Inclined photography model monomer extraction and white mold construction system based on AI driving

Abstract

The invention discloses an AI-driven oblique photography model monomer extraction and white mold construction system, and relates to the technical field of model monomer, a construction point cloud analysis module extracts three-dimensional point cloud data and corresponding geometric information and texture information from an oblique photography model, and generates initial point cloud data; segmenting the initial point cloud data based on an AI semantic segmentation module to generate building point cloud regions, and performing clustering analysis on the building point cloud regions to generate single building point cloud clusters; a geometric contour recognition module is used for analyzing the projection outer contour and elevation parameters of the single building, and a three-dimensional white pattern corresponding to the single building point cloud cluster is generated based on a white pattern construction module by using the projection outer contour and the elevation parameters; according to the method, the precision, the efficiency and the structuring degree of three-dimensional modeling of the oblique photography model are remarkably improved by constructing a full-process automatic processing chain from the oblique photography model to the parameterized and componentized white model.

Description

Technical Field

[0001] This invention relates to the field of model individualization technology, specifically to an AI-driven oblique photogrammetry model individualization extraction and white model construction system. Background Technology

[0002] With the development of UAV oblique photogrammetry and 3D modeling technology, urban 3D models, such as oblique photogrammetry models, have been widely used in smart cities, land surveying, BIM, digital twins, urban planning and other fields. However, existing oblique photogrammetry 3D models are mostly monolithic texture models, lacking effective separation and semantic information of individual buildings. This leads to shortcomings in subsequent applications such as 3D GIS, spatial analysis, and urban renewal simulation. Existing methods mostly rely on manual annotation or simple rules based on height thresholds for building segmentation. This approach not only has low accuracy and weak generalization ability but also struggles to adapt to changes in building morphology in complex urban scenes. Furthermore, in terms of building boundary extraction, existing projection algorithms often fail to accurately identify the projected outer contours of buildings due to the effects of occlusion, reflection, and noise on oblique photogrammetry point clouds. This results in morphological distortion in the subsequently constructed white model, failing to accurately reflect the actual geometric features of the building. In addition, traditional processing workflows involve numerous manual intervention steps, including manual vectorization, modeling, and annotation, making the entire process inefficient and costly, unable to meet the needs of automated processing of city-level models. Secondly, existing technologies lack the ability to deeply understand the semantics of oblique photogrammetry data. Most methods cannot perform fine-grained identification and separation of building components in complex scenes, resulting in 3D models lacking rich semantic information. Therefore, it is of great significance to develop an AI-driven oblique photogrammetry model individualization extraction and white model construction system. Summary of the Invention

[0003] The purpose of this invention is to provide an AI-driven oblique photogrammetry model individualization extraction and white model construction system to solve the problems in the background technology.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an AI-driven oblique photogrammetry model individualization extraction and white model construction system, comprising: Point cloud parsing module: used to extract 3D point cloud data and corresponding geometric and texture information from oblique photogrammetry models, and generate initial point cloud data; AI semantic segmentation module: Connected to the point cloud parsing module, it enhances the initial point cloud data to generate enhanced point cloud data, pre-trains a deep learning model using the enhanced point cloud data, and uses the trained deep learning model to perform semantic segmentation on the initial point cloud data to generate building point cloud regions. Individual building clustering module: Connected to the AI ​​semantic segmentation module, it performs spatial clustering analysis on building point cloud regions, aggregating point clouds belonging to the same building entity into independent clusters, generating individual building point cloud clusters; Geometric contour recognition module: connected to the individual building clustering module, it calculates the projected outer contour on the horizontal plane for each individual building point cloud cluster and extracts the elevation parameters of the individual building. White model construction module: Connected to the geometric contour recognition module, it generates the corresponding 3D white model based on the projected outer contour and elevation parameters through a programmed modeling method.

[0005] In a preferred embodiment, the point cloud parsing module includes: The data reading unit is used to read the oblique photogrammetry model file and obtain vertex coordinates, patch indexes, and texture mapping information. The point cloud information extraction unit is used to extract the spatial coordinate data of vertices from the oblique photogrammetry model file, construct the basic geometric structure of the point cloud, obtain the texture information corresponding to the vertices, and generate a point cloud with texture information. The data encapsulation unit is used to fuse the extracted basic geometric structure and color texture information, convert it into standard point cloud data containing spatial coordinates and color texture information, and encapsulate the standard point cloud data to output initial point cloud data containing geometric and texture information.

[0006] In a preferred embodiment, the AI ​​semantic segmentation module includes: The data augmentation unit is used to perform data augmentation operations on the initial point cloud data to generate augmented point cloud data. The semantic segmentation unit, based on a deep learning network model pre-trained on an augmented training dataset, performs semantic segmentation on the input initial point cloud data, identifies point clouds belonging to the building category, and outputs the building point cloud region.

[0007] In a preferred embodiment, the individual building clustering module includes: Pre-clustering units are used to perform spatial clustering analysis on building point cloud regions based on density, and to identify continuous point cloud regions with spatial continuity as a set of pre-clustering blocks. The boundary candidate region detection unit is used to extract the boundary shell point cloud of each pre-clustered block, calculate the spatial distance between the boundary shell point clouds of different pre-clustered blocks, identify the neighboring point pair regions with a spatial distance less than a preset threshold, and generate boundary candidate regions. The boundary point cloud attribution module is used to calculate the multi-dimensional features of point clouds in each boundary candidate region, including spatial coordinates, normal vector features, and color texture features. The feature affinity between each point cloud and its neighboring points in the boundary candidate region is calculated based on multi-dimensional features. The feature affinity includes spatial distance factor, normal direction consistency and color and texture similarity. The iterative attribution allocation sub-unit is used to segment the point cloud in the boundary candidate region through multiple rounds of iteration. Establish an attribution weight vector for each point cloud within the boundary candidate region, and record the initial attribution probability of the point cloud to each adjacent pre-clustered block; In each iteration, each point cloud sends a home weight update message to its neighboring points based on its feature affinity relationship with neighboring points; Each point aggregates all received weight update messages, combines them with its own feature information, and updates the weight assigned to each pre-cluster block. The iteration process terminates when the change in the weight of all points is less than a preset threshold or when the maximum number of iterations is reached. Based on the iteratively generated attribution weight vector, each point cloud in the boundary candidate region is assigned to the pre-clustered block with the largest weight; The result generation unit merges the boundary candidate area point clouds that have completed the assignment into the corresponding pre-clustered blocks to generate a single building point cloud cluster containing geometric boundaries. Simultaneously, based on the segmentation standard point cloud pairs generated during the iteration process, they are stored in the segmentation standard library.

[0008] In a preferred embodiment, the step of simultaneously generating segmentation standard point cloud pairs based on the iteration process and storing them in the segmentation standard library is as follows: During the iterative attribution assignment process, successfully segmented boundary point cloud pairs are identified, and multi-dimensional feature combinations of the boundary point cloud pairs are extracted as segmentation feature templates. The multidimensional feature combination includes spatial distribution features, normal direction relationship, color texture features, and local geometric context features; The segmentation feature templates and their corresponding segmentation schemes are stored in the segmentation standard library, and a multi-dimensional index structure is built for the segmentation feature templates. The multi-dimensional index structure is constructed based on building type, boundary morphology features and geospatial location. When processing new boundary candidate regions, the feature affinity of the current boundary candidate region point cloud is matched with the segmentation feature templates in the segmentation standard library. When the matching degree exceeds the preset threshold, the corresponding segmentation scheme is directly called.

[0009] In a preferred embodiment, the geometric contour recognition module includes: The projection conversion unit is used to project the three-dimensional spatial point cloud of a single building point cloud cluster onto a horizontal plane to form a two-dimensional projection point set. The contour vectorization unit generates the projected outer contour of a single building by using a boundary extraction algorithm based on a two-dimensional projection point set; The elevation calculation unit calculates the building footprint, altitude, and building height of a single building as elevation parameters based on a two-dimensional projection point set. The data encapsulation unit is used to associate the projected outer contour with elevation parameters to generate standardized vector data.

[0010] In a preferred embodiment, the white model construction module includes: The architectural syntax parsing unit is used to receive standardized vector data, parse the building structure type based on the projected outer contour and elevation parameters, and deconstruct the building into a parameterized sequence of building components. The parameterized component generation unit generates geometric models of building components containing semantic information based on a sequence of building components. The component assembly unit is used to assemble the generated building components according to the building structure logic to form a complete three-dimensional white model.

[0011] In a preferred embodiment, the architectural syntax parsing unit includes: The building structure analysis sub-unit is used to analyze the geometric features and elevation parameters of the projected outer contour in standardized vector data, and classifies individual buildings into multiple basic geometric types through contour shape recognition algorithms. The component storage sub-unit establishes a component association network based on multiple basic geometric types of a single building and records the spatial constraint relationships between components; The component sequence generates sub-units, and based on the component association network and spatial constraints, generates a component assembly sequence that conforms to the building construction logic.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention achieves precise control over building point cloud segmentation from coarse-grained to fine-grained by constructing a dual analysis mechanism of pre-clustered block set and boundary candidate region. It adopts a multi-dimensional feature affinity calculation and iterative attribution allocation strategy, and establishes a dynamic weight update mechanism in the boundary candidate region, which effectively solves complex scenarios such as adhered buildings and shared wall buildings that are difficult to handle by traditional clustering algorithms. At the same time, after each successful segmentation of the boundary point cloud, the system extracts the multi-dimensional feature combination of the corresponding boundary point cloud pair to form a reusable segmentation feature template, and constructs a multi-dimensional index structure through building type, boundary morphology and geospatial location. This mechanism enables the system to quickly match historical experience when processing new data, which greatly improves processing efficiency. The design of the dual analysis mechanism and feature reuse mechanism enables the system to significantly improve efficiency while maintaining processing accuracy, achieving simultaneous improvement in segmentation quality and processing efficiency. 2. This invention achieves intelligent deconstruction and reorganization of building structures through a building grammar parsing unit. This rule-based structural parsing method can not only accurately identify key components such as the main structure and roof form of a building, but also establish a spatial constraint relationship network between components, ensuring that the generated component sequence conforms to the logical rules of building construction. On this basis, the system realizes the continuous accumulation and efficient reuse of building knowledge through component storage and sequence generation mechanisms. Based on self-learning capabilities and a multi-level index structure, it continuously optimizes the component library content and assembly scheme, effectively improving processing efficiency and ensuring the consistency of modeling quality. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0014] Figure 1 This is a system flowchart of the present invention.

[0015] Figure 2 This is a logic block diagram of the present invention. Detailed Implementation

[0016] 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 only some embodiments of the present invention, 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.

[0017] Example 1, please refer to Figure 1 and Figure 2 As shown in this embodiment, the AI-driven oblique photogrammetry model individualization extraction and white model construction system includes... Point cloud parsing module: used to extract 3D point cloud data and corresponding geometric and texture information from oblique photogrammetry models, and generate initial point cloud data; AI semantic segmentation module: Connected to the point cloud parsing module, it enhances the initial point cloud data to generate enhanced point cloud data, pre-trains a deep learning model using the enhanced point cloud data, and uses the trained deep learning model to perform semantic segmentation on the initial point cloud data to generate building point cloud regions. Individual building clustering module: Connected to the AI ​​semantic segmentation module, it performs spatial clustering analysis on building point cloud regions, aggregating point clouds belonging to the same building entity into independent clusters, generating individual building point cloud clusters; Geometric contour recognition module: connected to the individual building clustering module, it calculates the projected outer contour on the horizontal plane for each individual building point cloud cluster and extracts the elevation parameters of the individual building. White model construction module: Connected to the geometric contour recognition module, it generates the corresponding 3D white model based on the projected outer contour and elevation parameters through a programmatic modeling method; Furthermore, with the development of UAV oblique photogrammetry and 3D modeling technology, urban 3D models, such as oblique photogrammetry models, have been widely used in smart cities, land surveying, BIM, digital twins, urban planning and other fields. However, existing oblique photogrammetry 3D models are mostly monolithic texture models, lacking effective separation and semantic information of individual buildings. This leads to shortcomings in subsequent applications such as 3D GIS, spatial analysis, and urban renewal simulation. Existing methods mostly rely on manual annotation or simple rules based on height thresholds for building segmentation. This approach not only has low accuracy and weak generalization ability but also struggles to adapt to changes in building morphology in complex urban scenes. Furthermore, in terms of building boundary extraction, existing projection algorithms often fail to accurately identify the projected outer contours of buildings due to the effects of occlusion, reflection, and noise on oblique photogrammetry point clouds. This results in morphological distortion in the subsequently constructed white model, failing to accurately reflect the actual geometric features of the building. In addition, traditional processing workflows involve numerous manual intervention steps, including manual vectorization, modeling, and annotation, making the entire process inefficient and costly, unable to meet the needs of automated processing of city-level models. Secondly, existing technologies lack the ability to deeply understand the semantics of oblique photogrammetry data. Most methods cannot perform fine-grained identification and separation of building components in complex scenes, resulting in 3D models lacking rich semantic information. This invention achieves precise control over building point cloud segmentation from coarse-grained to fine-grained by constructing a dual analysis mechanism of pre-clustered block sets and boundary candidate regions. It employs a multi-dimensional feature affinity calculation and iterative attribution allocation strategy, and establishes a dynamic weight update mechanism in the boundary candidate regions. This effectively solves complex scenarios such as adhered buildings and buildings with shared walls that are difficult for traditional clustering algorithms to handle. At the same time, after each successful segmentation of the boundary point cloud, the system extracts the multi-dimensional feature combination of the corresponding boundary point cloud pairs to form a reusable segmentation feature template. It also constructs a multi-dimensional index structure through building type, boundary morphology, and geospatial location. This mechanism enables the system to quickly match historical experience when processing new data, greatly improving processing efficiency. The design of the dual analysis mechanism and feature reuse mechanism enables the system to significantly improve efficiency while maintaining processing accuracy, achieving a simultaneous improvement in segmentation quality and processing efficiency. Meanwhile, the system achieves intelligent deconstruction and reorganization of building structures through the architectural grammar parsing unit. This rule-based structural parsing method can not only accurately identify key components such as the main structure and roof form of the building, but also establish a spatial constraint relationship network between components to ensure that the generated component sequence conforms to the logical rules of building construction. On this basis, the system realizes the continuous accumulation and efficient reuse of architectural knowledge through component storage and sequence generation mechanisms. Based on self-learning ability and multi-level index structure, the system continuously optimizes the content of the component library and assembly scheme, effectively improving processing efficiency and ensuring the consistency of modeling quality.

[0018] In one embodiment, the point cloud parsing module includes: The data reading unit is used to read the oblique photogrammetry model file and obtain vertex coordinates, patch indexes, and texture mapping information. The point cloud information extraction unit is used to extract the spatial coordinate data of vertices from the oblique photogrammetry model file, construct the basic geometric structure of the point cloud, obtain the texture information corresponding to the vertices, and generate a point cloud with texture information. The data encapsulation unit is used to fuse the extracted basic geometric structure and color texture information, convert it into standard point cloud data containing spatial coordinates and color texture information, and encapsulate the standard point cloud data to output initial point cloud data containing geometric and texture information. Furthermore, the data reading unit, adhering to the format specifications of oblique photogrammetry model files, uses the trimesh and pyntcloud libraries to read common 3D model formats such as OBJ and PLY, accurately extracting vertex coordinate data, triangular facet index relationships, and corresponding texture mapping coordinates. Simultaneously, it decodes texture image files in formats such as JPEG and PNG using the OpenCV library, establishing a mapping table between geometric and texture data. The point cloud information extraction unit performs matrix processing on vertex coordinates based on NumPy array operations, constructing a point cloud spatial topology. Through the mapping relationship between UV texture coordinates and vertices, it uses bilinear interpolation to assign corresponding RGB color values ​​to each vertex and employs principal component analysis to calculate local surface normals of the point cloud, ultimately generating a complete point cloud dataset containing spatial coordinates, color information, and normal vectors. The data encapsulation unit utilizes the point cloud processing functions of the PCL or Open3D libraries to standardize and fuse the extracted geometric and texture attributes. During storage, it optimizes the data storage structure by constructing a KD-tree spatial index, reduces data volume using an octree compression algorithm, and establishes a fast data retrieval mechanism based on a hash table, ultimately outputting a point cloud data file format conforming to industry standards as the initial point cloud data.

[0019] In one embodiment, the AI ​​semantic segmentation module includes: The data augmentation unit is used to perform data augmentation operations on the initial point cloud data to generate augmented point cloud data. The semantic segmentation unit, based on a deep learning network model pre-trained on an augmented training dataset, performs semantic segmentation on the input initial point cloud data, identifies point clouds belonging to the building category, and outputs the building point cloud region. Furthermore, the data augmentation unit generates augmented point cloud data by performing a series of geometric transformations and attribute perturbations on the initial point cloud data. The data augmentation unit systematically preprocesses the initial point cloud data by constructing a multimodal augmentation pipeline. This pipeline includes geometric transformations and attribute perturbations. The geometric transformation steps involve multi-angle rotation and multi-scale scaling of the point cloud to enhance the model's spatial generalization ability. A random point cloud discarding mechanism simulates data characteristics under different occlusion conditions. Simultaneously, random noise conforming to a Gaussian distribution is injected into the point cloud coordinate data to simulate the inherent measurement errors and environmental disturbances during real data acquisition, improving the model's adaptability to measurement noise. At the attribute augmentation level, color... Techniques such as color space transformation and brightness adjustment enrich the diversity of texture features. Finally, cross-sample fusion technology is used to combine and reconstruct point cloud fragments from different sources to form an enhanced training dataset. This cross-sample fusion technology involves extracting point cloud fragments with complete structures from point cloud data from different acquisition areas or time phases, ensuring that each fragment contains complete geometric features and texture information. Next, a feature alignment stage is performed, calculating key geometric features between point cloud fragments, including principal orientation, scale, and point density distribution. Based on the feature matching results, spatial transformation is applied to the point cloud to ensure consistency in scale, orientation, and spatial distribution among fragments from different sources. In the data fusion stage, a spatial topology-based approach is employed. The relation fusion algorithm detects the geometric continuity of point cloud boundary regions, smoothly transitions overlapping areas, and ensures the logical structure of the stitched point cloud through semantic consistency verification. Finally, it checks the visual consistency and structural integrity of the generated new samples, discarding data with obvious stitching flaws and including qualified samples in the augmented training dataset. The semantic segmentation unit performs end-to-end semantic parsing of the input initial point cloud data using a pre-trained deep learning network model. First, it preprocesses the initial point cloud data, unifying it to a standard coordinate system through spatial coordinate normalization and constructing an octree-based spatial index structure to support efficient neighborhood queries. Then, it proceeds to... In the multi-level feature extraction stage, the network captures multi-scale geometric features of the initial point cloud data through a series of local feature learning modules. Each module sequentially performs nearest neighbor search to build a local graph structure, applies a shared-weight multilayer perceptron for point-by-point feature extraction, aggregates neighborhood features through max pooling, and dynamically adjusts feature weights using an attention mechanism to enhance the expressive power of key geometric structures. In the context information fusion stage, the network fuses shallow geometric details and deep semantic features across layers through a skip connection structure. At the same time, a global feature pooling module is introduced to extract the overall context descriptor of the initial point cloud data, and global context information is injected into the feature representation of each point through a feature broadcasting mechanism.Finally, a multilayer perceptron-based classifier is used to calculate the probability distribution of each point in the initial point cloud data belonging to the building category. A dynamic threshold strategy is used to binarize the initial classification results. Then, a region generation algorithm based on connected component analysis is used to aggregate discrete points into continuous regions. Finally, a boundary optimization algorithm is used to smooth the contours of the building point cloud regions and remove outliers. The final output is a building point cloud region with complete geometric structure and clear boundaries.

[0020] In one embodiment, the individual building clustering module includes: Pre-clustering units are used to perform spatial clustering analysis on building point cloud regions based on density, and to identify continuous point cloud regions with spatial continuity as a set of pre-clustering blocks. The boundary candidate region detection unit is used to extract the boundary shell point cloud of each pre-clustered block, calculate the spatial distance between the boundary shell point clouds of different pre-clustered blocks, identify the neighboring point pair regions with a spatial distance less than a preset threshold, and generate boundary candidate regions. The boundary point cloud attribution module is used to calculate the multi-dimensional features of point clouds in each boundary candidate region, including spatial coordinates, normal vector features, and color texture features. The feature affinity between each point cloud and its neighboring points in the boundary candidate region is calculated based on multi-dimensional features. The feature affinity includes spatial distance factor, normal direction consistency and color and texture similarity. The iterative attribution allocation sub-unit is used to segment the point cloud in the boundary candidate region through multiple rounds of iteration. Establish an attribution weight vector for each point cloud within the boundary candidate region, and record the initial attribution probability of the point cloud to each adjacent pre-clustered block; In each iteration, each point cloud sends a home weight update message to its neighboring points based on its feature affinity relationship with neighboring points; Each point aggregates all received weight update messages, combines them with its own feature information, and updates the weight assigned to each pre-cluster block. The iteration process terminates when the change in the weight of all points is less than a preset threshold or when the maximum number of iterations is reached. Based on the iteratively generated attribution weight vector, each point cloud in the boundary candidate region is assigned to the pre-clustered block with the largest weight; The result generation unit merges the boundary candidate area point clouds that have completed the assignment into the corresponding pre-clustered blocks to generate a single building point cloud cluster containing geometric boundaries. Simultaneously, based on the segmentation standard point cloud pairs generated during the iteration process, they are stored in the segmentation standard library; Furthermore, the pre-clustering unit uses a density-based spatial clustering algorithm to analyze the building point cloud region. By setting a neighborhood radius and a minimum point count threshold, it identifies point cloud regions with spatial continuity to form a pre-clustered block set. The boundary candidate region detection unit extracts the boundary shell point cloud by calculating the convex hull and alpha shape of each pre-clustered block. It uses KD trees to accelerate the calculation of the nearest neighbor distance between the boundary point clouds of different pre-clustered blocks, and filters out neighboring point pairs with a spatial distance less than a preset threshold to form boundary candidate regions. The steps are as follows: based on the convex hull algorithm, it constructs the minimum convex hull boundary of the point cloud using a fast incremental method, and based on the alpha shape algorithm, it controls the degree of boundary concavity by adjusting the alpha radius parameter to retain more geometric details, thereby accurately improving the boundary concavity. The boundary shell point clouds of pre-clustered blocks are extracted. Then, a KD-tree spatial index structure is constructed for each pre-clustered block's boundary shell point cloud. A binary tree partitioning strategy optimizes spatial retrieval efficiency. Based on the KD-tree spatial index structure, the nearest neighbor distance across pre-clustered blocks is calculated. For any two adjacent pre-clustered block boundary point clouds, the radius search method of the KD-tree is used to quickly find neighboring point pairs that satisfy the maximum search distance, and the Euclidean distance between these point pairs is accurately calculated. Finally, neighboring point pairs that meet the spatial distance conditions are filtered out using a preset distance threshold. The local regions containing these point pairs are merged into continuous boundary candidate regions, and the affiliation relationship between each boundary candidate region and its corresponding pre-clustered block is recorded. (Boundary point cloud affiliation module) Principal component analysis (PCA) is used to calculate the normal vector features of the point cloud in the boundary candidate region, and color texture features are extracted from the point cloud to construct a multi-dimensional feature structure containing spatial coordinates, normal vectors, and color texture. Based on the multi-dimensional features, a weighted fusion method is used to calculate the feature affinity between point pairs. The spatial distance factor uses an exponential decay function to process Euclidean distance, normal direction consistency is measured using vector dot product, and color texture similarity is calculated using cosine similarity. The steps are as follows: Normalization preprocessing is performed on the three-dimensional features, mapping the Euclidean distance to the zero-to-one interval, and normalizing the normal and color vectors. Then, the similarity of the three feature dimensions is calculated separately, and the spatial distance factor is converted from Euclidean distance to similarity using an exponential decay function. The similarity scores for spatial distance similarity, normal direction similarity, and color texture similarity are calculated by taking the dot product of the normal vectors of two points. Color texture similarity is obtained by treating the RGB color texture values ​​of points as vectors and calculating the cosine of the angle between them. A weighted fusion strategy is then used to combine the similarity scores of the three dimensions into a comprehensive feature affinity. Adjustable weight coefficients are assigned to spatial distance similarity, normal direction similarity, and color texture similarity, and the final feature affinity value is calculated by weighted summation. Finally, the calculated feature affinity is standardized to ensure its value range is between zero and one, and a minimum threshold is set to filter out weakly correlated point pairs, retaining significant feature associations for subsequent iterative assignment calculations.The iterative assignment subunit optimizes point cloud assignment through multiple rounds of iteration. During initialization, an assignment weight vector based on the number of adjacent pre-clustered blocks is established for each point. During iteration, each point sends assignment weight update messages to neighboring points based on feature affinity. Receiving points aggregate all messages through weighted summation and update their assignment weights based on their own features. The process terminates when all weight changes stabilize or reach the iteration limit. Finally, the point cloud is assigned to the corresponding pre-clustered block according to the maximum weight principle. The result generation unit merges the reassigned point clouds into the target pre-clustered block and generates individual building point cloud clusters with clear geometric boundaries through boundary smoothing. Simultaneously, the effective segmented point cloud pairs and their multi-dimensional feature combinations formed during the iteration process are stored in a segmentation standard library, establishing a fast retrieval mechanism based on feature hashing for standardized reuse in subsequent boundary segmentation.

[0021] In one embodiment, the step of simultaneously generating segmentation standard point cloud pairs based on the iteration process and storing them in the segmentation standard library is as follows: During the iterative attribution assignment process, successfully segmented boundary point cloud pairs are identified, and multi-dimensional feature combinations of the boundary point cloud pairs are extracted as segmentation feature templates. The multidimensional feature combination includes spatial distribution features, normal direction relationship, color texture features, and local geometric context features; The segmentation feature templates and their corresponding segmentation schemes are stored in the segmentation standard library, and a multi-dimensional index structure is built for the segmentation feature templates. The multi-dimensional index structure is constructed based on building type, boundary morphology features and geospatial location. When processing new boundary candidate regions, the feature affinity of the current boundary candidate region point cloud is matched with the segmentation feature templates in the segmentation standard library. When the matching degree exceeds the preset threshold, the corresponding segmentation scheme is directly called. Furthermore, during the iterative assignment process, when two pre-clustered blocks are successfully segmented, the system immediately captures the boundary point cloud pair at their intersection as a standard sample. It calculates the spatial distribution characteristics along the line connecting the two points in this point cloud pair, including distance distribution and density variation; extracts the directional relationship of the two normal vectors, including the angle and relative orientation; analyzes the differences in color and texture features, including RGB value differences and texture pattern variations; and captures local geometric context features, including the distribution pattern of neighboring point clouds and curvature variation patterns. These multi-dimensional features are combined to construct a standardized segmentation feature template. Then, this segmentation feature template is compared with its corresponding segmentation scheme, which includes the final assigned weight distribution and boundary line. The shapes are stored together in the segmentation standard library, and a multi-dimensional index structure based on a multi-hash table is established for each segmentation feature template. This index structure uses building type as the first-level hash key, boundary morphology features as the second-level hash key, and geospatial location as the third-level hash key to achieve fast hierarchical retrieval. When processing new boundary candidate areas, the system extracts multi-dimensional feature combinations of the current boundary candidate area point cloud in real time to generate query vectors. Feature matching is performed by calculating the cosine similarity between the query vector and the template vectors in the segmentation standard library. When the matching degree of a template exceeds a preset threshold, the system automatically calls the historical segmentation scheme corresponding to that template and directly maps the stored attribution weight distribution to the current boundary candidate area to complete fast segmentation.

[0022] In one embodiment, the geometric contour recognition module includes: The projection conversion unit is used to project the three-dimensional spatial point cloud of a single building point cloud cluster onto a horizontal plane to form a two-dimensional projection point set. The contour vectorization unit generates the projected outer contour of a single building by using a boundary extraction algorithm based on a two-dimensional projection point set; The elevation calculation unit calculates the building footprint, altitude, and building height of a single building as elevation parameters based on a two-dimensional projection point set. The data encapsulation unit is used to associate the projected outer contour with elevation parameters to generate standardized vector data. Furthermore, the projection transformation unit performs orthogonal projection processing on the 3D coordinates of the individual building point cloud clusters, keeping the X and Y coordinates of each point unchanged while directly ignoring the Z coordinate value, compressing all 3D point clouds onto a horizontal plane to form a 2D projection point set, and simultaneously establishing a corresponding mapping relationship between 3D points and 2D projection points; the contour vectorization unit uses Alpha based on the obtained 2D projection point set. The Shape boundary extraction algorithm constructs a triangular mesh by traversing the point set and dynamically adjusts the alpha radius parameter to control the boundary generation accuracy. When the alpha radius is small, it captures detailed features, while a larger alpha radius forms a smooth contour, ultimately extracting a closed polygon vector contour describing the building's external shape. The elevation calculation unit calculates the elevation parameters of the area enclosed by the two-dimensional projection point set based on the original 3D point cloud data. By statistically analyzing the elevation values ​​of all 3D points inside the projection polygon, it determines the bottom and top elevations of the building. The minimum bottom elevation is taken as the building's altitude benchmark, and the difference between the maximum top elevation and the minimum bottom elevation is taken as the building's height. At the same time, it records the projected area of ​​the building's footprint. The data encapsulation unit associates and encapsulates the generated projection outer contour vector polygon with the calculated altitude and building height parameters. It organizes the data structure according to the standard vector data format, storing geometric information as polygon features and elevation parameters as attribute fields, ultimately outputting standardized vector data containing complete geometric information and semantic attributes.

[0023] In one embodiment, the white model building module includes: The architectural syntax parsing unit is used to receive standardized vector data, parse the building structure type based on the projected outer contour and elevation parameters, and deconstruct the building into a parameterized sequence of building components. The parameterized component generation unit generates geometric models of building components containing semantic information based on a sequence of building components. The component assembly unit is used to assemble the generated building components according to the building structure logic to form a complete three-dimensional white model; Furthermore, the architectural grammar parsing unit analyzes the projected outer contour shape features and elevation parameters in standardized vector data, uses contour rectangularity and concavity / convexity detection algorithms to identify the basic structural types of buildings, and uses rule-based reasoning to deconstruct the building into a parameterized sequence of components. Based on contour complexity and aspect ratio, the building is classified into basic types such as slab buildings, tower buildings, or composite buildings, and a preset component decomposition rule is matched for each type. The parameterized component generation unit, based on the parsed building component sequence and corresponding type parameters, calls a predefined parameterized geometry generator to create geometric models of building components with semantic labels. The main wall components are generated using contour extrusion height parameters, the pitched roof components are constructed using a slope geometry algorithm based on roof type parameters, and auxiliary structural components are instantiated and modeled according to relative position parameters. All components carry semantic identifiers of building component types and spatial positioning information. The component assembly unit assembles the generated building components based on architectural structural logic. Spatial assembly involves positioning components correctly through coordinate transformation, handling geometric intersections between components using Boolean operations, implementing collision detection to ensure no conflicts between components, and finally integrating all components to form a structurally complete 3D white model rich in semantic information. Taking an L-shaped office building as an example, the architectural syntax parsing unit first analyzes its L-shaped projected outline and 30-meter height parameters. By identifying the shape complexity, it determines that the building is a structure composed of two rectangular blocks, and then decomposes it into three parameterized component sequences: the main building block, the podium block, and the pitched roof. The parameterized component generation unit then stretches the two rectangular outlines according to the sequence instructions to generate the 30-meter-high main wall component, and simultaneously constructs a sloping roof component covering the entire L-shape based on the slope parameters. All components carry type semantic identifiers and spatial positioning information. Finally, the component assembly unit precisely splices the main building block and the podium block according to the L-shaped spatial relationship and assembles the pitched roof component on the top of the building, ultimately generating a complete 3D white model of the L-shaped office building.

[0024] In one embodiment, the architectural syntax parsing unit includes: The building structure analysis sub-unit is used to analyze the geometric features and elevation parameters of the projected outer contour in standardized vector data, and classifies individual buildings into multiple basic geometric types through contour shape recognition algorithms. The component storage sub-unit establishes a component association network based on multiple basic geometric types of a single building and records the spatial constraint relationships between components; The component sequence generates sub-units, generating a component assembly sequence that conforms to the building construction logic based on the component association network and spatial constraints. Furthermore, the building structure analysis subunit calculates geometric feature parameters such as the rectangularity and convex hull area ratio of the projected outer contour, and combines them with the building height-to-width ratio and contour complexity index. It then uses a contour classifier based on a convolutional neural network to identify individual buildings into basic geometric types such as cubes, prisms, and cylinders. The component storage subunit queries the corresponding component composition template in a predefined building component knowledge graph based on the identified basic geometric types, establishes a component association network with the main structural components as root nodes and auxiliary components as child nodes, and uses a graph structure to record the spatial constraint relationships between components, including vertical support relationships, horizontal connection relationships, and coplanar alignment relationships. The component sequence generation subunit performs topological sorting analysis based on the component association network, prioritizing basic support components before processing upper-level components. It also considers spatial constraint relationships to determine the assembly priority between components, resolves dependency conflicts between components through a constraint propagation algorithm, and finally generates a component assembly sequence that conforms to the building construction logic.

[0025] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A system for individual model extraction and white model construction based on AI-driven oblique photogrammetry, characterized in that, include: Point cloud parsing module: used to extract 3D point cloud data and corresponding geometric and texture information from oblique photogrammetry models, and generate initial point cloud data; AI semantic segmentation module: Connected to the point cloud parsing module, it enhances the initial point cloud data to generate enhanced point cloud data, pre-trains a deep learning model using the enhanced point cloud data, and uses the trained deep learning model to perform semantic segmentation on the initial point cloud data to generate building point cloud regions. Individual building clustering module: Connected to the AI ​​semantic segmentation module, it performs spatial clustering analysis on building point cloud regions, aggregating point clouds belonging to the same building entity into independent clusters, generating individual building point cloud clusters; Geometric contour recognition module: connected to the individual building clustering module, it calculates the projected outer contour on the horizontal plane for each individual building point cloud cluster and extracts the elevation parameters of the individual building. White model construction module: Connected to the geometric contour recognition module, it generates the corresponding 3D white model based on the projected outer contour and elevation parameters through a programmed modeling method.

2. The AI-driven oblique photogrammetry model individualization extraction and white model construction system according to claim 1, characterized in that, The point cloud parsing module includes: The data reading unit is used to read the oblique photogrammetry model file and obtain vertex coordinates, patch indexes, and texture mapping information. The point cloud information extraction unit is used to extract the spatial coordinate data of vertices from the oblique photogrammetry model file, construct the basic geometric structure of the point cloud, obtain the texture information corresponding to the vertices, and generate a point cloud with texture information. The data encapsulation unit is used to fuse the extracted basic geometric structure and color texture information, convert it into standard point cloud data containing spatial coordinates and color texture information, and encapsulate the standard point cloud data to output initial point cloud data containing geometric and texture information.

3. The AI-driven oblique photogrammetry model individualization extraction and white model construction system according to claim 1, characterized in that, The AI ​​semantic segmentation module includes: The data augmentation unit is used to perform data augmentation operations on the initial point cloud data to generate augmented point cloud data. The semantic segmentation unit, based on a deep learning network model pre-trained on an augmented training dataset, performs semantic segmentation on the input initial point cloud data, identifies point clouds belonging to the building category, and outputs the building point cloud region.

4. The AI-driven oblique photogrammetry model individualization extraction and white model construction system according to claim 1, characterized in that, The individual building clustering module includes: Pre-clustering units are used to perform spatial clustering analysis on building point cloud regions based on density, and to identify continuous point cloud regions with spatial continuity as a set of pre-clustering blocks. The boundary candidate region detection unit is used to extract the boundary shell point cloud of each pre-clustered block, calculate the spatial distance between the boundary shell point clouds of different pre-clustered blocks, identify the neighboring point pair regions with a spatial distance less than a preset threshold, and generate boundary candidate regions. The boundary point cloud attribution module is used to calculate the multi-dimensional features of point clouds in each boundary candidate region, including spatial coordinates, normal vector features, and color texture features. The feature affinity between each point cloud and its neighboring points in the boundary candidate region is calculated based on multi-dimensional features. The feature affinity includes spatial distance factor, normal direction consistency and color and texture similarity. The iterative attribution allocation sub-unit is used to segment the point cloud in the boundary candidate region through multiple rounds of iteration. Establish an attribution weight vector for each point cloud within the boundary candidate region, and record the initial attribution probability of the point cloud to each adjacent pre-clustered block; In each iteration, each point cloud sends a home weight update message to its neighboring points based on its feature affinity relationship with neighboring points; Each point aggregates all received weight update messages, combines them with its own feature information, and updates the weight assigned to each pre-cluster block. The iteration process terminates when the change in the weight of all points is less than a preset threshold or when the maximum number of iterations is reached. Based on the iteratively generated attribution weight vector, each point cloud in the boundary candidate region is assigned to the pre-clustered block with the largest weight; The result generation unit merges the boundary candidate area point clouds that have completed the assignment into the corresponding pre-clustered blocks to generate a single building point cloud cluster containing geometric boundaries. Simultaneously, based on the segmentation standard point cloud pairs generated during the iteration process, they are stored in the segmentation standard library.

5. The AI-driven oblique photogrammetry model individualization extraction and white model construction system according to claim 4, characterized in that, The step of simultaneously generating segmentation standard point cloud pairs based on the iteration process and storing them in the segmentation standard library is as follows: During the iterative attribution assignment process, successfully segmented boundary point cloud pairs are identified, and multi-dimensional feature combinations of the boundary point cloud pairs are extracted as segmentation feature templates. The multidimensional feature combination includes spatial distribution features, normal direction relationship, color texture features, and local geometric context features; The segmentation feature templates and their corresponding segmentation schemes are stored in the segmentation standard library, and a multi-dimensional index structure is built for the segmentation feature templates. The multi-dimensional index structure is constructed based on building type, boundary morphology features and geospatial location. When processing new boundary candidate regions, the feature affinity of the current boundary candidate region point cloud is matched with the segmentation feature templates in the segmentation standard library. When the matching degree exceeds the preset threshold, the corresponding segmentation scheme is directly called.

6. The AI-driven oblique photogrammetry model individualization extraction and white model construction system according to claim 1, characterized in that, The geometric contour recognition module includes: The projection conversion unit is used to project the three-dimensional spatial point cloud of a single building point cloud cluster onto a horizontal plane to form a two-dimensional projection point set. The contour vectorization unit generates the projected outer contour of a single building by using a boundary extraction algorithm based on a two-dimensional projection point set; The elevation calculation unit calculates the building footprint, altitude, and building height of a single building as elevation parameters based on a two-dimensional projection point set. The data encapsulation unit is used to associate the projected outer contour with elevation parameters to generate standardized vector data.

7. The AI-driven oblique photogrammetry model individualization extraction and white model construction system according to claim 1, characterized in that, The white model construction module includes: The architectural syntax parsing unit is used to receive standardized vector data, parse the building structure type based on the projected outer contour and elevation parameters, and deconstruct the building into a parameterized sequence of building components. The parameterized component generation unit generates geometric models of building components containing semantic information based on a sequence of building components. The component assembly unit is used to assemble the generated building components according to the building structure logic to form a complete three-dimensional white model.

8. The AI-driven oblique photogrammetry model individualization extraction and white model construction system according to claim 7, characterized in that, The architectural syntax parsing unit includes: The building structure analysis sub-unit is used to analyze the geometric features and elevation parameters of the projected outer contour in standardized vector data, and classifies individual buildings into multiple basic geometric types through contour shape recognition algorithms. The component storage sub-unit establishes a component association network based on multiple basic geometric types of a single building and records the spatial constraint relationships between components; The component sequence generates sub-units, and based on the component association network and spatial constraints, generates a component assembly sequence that conforms to the building construction logic.

Images