Building live-action three-dimensional reconstruction system based on point cloud data

Abstract

The invention discloses a building live-action three-dimensional reconstruction system based on point cloud data, and relates to the technical field of building live-action three-dimensional reconstruction, the building live-action three-dimensional reconstruction system comprises a preprocessing module, a multistage adaptive noise filtering module, a semantic and geometric joint reasoning module and a three-dimensional model generation module which are connected in sequence; the preprocessing module is used for receiving original building point cloud data, performing registration and initial space division, and outputting a registered point cloud data set. According to the building live-action three-dimensional reconstruction system based on the point cloud data, by introducing a multi-stage adaptive noise filtering and semantic geometry joint reasoning cooperation mechanism, the contradiction that noise elimination is inaccurate and feature details are easy to lose in a complex building scene in an existing method is effectively solved. The system can intelligently distinguish the real discrete noise from the inherent complex geometric features of the building, and maintains the integrity of the details such as decorative members and curved roofs to the greatest extent while ensuring the denoising effect, thereby achieving the high-fidelity point cloud data preprocessing.

Description

Technical Field

[0001] This invention relates to the field of 3D reconstruction of architectural real-world scenes, specifically a 3D reconstruction system for architectural real-world scenes based on point cloud data. Background Technology

[0002] Architectural 3D reconstruction technology refers to the process of converting real-world buildings into high-precision, visualized 3D models using various digital methods. It is widely applied in urban planning, building preservation, intelligent operation and maintenance, and virtual display. Point cloud data acquired using technologies such as laser scanning and oblique photography can accurately reflect the geometric structure and surface features of buildings, providing a reliable data foundation for subsequent analysis, simulation, and display. However, in actual data acquisition, point cloud data is often affected by multiple noises, including environmental interference, equipment errors, and motion distortion, leading to problems such as burrs, holes, and blurred edges in the reconstructed model, severely impacting the accuracy and usability of the 3D model.

[0003] To improve the accuracy and robustness of 3D building reconstruction, various point cloud denoising and reconstruction methods have been proposed in existing technologies. For example, patent publication number "CN118918244A" describes "a point cloud-based method and system for 3D building reconstruction," which uses recursive nearest neighbor search to filter non-noise points, and then combines plane fitting to determine discrete points, ultimately achieving noise removal and 3D model generation. This technology can, to a certain extent, identify and remove isolated noise points, improving the smoothness of the reconstructed surface.

[0004] However, the above methods still have the following shortcomings: First, recursive nearest neighbor search relies on random point selection and fixed thresholds, which can easily lead to the deletion of effective points or the retention of noise points when the point cloud distribution is uneven or has a complex structure. Second, the method of judging discrete points by plane fitting has poor adaptability to non-planar regions and is prone to feature loss. Non-planar regions include curved surfaces, eaves, and decorative components. Third, the method does not fully consider the fusion and semantic information association of multi-source and multi-scale point clouds in the real building scene, resulting in the reconstruction results lacking structural hierarchy and realism, making it difficult to support refined modeling and subsequent applications.

[0005] Therefore, there is still a lack of point cloud reconstruction systems in the existing technology that can adapt to complex architectural scenes, maintain the integrity of geometric features, and have high robustness. There is an urgent need to propose a more intelligent, accurate, and practical 3D reconstruction solution for real-world architectural scenes. Summary of the Invention

[0006] The purpose of this invention is to provide a 3D reconstruction system for building scenes based on point cloud data, so as to solve the problems mentioned in the background art.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a 3D reconstruction system for building real scene based on point cloud data, comprising a preprocessing module, a multi-level adaptive noise filtering module, a semantic and geometric joint reasoning module, and a 3D model generation module connected in sequence; The preprocessing module is used to receive the original building point cloud data, perform registration and initial spatial division, and output the registered point cloud dataset. The multi-level adaptive noise filtering module is connected to the output of the preprocessing module and is used to perform a processing flow containing at least two different filtering mechanisms on the point cloud dataset to identify and label potential noise points. The first filtering mechanism filters based on the local geometric properties of the point cloud, and the second filtering mechanism filters based on the distribution consistency of the point cloud in the feature space. The semantic and geometric joint reasoning module is connected to the output of the multi-level adaptive noise filtering module. It is used to receive filtered point cloud data, simultaneously perform multi-type geometric primitive fitting and semantic category prediction, and perform secondary verification and correction on the potential noise points based on the correlation between the fitting results and the semantic prediction results, so as to generate denoised point cloud data with semantic labels. The 3D model generation module is connected to the output of the semantic and geometric joint reasoning module. It is used to generate a 3D model of the building scene with complete topological structure and semantic annotation based on the denoised and semantically labeled point cloud data, by fusing multi-scale geometric features and semantic context information.

[0008] Furthermore, the multi-level adaptive noise filtering module includes a primary filtering unit and a secondary filtering unit; The primary filtering unit is used to execute the first filtering mechanism, and its specific configuration is as follows: adaptively spatially divide the point cloud dataset into blocks, calculate the local point density and curvature value of each point in each block; mark points with local point density lower than a first dynamic threshold and curvature value higher than a second dynamic threshold as primary suspected noise points; the first dynamic threshold and the second dynamic threshold are adaptively determined according to the statistical distribution of the point cloud in the block. The secondary filtering unit, connected to the primary filtering unit, is used to execute the second filtering mechanism. Its specific configuration includes: extracting multidimensional feature vectors from the point cloud dataset that are not marked as primary suspected noise points; the multidimensional feature vectors at least include spatial coordinates, normal vectors, and color information; performing cluster analysis on the multidimensional feature vectors in the feature space; marking points belonging to clusters with fewer than a preset threshold number of cluster members, and points whose distance from the center of the main cluster exceeds a dynamic distance threshold, as secondary suspected noise points; the union of the primary suspected noise points and the secondary suspected noise points constitutes the potential noise point set.

[0009] Furthermore, the semantic and geometric joint reasoning module includes a geometric primitive fitting unit, a semantic prediction unit, and a joint decision-making unit; The geometric primitive fitting unit is used to perform parallel fitting of three types of geometric primitives—planar, cylindrical, and spherical—for each local region in the filtered point cloud data, and to calculate the residual of each point relative to each type of fitting primitive. The semantic prediction unit, based on a neural network model, performs semantic category prediction on each point in the filtered point cloud data and outputs the probability distribution of each point belonging to multiple predefined building component categories. The building component categories include at least walls, roofs, windows, doors, and decorative components. The joint decision-making unit is connected to the geometric primitive fitting unit and the semantic prediction unit, respectively, and is used to make decisions based on the residuals and the probability distribution. Its configuration is as follows: for a target point marked as a potential noise point, find several neighboring points within its spatial neighborhood; if the fitting residuals of points belonging to the same semantic category among the neighboring points for a certain type of geometric primitive are generally lower than the residual threshold, while the fitting residual of the target point for that type of geometric primitive is higher than the residual threshold, and the predicted semantic category probability distribution of the target point is inconsistent with the mainstream semantic category of the neighboring points, then the target point is confirmed as a final noise point; otherwise, the target point is removed from the potential noise point set.

[0010] Furthermore, the system also includes an iterative optimization module, which is connected between the multi-level adaptive noise filtering module and the semantic and geometric joint reasoning module to form a closed feedback loop; The iterative optimization module is configured to: after the joint decision-making unit outputs the final noise point set, feed the final noise point set back to the primary filtering unit; the primary filtering unit dynamically adjusts the determination strategy for the first dynamic threshold and the second dynamic threshold in the next round of processing based on the spatial distribution and attribute characteristics of the final noise points confirmed in this iteration; the system executes at least two complete filtering and inference processes that include the feedback loop.

[0011] Furthermore, the 3D model generation module includes a feature fusion unit and an implicit reconstruction unit; The feature fusion unit is used to extract multi-scale features from the denoised and semantically labeled point cloud data. The multi-scale features include geometric detail features calculated from local neighborhoods, structural features obtained by aggregating contextual information from a larger region, and semantic features output by the semantic prediction unit. The feature fusion unit uses an attention mechanism to weightedly fuse the geometric detail features, structural features, and semantic features to generate enhanced point cloud features. The implicit reconstruction unit, connected to the feature fusion unit, is used to take the enhanced point cloud features as input and generate a continuous, seamless three-dimensional mesh surface model by solving the implicit field function.

[0012] Furthermore, the 3D model generation module also includes a topology verification and repair unit; The topology verification and repair unit, connected to the implicit reconstruction unit, is used to perform automated topology analysis on the generated 3D mesh surface model, detect and identify non-manifold edges, isolated vertices, or self-intersecting faces in the model; based on the semantic tags output by the semantic and geometric joint reasoning module, it performs category-aware automatic repair of the identified topology errors. The repair includes at least bridging broken edges belonging to the same semantic component and filling small-scale holes belonging to non-structural semantic categories caused by noise removal.

[0013] Furthermore, the initial spatial division in the preprocessing module is as follows: based on the height distribution histogram of the building in the vertical direction extracted from the registered point cloud, obvious floor height clusters are identified; using the identified floor heights as a reference, combined with the point density changes in the horizontal plane, the point cloud data is adaptively divided into multiple three-dimensional blocks aligned with the building structure.

[0014] Furthermore, the adaptive spatial segmentation in the primary filtering unit adopts a non-uniform partitioning strategy: small segment sizes are used in areas with high point cloud density, and large segment sizes are used in areas with low point cloud density; the segment boundaries are generated through an iterative clustering method to maintain the consistency of point cloud geometric attributes within a single segment.

[0015] Furthermore, the neural network model used in the semantic prediction unit employs a composite loss function that integrates geometric loss and semantic loss during training. The geometric loss term encourages the network to learn features that are invariant to changes in the local geometric structure of the point cloud, while the semantic loss term supervises the network to accurately predict the semantic category of the points.

[0016] Furthermore, the system also includes an output module connected to the 3D model generation module; the output module is used to convert the generated 3D building scene model with complete topological structure and semantic annotation into a file that conforms to industrial basic standards or general 3D format, and to associate and store the semantic tags, geometric parameters and topological relationship information of each component in the model.

[0017] This invention provides a 3D reconstruction system for building scenes based on point cloud data. It has the following beneficial effects: This point cloud-based 3D reconstruction system for architectural scenes effectively addresses the contradiction between inaccurate noise removal and easy loss of feature details in existing methods in complex architectural scenes by introducing a multi-level adaptive noise filtering and semantic-geometric joint reasoning mechanism. The system intelligently distinguishes between real discrete noise and the inherent complex geometric features of the building, ensuring effective denoising while preserving the integrity of details such as decorative components and curved roofs to the greatest extent possible, thus achieving high-fidelity point cloud data preprocessing.

[0018] This point cloud-based 3D reconstruction system for building scenes generates 3D models with both geometric accuracy and rich semantic structure through deep reconstruction that integrates multi-scale features and semantic information, as well as topology-aware model optimization and standardized output. This achievement can be directly integrated into the Building Information Modeling (BIM) workflow, providing a high-quality and directly usable data foundation for downstream applications such as architectural design, digital archiving of cultural heritage, and smart city management, thereby improving the overall efficiency and reliability from data acquisition to engineering applications. Attached Figure Description

[0019] Figure 1 This is a flowchart of a multi-level adaptive noise filtering process for a three-dimensional reconstruction system of building real scene based on point cloud data according to the present invention. Figure 2 This invention provides a semantic and geometric joint reasoning decision graph for a three-dimensional reconstruction system of architectural real-world scenes based on point cloud data. Figure 3 This is an iterative optimization feedback loop diagram for a three-dimensional reconstruction system of building scenes based on point cloud data according to the present invention. Detailed Implementation

[0020] 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 only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figures 1 to 3 The present invention provides a technical solution: a 3D reconstruction system for building real scene based on point cloud data, comprising a preprocessing module, a multi-level adaptive noise filtering module, a semantic and geometric joint reasoning module, and a 3D model generation module connected in sequence; The preprocessing module is used to receive the raw building point cloud data, perform registration and initial spatial division, and output the registered point cloud dataset. A multi-level adaptive noise filtering module is connected to the output of the preprocessing module. It is used to perform a processing flow containing at least two different filtering mechanisms on the point cloud dataset to identify and label potential noise points. The first filtering mechanism filters based on the local geometric properties of the point cloud, and the second filtering mechanism filters based on the distribution consistency of the point cloud in the feature space. The semantic and geometric joint reasoning module is connected to the output of the multi-level adaptive noise filtering module. It is used to receive filtered point cloud data, simultaneously perform multi-type geometric primitive fitting and semantic category prediction, and perform secondary verification and correction on potential noise points based on the correlation between the fitting results and the semantic prediction results, so as to generate denoised point cloud data with semantic labels. The 3D model generation module, connected to the output of the semantic and geometric joint reasoning module, is used to generate a 3D model of the building scene with complete topological structure and semantic annotation based on denoised and semantically labeled point cloud data, by fusing multi-scale geometric features and semantic context information.

[0022] It should be further explained that the system first acquires the original building point cloud data collected by a 3D laser scanner or oblique photogrammetry through a preprocessing module. This module uses a registration algorithm based on feature point matching to unify the point cloud data from multiple stations and angles into the same coordinate system, forming a complete point cloud dataset. Subsequently, the system executes an adaptive noise filtering process consisting of two stages: In the first stage, the system adaptively divides the registered point cloud into spatial segments based on the inherent structural characteristics of the building. Instead of uniform grid division, it identifies potential floors in the vertical direction based on the aggregation characteristics of the point cloud's height distribution, and subdivides regions in the horizontal plane based on changes in point density, ensuring a good correspondence between each segmented block and building components. The process involves... The structural characteristics include the continuity of floor slabs and load-bearing walls. Within each block, the system calculates the local point density and curvature of each point in parallel, and dynamically sets a threshold based on the statistical distribution of the density and curvature of all points in the block. The statistical distribution includes the mean and variance. Points with significantly lower density than the neighborhood average and abnormally high curvature are marked as first-class potential noise points. In the second stage, the system extracts the spatial coordinates, normal vectors, and RGB color information of all points to form a multi-dimensional feature vector. It uses a density-based clustering algorithm to cluster in the feature space, and points whose cluster size is extremely small or whose deviation from the main cluster center exceeds the dynamic distance threshold are identified as second-class potential noise points. The first and second-class potential noise points together constitute the noise point set to be verified.

[0023] Next, the semantic and geometric joint reasoning module begins its work: this module contains a trained deep neural network, such as a variant of the PointNet++ architecture, to perform semantic segmentation on the filtered point cloud in real time, predicting the probability of each point belonging to categories such as walls, windows, roofs, doors, and decorative components; simultaneously, this module performs parallel fitting of various geometric primitives such as planes, cylinders, and spheres on the local neighborhood of the point cloud, and calculates the fitting residual of each point relative to each primitive; for each noise point to be verified, the system retrieves its K nearest neighbors, analyzes the mainstream semantic category of these neighboring points and their consistency with the low residual fitting of a certain geometric primitive, if the semantic category of the target point conflicts with the mainstream category, and its fitting residual to the same primitive is significantly higher than the average level of the neighboring points, then it is finally identified as a real noise point and removed; this process effectively distinguishes real noise from valid points located on complex curved surfaces, such as domes and reliefs.

[0024] Finally, the 3D model generation module receives the denoised and semantically labeled point cloud. First, it captures the local geometric details and global structural features of the point cloud through a multi-scale feature extraction network. Then, it uses an attention mechanism to weightedly fuse these geometric and semantic features. The fused features are input into a coordinate-based neural implicit field network, which learns a continuous function to map 3D spatial coordinates to signed distance values. By extracting zero isosurfaces, a smooth, complete, and hole-free 3D mesh model is obtained. The system also performs automated topology verification on the mesh model, identifies and guides the repair of non-manifold edges or micro-holes based on semantic labels, and finally outputs a 3D model of the building scene with both geometric accuracy and component-level semantic information, which can be directly used for BIM system integration or digital archiving.

[0025] The multi-level adaptive noise filtering module includes a primary filtering unit and a secondary filtering unit; The primary filtering unit is used to execute the first filtering mechanism. Its specific configuration is as follows: adaptively divide the point cloud dataset into spatial blocks, calculate the local point density and curvature value of each point in each block; mark the points with local point density lower than the first dynamic threshold and curvature value higher than the second dynamic threshold as primary suspected noise points; the first dynamic threshold and the second dynamic threshold are adaptively determined according to the statistical distribution of the point cloud in the block. The secondary filtering unit, connected to the primary filtering unit, is used to execute the second filtering mechanism. Its specific configuration is as follows: extracting multidimensional feature vectors from the point cloud dataset that are not marked as primary suspected noise points. The multidimensional feature vectors contain at least spatial coordinates, normal vectors, and color information. Performing cluster analysis on the multidimensional feature vectors in the feature space. Marking points belonging to clusters with fewer than a preset threshold number of cluster members, as well as points whose distance from the center of the main cluster exceeds a dynamic distance threshold, as secondary suspected noise points. The union of primary suspected noise points and secondary suspected noise points constitutes a potential noise point set.

[0026] It should be further explained that the specific implementation of the multi-level adaptive noise filtering module is as follows: the module includes a primary filtering unit and a secondary filtering unit connected in series, which work together to implement two levels of filtering based on different principles. The primary filtering unit first adaptively divides the preprocessed point cloud into spatial blocks. The block strategy is not a fixed-size grid, but dynamically adjusted based on the local density of the point cloud and prior knowledge of the building structure. Within each block, the unit calculates the local point density and curvature of each point in parallel. The local point density is obtained by counting the number of neighboring points within a spatial sphere of a given radius, and the curvature is calculated based on the eigenvalues ​​of the covariance matrix of the point and its neighbors. Subsequently, based on the statistical distribution of all point density and curvature values ​​within the block, the unit calculates the first and second dynamic thresholds using the mean-variance method. The first dynamic threshold is the mean of the local point density within the block minus 1.2 times the variance, and the second dynamic threshold is the mean of the curvature values ​​within the block plus 1.5 times the variance. This calculation method allows the thresholds to adaptively match the statistical distribution characteristics of the point cloud in different blocks, avoiding the problem of insufficient filtering accuracy caused by fixed thresholds.

[0027] For example, the density threshold is set to a certain percentage lower than the median density within the block, and the curvature threshold is set to a certain percentage higher than the median curvature within the block; to dynamically determine the first dynamic threshold and the second dynamic threshold; any point that simultaneously satisfies that its local point density is lower than the first dynamic threshold and its curvature is higher than the second dynamic threshold is marked as a primary suspected noise point.

[0028] Subsequently, the secondary filtering unit begins its work, taking the original point cloud dataset as input, but paying particular attention to the set of points not labeled as primary suspected noise points. For these points, the unit extracts a multidimensional feature vector for each point, which consists of at least three elements: three-dimensional spatial coordinates, a normal vector based on local surface fitting, and RGB color information, forming a nine-dimensional or higher-dimensional feature representation. Then, the unit uses a density-based clustering algorithm to perform cluster analysis on the points in this high-dimensional feature space. In the secondary filtering unit of the multi-level adaptive noise filtering module, the density-based clustering analysis adopts the DBSCAN algorithm. The preset number threshold is set to one-thousandth of the total number of point clouds in the block. The dynamic distance threshold is determined by calculating 1.5 times the standard deviation of the main cluster. At the same time, it is adaptively adjusted in combination with the average density of the point clouds in the block to ensure that abnormal clusters can be accurately identified for point clouds with different density distributions. After clustering is completed, the unit identifies two types of outliers: first, all points in clusters whose total number of members is less than a preset threshold (the preset threshold is less than one-thousandth of the total number of points); second, points that, although belonging to a large cluster, have a Mahalanobis distance from their feature vector to the center of the cluster that exceeds a dynamic distance threshold calculated based on the distribution characteristics of the cluster. These two types of points are marked as secondary suspected noise points.

[0029] Finally, the outputs of the primary filtering unit and the secondary filtering unit are combined to form a complete set of potential noise points.

[0030] This implementation combines rapid filtering based on local geometric attributes with in-depth analysis based on the overall distribution consistency of high-dimensional feature space. Local geometric attributes include density and curvature. The two are not simply executed sequentially, but rather a secondary filtering unit focuses on performing supplementary in-depth checks on points missed by the primary filter, forming a composite filtering mechanism that covers different noise types. This effectively solves the problem of single geometric thresholds failing in the face of complex surfaces or feature noise.

[0031] The semantic and geometric joint reasoning module includes a geometric primitive fitting unit, a semantic prediction unit, and a joint decision-making unit; The geometric primitive fitting unit is used to perform parallel fitting of three types of geometric primitives—planar, cylindrical, and spherical—for each local region in the filtered point cloud data, and to calculate the residual of each point relative to each type of fitted primitive. The semantic prediction unit, based on a neural network model, performs semantic category prediction on each point in the filtered point cloud data and outputs the probability distribution of each point belonging to multiple predefined building component categories. The building component categories include at least walls, roofs, windows, doors, and decorative components. The joint decision-making unit, connected to the geometric primitive fitting unit and the semantic prediction unit respectively, is used to make decisions based on residuals and probability distributions. Its configuration is as follows: for a target point marked as a potential noise point, find several neighboring points in its spatial neighborhood; if the fitting residuals of the neighboring points belonging to the same semantic category to a certain type of geometric primitive are generally lower than the residual threshold, while the fitting residual of the target point to that type of geometric primitive is higher than the residual threshold, and the probability distribution of the predicted semantic category of the target point is inconsistent with the mainstream semantic category of the neighboring points, then the target point is confirmed as the final noise point; otherwise, the target point is removed from the set of potential noise points.

[0032] It should be further explained that the specific implementation of the semantic and geometric joint reasoning module is as follows: This module consists of three core parts that work together: a geometric primitive fitting unit, a semantic prediction unit, and a joint decision-making unit. The geometric primitive fitting unit receives the filtered point cloud. For each target point in the point cloud, it selects all points in its spatial neighborhood to form a local point set. For this local point set, the unit performs three different types of geometric primitive model fitting in parallel: fitting a planar model using the least squares method, fitting a cylindrical model using the cylindrical equation, and fitting a spherical model using the spherical equation. For each fitting, the vertical distance from each point in the local point set to the fitted geometric primitive surface is calculated as the fitting residual for that point to that primitive.

[0033] The semantic prediction unit employs a deep neural network model pre-trained on a building point cloud dataset with component category annotations. This network takes the spatial and feature information of each point in the aforementioned point cloud and its surrounding neighborhood points as input, directly outputting a probability distribution vector indicating whether the point belongs to a predefined building component category. Categories include at least walls, roofs, windows, doors, and decorative components. The semantic prediction unit uses a PointNet++ variant neural network model with a 9-dimensional input dimension, encompassing 3D spatial coordinates, 3D normal vectors, and 3D color information. The model includes 4 sampling layers, 4 grouping layers, and 3 fully connected layers. Skip connections between layers enhance feature propagation, and the ReLU activation function is uniformly used. The dataset used for model training comprises no fewer than 100 sets, covering different building types such as residential, commercial, and ancient buildings. Each set of data includes complete point cloud data and corresponding component-level semantic labels, ensuring the model has good generalization ability across various building scenarios.

[0034] The joint decision-making unit receives various residual data from the geometric fitting unit and probability distribution data from the semantic prediction unit. For each target point marked as a "potential noise point", the decision-making unit performs the following operations: First, it determines the K spatial nearest neighbors of the target point; second, it analyzes the semantic prediction results of these neighboring points and determines the category with the highest probability as the "mainstream semantic category" of the neighborhood; at the same time, it checks the fitting residuals of these neighboring points for three types of primitives: plane, cylinder, and sphere. If it is found that the fitting residuals of the vast majority of neighboring points for a specific type of primitive, such as a plane, are all below a preset residual threshold, then it is determined that there is a primitive with good geometric consistency in the neighborhood.

[0035] At this point, the decision-making unit makes a joint judgment: if the fitting residual of the target point to the aforementioned well-consistent primitives is significantly higher than the residual threshold, and the predicted semantic category of the target point itself, i.e. the category with the highest probability, is different from the mainstream semantic category of the neighborhood, then the joint decision-making unit finally confirms that the target point is a real noise point; otherwise, if the semantic category of the target point is consistent with the mainstream category, or if its high residual is due to being located on a non-planar surface such as a cylinder or sphere, i.e., the primitive type with the lower residual is consistent with the neighborhood, then the decision-making unit determines that the target point is a valid point and removes it from the set of potential noise points.

[0036] This implementation method overcomes the limitations of traditional methods that rely solely on geometric or semantic information for discrimination. By establishing a cross-validation mechanism that considers geometric fit consistency, semantic context consistency, and the target point's own characteristics, it can accurately distinguish between real discrete noise and valid points located on complex geometric features. This achieves a balance between denoising accuracy and feature integrity while preserving the architectural details to the greatest extent possible. The complex geometric features include Roman columns and domes. The specific fitting algorithm, network model inputs and outputs, decision logic, and thresholds are all clearly defined.

[0037] The system also includes an iterative optimization module, which connects the multi-level adaptive noise filtering module and the semantic and geometric joint reasoning module to form a closed feedback loop; The iterative optimization module is configured as follows: after the joint decision-making unit outputs the final noise point set, the final noise point set is fed back to the primary filtering unit; the primary filtering unit dynamically adjusts the determination strategy of the first dynamic threshold and the second dynamic threshold in the next round of processing based on the spatial distribution and attribute characteristics of the final noise points confirmed in this iteration; the system executes at least two complete filtering and inference processes containing feedback loops.

[0038] It should be further explained that the specific implementation of the iterative optimization module is as follows: The module is implemented in hardware or software logic as an independent unit with data storage and feedback control functions. Its input end is connected to the output end of the joint decision-making unit in the semantic and geometric joint reasoning module to receive the final confirmed noise point set, and its output end is connected back to the threshold setting sub-module of the primary filtering unit in the multi-level adaptive noise filtering module, thereby forming a closed loop of cyclical data processing and parameter optimization.

[0039] After the system completes the process from primary filtering, secondary filtering to joint inference for the first time, the iterative optimization module begins to work: it collects and analyzes the final set of noise points generated in the current round, extracts the spatial distribution characteristics of these noise points in the original point cloud and their local geometric properties, such as the actual numerical distribution of density and curvature; based on these analysis results, the module generates adjustment instructions for the determination strategies of the first dynamic threshold and the second dynamic threshold in the primary filtering unit.

[0040] For example, if the module finds that most of the points ultimately identified as noise are concentrated on a specific type of material surface, such as a glass curtain wall, characterized by low density and gentle curvature changes, it will instruct the primary filtering unit to use a more lenient density threshold and a more stringent curvature threshold combination for the point cloud blocks corresponding to this type of region in the next round of processing, in order to avoid excessive deletion of valid points. The adjusted primary filtering unit runs again on the new subset of point cloud data, generating a new set of potential noise points, which are then further identified through subsequent secondary filtering and joint inference; where the subset of point cloud data refers to the result after the previous round of denoising.

[0041] The system is pre-programmed to execute this complete process, which includes a feedback loop, at least twice. The iteration termination condition of the iterative optimization module is set to automatically stop iteration when the change rate of the final noise point set is less than 5% in two consecutive iterations, or when the number of iterations reaches 5. This setting ensures the optimization effect of noise filtering while effectively controlling computational costs and avoiding efficiency waste caused by excessive iteration. Through this multi-round self-verification and parameter fine-tuning, the noise filtering standard can adapt to the dynamic evolution of the point cloud characteristics of specific buildings and gradually converge to a better filtering state.

[0042] This implementation breaks through the limitation of the traditional point cloud denoising process of "one-way one-time" processing. By establishing a direct feedback channel from the decision result to the filtering parameters and using real noise samples generated in the previous iteration as training information, the system has online self-optimization capability, thereby improving the long-term stability and accuracy of distinguishing noise and feature points on complex and non-uniform building point clouds.

[0043] The 3D model generation module includes a feature fusion unit and an implicit reconstruction unit; The feature fusion unit is used to extract multi-scale features from denoised point cloud data with semantic labels. The multi-scale features include geometric detail features calculated from local neighborhoods, structural features obtained by aggregating contextual information from a larger region, and semantic features output by the semantic prediction unit. The feature fusion unit uses an attention mechanism to weightedly fuse geometric detail features, structural features, and semantic features to generate enhanced point cloud features. The implicit reconstruction unit, connected to the feature fusion unit, is used to take the enhanced point cloud features as input and generate a continuous, seamless 3D mesh surface model by solving the implicit field function.

[0044] The implicit field function in the implicit reconstruction unit adopts a neural radiation field architecture based on a multilayer perceptron. The multilayer perceptron contains 6 hidden layers, each with 256 neurons. The activation function uses the Swish function to improve the nonlinear fitting ability of the model. The sampling density of the moving cube algorithm is set to 1 sampling point per cubic centimeter, and the isosurface extraction threshold is set to 0.01. This parameter configuration ensures that the generated 3D mesh surface model has both continuous and seamless characteristics and can accurately restore the geometric details of the building.

[0045] It should be further explained that the specific implementation of the feature fusion unit and implicit reconstruction unit in the 3D model generation module is as follows: The feature fusion unit receives a denoised point cloud with semantic labels from the semantic and geometric joint reasoning module. This unit first processes the input data through three parallel feature extraction paths: The first path operates in the local neighborhood of the point cloud through a graph convolutional network, aggregating the coordinates, normal vectors, and curvature information of the points to generate geometric detail features that characterize subtle geometric changes; The second path uses a point cloud Transformer network module with a larger receptive field to capture long-term dependencies across local regions in the point cloud, such as the overall outline of a building and the alignment relationship between floors, thereby generating structural features that express the macro layout; The third path directly uses the semantic category label assigned to each point and converts it into a dense semantic feature vector through a learnable embedding layer.

[0046] Subsequently, the feature fusion unit employs a multi-head attention mechanism to fuse the three types of features mentioned above. This mechanism takes geometric detail features, structural features, and semantic features as inputs for query, key, and value, respectively. By calculating attention weights, it dynamically determines the contribution of each type of feature when reconstructing different regions. For example, it assigns higher weights to geometric detail features when reconstructing regular wall surfaces, and enhances the guiding role of semantic features when reconstructing door and window openings with clear semantic boundaries. Finally, it outputs a set of enhanced point cloud features that integrate multi-scale information and contextual semantics.

[0047] The data structure for enhancing point cloud features is a 128-dimensional vector, of which 32 dimensions are geometric detail features calculated from local neighborhoods, 48 ​​dimensions are structural features obtained by aggregating contextual information from a larger region, and 48 dimensions are semantic features output by the semantic prediction unit. During feature fusion, the attention mechanism calculates the correlation between various features and the current reconstruction task to obtain weight coefficients. The weight coefficients range from 0 to 1, and the sum of the weight coefficients of all features is 1. Finally, the enhanced point cloud features are generated through weighted summation, ensuring that various features adaptively play their role according to the reconstruction requirements.

[0048] The implicit reconstruction unit constructs a neural network that takes coordinate points as input to represent a continuous implicit scalar field. This network takes the aforementioned enhanced point cloud features as conditional input. Specifically, for any point in space, its enhanced features are obtained from the features of its nearest few point cloud points through trilinear interpolation and input together with the spatial coordinates of the point into a multilayer perceptron. The perceptron outputs a signed distance value representing whether the point is inside or outside the object. By performing dense sampling and computation throughout the three-dimensional space and applying the moving cube algorithm to extract the zero isosurface, a smooth, continuous, and seamless three-dimensional triangular mesh surface model is generated.

[0049] This implementation method abandons the traditional point cloud reconstruction method that relies on a single geometric feature or directly performs surface triangulation. By introducing multi-path feature extraction and adaptive fusion based on attention mechanism, it ensures that all levels of information, from local details to global structure and high-level semantics, can effectively and collaboratively guide the reconstruction process. In particular, the introduction of semantic features enables the reconstruction process to "understand" the boundaries of building components, thereby generating more accurate geometric results in complex junction areas, including window-wall joints. The conditional implicit field reconstruction method fundamentally guarantees the integrity and watertightness of the output model.

[0050] The 3D model generation module also includes a topology verification and repair unit; The topology verification and repair unit, connected to the implicit reconstruction unit, is used to perform automated topology analysis on the generated 3D mesh surface model, detect and identify non-manifold edges, isolated vertices, or self-intersecting faces in the model; based on the semantic labels output by the semantic and geometric joint reasoning module, it performs category-aware automatic repair of the identified topology errors. The repair includes at least bridging broken edges belonging to the same semantic component and filling small-scale holes belonging to non-structural semantic categories caused by noise removal.

[0051] In the topology verification and repair unit, the minimum spanning tree algorithm is used for bridging broken edges. By calculating the Euclidean distance between the endpoints of the broken edges, a minimum spanning tree is constructed to achieve smooth bridging of the edges. During the bridging process, the continuity and geometric consistency of the edge connections are ensured. The area threshold for filling small-scale holes is set to 0.1 square meters. For holes with an area smaller than this threshold and belonging to the non-structural semantic category, a triangular patch filling algorithm is used. Based on the topological relationship and geometric features of the hole boundary, the optimal triangular patch is generated to complete the hole filling while maintaining a smooth transition of the surface.

[0052] It should be further explained that the specific implementation of the topology verification and repair unit is as follows: This unit directly receives the 3D mesh model generated by the implicit reconstruction unit and the semantic label information associated with each vertex, and executes an automated, phased processing flow. First, the unit performs topological analysis on the mesh model. By traversing all mesh edges, vertices, and faces, it applies standard computational geometry algorithms to detect and identify four main types of defects: non-manifold edges, isolated vertices, self-intersecting faces, and small-scale holes with a size smaller than a preset area threshold caused by excessive denoising or insufficient sampling. Non-manifold edges: refer to edges shared by three or more faces; isolated vertices: refer to vertices that do not belong to any valid face; self-intersecting faces: refer to self-overlapping of the model surface at non-boundary locations.

[0053] Subsequently, the unit enters the category-aware automatic repair stage, where the repair strategy is strictly guided by the input semantic information. In category-aware automatic repair, the core logic of "category awareness" is to select an appropriate repair strategy based on the semantic label of the location of the topological error: for structural semantic categories such as walls, roofs, doors, and windows, a high-precision repair strategy is adopted, prioritizing structural integrity and geometric accuracy during the repair process; for example, the accuracy of bridging fracture edges is controlled within 0.005 meters. For non-structural semantic categories such as decorative components, an efficient repair strategy is adopted, improving repair efficiency while ensuring visual appeal; edge bridging accuracy is controlled within 0.01 meters, achieving a balance between repair effectiveness and efficiency.

[0054] For non-manifold edges and self-intersecting surfaces, the element is not simply deleted or merged geometrically. Instead, the semantic labels associated with the vertices or edges involved are first checked: if they belong to the same building component category, such as both being "walls", the system determines that they are imperfect connections within the same component and performs local mesh re-division and smoothing operations to eliminate defects while maintaining the overall geometry of the area; if they belong to different component categories, such as the boundary between "walls" and "windows", the system will retain the boundary and only perform minimal compliance processing on non-manifold states to ensure that the distinction between components is preserved.

[0055] For small-scale holes, the unit also determines the repair method based on the mainstream semantic labels of the hole boundary vertices: if the label belongs to the structural or major component category, such as "wall" or "roof", then smooth completion is performed based on the Poisson equation of the surrounding mesh to ensure the continuity of the surface; if the label belongs to the decorative or non-critical category, such as "decorative component", and the hole area is less than a certain threshold, then a simpler triangular patch filling algorithm is allowed to improve processing efficiency.

[0056] This implementation combines traditional post-retro geometric repair with high-level semantic understanding, transforming the repair process from blind geometric optimization to semantically based intelligent decision-making. This effectively avoids blurring or destroying important geometric boundaries between different building components while correcting topological errors, ensuring that the final 3D model has both strict topological correctness and maintains clear component division and realistic detail representation.

[0057] The initial spatial division in the preprocessing module is as follows: based on the height distribution histogram of the building in the vertical direction extracted from the registered point cloud, obvious clusters of floor heights are identified; using the identified floor heights as a reference, combined with the changes in point density in the horizontal plane, the point cloud data is adaptively divided into multiple three-dimensional blocks aligned with the building structure.

[0058] It should be further explained that the specific implementation method of the initial spatial division in the preprocessing module is as follows: After completing the coordinate system registration of the multi-site cloud data, the system first analyzes the height values ​​of all points along the approximate vertical direction of the building, usually the Z-axis of the world coordinate system. By constructing a height distribution histogram and applying a peak detection algorithm, the system identifies the intervals with obvious bulges in the histogram. These intervals correspond to the dense presentation of horizontal structures such as floor slabs and roofs in the point cloud, and are thus determined as potential floor heights. After obtaining these height intervals, the system expands a preset buffer range above and below each identified floor height interval as the center, forming a three-dimensional horizontal slice layer.

[0059] Subsequently, within each horizontal slice layer, the system performs adaptive subdivision in the two-dimensional plane: by calculating the two-dimensional local point density of each point on the horizontal plane, and using a density-based clustering algorithm such as DBSCAN to cluster the horizontal coordinates, the point cloud is divided into multiple continuous regions on the horizontal plane. The point density within each region is relatively uniform and usually corresponds to a complete building component, such as a wall or a room. Finally, the vertical floor slices are combined with the density clustering regions in the horizontal plane to form multiple three-dimensional blocks that are naturally aligned with the building structure.

[0060] This implementation method abandons the traditional method of uniform bounding box division or random sampling. By actively mining the hidden architectural structure patterns in point cloud data, namely the vertical layering of floors and the horizontal aggregation of components, it achieves a high degree of consistency between the division results and the actual physical structure of the building. This provides a more accurate and meaningful computing unit for all subsequent local space-based processing, such as noise filtering and feature extraction, fundamentally improving the accuracy and efficiency of subsequent processes.

[0061] The adaptive spatial partitioning in the primary filtering unit adopts a non-uniform partitioning strategy: small partition sizes are used in areas with high point cloud density, and large partition sizes are used in areas with low point cloud density; the partition boundaries are generated by an iterative clustering method to maintain the consistency of point cloud geometric attributes within a single partition.

[0062] It should be further explained that the specific implementation of adaptive spatial partitioning in the primary filtering unit is as follows: The unit first calculates the 3D spatial bounding box of the input point cloud and divides it into uniform grids based on an initial reference size; then, the unit counts the number of points contained in each grid cube to evaluate its local point cloud density, and performs non-uniform refinement and merging operations based on this: for grids with more than a high density threshold, the unit further subdivides them in three dimensions to ensure that high detail areas can be covered by smaller, finer partitions; high detail areas include richly decorated facades and window frames; conversely, for multiple adjacent grids with fewer than a low density threshold, the unit merges them into a larger partition to cover large areas with sparse but structurally coherent point clouds, such as flat walls and roofs.

[0063] To achieve natural alignment of block boundaries with point cloud distribution and maintain geometric consistency within blocks, the unit employs an iterative clustering method to optimize the boundaries of the initially divided blocks: based on the point set within each block, the statistical characteristics of the principal direction or curvature of its normal vector distribution are calculated as clustering features. Points located at the edge of the current block whose feature values ​​are closer to the central features of adjacent blocks are redistributed. Through multiple iterations until the block boundaries stabilize, a series of geometric attributes of different sizes and irregular shapes but with consistent internal point cloud properties are finally formed, such as three-dimensional partitions with consistent surface orientation and curvature changes.

[0064] The segmentation strategy of this implementation is not static or based solely on simple rules. Instead, it drives the adaptive subdivision and merging of the mesh by dynamically evaluating local density. Combined with iterative clustering optimization of the boundary based on geometric attributes, each processing unit corresponds to a geometrically homogeneous physical surface segment to the greatest extent possible. This provides a highly reliable analytical basis for subsequent threshold calculations based on local statistics, such as density and curvature thresholds, thereby improving the accuracy and reliability of primary filtering in complex architectural scenes with extremely uneven point cloud density distribution.

[0065] The neural network model used in the semantic prediction unit employs a composite loss function that integrates geometric loss and semantic loss during training. The geometric loss term encourages the network to learn features that are invariant to changes in the local geometry of the point cloud, while the semantic loss term supervises the network to accurately predict the semantic category of the points.

[0066] It should be further explained that the specific implementation of the neural network training process in the semantic prediction unit is as follows: The deep neural network model used for semantic prediction does not only use the standard classification cross-entropy loss during its training process, but also adopts a composite loss function composed of a weighted sum of a geometric consistency loss term and a semantic classification loss term, in order to simultaneously optimize the network's ability to understand the local geometric structure of the point cloud and its ability to distinguish semantic categories.

[0067] In the composite loss function, the weighting coefficients for the geometric loss term and the semantic loss term are set to 0.3 and 0.7, respectively. When calculating the geometric loss term, the intermediate layer feature vectors before and after the transformation are first subjected to L2 normalization to normalize the vector magnitude to 1. Then, the cosine similarity value is obtained by calculating the dot product of the two sets of normalized vectors. Finally, the negative value of this value is taken as the result of the geometric loss term. The semantic loss term adopts weighted cross-entropy loss. The weight allocation adopts the inverse proportional allocation principle according to the sample quantity ratio of each building component category. The category with fewer samples has a larger weight coefficient, so as to balance the training weights of different categories of samples and avoid model prediction bias due to sample imbalance.

[0068] The geometric loss term is designed to force the network to learn feature representations that are invariant to a certain degree to rigid transformations common in local regions of point clouds, such as translation, rotation, and non-rigid deformation. In the specific implementation, for each point cloud sample in the training batch, after obtaining the intermediate layer feature vector of each point through the feature extraction layer of the network, a random spatial transformation, such as a small-amplitude rotation or elastic deformation, is applied to the point cloud, and features are extracted again. The geometric loss term calculates the negative cosine similarity between the intermediate layer feature vectors of corresponding points of the same original point cloud before and after the transformation. By minimizing this loss, the features extracted by the network are encouraged to remain stable to the above geometric perturbations.

[0069] The semantic classification loss term employs standard weighted cross-entropy loss, which directly supervises the consistency between the semantic category probability distribution output by the network and the manually labeled real labels. The weights are used to balance the imbalance in the frequency of different semantic categories in the dataset. During training, the geometric loss term and the semantic classification loss term are linearly combined according to a preset fixed ratio to form the overall training objective. All network parameters are jointly optimized through backpropagation, ensuring that after training, the extracted point features not only contain discriminative semantic information but also encode robust local geometric context. This enables the network to output more accurate and robust semantic prediction results when faced with noise, occlusion, and density variations common in real-world scanning.

[0070] This implementation deeply integrates the self-supervised geometric consistency learning paradigm with the supervised semantic learning paradigm in point cloud segmentation tasks. By introducing geometric loss as an intrinsic, data-driven regularization constraint, it effectively improves the quality and generalization ability of network features, enabling it to better distinguish components with similar geometric structures but different semantics, such as distinguishing between walls and floors, as well as components with the same semantics but different geometric representations, such as windows of different shapes. This enhances the semantic parsing accuracy of the entire system for complex architectural scenes.

[0071] The system also includes an output module connected to the 3D model generation module. The output module is used to convert the generated 3D building scene model with complete topological structure and semantic annotation into a file that conforms to industrial basic standards or general 3D format, and to associate and store the semantic labels, geometric parameters and topological relationship information of each component in the model.

[0072] Further explanation is needed regarding the specific implementation of the output module: This module receives the final output from the 3D model generation module, namely a 3D architectural model with complete triangular mesh geometry and vertex-level semantic tags. The module's function is to perform a structured data conversion and encapsulation operation to generate standardized files that can be directly recognized and used by downstream professional software. The conversion process first analyzes the input mesh model, extracting all triangular faces, vertex coordinates, and the connection relationships between faces and vertices. Based on the semantic tags carried by each vertex, it automatically groups the triangular face sets belonging to the same continuous semantic region, identifying them as independent architectural component instances. Subsequently, the module provides at least two output paths: the first path is to construct a data model according to industrial basic standards, and for each identified... The first path involves creating a corresponding IFC entity for each component instance. For example, a group of components with the semantic label "wall" is mapped to an IfcWall entity, and "window" is mapped to an IfcWindow entity. Following the IFC pattern requirements, the geometric representation attributes, spatial location attributes, and connection relationship attributes with other components of this entity are accurately set, ultimately generating a complete IFC file. The second path involves generating a common 3D exchange format file, such as OBJ or FBX. In these files, by defining custom user data blocks or utilizing the extended attribute mechanism supported by the format, the component semantic labels to which each triangular facet belongs, the component geometric parameters obtained based on mesh calculations (such as area, volume, bounding box size), and the topological adjacency relationships between components are tightly associated with and embedded into the file in the form of key-value pairs or metadata.

[0073] This implementation method significantly extends the traditional 3D reconstruction process by stopping at the geometric mesh output node. By designing automated semantic mapping, geometric parameter calculation, and standardized encapsulation pipelines, it achieves direct conversion from internal data consisting of "geometry + semantics" to standard exchange files "rich in structured engineering information." This makes the system's output no longer just a model for visual display, but an engineering data carrier that can be immediately used for building information model analysis, quantity surveying, digital asset management, and maintenance, thus breaking down the key barriers between real-world 3D reconstruction results and the information technology practices of the construction industry.

[0074] This system effectively addresses the contradiction between inaccurate noise removal and easy loss of feature details in complex architectural scenes by introducing a multi-level adaptive noise filtering and semantic-geometric joint reasoning mechanism. The system can intelligently distinguish between real discrete noise and the inherent complex geometric features of buildings, ensuring denoising effect while preserving the integrity of details such as decorative components and curved roofs to the greatest extent, thus achieving high-fidelity point cloud data preprocessing.

[0075] The system generates 3D models with both geometric accuracy and rich semantic structure through deep reconstruction that integrates multi-scale features and semantic information, as well as topology-aware model optimization and standardized output. This achievement can be directly integrated into the Building Information Modeling (BIM) workflow, providing a high-quality and directly usable data foundation for downstream applications such as architectural design, digital archiving of cultural heritage, and smart city management, thereby improving the overall efficiency and reliability from data acquisition to engineering applications.

[0076] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A 3D reconstruction system for building scenes based on point cloud data, characterized in that, It includes a preprocessing module, a multi-level adaptive noise filtering module, a semantic and geometric joint reasoning module, and a 3D model generation module connected in sequence; The preprocessing module is used to receive the original building point cloud data, perform registration and initial spatial division, and output the registered point cloud dataset. The multi-level adaptive noise filtering module is connected to the output of the preprocessing module and is used to perform a processing flow containing at least two different filtering mechanisms on the point cloud dataset to identify and label potential noise points. The first filtering mechanism filters based on the local geometric properties of the point cloud, and the second filtering mechanism filters based on the distribution consistency of the point cloud in the feature space. The semantic and geometric joint reasoning module is connected to the output of the multi-level adaptive noise filtering module. It is used to receive filtered point cloud data, simultaneously perform multi-type geometric primitive fitting and semantic category prediction, and perform secondary verification and correction on the potential noise points based on the correlation between the fitting results and the semantic prediction results, so as to generate denoised point cloud data with semantic labels. The 3D model generation module is connected to the output of the semantic and geometric joint reasoning module. It is used to generate a 3D model of the building scene with complete topological structure and semantic annotation based on the denoised and semantically labeled point cloud data, by fusing multi-scale geometric features and semantic context information.

2. The architectural real-scene 3D reconstruction system based on point cloud data according to claim 1, characterized in that: The multi-level adaptive noise filtering module includes a primary filtering unit and a secondary filtering unit; The primary filtering unit is used to execute the first filtering mechanism, and its specific configuration is as follows: adaptively spatially divide the point cloud dataset into blocks, calculate the local point density and curvature value of each point in each block; mark points with local point density lower than a first dynamic threshold and curvature value higher than a second dynamic threshold as primary suspected noise points; the first dynamic threshold and the second dynamic threshold are adaptively determined according to the statistical distribution of the point cloud in the block. The secondary filtering unit is connected to the primary filtering unit and is used to execute the second filtering mechanism. Its specific configuration is as follows: extracting multi-dimensional feature vectors from the point cloud dataset that are not marked as primary suspected noise points. The multi-dimensional feature vectors contain at least spatial coordinates, normal vectors, and color information. Cluster analysis is performed on the multidimensional feature vectors in the feature space; points belonging to clusters with fewer than a preset threshold number of members, and points whose distance from the center of the primary cluster exceeds a dynamic distance threshold, are marked as secondary suspected noise points; the union of the primary suspected noise points and the secondary suspected noise points constitutes the potential noise point set.

3. The architectural real-scene 3D reconstruction system based on point cloud data according to claim 2, characterized in that: The semantic and geometric joint reasoning module includes a geometric primitive fitting unit, a semantic prediction unit, and a joint decision-making unit; The geometric primitive fitting unit is used to perform parallel fitting of three types of geometric primitives—planar, cylindrical, and spherical—for each local region in the filtered point cloud data, and to calculate the residual of each point relative to each type of fitting primitive. The semantic prediction unit, based on a neural network model, performs semantic category prediction on each point in the filtered point cloud data and outputs the probability distribution of each point belonging to multiple predefined building component categories. The building component categories include at least walls, roofs, windows, doors, and decorative components. The joint decision-making unit is connected to the geometric primitive fitting unit and the semantic prediction unit, respectively, and is used to make decisions based on the residuals and the probability distribution. Its configuration is as follows: for a target point marked as a potential noise point, find several neighboring points within its spatial neighborhood; if the fitting residuals of points belonging to the same semantic category among the neighboring points for a certain type of geometric primitive are generally lower than the residual threshold, while the fitting residual of the target point for that type of geometric primitive is higher than the residual threshold, and the predicted semantic category probability distribution of the target point is inconsistent with the mainstream semantic category of the neighboring points, then the target point is confirmed as a final noise point; otherwise, the target point is removed from the potential noise point set.

4. The architectural real-scene 3D reconstruction system based on point cloud data according to claim 3, characterized in that: The system also includes an iterative optimization module, which is connected between the multi-level adaptive noise filtering module and the semantic and geometric joint reasoning module to form a closed feedback loop; The iterative optimization module is configured to: after the joint decision-making unit outputs the final noise point set, feed the final noise point set back to the primary filtering unit; the primary filtering unit dynamically adjusts the determination strategy for the first dynamic threshold and the second dynamic threshold in the next round of processing based on the spatial distribution and attribute characteristics of the final noise points confirmed in this iteration; the system executes at least two complete filtering and inference processes that include the feedback loop.

5. The architectural real-scene 3D reconstruction system based on point cloud data according to claim 3, characterized in that: The 3D model generation module includes a feature fusion unit and an implicit reconstruction unit; The feature fusion unit is used to extract multi-scale features from the denoised and semantically labeled point cloud data. The multi-scale features include geometric detail features calculated from local neighborhoods, structural features obtained by aggregating contextual information from a larger region, and semantic features output by the semantic prediction unit. The feature fusion unit uses an attention mechanism to weightedly fuse the geometric detail features, structural features, and semantic features to generate enhanced point cloud features. The implicit reconstruction unit, connected to the feature fusion unit, is used to take the enhanced point cloud features as input and generate a continuous, seamless three-dimensional mesh surface model by solving the implicit field function.

6. The architectural real-scene 3D reconstruction system based on point cloud data according to claim 5, characterized in that: The 3D model generation module also includes a topology verification and repair unit; The topology verification and repair unit, connected to the implicit reconstruction unit, is used to perform automated topology analysis on the generated 3D mesh surface model, detect and identify non-manifold edges, isolated vertices, or self-intersecting faces in the model; based on the semantic tags output by the semantic and geometric joint reasoning module, it performs category-aware automatic repair of the identified topology errors. The repair includes at least bridging broken edges belonging to the same semantic component and filling small-scale holes belonging to non-structural semantic categories caused by noise removal.

7. The architectural real-scene 3D reconstruction system based on point cloud data according to claim 1, characterized in that: The initial spatial division in the preprocessing module is as follows: based on the height distribution histogram of the building in the vertical direction extracted from the registered point cloud, obvious clusters of floor heights are identified; using the identified floor heights as a reference, and combined with the changes in point density in the horizontal plane, the point cloud data is adaptively divided into multiple three-dimensional blocks aligned with the building structure.

8. A three-dimensional reconstruction system for building scenes based on point cloud data according to claim 2, characterized in that: The adaptive spatial segmentation in the primary filtering unit adopts a non-uniform partitioning strategy: small segment sizes are used in areas with high point cloud density, and large segment sizes are used in areas with low point cloud density.

9. A three-dimensional reconstruction system for building scenes based on point cloud data according to claim 3, characterized in that: The neural network model used in the semantic prediction unit employs a composite loss function that integrates geometric loss and semantic loss during training. The geometric loss term encourages the network to learn features that are invariant to changes in the local geometric structure of the point cloud, while the semantic loss term supervises the network to accurately predict the semantic category of the points.

10. A three-dimensional reconstruction system for building scenes based on point cloud data according to claim 1, characterized in that: The system also includes an output module connected to the 3D model generation module; the output module is used to convert the generated 3D building scene model with complete topological structure and semantic annotation into a file that conforms to industrial basic standards or general 3D format, and to associate and store the semantic tags, geometric parameters and topological relationship information of each component in the model.

Images