Indoor bim automatic reverse modeling method based on point cloud semantic analysis

Abstract

The application discloses an indoor BIM automatic inverse modeling method based on point cloud semantic analysis, relates to the technical field of building information modeling, and comprises the following steps: acquiring a quality score through local point cloud density estimation and implementing adaptive voxel downsampling, constructing a multi-level progressive registration framework to obtain point cloud data under a unified coordinate system; inputting the point cloud data into a multi-scale feature fusion instance segmentation network, combining an attention mechanism guided by building prior and a visual language model multi-modal semantic enhancement, and realizing accurate identification of component instances; comprehensively adopting a multi-constraint fusion algorithm to extract geometric parameters and inputting the geometric parameters into a BIM platform for preliminary parameterized reconstruction; establishing a quality evaluation system covering multi-dimensional indexes to calculate fitting errors, triggering a closed-loop iterative feedback mechanism according to deviations to dynamically adjust front-end parameters until the accuracy requirements are met. The application effectively solves the problems of data redundancy and occlusion analysis, realizes high-precision reconstruction and topological consistency in a complex scene, and significantly improves the modeling efficiency.

Description

Technical Field

[0001] This invention relates to the field of building information modeling, and specifically to an automatic inverse modeling method for indoor BIM based on point cloud semantic parsing. Background Technology

[0002] With the widespread application of Building Information Modeling (BIM) technology, the digital reconstruction of existing buildings has become a key aspect of urban renewal and smart operation. 3D laser scanning technology, by acquiring high-density indoor spatial point cloud data, provides core data support for constructing high-precision digital twin models. This process requires the system to efficiently process massive amounts of spatial point cloud information and accurately extract building components with rich semantic information, thereby achieving precise mapping and efficient transformation of physical space into a parametric BIM model.

[0003] Among them, the automatic inverse modeling method based on point cloud semantic parsing integrates deep learning, feature aggregation, and geometric inference algorithms to solve the problem of automatic mapping of point cloud data to structured components. This technical approach typically involves core steps such as spatial point cloud preprocessing, multi-scale instance segmentation, and iterative fitting of geometric parameters. An ideal inverse modeling process should have adaptive data processing capabilities, capable of automatically identifying and reconstructing building component models that conform to industry standards and topological constraints while preserving key spatial geometric details.

[0004] However, existing technologies generally face technical bottlenecks when dealing with complex indoor scenes, including severe redundancy in raw point cloud data and incomplete semantic information parsing. This means that the modeling process still largely relies on manual intervention. Traditional preprocessing methods often use fixed parameter configurations, making it difficult to dynamically adjust processing strategies based on point cloud quality. Furthermore, the registration, segmentation, and modeling modules operate in an independent open-loop state, lacking a global feedback loop. Simultaneously, existing segmentation models have insufficient generalization ability when dealing with non-Manhattan structures or severely occluded scenes, resulting in low accuracy in fine-grained component recognition and biases in geometric parameter extraction. Moreover, the lack of a systematic quality evaluation index system and a full-process feedback mechanism makes it difficult to effectively guarantee the topological integrity and geometric consistency of the modeling results. Summary of the Invention

[0005] The present invention provides an automatic inverse modeling method for indoor BIM based on point cloud semantic parsing, which can solve the above-mentioned problems.

[0006] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0007] An automatic inverse modeling method for indoor BIM based on point cloud semantic parsing includes the following steps:

[0008] Step 1: Obtain a quality score through local point cloud density estimation, perform adaptive voxel downsampling based on the quality score, and construct a multi-level progressive registration framework from the main axis of the building structure to planar features and then to the key points of the components to obtain point cloud data in a unified coordinate system.

[0009] Step 2: Input the point cloud data under the unified coordinate system into the multi-scale feature fusion instance segmentation network, extract the local detail features and global context features of the components, and enhance the multimodal semantics by introducing the building prior-guided attention mechanism and the spatial hierarchical instance clustering module, combined with the visual language model to perform multimodal semantic enhancement on the auxiliary annotation information of the two-dimensional projection image, so as to realize the recognition of indoor building components and obtain component instances.

[0010] Step 3: For the component instance, the geometric parameters are extracted by comprehensively using the random sampling consistency plane fitting algorithm, principal component analysis method and region growing algorithm. The geometric parameters are then input into the building information modeling platform according to the topological constraints of the indoor building components to realize the preliminary parametric reconstruction of the indoor building components and generate a preliminary digital model.

[0011] Step 4: Establish a multi-dimensional quality evaluation system covering geometric accuracy indicators, semantic integrity indicators, topological consistency indicators, and efficiency indicators. Calculate the fitting error between the preliminary digitization model and the original scanned point cloud using this multi-dimensional quality evaluation system. Trigger an iterative refinement process based on the deviation between the evaluation results and a preset evaluation threshold. Iterate through the error backpropagation mechanism to the multi-level progressive registration framework, the multi-scale feature fusion instance segmentation network, or the preliminary parametric reconstruction feedback adjustment instruction until the fitting error meets the set requirements.

[0012] This invention upgrades the traditional unidirectional, open-loop modeling process into a data-driven feedback closed-loop system by implementing quality-aware dynamic preprocessing and multi-level progressive registration, combined with multimodal enhanced instance segmentation and multi-constraint fusion geometric extraction, and finally establishing a full-process multi-dimensional closed-loop feedback mechanism. This method can automatically identify and reconstruct building component models that conform to industry standards and topological constraints while preserving key spatial geometric details, significantly reducing the proportion of manual intervention and improving the modeling efficiency and geometric reconstruction accuracy of digital transformation of complex interior scenes.

[0013] Specifically, in step 1, the method for obtaining the quality score includes: constructing a spatial index structure to realize neighborhood query of the original point cloud data; counting the number of sampling points in the neighborhood of each sampling point within a preset search radius; and using a Gaussian kernel to weight the contribution of the neighborhood points to obtain the local density; measuring the local geometric complexity by analyzing the distribution entropy of the sampling points in the neighborhood; and calculating the quality score by combining the number of sampling points with the eigenvalue distribution terms of the local covariance matrix.

[0014] In step 1, the method for implementing adaptive voxel downsampling based on the quality score includes:

[0015] Based on a lookup table containing multi-level voxel mesh sizes, the quality score is compared with a preset quality threshold.

[0016] If the quality score is higher than the preset quality threshold, the corresponding area is determined to be a component edge area or corner area, and the voxel mesh is set to the first preset size for downsampling.

[0017] If the quality score is lower than the preset quality threshold, the corresponding area is determined to be a flat area, and the voxel grid is set to a second preset size for downsampling, wherein the second preset size is greater than the first preset size.

[0018] This invention introduces a quality-aware dynamic preprocessing mechanism, which obtains a quality score by calculating local point cloud density and distribution entropy, and implements adaptive voxel downsampling. While preserving key geometric details such as component edges and intersections, it effectively filters out a large amount of redundant data in flat areas, significantly reducing the computational burden on subsequent processes.

[0019] Specifically, the execution process of the multi-level progressive registration framework includes:

[0020] Phase 1: Extract the orientation of the main wall and the ground normal vector using principal component analysis or global normal vector histogram, construct global rotation constraints based on the Manhattan world hypothesis, and achieve initial alignment by searching for matching points in the rotation space;

[0021] The second stage involves identifying planar regions in the initially aligned point cloud using a region growing algorithm or a random sampling consensus algorithm. By calculating the distance and angle deviations between planar regions with the same name in different scanning stations, a cost function that minimizes the distance between the planes is constructed. The transformation matrix is ​​then iteratively optimized using a nonlinear least squares method for fine registration.

[0022] The third stage involves extracting feature descriptors and constructing corresponding point sets for significant geometric feature points, including doorway corners, window frame vertices, and beam-column intersections. Local coordinate fine-tuning is then performed using the local iterative nearest point algorithm.

[0023] This invention constructs a multi-level progressive registration framework, from coarse registration of the global building structure main axis to fine registration based on planar features, and then to local optimization based on key points. It effectively solves the problem that traditional registration methods are prone to getting stuck in local optima in long corridors or repetitive structural scenarios, and enables the overall stitching accuracy of spatial point clouds to reach the millimeter level.

[0024] Specifically, in step 2, the internal execution logic of the multi-scale feature fusion instance segmentation network includes:

[0025] Different preset search radii are set in the feature extraction layer. The first preset search radius is used to capture micro-detail features, and the second preset search radius is used to capture macro-semantic features. A composite feature vector is generated by multi-level feature concatenation and fusion, wherein the first preset search radius is smaller than the second preset search radius.

[0026] The specific execution process of the building prior-guided attention mechanism includes: converting priors such as the orthogonality of the wall normal vector and the gravity direction, the parallelism of the ground normal vector and the gravity direction, and the embedding relationship priors of the door and window point clouds being surrounded by the wall plane into constraint weights in the feature learning process; and enhancing the corresponding attention weights of point cloud clusters whose geometric distribution conforms to the orthogonality prior, parallelism prior, or embedding relationship prior in the attention matrix calculation.

[0027] Step 2 also includes a synthetic data augmentation and domain adaptive transfer learning process, the specific process of which includes:

[0028] Using a standard building information modeling library, a synthetic point cloud dataset with semantic annotations is generated through virtual scanning technology, and the noise distribution, occlusion, and motion blur features of laser scanning are simulated during the generation process.

[0029] Domain adaptive transfer is performed through an adversarially trained network comprising a feature extractor, a component classifier, and a domain discriminator, wherein the domain discriminator is used to distinguish the source of input features, and the feature extractor aligns the feature distribution of the synthetic point cloud data with the feature distribution of the real scanned point cloud data in space through adversarial game with the domain discriminator.

[0030] Specifically, the execution process of the multimodal semantic enhancement and spatial hierarchical instance clustering module includes:

[0031] Multimodal semantic enhancement: Projecting three-dimensional point cloud data onto a two-dimensional plane with multiple preset viewpoints to generate a two-dimensional image containing intensity information, depth information or normal vector information. Using a pre-trained visual language model, the two-dimensional image is semantically annotated, and the two-dimensional semantic labels on the two-dimensional image are back-mapped to the three-dimensional point cloud space using projection geometric transformation relationships.

[0032] Spatial hierarchical clustering: identify horizontal planes to divide floor areas, identify vertical planes to divide room areas, and assign independent instance numbers to point cloud clusters with the same semantic labels and spatial connectivity within each room area.

[0033] Topology reasoning verification: When a door instance or window instance is detected to be spatially isolated from a wall instance, it is merged into the nearest logical cluster of walls through nearest neighbor search, and the geometric position of the door instance or window instance is corrected.

[0034] This invention employs a multi-scale feature fusion instance segmentation network, which can simultaneously take into account the micro-features and macro-topological semantics of components; it accelerates model convergence by introducing an attention mechanism guided by architectural priors; and it greatly overcomes the semantic fragmentation problem caused by data loss and severe occlusion in real indoor scenes by using adversarial training execution domain adaptive transfer learning and visual language model for 2D to 3D multimodal mapping enhancement, thus significantly improving the component classification accuracy.

[0035] Specifically, the strategy for extracting geometric parameters for the component instance in step 3 includes:

[0036] For wall components: the random sampling consistency plane fitting algorithm is used to extract the wall plane equation, the principal component analysis method is combined to determine the wall thickness centerline, and orthogonal constraints and coplanar constraints are introduced. When the included angle between two adjacent walls is within the preset right angle error range, the normal vector of the wall is adjusted to be perpendicular.

[0037] For door and window components: the coordinates of the four corner vertices of the opening are located by calculating the curvature change characteristics at the boundary, and the width, height, thickness and center position parameters are extracted by using a multi-segment line fitting strategy;

[0038] For floor slabs and ceiling components: A region growing algorithm is used to identify connected horizontal planes. The identified enclosing walls are used as horizontal boundary constraints. By calculating the intersection of the wall base and the horizontal plane, the geometry of the occluded area is completed.

[0039] The strategy for extracting geometric parameters for non-Manhattan structural components in the component instance in step 3 includes:

[0040] For inclined wall components: Non-orthogonal wall instances are identified by cluster analysis of the wall surface normal vectors, and the tilt angle parameter is calculated by using the slope of the intersection line of adjacent walls in three-dimensional space;

[0041] For curved components: Curvature analysis of sampling points and feature point detection are performed to extract point cloud data on the arc segment. The least squares optimization algorithm is used to fit the arc equation and extract the center coordinates, radius and initial arc length parameters.

[0042] For irregular curved surface components: a non-uniform rational B-spline surface reconstruction method is adopted to extract the control point network from the point cloud cluster and generate the irregular curved surface model within the building information modeling platform.

[0043] This invention employs a multi-constraint fusion geometric parameter extraction strategy, effectively solving the interference of noisy point clouds on plane fitting and ensuring that the generated rooms meet orthogonal and other building code requirements. In particular, for non-Manhattan structures such as sloping walls, curved components, and complex irregular curved surfaces, it achieves automatic parametric expression of complex irregular components through curvature analysis, least squares optimization, and B-spline surface reconstruction technology, avoiding the dilemma of forced vertical correction and a large amount of manual modeling.

[0044] Specifically, in step 4, the logic of the error backpropagation mechanism feeding back the adjustment command includes:

[0045] If the geometric accuracy index fails to meet the standard, the adjustment instruction will be sent to step 3 to adjust the distance threshold of the random sampling consistency algorithm or replace the high-order fitting model for secondary parameter extraction.

[0046] If the semantic integrity index is abnormal, the adjustment instruction is sent to step 2 to lower the confidence threshold of the multi-scale feature fusion instance segmentation network.

[0047] If the topology consistency index is abnormal, restart the topology reasoning module to perform extension, alignment or merging operations on the logically erroneous component instances and correct the spatial connection relationship between components.

[0048] If the efficiency index is lower than the preset efficiency standard, the adjustment instruction is sent to step 1 to increase the downsampling step size of adaptive voxel downsampling.

[0049] This invention establishes a multi-dimensional quality evaluation system covering geometry, semantics, topology, and efficiency, and transforms the evaluation results into specific feedback adjustment instructions. This automatic iterative refinement mechanism ensures that the generated BIM model and the original point cloud always maintain a high level of fit, and fully comply with building industry standards in terms of visual reproduction and logical topology, effectively reducing the cost of secondary correction after modeling.

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, embodiments of the present invention are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1This is a schematic diagram of the overall technical solution architecture of the indoor BIM automatic inverse modeling method based on point cloud semantic parsing proposed in this invention;

[0053] Figure 2 This is a schematic diagram illustrating the core principle framework of fine-grained instance segmentation and multimodal semantic enhancement in this invention;

[0054] Figure 3 This is a flowchart outlining the main stages of dynamic preprocessing and multi-level progressive registration for quality perception in this invention.

[0055] Figure 4 This is a flowchart outlining the main stages of geometric parameter extraction and parametric modeling for multi-constraint fusion in this invention.

[0056] Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow of the whole-process closed-loop feedback optimization and multi-dimensional quality evaluation in this invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0058] Example 1

[0059] Before providing a detailed description of the embodiments of the present invention, the execution subject of the embodiments of the present invention will first be described. The method provided by the embodiments of the present invention can be executed by an electronic device, which may be a server with powerful computing capabilities, a high-performance computing workstation, a cloud computing cluster, or a terminal device with BIM processing software installed. In the following description of the steps, the electronic device will be referred to as the system or the execution subject.

[0060] This invention proposes an automatic inverse modeling method for indoor building information models based on point cloud semantic parsing, such as... Figure 1 As shown, its core lies in establishing a closed-loop feedback mechanism across the entire chain, from bottom-level raw data processing to high-level semantic recognition and then to parametric model generation, to achieve high-precision digital reconstruction of complex indoor environments. This method ensures that the generated building information model meets the high standards required for engineering applications in geometric, semantic, and topological dimensions by introducing quality-aware preprocessing, multimodal enhanced instance segmentation, multi-constraint fusion geometric extraction, and a four-dimensional quality evaluation system.

[0061] In the above-mentioned automatic inverse modeling method for indoor building information model based on point cloud semantic parsing, step 1, quality-aware dynamic preprocessing and multi-level registration, serves as the data foundation for the entire modeling process and undertakes the key task of extracting high-quality, consistent spatial information from massive, messy original scanned point clouds.

[0062] Specifically, the quality-aware dynamic preprocessing process in step 1 begins with the estimation of the local point cloud density of the original indoor space point cloud.

[0063] Specifically, in combination Figure 3 As shown, the quality-aware dynamic preprocessing flow in step 1 begins with local point cloud density estimation of the original scanned point cloud. For example... Figure 3 The process shown involves the system sequentially estimating local point cloud density, calculating a quality score (including distribution entropy / curvature terms), and performing adaptive voxel downsampling (V1 / V2 size) based on the score. The system achieves fast neighborhood queries on large-scale point cloud data by constructing an efficient spatial index structure, such as a Kd-tree or octree. Within a preset search radius... Within, statistics are collected for each sampling point. Number of points in the neighborhood To obtain a more robust density representation, a density estimation algorithm based on kernel functions is adopted, which uses a Gaussian kernel to weight the contributions of neighboring points, thereby smoothing out local density fluctuations caused by noise.

[0064] After completing the density estimation, the system calculates the quality score for each region based on the point cloud density distribution. The quality score here reflects not only the completeness of the data sampling but also the complexity of the object's geometric features. In areas rich in geometric features, such as wall corners, column edges, door and window openings, and the junction of ceilings and walls, drastic changes in point cloud normal vectors lead to specific local density distribution patterns. The system identifies these key geometric regions by calculating the eigenvalue distribution of the local covariance matrix and assigns them high quality weight coefficients.

[0065] Based on quality score The system implements an adaptive voxel downsampling strategy. Specifically, the system predefines a lookup table containing multi-level voxel mesh sizes. For regions with quality scores higher than a preset quality threshold, i.e., point clouds identified as component edges or corners, the voxel mesh is set to the first preset size. (e.g., 5mm to 10mm) to ensure accurate capture of subtle geometric transitions. For flat, geometrically simple wall or floor areas, the quality score is lower, and the system sets the voxel mesh to a second preset size. (For example, 30mm to 50mm), significantly reducing redundant data while maintaining the flatness of the area. This process effectively balances the trade-off between computational efficiency and feature preservation.

[0066] In step 1 above, the multi-level progressive registration framework is key to achieving unified data from multiple stations. The first stage is coarse registration based on the principal axes of the building structure. The system first extracts the orientation of the main walls and the ground normal in the interior scene through principal component analysis (PCA) or global normal vector histogram. Considering that interior buildings often follow the Manhattan world assumption, i.e., the main structural surfaces are perpendicular or parallel to each other, the system uses these principal direction vectors to construct global rotation constraints. By searching for the best matching point in the rotation space, the initial alignment of the point clouds of different scanning stations is achieved. This stage mainly addresses the problems of large-scale spatial displacement and angular rotation.

[0067] The second stage involves fine-grained registration based on planar features. The system utilizes region growing algorithms or random sample consensus algorithms to identify large planar regions, such as the ground, ceiling, and main load-bearing walls, within the initially aligned point cloud. By calculating the distance and angular deviations between corresponding planes (i.e., the same building surface at different stations), a cost function that minimizes the distance between planes is constructed. The system employs nonlinear least squares methods, such as the Levenberg-Marquardt algorithm, to iteratively optimize the transformation matrix for each station, enabling millimeter-level precise translation and rotation of spatial positions.

[0068] The third stage is local optimization registration based on key component points. For local areas with significant geometric features, such as doorway corners, window frame vertices, and beam-column intersections, the system extracts their feature descriptors. A corresponding point set is constructed using these significant feature points, and coordinate fine-tuning is performed using the Iterative Closest Point (ICP) algorithm. The goal of this stage is to eliminate any minor misalignments that may remain from the first two stages, especially at component connections, ensuring that the overlap of the registered point cloud meets a preset engineering error range (e.g., less than 3mm).

[0069] After the first stage (coarse registration based on the principal axis of the building structure), the second stage (fine registration based on planar features), and the third stage (local optimization based on key points of components), high-quality, consistent spatial point cloud data is finally output, such as... Figure 3 As shown.

[0070] In the above-mentioned automatic inverse modeling method for indoor building information model based on point cloud semantic parsing, step 2, fine-grained instance segmentation of components and multimodal semantic enhancement, realizes the essential transformation from "point stack" to "component entity".

[0071] Specifically, such as Figure 2As shown, the multi-scale feature fusion instance segmentation network used in step 2 has a core component called the multi-scale neighborhood feature aggregation module (MFA). This module captures component features by setting multiple different preset search radii in the feature extraction layer of the neural network. The first preset search radius... The design aims to capture minute details, such as the cross-sectional shape of window frame profiles and the protrusions of door handles. Second preset search radius. The design aims to capture macroscopic semantics, such as the extension direction of an entire wall and the overall enclosure relationship of a room. The network uses multi-level feature concatenation and fusion to ensure that each point contains a composite feature vector from local details to global context.

[0072] To further improve recognition accuracy, the network incorporates an architectural prior-guided attention mechanism. This mechanism transforms architectural expertise into constraint weights during the feature learning process. The system pre-defines prior knowledge, including the verticality of walls (normal vectors are orthogonal to the direction of gravity), the horizontality of the ground (normal vectors are parallel to the direction of gravity), and the embedding relationship of doors and windows within the walls (door and window point clouds are surrounded by the wall plane). In the deep learning attention matrix calculation, if the geometric distribution of a point cloud cluster conforms to these prior features, its corresponding attention weights will be enhanced, thereby guiding the network to converge more quickly to the correct component classification result.

[0073] In step 2 above, to address common issues in real indoor scanning data such as occlusion (e.g., furniture blocking the wall base) and voids, this invention introduces Building Information Modeling (BIM) synthetic data augmentation and domain-adaptive transfer learning. First, a large-scale synthetic point cloud dataset with accurate and clean semantic annotations is automatically generated using existing standard BIM libraries (e.g., IFC format model rooms). During the generation process, the system simulates a real laser scanning process using virtual scanning technology, artificially adding different levels of noise, occlusion, and motion blur to make the synthetic data as close as possible to real-world conditions.

[0074] Subsequently, the system designed a domain adaptation module based on adversarial training. This module includes a feature extractor, a component classifier, and a domain discriminator. The domain discriminator aims to distinguish whether the input features originate from synthetic data or real scanned data, while the feature extractor attempts to deceive the domain discriminator, aligning the features of synthetic and real data in spatial distribution. Through this transfer learning strategy, the segmentation model can effectively transfer knowledge about the complete structure of components learned from synthetic data to real, incomplete scanned point clouds, thereby achieving robust recognition of incomplete point clouds.

[0075] Furthermore, step 2 utilizes multimodal semantic enhancement technology. The system projects 3D point cloud data onto multiple preset 2D planes, generating high-resolution intensity maps, depth maps, or normal vector maps. A visual language model (VLM) pre-trained on a large-scale 2D image dataset is used to semantically annotate these 2D projections. The 2D model can identify visual features that are difficult to distinguish in the 3D point cloud, such as the transparent boundaries of glass or complex decorative moldings. Subsequently, the system uses precise projection geometric transformation relationships to back-map the semantic labels on the 2D images back into the 3D point cloud space, achieving complementarity of multi-source information.

[0076] In the final stage of instance segmentation, the system utilizes a spatially hierarchical instance clustering module to physically separate the semantically labeled point clouds. The clustering process follows a hierarchical architectural logic of "floor → room → component." First, horizontal planes (ground surfaces) are identified to divide floors, then vertical planes (main walls) are identified to divide room areas. Finally, within each room, point cloud clusters with the same semantic labels and spatial connectivity are independently numbered. For the segmentation results, a topology inference module performs post-processing refinement, automatically verifying logical relationships such as whether a door is attached to a wall and whether a window is located in a hollow part of the wall. If a door instance is found to be spatially isolated from a wall, the system will merge it into the nearest wall logical cluster through nearest neighbor search and forcibly correct its geometric position.

[0077] In the above-mentioned automatic inverse modeling method for indoor building information models based on point cloud semantic parsing, please refer to Figure 4 Step 3, the extraction and parametric modeling of geometric parameters through multi-constraint fusion, completes the crucial leap from discrete point clouds to regularized building models. For example... Figure 4 As shown, after receiving the segmented component point cloud clusters, the system adopts customized parameter extraction strategies for different components.

[0078] Specifically, step 3 employs customized parameterization strategies for different components. For wall components, the system does not rely solely on a single plane fitting algorithm, but rather uses a composite strategy combining the Random Sample Consensus (RANSAC) plane fitting algorithm with Principal Component Analysis (PCA). The system first uses RANSAC to extract the principal plane equations of the wall surface, and then uses PCA to analyze the principal directions of the wall point cloud clusters, thereby determining the wall's thickness centerline. To ensure that the modeling results comply with building codes, the system introduces orthogonal and coplanar constraints. If the angle between two adjacent walls is within the range of 90±3 degrees, the system forces their normal vectors to be strictly perpendicular during the parameter extraction stage, ensuring the squareness of the room in the generated BIM model.

[0079] For opening components such as doors and windows, the system focuses on boundary extraction. Using the component point cloud obtained from instance segmentation, the system calculates the curvature variation characteristics of the boundary points. A feature point detection algorithm is used to locate the coordinates of the four corner vertices of the opening. For complex doors and windows with decorative lines, the system employs a multi-segment line fitting strategy to extract their width, height, thickness, and center position parameters. For hinged and sliding doors, the system further categorizes them by identifying the angle between the door leaf and the wall, and incorporates this attribute information as part of the parametric model.

[0080] For floor slabs and ceilings, the system uses a region growing algorithm to identify connected horizontal planes. Considering the complexity of the indoor environment, floor slabs are often obscured by furniture, so the system combines the identified room enclosure walls as horizontal boundary constraints. By calculating the intersection of the wall base and the ground, the system automatically infers and completes the geometric range of the floor slab in the obscured area, ensuring that the floor slab components can be completely closed in the BIM model.

[0081] For non-Manhattan structures, such as sloping walls or curved components, this invention provides a specialized processing path. For sloping walls, the system identifies wall instances that do not conform to orthogonal constraints by performing cluster analysis on the wall surface normal vectors. Using the slope of the intersection line between adjacent walls in three-dimensional space, the tilt angle parameters are automatically calculated, avoiding forced correction to make them vertical walls. For curved components, such as curved corridors or curved walls, the system employs feature point detection based on curvature analysis. The system extracts sampling points on the arc segment and fits the arc equation using a least-squares optimization algorithm to obtain accurate center coordinates, radius, and initial arc length parameters.

[0082] In step 3 above, parametric modeling is not a one-time generation process, but rather an iterative process that includes error feedback and refinement. The extracted geometric parameters are input in real-time into a BIM platform (such as Revit / Dynamo) to generate a digital model. Subsequently, the system calculates the fitting error between the generated digital model surface and the original scanned point cloud. The error evaluation metrics here... The vertical distance from a point to a surface is used as the metric. If the global average error exceeds a first preset error threshold (e.g., 15mm), or the local deviation of a specific component (e.g., a structural column) exceeds a second preset error threshold (e.g., 10mm), the system automatically triggers an iterative refinement process. The system calculates the fitting error and then evaluates it: Figure 4 As shown, when the error is less than or equal to the threshold, the regularized geometric components are directly output; when the error is greater than the threshold, the iterative refinement process is triggered, and the parameter extraction module in step 3 is called back to automatically adjust the RANSAC distance threshold or try a higher-order fitting model (such as using quadratic surface fitting instead of plane fitting). If the data quality cannot support the accuracy requirements, the system will issue a supplementary scan suggestion to the operator through the graphical interface, pointing out the local coordinate areas that need high-density scanning.

[0083] In the above-mentioned automatic inverse modeling method for indoor building information models based on point cloud semantic parsing, step 4, full-process closed-loop feedback optimization and quality evaluation, is the core to ensure the intelligence and automation of the modeling system.

[0084] Specifically, such as Figure 5 As shown, step 4 establishes a multi-dimensional quality evaluation index system covering geometric accuracy, semantic integrity, topological consistency, and efficiency indicators. The data input for this system includes not only the generated BIM model and point cloud data but also a standardized benchmark dataset as a reference. In terms of geometric accuracy, the system automatically calculates the positional deviation, key dimension (such as wall thickness and floor height) deviation, and overall root mean square error of all components. In terms of semantic integrity, the system compares the identified components with a pre-set building component library, calculating the component detection rate, classification accuracy, missed detection rate, and false detection rate. In terms of topological consistency, the system focuses on checking whether the connections between walls and panels are seamless, whether doors and windows cause spatial conflicts, and whether rooms form enclosed topological spaces. In terms of efficiency, the system records the end-to-end processing time and the manual intervention time for each module as a basis for evaluating system stability.

[0085] The closed-loop feedback mechanism in step 4 transforms the evaluation results into specific control commands, such as... Figure 5 The feedback branches below show the following: If the geometric accuracy index fails to meet the standard, the feedback signal triggers the iterative fitting of the parameter extraction module; if the semantic integrity index is abnormal, the system automatically lowers the confidence threshold of the instance segmentation network, forcing the network to output more candidate instances, or reminds the user to perform auxiliary annotation of key nodes; if a red item appears in the topology consistency section, the system starts the topology reasoning module to relocate the logically erroneous components; if the processing efficiency is lower than the preset standard, the system automatically increases the downsampling step size in the preprocessing stage to sacrifice the accuracy of some non-core areas in exchange for overall speed.

[0086] Furthermore, to support the continuous evolution of the entire system, this invention also involves the construction of a standardized benchmark dataset. This dataset covers various scenarios, including residential buildings, large office spaces, commercial complexes, and historical buildings with complex and irregular structures. Each scenario includes a high-precision ground truth model for quantitatively calibrating the performance of the aforementioned modules.

[0087] Example 2

[0088] In another preferred embodiment, for complex indoor scenarios such as large industrial plants with highly repetitive components and a large number of pipeline interference, the indoor BIM automatic inverse modeling method based on point cloud semantic parsing described in this invention has been specifically enhanced in the execution details of each step to cope with the challenges of data explosion and semantic recognition difficulties in large scenarios.

[0089] In the quality-aware dynamic preprocessing of step 1, for slender components such as metal pipes and supports in industrial scenarios, the density estimation module introduces a local geometric feature enhancement algorithm based on tensor voting. When calculating the point cloud quality score, the system considers not only density but also the curvature tensor features of the points. For pipeline areas exhibiting linear features, extremely high weights are assigned to prevent them from being filtered out during downsampling. The first preset size of the adaptive voxel mesh is set to a finer 2mm to capture small components such as industrial flanges and valves. In the vast, flat industrial ground, the second preset size can be relaxed to 100mm. In the multi-level registration stage, the coarse registration module uses standard steel columns in the factory as feature anchor points and constructs global coordinate constraints by extracting the central axis of the columns. The fine registration module focuses on the planar features of beam-column nodes, solving the registration drift problem in large-span spaces by minimizing the distance error of the web plane of the steel beams between multiple stations.

[0090] In the instance segmentation stage of step 2, considering the diverse types of industrial components (such as pumps, distribution boxes, and pipelines of various diameters), the multi-scale feature fusion network expands the search hierarchy of the multi-scale neighborhood feature aggregation module. The system adds a third preset search radius. Specifically designed to capture long-distance pipeline topology features, the building-prior-guided attention mechanism is expanded to include industrial-prior-guided features, adding industrial topology constraints such as horizontal / vertical pipeline routing, pumps fixed to foundations, and cable trays suspended from beams. In this embodiment, the multimodal semantic enhancement module utilizes the plant's process flow diagram (PID diagram) as auxiliary information. The system extracts equipment numbers and connections from the PID diagram using natural language processing technology, matching them with the two-dimensional annotation results output by the visual language model, significantly improving the recognition accuracy for heavily occluded industrial equipment.

[0091] In the parameter extraction stage of step 3, for circular or rectangular pipelines commonly found in industrial scenarios, the system introduces cylindrical and cuboid fitting models. Utilizing the normal vector clustering features of the boundary point cloud, the pipeline's orientation is identified through Gaussian sphere projection, and then nonlinear optimization is used to solve for the pipeline's radius and length parameters. For industrial diagonal bracing beams, the system automatically generates a BIM model of the variable cross-section member in the non-Manhattan structure extraction module by extracting the profile's cross-sectional features and scanning along the axis.

[0092] In the closed-loop feedback of step 4, the quality evaluation system adds an evaluation index for the continuity of pipeline connections. If a logical break in the pipeline is detected in the BIM model, the feedback mechanism will instruct the segmentation network in step 2 to rescan the broken area to determine whether there are any occluded small pipes. Simultaneously, to improve the efficiency of processing trillions of point clouds, the feedback module will dynamically adjust the number of parallel computing cores and memory allocation weights in the preprocessing stage based on the current system memory pressure.

[0093] Example 3

[0094] In another implementation plan for the protection and restoration of historical buildings, the method described in this invention focuses on the extremely high-precision restoration and topological analysis of irregular artistic components (such as carved beams, arches, and relief walls).

[0095] In step 1, the point cloud data is noisy and uneven due to weathering and peeling of historical building materials. The system employs a robust local plane (RLP)-based denoising algorithm in the quality-aware preprocessing. When calculating the quality score, outliers are removed by analyzing the residual distribution within the local neighborhood. In the third stage (based on component key points), the registration framework introduces feature matching based on feature histograms (FPFH). Through meticulous alignment of carved details, it ensures that complex ancient building components do not produce artifacts at the joints.

[0096] In step 2, for complex wooden components unique to ancient buildings, such as dougong (bracket sets) and queti (steles), the instance segmentation network integrates a topology learning module based on graph convolutional neural networks (GCN). The system uses the construction methods of the historical wooden structure as a topological prior and employs an attention mechanism to constrain the modular relationships between components. When performing multimodal semantic enhancement, the system introduces historical survey drawings of ancient buildings as a reference and inputs the classification criteria of ancient buildings into the visual language model through text feature vectors, enabling the system to accurately distinguish minute wooden components such as "dou," "sheng," and "ang."

[0097] In step 3, the non-Manhattan structural parameter extraction module played a crucial role in addressing the numerous curved components and irregular walls in historical buildings. For curved roofs, the system abandoned simple geometric shape fitting and instead adopted non-uniform rational B-spline surface reconstruction technology. By automatically extracting the control point network from the point cloud cluster, a high-fidelity irregular curved surface model was generated within the BIM platform. The error feedback refinement threshold during the parametric modeling process was set to an extremely low level (e.g., 2mm) to meet the extreme precision requirements of ancient building restoration.

[0098] In step 4, the multidimensional quality evaluation system adds a cultural heritage authenticity score, which evaluates the restoration effect by comparing the subtle texture deviations between the generated model and the original point cloud. If the closed-loop feedback optimization mechanism detects the loss of certain key artistic details (such as the depth of carvings), it will trigger a higher-precision local fitting algorithm.

[0099] In this embodiment, when extracting geometric parameters, the system also establishes a topological connection diagram based on the geometric contact relationships between components. For example, by analyzing the spatial overlap between columns and beams, and between beams and purlins, the system automatically generates a complete set of node connection tables while outputting the BIM model. This process in this embodiment greatly assists in the subsequent structural stability analysis of ancient buildings.

[0100] In the above embodiments, in order to more clearly describe the mathematical process of parameter extraction and error evaluation, two key calculation formulas are introduced.

[0101] The first formula is used to calculate the local point cloud quality score in step 1. The system measures local geometric complexity by analyzing the distribution entropy of points within the neighborhood:

[0102]

[0103] in, For the first Quality score for each sampling point The number of points within the search radius. These are the three eigenvalues ​​of the local covariance matrix. By combining the density term and the curvature term (eigenvalue ratio), this formula can automatically assign low scores to flat regions (where eigenvalue differences are large) and high scores to intersecting or edge regions (where eigenvalues ​​tend to be balanced), thereby guiding subsequent adaptive sampling.

[0104] The second formula is used for the quantitative evaluation of the fitting error in steps 3 and 4, guiding the closed-loop feedback process of the system:

[0105]

[0106] in, The root mean square error for a specific component or region. The total number of sampling points participating in the evaluation. These are the sampling points in the original point cloud. This represents the projection of the point onto the geometric surface of the generated BIM component. The system determines in real time whether the modeling result exceeds a preset accuracy threshold by calculating the mean of the sum of the squares of the Euclidean distances from the point to the surface.

[0107] Furthermore, in the implementation of this invention, all computational processes are parallelized on high-performance computing workstations. Point cloud data reading and writing employs a streaming architecture to support loading hundreds of millions of point clouds. In the BIM automatic generation stage, the system encapsulates the extracted geometric parameters (length, width, height, positioning points, rotation vectors) into specific types of component objects through customized API interfaces (such as Revit API or IFC.js). For non-standard components, the system utilizes adaptive families or mass models to ensure that the generated model not only has a geometric shape but also contains complete semantic attribute information.

[0108] The end-to-end closed-loop feedback mechanism enables communication between modules through a message queue. When the evaluation module in step 4 generates a feedback instruction, the instruction contains a specific error code and the ID of the affected component. For example, the error code "E002-Topology-Disconnect" triggers the topology inference module to extend and align the wall with the specific ID. This data-driven closed-loop logic allows the system to continuously approximate the true physical state of the indoor scene through autonomous iteration with minimal human intervention.

[0109] This invention also fully considers the impact of various indoor lighting conditions and sensor noise. In the multimodal enhancement stage, the system performs histogram equalization on the input intensity map to eliminate the influence of uneven lighting on 2D recognition. In the point cloud registration stage, a robust weighting function based on median filtering is introduced to effectively reduce the interference of glass reflection points and dynamic pedestrian points on registration accuracy.

[0110] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An automatic inverse modeling method for indoor BIM based on point cloud semantic parsing, characterized in that, Includes the following steps: Step 1: Obtain a quality score through local point cloud density estimation, perform adaptive voxel downsampling based on the quality score, and construct a multi-level progressive registration framework from the main axis of the building structure to planar features and then to the key points of the components to obtain point cloud data in a unified coordinate system. The quality score is used to reflect the complexity of the geometric features of the object. Step 2: Input the point cloud data under the unified coordinate system into the multi-scale feature fusion instance segmentation network, extract the local detail features and global context features of the components, and enhance the multimodal semantics by introducing the building prior-guided attention mechanism and the spatial hierarchical instance clustering module, combined with the visual language model to perform multimodal semantic enhancement on the auxiliary annotation information of the two-dimensional projection image, so as to realize the recognition of indoor building components and obtain component instances. Step 3: For the component instance, the geometric parameters are extracted by comprehensively using the random sampling consistency plane fitting algorithm, principal component analysis method and region growing algorithm. The geometric parameters are then input into the building information modeling platform according to the topological constraints of the indoor building components to realize the preliminary parametric reconstruction of the indoor building components and generate a preliminary digital model. Step 4: Establish a multi-dimensional quality evaluation system covering geometric accuracy indicators, semantic integrity indicators, topological consistency indicators, and efficiency indicators. Calculate the fitting error between the preliminary digitization model and the original scanned point cloud using this multi-dimensional quality evaluation system. Trigger an iterative refinement process based on the deviation between the evaluation results and a preset evaluation threshold. Iterate through the error backpropagation mechanism to the multi-level progressive registration framework, the multi-scale feature fusion instance segmentation network, or the preliminary parametric reconstruction feedback adjustment instruction until the fitting error meets the set requirements.

2. The indoor BIM automatic inverse modeling method based on point cloud semantic parsing according to claim 1, characterized in that, In step 1, the method for obtaining the quality score includes: constructing a spatial index structure to realize neighborhood query of the original point cloud data; counting the number of sampling points in the neighborhood of each sampling point within a preset search radius; and using a Gaussian kernel to weight the contribution of the neighborhood points to obtain the local density; measuring the local geometric complexity by analyzing the distribution entropy of the sampling points in the neighborhood; and calculating the quality score by combining the number of sampling points with the eigenvalue distribution terms of the local covariance matrix.

3. The indoor BIM automatic inverse modeling method based on point cloud semantic parsing according to claim 2, characterized in that, In step 1, the method for implementing adaptive voxel downsampling based on the quality score includes: Based on a lookup table containing multi-level voxel mesh sizes, the quality score is compared with a preset quality threshold. If the quality score is higher than the preset quality threshold, the corresponding area is determined to be a component edge area or corner area, and the voxel mesh is set to the first preset size for downsampling. If the quality score is lower than the preset quality threshold, the corresponding area is determined to be a flat area, and the voxel grid is set to a second preset size for downsampling, wherein the second preset size is greater than the first preset size.

4. The indoor BIM automatic inverse modeling method based on point cloud semantic parsing according to claim 1, characterized in that, The specific execution process of the multi-level progressive registration framework includes: Phase 1: Extract the orientation of the main wall and the ground normal vector using principal component analysis or global normal vector histogram, construct global rotation constraints based on the Manhattan world hypothesis, and achieve initial alignment by searching for matching points in the rotation space; The second stage involves identifying planar regions in the initially aligned point cloud using a region growing algorithm or a random sampling consensus algorithm. By calculating the distance and angle deviations between planar regions with the same name in different scanning stations, a cost function that minimizes the distance between the planes is constructed. The transformation matrix is ​​then iteratively optimized using a nonlinear least squares method for fine registration. The third stage involves extracting feature descriptors and constructing corresponding point sets for significant geometric feature points, including doorway corners, window frame vertices, and beam-column intersections. Local coordinate fine-tuning is then performed using the local iterative nearest point algorithm.

5. The indoor BIM automatic inverse modeling method based on point cloud semantic parsing according to claim 1, characterized in that, In step 2, the internal execution logic of the multi-scale feature fusion instance segmentation network includes: Different preset search radii are set in the feature extraction layer. The first preset search radius is used to capture micro-detail features, and the second preset search radius is used to capture macro-semantic features. A composite feature vector is generated by multi-level feature concatenation and fusion, wherein the first preset search radius is smaller than the second preset search radius. The specific execution process of the building prior-guided attention mechanism includes: converting priors such as the orthogonality of the wall normal vector and the gravity direction, the parallelism of the ground normal vector and the gravity direction, and the embedding relationship priors of the door and window point clouds being surrounded by the wall plane into constraint weights in the feature learning process; and enhancing the corresponding attention weights of point cloud clusters whose geometric distribution conforms to the orthogonality prior, parallelism prior, or embedding relationship prior in the attention matrix calculation.

6. The indoor BIM automatic inverse modeling method based on point cloud semantic parsing according to claim 1, characterized in that, Step 2 also includes a synthetic data augmentation and domain adaptive transfer learning process, the specific process of which includes: Using a standard building information modeling library, a synthetic point cloud dataset with semantic annotations is generated through virtual scanning technology, and the noise distribution, occlusion, and motion blur features of laser scanning are simulated during the generation process. Domain adaptive transfer is performed through an adversarially trained network comprising a feature extractor, a component classifier, and a domain discriminator, wherein the domain discriminator is used to distinguish the source of input features, and the feature extractor aligns the feature distribution of the synthetic point cloud data with the feature distribution of the real scanned point cloud data in space through adversarial game with the domain discriminator.

7. The indoor BIM automatic inverse modeling method based on point cloud semantic parsing according to claim 1, characterized in that, The specific execution process of the multimodal semantic enhancement and spatial hierarchical instance clustering module includes: Multimodal semantic enhancement: Projecting three-dimensional point cloud data onto a two-dimensional plane with multiple preset viewpoints to generate a two-dimensional image containing intensity information, depth information or normal vector information. Using a pre-trained visual language model, the two-dimensional image is semantically annotated, and the two-dimensional semantic labels on the two-dimensional image are back-mapped to the three-dimensional point cloud space using projection geometric transformation relationships. Spatial hierarchical clustering: identify horizontal planes to divide floor areas, identify vertical planes to divide room areas, and assign independent instance numbers to point cloud clusters with the same semantic labels and spatial connectivity within each room area. Topology reasoning verification: When a door instance or window instance is detected to be spatially isolated from a wall instance, it is merged into the nearest logical cluster of walls through nearest neighbor search, and the geometric position of the door instance or window instance is corrected.

8. The indoor BIM automatic inverse modeling method based on point cloud semantic parsing according to claim 1, characterized in that, The strategy for extracting geometric parameters for the component instance in step 3 includes: For wall components: the random sampling consistency plane fitting algorithm is used to extract the wall plane equation, the principal component analysis method is combined to determine the wall thickness centerline, and orthogonal constraints and coplanar constraints are introduced. When the included angle between two adjacent walls is within the preset right angle error range, the normal vector of the wall is adjusted to be perpendicular. For door and window components: the coordinates of the four corner vertices of the opening are located by calculating the curvature change characteristics at the boundary, and the width, height, thickness and center position parameters are extracted by using a multi-segment line fitting strategy; For floor slabs and ceiling components: A region growing algorithm is used to identify connected horizontal planes. The identified enclosing walls are used as horizontal boundary constraints. By calculating the intersection of the wall base and the horizontal plane, the geometry of the occluded area is completed.

9. The indoor BIM automatic inverse modeling method based on point cloud semantic parsing according to claim 1, characterized in that, The strategy for extracting geometric parameters for non-Manhattan structural components in the component instance in step 3 includes: For inclined wall components: Non-orthogonal wall instances are identified by cluster analysis of the wall surface normal vectors, and the tilt angle parameter is calculated by using the slope of the intersection line of adjacent walls in three-dimensional space; For curved components: Curvature analysis of sampling points and feature point detection are performed to extract point cloud data on the arc segment. The least squares optimization algorithm is used to fit the arc equation and extract the center coordinates, radius and initial arc length parameters. For irregular curved surface components: a non-uniform rational B-spline surface reconstruction method is adopted to extract the control point network from the point cloud cluster and generate the irregular curved surface model within the building information modeling platform.

10. The indoor BIM automatic inverse modeling method based on point cloud semantic parsing according to claim 1, characterized in that, In step 4, the logic of the error backpropagation mechanism for feeding back adjustment instructions includes: If the geometric accuracy index fails to meet the standard, the adjustment instruction will be sent to step 3 to adjust the distance threshold of the random sampling consistency algorithm or replace the high-order fitting model for secondary parameter extraction. If the semantic integrity index is abnormal, the adjustment instruction is sent to step 2 to lower the confidence threshold of the multi-scale feature fusion instance segmentation network. If the topology consistency index is abnormal, restart the topology reasoning module to perform extension, alignment or merging operations on the logically erroneous component instances and correct the spatial connection relationship between components. If the efficiency index is lower than the preset efficiency standard, the adjustment instruction is sent to step 1 to increase the downsampling step size of adaptive voxel downsampling.

Images