

[0119] In the multi-source data fusion processing stage, a seven-parameter coordinate transformation was first performed to convert the UAV point cloud to the ground scan point cloud coordinate system, with a transformation error of 0.008m. Then, ICP fine registration was performed, with a root mean square error of 0.004m. Noise removal eliminated approximately 8 million outliers, accounting for 2.9% of the original point cloud. Adaptive downsampling reduced the number of point clouds from 280 million to 30 million, a compression ratio of 89.3%.

[0120] In the building outline and structural feature recognition stage, the accuracy rate of wall recognition using deep neural network semantic segmentation reached 96.2%, ground recognition 98.1%, roof recognition 94.7%, column recognition 91.3%, and beam recognition 89.8%. The number of boundary feature points extracted was 4.2 million, accounting for 14.0% of the preprocessed point cloud.

[0121] During the floor separation detection phase, the height distribution histogram clearly showed six peaks corresponding to the locations of the six floor slabs, with peak-to-valley ratios all greater than 0.45, meeting the separation significance threshold requirement. The detected floor heights were 0.0m, 3.2m, 6.4m, 9.6m, 12.8m, 16.0m, and 19.2m (roof), with errors less than 0.05m compared to the actual on-site measurements. A total of 36 door openings and 72 window openings were identified during the door and window opening detection phase, achieving an accuracy rate of 92.6% after manual verification. Missed detections mainly occurred in the obscured rear facade area.

[0122] In the stage of automatically generating component models based on the BIM component template library, a total of 48 exterior wall component templates, 7 floor slab component templates, 36 door component templates (including 28 standard casement doors and 8 sliding doors), and 72 window component templates (including 24 fixed windows, 32 sliding windows, and 16 casement windows) were matched. The parametric instantiation process took approximately 15 minutes and generated 163 BIM component models.

[0123] During the human-computer interaction correction phase, operators reviewed the automatic modeling results through a visual interface, correcting 12 component positions, 8 component dimensions, deleting 3 falsely detected components, and adding 5 missing components. The total interactive operation time was approximately 2 hours. The attribute supplementation phase involved entering attribute information such as material type, usage function, and maintenance status for each component. The total attribute entry time was approximately 1 hour.

[0124] During the closed-loop quality verification phase, the overall fitting error in the initial verification was 0.042m, and the standard deviation was 0.051m, exceeding the preset threshold. The system automatically generated parameter correction instructions, adjusting the parameters of 6 exterior wall components and 2 door components with large deviations. After the second iteration, the overall fitting error decreased to 0.035m, and the standard deviation decreased to 0.028m, meeting the preset accuracy requirements. The total iteration process took approximately 20 minutes.

[0125] According to comprehensive statistics, the total time required to complete the BIM modeling of the existing building using the method of this invention is approximately 24 hours, including 6 hours for data acquisition, 8 hours for data processing, 0.5 hours for automatic modeling, 3 hours for human-computer interaction, 0.5 hours for quality verification, and 6 hours for results processing. Compared to the traditional manual modeling method, which requires 15 working days (120 hours), the modeling efficiency of the method of this invention is improved by approximately 80%, while the modeling accuracy reaches 0.035m, which is superior to the accuracy level of 0.05m to 0.10m of the traditional method.

[0126] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A rapid modeling method for existing buildings based on the fusion of point cloud and BIM, characterized in that, Includes the following steps: The process involves acquiring and preprocessing multi-source heterogeneous point cloud data. The original point cloud data of the existing building is acquired through a scanning device. Coordinate transformation and registration are performed on the original point cloud data to obtain a global point cloud dataset. Noise removal and downsampling processing are then performed on the global point cloud dataset to obtain a preprocessed point cloud dataset. The automatic identification steps for building outline and structural features are performed. For each point in the preprocessed point cloud dataset, a local geometric feature descriptor is calculated. A segmentation method combining region growing and deep neural networks is used to perform semantic segmentation processing to extract the building outline point cloud and the building structure boundary feature points. The process involves performing a collaborative detection step for floor division and door / window openings. Based on the point cloud of the building's main outline, a height distribution histogram is constructed along the vertical direction to identify the floor division location. Planar detection and edge extraction are performed on each floor area to identify candidate areas for door / window openings. Perform parametric model matching and conversion steps based on the BIM component template library, retrieve candidate component templates from the building component parametric template library and calculate the matching degree score, and perform parametric instantiation processing on the target template according to the extracted geometric parameters to generate a BIM component model; The system performs local correction and attribute supplementation steps driven by human-computer interaction, displays the BIM component model and point cloud data overlaid through a visual interface, receives user correction commands to adjust the model, and receives user attribute information to generate a three-dimensional information model of the existing building. The closed-loop quality verification and model consistency calibration steps are performed to calculate the geometric deviation distribution between the existing building 3D information model and the preprocessed point cloud dataset. When the geometric deviation exceeds the preset threshold, it is fed back to the parameterized model matching and conversion steps to perform parameter correction. The process is iterated until the preset accuracy requirements are met, and then the final model is output.

2. The method for rapid modeling of existing buildings based on point cloud and BIM fusion as described in claim 1, characterized in that, The noise removal process based on the combination of statistical filtering and radius filtering includes: setting the threshold for the number of neighborhood points for statistical filtering to 20 to 50 points, setting the threshold for the standard deviation multiple of statistical filtering to 1.0 to 2.0, setting the search radius for radius filtering to 0.05m to 0.15m, and setting the minimum number of neighborhood points for radius filtering to 5 to 10 points.

3. The method for rapid modeling of existing buildings based on point cloud and BIM fusion according to claim 1, characterized in that, In the aforementioned voxel-based adaptive downsampling process, the voxel grid size is adaptively set according to the original point cloud density. When the original point cloud density is greater than 10,000 points per cubic meter, the voxel grid size is set to 0.02m to 0.05m. When the original point cloud density is 1,000 to 10,000 points per cubic meter, the voxel grid size is set to 0.05m to 0.10m.

4. The method for rapid modeling of existing buildings based on point cloud and BIM fusion according to claim 1, characterized in that, The calculation method of the local geometric feature descriptor includes: constructing a spherical neighborhood with a radius of 0.3m to 0.5m centered on the target point; calculating the covariance matrix of the point set within the spherical neighborhood and performing eigenvalue decomposition to obtain three eigenvalues; calculating the eigenvector of the corresponding minimum eigenvalue based on the three eigenvalues; calculating the curvature based on the ratio of the minimum eigenvalue to the sum of the three eigenvalues; and calculating the linearity based on the ratio of the difference between the maximum and the second largest eigenvalue to the maximum eigenvalue.

5. The method for rapid modeling of existing buildings based on point cloud and BIM fusion according to claim 1, characterized in that, The segmentation method based on the combination of region growing and deep neural network includes: using a region growing algorithm based on normal vector consistency constraints to perform preliminary segmentation on the preprocessed point cloud dataset to obtain several candidate regions; inputting the point cloud of each candidate region into a pre-trained point cloud semantic segmentation neural network for classification and recognition; the output categories of the point cloud semantic segmentation neural network include wall, ground, roof, column and beam.

6. The method for rapid modeling of existing buildings based on point cloud and BIM fusion according to claim 1, characterized in that, The method for constructing the height distribution histogram includes: dividing the point cloud of the building's main outline into several height intervals with a height interval of 0.1m to 0.3m along the vertical direction; counting the number of points in each height interval to form a histogram distribution; performing Gaussian smoothing on the histogram distribution to eliminate local fluctuations; and identifying the valley positions in the histogram distribution where the number of points decreases significantly as candidate positions for floor separation.

7. The method for rapid modeling of existing buildings based on point cloud and BIM fusion according to claim 1, characterized in that, The method for identifying candidate areas for door and window openings includes: performing RANSAC planar detection on the point cloud of the facade of each floor area to obtain the main wall surface; extracting areas on the main wall surface with significantly lower point density than the surrounding areas as candidate areas for openings; performing edge detection on the candidate areas for openings to obtain the boundary contours of the openings; and distinguishing door openings from window openings based on the geometric features of the boundary contours.

8. The method for rapid modeling of existing buildings based on point cloud and BIM fusion according to claim 1, characterized in that, The method for calculating the matching score between the candidate component template and the extracted geometric features includes: calculating the Euclidean distance between the geometric parameters of the candidate component template and the extracted geometric parameters as a geometric distance score; calculating the consistency between the semantic category of the candidate component template and the identified semantic category as a semantic consistency score; and weighted summing the geometric distance score and the semantic consistency score to obtain the matching score.

9. The method for rapid modeling of existing buildings based on point cloud and BIM fusion according to claim 1, characterized in that, The method for calculating the geometric deviation distribution includes: uniformly sampling the surface of the existing building's three-dimensional information model to obtain a model sampling point set; calculating the distance from each point in the model sampling point set to the nearest point in the preprocessed point cloud dataset as a point-level deviation; and statistically analyzing the mean and standard deviation of all point-level deviations as a metric for the geometric deviation distribution.

10. The method for rapid modeling of existing buildings based on point cloud and BIM fusion according to claim 1, characterized in that, The preset thresholds are set as follows: the average point-level deviation is no more than 0.02m, and the standard deviation of the point-level deviation is no more than 0.03m; the preset accuracy requirements are set as follows: the overall fitting error between the model and the point cloud is no more than 0.05m, and the size error of the door and window openings is no more than 0.02m.

Images