Three-dimensional grid topology reconstruction method based on geometric-semantic-topological cooperation

Abstract

This application provides a 3D mesh topology reconstruction method based on geometric-semantic-topology collaboration. The method constructs an initial 3D mesh topology based on the spatial continuity of a 3D scanned point cloud, generating geometric shape representations, object semantic representations, and topological connectivity representations. Based on the consistency relationship among these three elements, collaborative reconstruction constraints are generated and iteratively corrected to obtain an optimized 3D mesh topology. The method identifies the mesh region to be reconstructed and reconstructs the local 3D mesh topology using the corresponding local point cloud, obtaining a corrected 3D mesh topology. Based on the corrected 3D mesh topology, 3D surfaces are generated and organized to obtain a parametric 3D mesh model. This method can reduce the impact of error propagation at each level in serial reconstruction on mesh shape and topological relationships.

Description

Technical Field

[0001] This application relates to the field of 3D computer graphics modeling technology, and more specifically, to a 3D mesh topology reconstruction method based on geometry-semantics-topology collaboration. Background Technology

[0002] With the development of 3D scanning equipment and 3D computer graphics technology, the recovery of the surface morphology, object category, and spatial connectivity of real objects through 3D scanning point clouds has been applied in fields such as digitization of existing buildings, reconstruction of completed scenes, and construction of digital twin scenes. Due to the presence of occlusion, local point density variations, blurred boundaries between adjacent objects, and similar surface morphologies in the scanned objects, the 3D reconstruction process not only needs to restore the geometric shape of the objects but also needs to maintain the spatial relationships between different objects so that the generated 3D data can represent the scanned scene relatively completely.

[0003] Existing 3D reconstruction schemes typically employ a sequential processing approach. First, the scanned point cloud is segmented and geometrically fitted to form multiple geometric entities. Then, the object category is determined based on the surface features of each geometric entity. Finally, using the determined geometric entities and object categories, the connection relationships between objects are inferred according to distance, enclosing range, or boundary overlap, and the 3D result is output accordingly.

[0004] However, in the aforementioned sequential processing, the geometric reconstruction results, object category determination results, and spatial relationship inference results are generated unidirectionally by different processing stages. Deviations from previous stages directly serve as input for subsequent stages. When local point clouds are affected by occlusion, missing data, or blurred boundaries, geometric reconstruction deviations may further cause deviations in object category determination and spatial relationship inference, and subsequent processing results are difficult to return to previous processing stages for mutual correction. Furthermore, local anomalies typically require either continuing processing with existing intermediate results or reconstructing the entire point cloud, making it difficult to reuse original scan data for local reconstruction of unstable areas. Consequently, inconsistencies may exist between the resulting surface shapes, object attributes, and object connectivity relationships, reducing the shape integrity and topological consistency of the final 3D result. Summary of the Invention

[0005] This application provides a three-dimensional mesh topology reconstruction method based on geometry-semantics-topology collaboration, which at least alleviates the technical problems in the existing three-dimensional reconstruction process, such as unidirectional transmission of different processing results, gradual accumulation of local deviations, difficulty in recovering local unstable regions, and inconsistency between the final surface shape and the object connection relationship.

[0006] A 3D mesh topology reconstruction method based on geometry-semantics-topology synergy includes: A 3D scanned point cloud is acquired, an initial 3D mesh topology is constructed based on the spatial continuity of the 3D scanned point cloud, and geometric shape representation, object semantic representation, and topological connection representation are generated based on the 3D scanned point cloud and the initial 3D mesh topology. Based on the consistency relationship between the geometric shape representation, the object semantic representation, and the topological connection representation, a cooperative reconstruction constraint is generated. The cooperative reconstruction constraint is used to perform cyclic correction on the geometric shape representation, the object semantic representation, and the topological connection representation, thereby updating the initial 3D mesh topology and obtaining an optimized 3D mesh topology. The mesh region to be reconstructed is identified based on the consistency of the geometric shape representation, the object semantic representation, and the topological connection representation in the optimized 3D mesh topology. The local point cloud corresponding to the mesh region to be reconstructed is extracted from the 3D scan point cloud. The 3D mesh topology of the mesh region to be reconstructed is reconstructed using the local point cloud to update the optimized 3D mesh topology and obtain the corrected 3D mesh topology. A three-dimensional surface is generated based on the geometric shape representation in the corrected three-dimensional mesh topology. The three-dimensional surface is then organized according to the object semantic representation and the topological connection representation to obtain a parametric three-dimensional mesh model.

[0007] This application addresses the technical problem in existing 3D reconstruction schemes where point cloud segmentation, shape restoration, object category determination, and spatial connectivity inference are performed separately, resulting in a lack of a unified basis for the corresponding processing results. It constructs an initial 3D mesh topology based on the spatial continuity of the 3D scanned point cloud, and generates geometric shape representations, object semantic representations, and topological connectivity representations based on the 3D scanned point cloud and the initial 3D mesh topology. These geometric shape representations, object semantic representations, and topological connectivity representations are generated from the same 3D scanned point cloud and the same initial 3D mesh topology, ensuring that the surface morphology, object attributes, and spatial connectivity status of an object can all correspond to their respective positions within the initial 3D mesh topology. Compared to traditional schemes that separately store geometric restoration results, category determination results, and connectivity inference results, and then stitch the results together in subsequent stages, this application reduces misalignment caused by inconsistencies in object segmentation or spatial location correspondence between different processing results, providing a unified 3D data foundation for subsequent mutual correction based on the relationships between the three representations.

[0008] This application addresses the technical problem of traditional sequential reconstruction schemes, which use the results of the previous stage as fixed inputs for the next stage, leading to the sequential propagation and accumulation of local shape deviations, object attribute deviations, and connection relationship deviations. It generates collaborative reconstruction constraints based on the consistency relationship between geometric shape representation, object semantic representation, and topological connection representation, and uses these constraints to perform cyclical corrections on these representations. The collaborative reconstruction constraints are not generated from only one type of representation but originate from the consistency relationship among the three, thus allowing the use of inconsistencies between one representation and the others to control subsequent corrections. After each correction, the geometric shape representation, object semantic representation, and topological connection representation continue to participate in the next cyclical correction, enabling the object semantic representation and topological connection representation to have a reverse effect on inconsistent geometric shape representations, object semantic representations, and topological connection representations, and vice versa. Compared to traditional methods that rely solely on geometric reconstruction results to drive object category determination, and then use the object category determination results to drive connection relationship inference, this application breaks the unidirectional transmission relationship between different processing stages, reducing the degree to which local deviations in a certain stage are continuously amplified in subsequent stages.

[0009] This application utilizes cyclically corrected geometric shape representation, object semantic representation, and topological connectivity representation to update the initial 3D mesh topology, resulting in an optimized 3D mesh topology. This optimized 3D mesh topology not only represents the surface shape of the point cloud but also collectively reflects the object attributes and topological connectivity states after consistency relation correction. Therefore, the cyclic correction results of geometric shape representation, object semantic representation, and topological connectivity representation are incorporated into the initial 3D mesh topology, rather than forming independent data separate from it. Compared to traditional methods that add category and connectivity labels after the geometric entities are fixed, this application updates the initial 3D mesh topology through three representations, which helps maintain a high degree of consistency between the mesh shape, object attributes, and connectivity structure in the optimized 3D mesh topology.

[0010] This application addresses the technical problem that occlusion, local point density variations, or blurred boundaries can easily lead to unstable results during the overall reconstruction process, and traditional methods struggle to specifically revert to the original scan data to reconstruct local areas. It identifies the mesh region to be reconstructed based on the consistency of geometric shape representation, object semantic representation, and topological connectivity representation in the optimized 3D mesh topology. Since the region to be reconstructed is determined by the consistency of these three representations, it points to mesh locations where inconsistencies exist among them, rather than relying solely on a single geometric deviation or object category change. Subsequently, the local point cloud corresponding to the mesh region to be reconstructed is extracted from the 3D scan point cloud. This local point cloud is then used to reconstruct the 3D mesh topology of the region to be reconstructed, allowing the local reconstruction process to be based on the original scan data again. This reduces the likelihood of repeatedly processing intermediate results and perpetuating inappropriate outcomes.

[0011] This application utilizes the reconstructed 3D mesh topology of the region to be reconstructed to update and optimize the 3D mesh topology, resulting in a corrected 3D mesh topology. This preserves the 3D mesh topology of regions not identified as needing reconstruction, while allowing the reconstruction of local regions with inconsistent characterizations. Compared to the traditional approach of reprocessing all 3D scanned point clouds after discovering local anomalies, this application focuses point cloud extraction and 3D mesh topology reconstruction on the region to be reconstructed, which helps reduce the amount of redundant processing of irrelevant mesh regions. Compared to directly using existing mesh results from local anomaly regions for further inference, this application reuses local point clouds to recover the corresponding regions, which helps improve the consistency between local mesh morphology, object attributes, and topological connectivity.

[0012] This application addresses the technical problem of traditional methods that, after forming geometric surfaces, separately add object attributes and connection relationships, easily leading to a mismatch between the final surface shape and the object's organizational structure. It generates 3D surfaces based on the geometric shape representation in a corrected 3D mesh topology and organizes the 3D surfaces according to object semantic representation and topological connection representation. The shape of the 3D surface originates from the geometric shape representation in the corrected 3D mesh topology; the object affiliation of the 3D surface is determined by the object semantic representation within the same corrected 3D mesh topology; and the connection method between 3D surfaces is determined by the topological connection representation within the same corrected 3D mesh topology. Therefore, the generation of 3D surfaces, object organization, and topological connections all share a common corrected 3D mesh topology source, reducing the possibility of mismatches between surface shapes, object attributes, and spatial connection relationships generated separately. Attached Figure Description

[0013] Figure 1 This application provides an embodiment of a 3D mesh topology reconstruction scene based on geometry-semantics-topology collaboration. Figure 2This application provides an embodiment of a 3D mesh topology reconstruction method based on geometry-semantics-topology collaboration. Figure 3 This application provides an embodiment of a three-dimensional mesh topology reconstruction device based on geometry-semantics-topology collaboration. Figure 4 An electronic device is described in an embodiment of this application; Figure 5 This application provides an embodiment of a computer-readable storage medium. Figure 6 This application provides an embodiment of a three-dimensional mesh topology reconstruction system based on geometric-semantic-topology collaboration. Detailed Implementation

[0014] like Figure 1 As shown, this is a 3D mesh topology reconstruction scene based on geometry-semantics-topology collaboration according to an embodiment of this application; as Figure 2 The illustration shows a 3D mesh topology reconstruction method based on geometry-semantics-topology collaboration according to an embodiment of this application. The method includes: acquiring a 3D scanned point cloud; constructing an initial 3D mesh topology based on the spatial continuity of the 3D scanned point cloud; generating geometric shape representations, object semantic representations, and topological connectivity representations based on the 3D scanned point cloud and the initial 3D mesh topology; generating collaborative reconstruction constraints based on the consistency relationships among the geometric shape representations, object semantic representations, and topological connectivity representations; and using the collaborative reconstruction constraints to iteratively correct the geometric shape representations, object semantic representations, and topological connectivity representations, thereby updating the initial 3D mesh. The topology is optimized to obtain a 3D mesh topology. Based on the consistency of the geometric shape representation, object semantic representation, and topological connection representation in the optimized 3D mesh topology, the mesh region to be reconstructed is identified. The local point cloud corresponding to the mesh region to be reconstructed is extracted from the 3D scan point cloud. The 3D mesh topology of the mesh region to be reconstructed is reconstructed using the local point cloud to update the optimized 3D mesh topology, resulting in a corrected 3D mesh topology. A 3D surface is generated based on the geometric shape representation in the corrected 3D mesh topology. The 3D surface is organized according to the object semantic representation and the topological connection representation to obtain a parametric 3D mesh model.

[0015] Specifically, the 3D scan point cloud in this application can be read from ground laser scanning data that has undergone multi-station registration. Each point data in the 3D scan point cloud records at least its spatial position in a unified 3D coordinate space. Before the point data collected by multiple scanning stations enters the reconstruction process, the same building surface collected by different scanning stations is first identified in the overlapping scanning area of ​​adjacent scanning stations. Then, coordinate alignment is completed based on the differences in the spatial positions of the points on the same building surface, so that the point data in different scanning stations can jointly represent the same existing building scanning scene. As one implementation, the 3D scan point cloud can originate from the indoor scanning results of an existing office building. The scanning stations are arranged along building corridors, room entrances, and major corner areas. The overlapping scanning area is retained between adjacent scanning stations, and point cloud registration is then completed using the overlapping scanning area. The resulting 3D scan point cloud can jointly represent the surface morphology of walls, columns, beam bottoms, floor slabs, door and window openings, and the intersection areas of adjacent components. In this application, spatial continuity refers to the degree of continuity of adjacent point data in terms of point spatial position, local surface direction, and local surface change trend. For adjacent point data belonging to the same continuous surface, the spatial spacing between the points is within a similar range, and the change in the local surface direction is small. For adjacent point data located on different surfaces, the local surface direction or the trend of local surface change will change significantly at the junction boundary. Utilizing this spatial continuity, the 3D scanned point cloud can be divided into multiple curved mesh fragments with continuous internal surface changes and boundary connections between them. These curved mesh fragments are then used as the local surface regions corresponding to the fragment nodes of the initial 3D mesh topology, and the fragment adjacency relationships between the curved mesh fragments are converted into node adjacency records in the initial 3D mesh topology.

[0016] The initial 3D mesh topology is used to uniformly record local surface regions in the 3D scanned point cloud and the spatial relationships between these local surface regions. Each fragment node in the initial 3D mesh topology corresponds to a surface mesh fragment, and the node adjacency records between fragment nodes correspond to the fragment adjacency relationships between surface mesh fragments. The geometric shape representation is used to describe the surface morphology, spatial fitting state, and uncertainty of the geometric shape representation corresponding to the surface mesh fragment; the object semantic representation is used to describe the candidate judgment of the 3D object type corresponding to the surface mesh fragment, the semantic credibility state, and the uncertainty of the object semantic representation; the topological connection representation is used to describe the candidate topological connection, topological connection form, topological consistency state, and uncertainty of the topological connection representation between adjacent surface mesh fragments. The geometric shape representation, the object semantic representation, and the topological connection representation are all registered to the corresponding fragment node or the node adjacency record in the initial 3D mesh topology, so that the surface morphology, 3D object type, and topological connection form of the same local surface region maintain a unified spatial correspondence. In this application, the consistency relationship is not merely a comparison of whether numerical values ​​are the same, but rather a determination of whether the geometric shape representation and the object semantic representation satisfy a geometric semantic correspondence, whether the object semantic representation and the topological connection representation satisfy a semantic-topological correspondence, and whether the geometric shape representation and the topological connection representation satisfy a geometric-topological correspondence. For example, when the geometric shape representation of the surface mesh fragment represents an approximate vertical plane, and the object semantic representation determines the candidate object semantic category corresponding to the surface mesh fragment as a wall-type 3D object, the geometric shape representation and the object semantic representation can be in a state that satisfies the geometric semantic correspondence; when the object semantic representations of two adjacent surface mesh fragments both represent wall-type 3D objects, and the topological connection representation indicates that the two adjacent surface mesh fragments are disconnected in the coplanar direction, then the topological connection representation needs to be corrected based on the surface boundaries of the two adjacent surface mesh fragments.

[0017] The collaborative reconstruction constraints record the locations where the consistency relationship is not satisfied among the geometric representation, the object semantic representation, and the topological connection representation, the representations to be corrected involved in the unsatisfied state, and the supporting representations used to correct the representations to be corrected. During iterative correction of the geometric representation, the object semantic representation, and the topological connection representation, each time one type of representation is determined as the representation to be corrected, and the representation with lower uncertainty and a consistency relationship with the representation to be corrected is determined as the supporting representation. After the representation to be corrected completes one correction, the corresponding correction result re-participates in the determination of the consistency relationship, thereby forming the collaborative reconstruction constraints used in the next correction. The iteratively corrected geometric representation, object semantic representation, and topological connection representation are re-registered to the initial 3D mesh topology to obtain the optimized 3D mesh topology. In one embodiment, when there are still mesh regions in the optimized 3D mesh topology where the geometric representation, object semantic representation, and topological connection representation change repeatedly or where the consistency is low, these mesh regions are identified as the mesh regions to be reconstructed. Subsequently, based on the position of the mesh region to be reconstructed in the unified 3D coordinate space, the local point cloud corresponding to the mesh region to be reconstructed is extracted from the 3D scanned point cloud, and then the 3D mesh topology of the mesh region to be reconstructed is rebuilt based on the local point cloud. The corresponding region in the optimized 3D mesh topology is updated with the reconstructed 3D mesh topology of the mesh region to be reconstructed to form the corrected 3D mesh topology. Thus, when the mesh region to be reconstructed is affected by occlusion, changes in point density, or blurred boundary, the surface morphology of the mesh region to be reconstructed can be restored by returning to the 3D scanned point cloud, without relying solely on the already formed geometric shape representation, object semantic representation, and topological connection representation for further inference. The geometric shape representation in the corrected 3D mesh topology is used to determine the spatial position, surface orientation, and surface morphology of the 3D surface; the object semantic representation is used to determine the 3D object type corresponding to each 3D surface; and the topological connection representation is used to determine the topological connection form between 3D surfaces belonging to the same 3D object type and between 3D surfaces belonging to different 3D object types. After organizing the three-dimensional surface according to the object semantic representation and the topological connection representation, the parametric three-dimensional mesh model is obtained.

[0018] Optionally, constructing an initial 3D mesh topology based on the spatial continuity of the 3D scanned point cloud includes: The neighborhood distribution of point data in the 3D scanned point cloud is checked to generate a neighborhood distribution check result. Outlier point data is determined based on the neighborhood distribution check result. The outlier point data is removed from the 3D scanned point cloud to obtain a cleaned 3D point cloud. A pre-configured point cloud processing resolution is obtained, and the cleaned 3D point cloud is spatially regularized according to the point cloud processing resolution to obtain a regularized 3D point cloud. The spatial positions of adjacent point data in the regularized 3D point cloud are extracted, and a point normal representation is generated based on the point spatial positions. A surface continuity representation is generated based on the changes in the point normal representations of adjacent point data. Surface mesh fragments are generated based on the point spatial positions, the point normal representations, and the surface continuity representations. Fragment adjacency relationships are generated based on the boundary continuity between adjacent surface mesh fragments. The initial 3D mesh topology is constructed using the surface mesh fragments and the fragment adjacency relationships.

[0019] Specifically, in this application, the neighborhood distribution test is used to determine whether the spatial distance between each point data and its surrounding neighboring point data conforms to the density distribution of the local point cloud. When performing the neighborhood distribution test, firstly, based on the spatial position of each point data, multiple neighboring point data that are relatively close to the point data are searched from the 3D scanned point cloud. The number of neighboring point data can be configured according to the acquisition density of the 3D scanned point cloud; in an exemplary embodiment, a fixed number of neighboring point data (e.g., 50) can be selected for each point data, thereby avoiding the direct change in the neighborhood range of the neighborhood distribution test due to point density variations in different areas of the existing building scanning scene caused by different scanning distances. For each point data, the spatial distance between the point data and each of the neighboring point data is calculated, and then the mean neighborhood distance and the dispersion of the neighborhood distance are calculated based on the multiple spatial distances. The mean neighborhood distance reflects the normal spacing between the neighboring point data surrounding the point data, and the dispersion of the neighborhood distance reflects the degree of fluctuation of each spatial distance relative to the mean neighborhood distance. The average neighborhood distance corresponding to the point data is then calculated, and compared with a distance limit determined by the mean neighborhood distance and the dispersion of the neighborhood distance to generate the neighborhood distribution test result. When the average neighborhood distance exceeds the distance limit, the neighborhood distribution test result identifies the point data as outlier data; when the average neighborhood distance does not exceed the distance limit, the neighborhood distribution test result identifies the point data as retained point data. The retained point data is used to form the cleaned-up 3D point cloud. As one implementation, the distance limit can be obtained by adding a predetermined multiple of the dispersion of the neighborhood distance to the mean neighborhood distance (e.g., adding 2.5 times the dispersion of the neighborhood distance to the mean neighborhood distance). Outlier data is removed from the 3D scan point cloud, and the retained point data is retained to obtain the cleaned-up 3D point cloud. The cleaned-up 3D point cloud retains the retained point data continuously distributed near the building surface and reduces outlier data formed by moving objects, scan reflection anomalies, or multi-station registration edges. The purified 3D point cloud then enters the spatial regularization process. Therefore, the outlier point data will no longer participate in the point normal representation, the surface continuity representation, and the generation of the surface mesh fragments.

[0020] The point cloud processing resolution is configured before the spatial normalization is performed. The point cloud processing resolution represents the target spatial spacing between adjacent retained point data in the normalized 3D point cloud. The configuration of the point cloud processing resolution is based on the original point density of the 3D scanned point cloud and the minimum surface boundary scale to be retained. The point cloud processing resolution should not be greater than the narrow surface size to be retained in the object to be reconstructed; otherwise, the point data on both sides of the narrow surface will enter the same spatial range. The point cloud processing resolution should also not be significantly smaller than the original point spacing of the 3D scanned point cloud; otherwise, it will be difficult to unify the point cloud density in different regions of the 3D scanned point cloud. In an exemplary embodiment, for indoor building point clouds, the point cloud processing resolution can be configured to the centimeter level (e.g., 0.02 meters). When performing the spatial normalization on the cleaned 3D point cloud according to the point cloud processing resolution, the 3D space where the cleaned 3D point cloud is located is divided into multiple voxel units, and the side length of each voxel unit is determined by the point cloud processing resolution. For each voxel cell containing the point data, the voxel centroid position is calculated based on the spatial positions of all the point data within the voxel cell, and the point data in the voxel cell is represented by regularized point data located at the voxel centroid position. Multiple regularized point data collectively form the regularized 3D point cloud. Through spatial regularization, regions with high original point density will not have excessive data weights on subsequent surface determination due to containing too much point data, and regions with low original point density will not adopt different neighborhood scales from other regions due to unfounded changes in processing scale.

[0021] For each regularized point data in the regularized 3D point cloud, neighboring point data are searched according to the point's spatial position, and the centroid position of the neighborhood is calculated based on the spatial position of the neighboring point data. Then, the spatial offset of each neighboring point data relative to the centroid position of the neighborhood is calculated, and the degree of change of the spatial offset in three spatial directions and the correlation between different spatial directions are statistically analyzed to determine three mutually orthogonal spatial change directions. The spatial change direction with the smallest distribution change of neighboring point data is approximately perpendicular to the local surface, and this spatial change direction and its direction are registered as the point normal representation. To reduce the direction reversal of adjacent point normal representations, a point normal representation is selected from the regularized 3D point cloud as the starting point for normal propagation, and the point normal representations are compared step by step along the spatial adjacency order of the neighboring point data. When two adjacent point normal representations point to opposite half-spaces of the local surface, the direction of the latter point normal representation is reversed, so that the adjacent point normal representations maintain continuous pointing. The point normal representation after directional propagation is used to generate the continuous representation of the surface and to determine the boundaries of the surface mesh fragments.

[0022] The surface continuity representation reflects whether adjacent point data lie on the same continuous surface. When generating the surface continuity representation, the spatial position change and normal representation change between adjacent point data are calculated respectively. When the spatial distance between adjacent point data does not exceed the distance range corresponding to the point cloud processing resolution, and the normal representation change between adjacent point data does not exceed the surface direction change range, the adjacent point data are registered as a continuous relationship; when the spatial distance or the normal representation change exceeds the corresponding range, the adjacent point data are registered as a boundary relationship. The continuous relationship and the boundary relationship together form the surface continuity representation. The continuous relationship is used to generate the surface mesh fragments, and the boundary relationship is used to determine the boundaries of the surface mesh fragments. Further, point data that have not yet been assigned to the surface mesh fragments are selected from the regularized 3D point cloud as starting point data, and adjacent point data with the continuous relationship with the starting point data are gradually absorbed according to the surface continuity representation. For each neighboring point data absorbed, the system continues to judge other point data around the neighboring point data based on the continuous surface representation of the neighboring point data, until all neighboring point data are determined to be part of the boundary relationship or have been assigned to other surface mesh fragments, thereby generating one surface mesh fragment. The remaining point data in the regularized 3D point cloud are processed in the same way to generate multiple surface mesh fragments.

[0023] For any two surface mesh fragments, the boundary point data registered as the boundary relationship by the continuous surface representation in the two surface mesh fragments are extracted, and the spatial distance and continuity length along the boundary direction between the boundary point data of the two surface mesh fragments are compared. When the boundary point data of the two surface mesh fragments are close to each other in a local spatial range and form a continuous correspondence along the boundary direction, the two surface mesh fragments are registered as the fragment adjacency relationship. Each surface mesh fragment corresponds to a fragment node in the initial three-dimensional mesh topology, and each fragment adjacency relationship is used to generate a node adjacency record in the initial three-dimensional mesh topology, thereby constructing the initial three-dimensional mesh topology. In a preferred embodiment, the surface mesh fragments can be generated using a voxelized surface over-segmentation method. First, voxel units are formed according to the point cloud processing resolution, and then adjacent voxel units are aggregated according to the point spatial position, point normal representation, and continuous surface representation in the voxel units, so that each surface mesh fragment contains a local continuous surface. The aggregation scale used for region aggregation can be larger than the point cloud processing resolution (e.g., an aggregation scale on the order of 0.15 meters) in order to reduce the number of fragment nodes required for the subsequent collaborative processing of the geometric shape representation, the object semantic representation, and the topological connection representation while preserving the wall-column junctions, door and window openings, and component corner boundaries.

[0024] Optionally, the step of generating geometric shape representation, object semantic representation, and topological connectivity representation based on the 3D scanned point cloud and the initial 3D mesh topology includes: Fragment nodes are generated based on each surface mesh fragment, and a fragment node mapping record representing the correspondence between the surface mesh fragment and the fragment node is generated. The fragment node mapping record is used to associate the fragment adjacency relationship with the corresponding fragment node to generate a node adjacency record. The point data, point spatial position, point normal representation, and surface continuity representation in the surface mesh fragment are registered to the corresponding fragment node to generate a fragment basic description. In each fragment node, a geometric shape representation position, an object semantic representation position, and a topological connection representation position are configured. The fragment basic description is associated with the geometric shape representation position and the object semantic representation position, respectively, and the fragment basic description and the node adjacency record are associated with the topological connection representation position to generate the geometric shape representation, the object semantic representation, and the topological connection representation, respectively.

[0025] Specifically, in this application, the fragment node is a node record in the initial 3D mesh topology used to register a surface mesh fragment and its related data. For each surface mesh fragment, a fragment node corresponding one-to-one with the surface mesh fragment is created, and mutually associated node numbers are configured for the surface mesh fragment and the fragment node. The point data index in the surface mesh fragment, the node number corresponding to the surface mesh fragment, and the storage location of the fragment node in the initial 3D mesh topology together form the fragment node mapping record. The fragment node mapping record is used for bidirectional positioning between the local point data in the 3D scanned point cloud and the fragment nodes in the initial 3D mesh topology. When subsequent processing needs to return the 3D scanned point cloud from the fragment node, the point data in the corresponding surface mesh fragment can be retrieved according to the fragment node mapping record; when subsequent processing determines local point data from the 3D scanned point cloud, the fragment node to which the local point data belongs can be determined according to the fragment node mapping record. Therefore, the fragment node mapping record is not only used to establish the initial 3D mesh topology, but also used to extract the local point cloud corresponding to the mesh region to be reconstructed.

[0026] The fragment adjacency relationship initially records the boundary continuity relationships between the surface mesh fragments. Using the fragment node mapping record, the surface mesh fragments at both ends of each fragment adjacency relationship are converted into corresponding fragment nodes. Then, the node number, boundary continuity position, and boundary continuity range of the two fragment nodes are registered as node adjacency records. These node adjacency records serve as fragment node association data in the initial 3D mesh topology, and are subsequently used to generate the candidate topology connections and to retrieve the basic fragment descriptions of adjacent fragment nodes when generating the object semantic representation. The basic fragment descriptions are the common inputs used by the surface mesh fragments when participating in the generation of the geometric shape representation, the object semantic representation, and the topology connection representation. When generating the basic fragment descriptions, the point data indexes, point spatial positions, point normal representations, and surface continuity representations in the surface mesh fragments are merged according to the same fragment node. The point data index is used to read the point data from the 3D scanned point cloud. The point spatial position is used to calculate the spatial fit relationship between the surface primitive and the point data. The point normal representation is used to determine the local surface direction. The surface continuity representation is used to represent the continuity relationship and the boundary relationship inside the surface mesh fragment.

[0027] The geometric shape representation position, object semantic representation position, and topological connection representation position in each fragment node can use data fields that are independent of each other but share the same node number. The geometric shape representation position registers the surface primitive description, geometric deviation record, and geometric reliability record; the object semantic representation position registers the candidate object semantic category and semantic reliability record; the topological connection representation position registers the candidate topological connection, topological connection form, and topological consistency record. The geometric shape representation position, object semantic representation position, and topological connection representation position are all associated with the fragment basic description through the same fragment node, thereby making the geometric shape representation and the object semantic representation correspond to the same surface mesh fragment, and making the topological connection representation correspond to the node adjacency record between related fragment nodes. In one embodiment, the fragment node is stored in a field-based manner, and the data fields in the fragment node may include the node number, the point data index, the point spatial location range, the point normal representation statistics, the surface continuity representation, the adjacent node number, the geometric shape representation position, the object semantic representation position, and the topological connection representation position. The node adjacency record and the fragment node are stored separately. The node adjacency record includes at least the node number of the two fragment nodes, the boundary continuation position, and the boundary continuation range. With the above storage method, the cyclic correction only needs to update the corresponding geometric shape representation position, the object semantic representation position, or the topological connection representation position, without changing the correspondence between the surface mesh fragment and the fragment node.

[0028] Optionally, generating the geometric representation includes: For each surface mesh fragment, a candidate description of a surface primitive corresponding to a pre-configured surface primitive type is generated using the point data and spatial positions of the points in the surface mesh fragment. A spatial fit check is performed on each candidate description of the surface primitive and the point data to generate a spatial fit record. The surface primitive description is determined based on the spatial fit record, and a geometric deviation record is generated based on the spatial deviation between the surface primitive description and the point data. A geometric reliability record is generated based on the geometric deviation record. The surface primitive description, the geometric deviation record, and the geometric reliability record are registered at the corresponding geometric shape representation position to obtain the geometric shape representation. The surface primitive description is used to participate in the generation of the collaborative reconstruction constraints, and the geometric deviation record and the geometric reliability record are used to determine the degree of uncertainty of the geometric shape representation.

[0029] Specifically, in this application, the surface primitive type is configured before generating the geometric shape representation. The surface primitive type represents a parametric surface category capable of describing the local surface morphology of the scanned object. For existing building scanning scenarios, the surface primitive type may include planes, cylindrical surfaces, and conical surfaces; for scanning scenarios containing free-form surface components, a free-form surface type represented by control points and surface order can also be added. Each surface primitive type corresponds to a set of surface parameter generation rules and point-to-surface distance calculation rules, enabling different surface primitive candidate descriptions to be compared using the same spatial fit verification standard. For each surface mesh fragment, the point data in the surface mesh fragment are subjected to random sampling consistency processing according to each pre-configured surface primitive type. Each time the random sampling consistency processing is performed, point data that can determine the surface parameters corresponding to the current surface primitive type are selected from the surface mesh fragment, a set of surface parameters is generated based on the selected point data, and the corresponding candidate surface is determined using the surface parameters. Subsequently, the spatial distance from the remaining point data in the surface mesh fragment to the candidate surface is calculated. Point data whose spatial distance does not exceed the fitting distance range are registered as interior point data. The proportion of interior point data to all point data in the surface mesh fragment and the spatial distance from the interior point data to the candidate surface are registered in the fitting process record.

[0030] The random sampling consistency process is repeatedly performed on the same surface primitive type. Based on the fitting process records, candidate surfaces with a high proportion of interior point data and a low spatial distance from the interior point data to the candidate surface are selected to generate candidate surface primitive descriptions corresponding to that surface primitive type. Each candidate surface primitive description includes at least the surface primitive type and the corresponding surface parameters. For example, a candidate surface primitive description of a planar type may include the plane's normal direction and the plane's position relative to the origin; a candidate surface primitive description of a cylindrical surface type may include the axial direction, the position the axial direction passes through, and the radius; a candidate surface primitive description of a conical surface type may include the axial direction, vertex position, and opening degree. The spatial fit test is used to determine the surface primitive description that best matches the surface mesh fragment among candidate surface primitive descriptions of different surface primitive types. For each candidate surface primitive description, the spatial distance from the point data to the corresponding candidate surface, the proportion of the interior point data, and the distribution of the spatial distance are statistically analyzed, and these parameters are written into the spatial fitting record. The proportion of the interior point data in the spatial fitting record reflects the degree to which the corresponding candidate surface covers the surface mesh fragment, and the distribution of the spatial distance reflects the degree of deviation of the point data relative to the corresponding candidate surface.

[0031] When determining the surface primitive description based on the spatial fitting record, candidate surface primitive descriptions with an interior point data ratio lower than a pre-configured interior point coverage threshold are first excluded. Then, the candidate surface primitive description with a lower spatial distance is selected from the remaining candidate surface primitive descriptions as the surface primitive description. The interior point coverage threshold is configured based on the proportion of local missing points allowed in the surface mesh fragment. By adopting the above determination order, the number of candidate surfaces that only cover a small amount of the point data but have a low local spatial distance can be determined as the surface primitive description. In an exemplary embodiment, the fitting distance range of the random sampling consistency processing can be configured to be on the order of centimeters (e.g., 0.02 meters) according to the scan resolution, and the number of repeated samplings can be set according to the number of point data in the surface mesh fragment (e.g., no more than 1000 times). After determining the surface primitive description, the spatial deviation from the point data in the surface mesh fragment to the surface represented by the surface primitive description is calculated. The squares of the multiple spatial deviations are averaged, and the overall deviation degree is determined based on the average result to generate the geometric deviation record. The geometric deviation record may also include the proportion of interior point data and the proportion of point data where the spatial deviation exceeds the fitting distance range, so that the geometric deviation record simultaneously reflects the overall fit state and local anomaly state of the surface primitive description to the surface mesh fragment.

[0032] The geometrically reliable record is obtained by converting the geometric deviation record. During the conversion, the overall deviation degree represented by the geometric deviation record is compared with a pre-configured geometric deviation scale. Surface primitive descriptions with lower overall deviation degrees correspond to higher geometric reliability, while surface primitive descriptions with higher overall deviation degrees correspond to lower geometric reliability. As one implementation, the geometrically reliable record can be generated according to a mapping relationship that gradually decreases as the overall deviation degree increases. The geometric deviation scale can be configured based on the scan resolution (e.g., 0.015 meters). When the overall deviation degree exceeds a pre-configured geometric high uncertainty threshold, a geometric high uncertainty state is written into the geometrically reliable record (e.g., writing the geometric high uncertainty state when the overall deviation degree exceeds 0.02 meters). The surface primitive description, the geometric deviation record, and the geometrically reliable record together form the geometric shape representation. The surface primitive description is used to subsequently generate geometric semantic consistency records and geometric topological consistency records; the geometric deviation record and the geometric confidence record are used to determine the uncertainty of the geometric shape representation and to determine whether the geometric shape representation should be used as the representation to be corrected or the supporting representation in the cooperative reconstruction constraints. Thus, the geometric shape representation not only describes the geometric shape of the surface mesh fragment but also carries the uncertainty of the geometric shape relative to the geometric shape representation of the 3D scanned point cloud. In a preferred embodiment, for surface mesh fragments whose geometric confidence records are in a state of high geometric uncertainty, the determined surface primitive description is not immediately deleted; instead, multiple candidate surface primitive descriptions and the spatial fitting records corresponding to each candidate surface primitive description are retained. The retained multiple candidate surface primitive descriptions can re-participate in the determination of the surface primitive description after the object semantic representation or the topological connectivity representation provides new cooperative reconstruction constraints, reducing the possibility of prematurely fixing a single surface primitive type in the initial geometric fitting.

[0033] Optionally, generating the semantic representation of the object includes: Fragment semantic features are generated using the point normal representation, the surface continuity representation, and the boundary morphology of the surface mesh fragments; object semantic category indexes pre-configured according to 3D object types are obtained, and candidate object semantic categories are filtered from the object semantic category index based on the fragment semantic features; a semantically reliable record for each candidate object semantic category is generated by combining the geometric shape representation, the fragment basic description of adjacent fragment nodes, and the fragment semantic features; the candidate object semantic categories and the semantically reliable records are registered to the corresponding object semantic representation positions to obtain the object semantic representation; the candidate object semantic categories are used to participate in the generation of the collaborative reconstruction constraints, and the degree of uncertainty of the object semantic representation is determined using the semantically reliable records.

[0034] Specifically, in this application, the fragment semantic features are used to describe the local attributes of the surface mesh fragments related to the 3D object type in terms of surface orientation, surface variation, and boundary morphology. When generating the fragment semantic features, firstly, the main surface orientation of the surface mesh fragment is determined based on the directional distribution of multiple point normal representations within the surface mesh fragment; then, the degree of surface variation within the surface mesh fragment is determined based on the continuity relationship and boundary relationship in the continuous surface representation; subsequently, the size ratio and boundary extension direction of the surface mesh fragment in three spatial directions are determined based on the boundary point data of the surface mesh fragment. The main surface orientation, the degree of surface variation, the size ratio, and the boundary extension direction are written into the fragment semantic features in a fixed field order. The fragment semantic features then participate in the screening of candidate object semantic categories and the generation of semantically reliable records. The object semantic category index is configured before generating the object semantic representation. When configuring the object semantic category index, a category number is configured for each 3D object type, and the 3D object type, the category number, and the surface morphology typically corresponding to the 3D object type are written into the same index record. In existing building scanning scenarios, the 3D object types can include walls, columns, beams, slabs, doors, windows, and staircases. The above-described 3D object types are merely examples; the object semantic category index can also add other 3D object types that can be distinguished by the 3D scan point cloud, based on the scene to be reconstructed.

[0035] In one embodiment, the selection of candidate object semantic categories from the object semantic category index based on the fragment semantic features can employ a graph attention classification structure consisting of an input mapping layer, a graph attention encoding layer, and a category output layer. The input mapping layer maps the main surface direction, surface variation degree, size ratio, and boundary extension direction in the fragment semantic features to a unified data dimension, enabling comparison of the fragment semantic features of different fragment nodes using the same data dimension. For each fragment node, the graph attention encoding layer reads the fragment semantic features of the current fragment node and the basic fragment descriptions of the adjacent fragment nodes pointed to by the node adjacency record. The graph attention encoding layer calculates the correlation between the current fragment node and each adjacent fragment node in the main surface direction, surface variation degree, and boundary extension direction, and assigns adjacency weights to each adjacent fragment node based on the correlation degree. Subsequently, the basic fragment descriptions of adjacent fragment nodes are summarized according to the adjacency weights, and the summarized basic fragment descriptions are combined with the semantic features of the current fragment node to form the semantic encoding result of the current fragment node. The graph attention encoding layer can be set in multiple layers (e.g., two layers). The semantic encoding result of the previous graph attention encoding layer serves as the input to the next graph attention encoding layer, enabling the current fragment node to gradually absorb the spatial association information of adjacent surface mesh fragments.

[0036] The category output layer maps the semantic encoding result to the category probability corresponding to each of the three-dimensional object types in the object semantic category index. The graph attention classification structure can be trained using building point clouds that have been labeled with the three-dimensional object types. During training, the labeled categories are compared with the category probabilities generated by the category output layer, and the cross-entropy between the labeled categories and the category probabilities is used as the training error. The parameters in the input mapping layer, the graph attention encoding layer, and the category output layer are adjusted according to the training error. As an example, the training data may include labeled point clouds of multiple existing buildings (e.g., 42 buildings, approximately 310 million labeled points). The internal data dimension of the graph attention encoding layer can be set to the hundreds (e.g., 128 dimensions). Each graph attention encoding layer can set multiple adjacency weight calculation branches (e.g., 4). Each adjacency weight calculation branch generates a set of adjacency weights, and then the semantic encoding results corresponding to multiple sets of adjacency weights are combined. When filtering the candidate object semantic categories, the category probabilities are matched with the 3D object types in the object semantic category index in descending order, and the 3D object types whose category probabilities are within the candidate object semantic category filtering range are retained to form the candidate object semantic categories. The candidate object semantic category filtering range can be configured using a probability range or a sorting quantity. When using the probability range, an object semantic candidate threshold is pre-configured, and 3D object types with category probabilities not lower than the object semantic candidate threshold are retained; when using the sorting quantity, a candidate category retention quantity is pre-configured, and 3D object types whose category probabilities rank high and whose quantity does not exceed the candidate category retention quantity are retained. Thus, at the surface mesh fragment where semantic judgment is unstable, only one unadjustable 3D object type will not be retained in the initial judgment stage.

[0037] When generating the semantically reliable record, the semantic categories of each candidate object are re-evaluated by combining the geometric shape representation, the fragment basic descriptions of adjacent fragment nodes, and the fragment semantic features. The surface primitive descriptions in the geometric shape representation are used to determine whether the semantic category of the candidate object is consistent with the local surface morphology; the fragment basic descriptions of adjacent fragment nodes are used to determine whether the semantic category of the candidate object is consistent with the spatial distribution of adjacent surfaces; and the fragment semantic features are used to represent the category distinction information of the current surface mesh fragment itself. The above evaluation results and the category probabilities are written together into the semantically reliable record. The semantically reliable record can use a semantic reliability level between 0 and 1 to represent the reliability state of the candidate object's semantic category, and records the candidate object's semantic category with the highest category probability and its corresponding semantic reliability level. When the highest semantic reliability level is lower than a pre-configured semantic high uncertainty threshold, a semantic high uncertainty state is written into the semantically reliable record (e.g., when the semantic reliability level is lower than 0.6, a semantic high uncertainty state is written). The candidate object's semantic category and the semantically reliable record are jointly registered at the object semantic representation position to form the object semantic representation. The candidate object semantic categories in the object semantic representation are used to generate the geometric semantic consistency record and the semantic topology consistency record. The semantic reliability record is used to determine the degree of uncertainty of the object semantic representation. The degree of uncertainty of the object semantic representation is also used to subsequently determine whether the object semantic representation is the representation to be corrected or the supporting representation in the cooperative reconstruction constraints. Thus, the object semantic representation retains the adjustable space of the candidate object semantic categories, allowing the geometric shape representation and the topological connection representation to change the arrangement order of the candidate object semantic categories in the subsequent cyclic correction.

[0038] Optionally, generating the topological connectivity representation includes: Candidate topological connections between adjacent fragment nodes are generated based on the node adjacency record. A topological connection form is generated by combining the candidate topological connections and the geometric shape representation. Topological connection rules pre-configured according to allowed connection methods between 3D objects are obtained. The topological connection form is verified using the object semantic representation and the topological connection rules, generating a topological consistency record. The candidate topological connections, the topological connection form, and the topological consistency record are registered to the corresponding topological connection representation positions to obtain the topological connection representation. The candidate topological connections and the topological connection form are used to participate in the generation of the collaborative reconstruction constraints, and the uncertainty of the topological connection representation is determined using the topological consistency record.

[0039] Specifically, in this application, the candidate topological connection represents the spatial basis for forming an internal connection within the same 3D object or a connection between different 3D objects between two adjacent fragment nodes. However, the candidate topological connection does not directly indicate that the two adjacent fragment nodes have formed a valid connection when initially generated. When generating the candidate topological connection, the two end fragment nodes, the boundary continuation position, and the boundary continuation range in the node adjacency record are read, and the two fragment nodes corresponding to the node adjacency record are registered as a group of candidate topological connections. Subsequently, the topological connection form is generated by combining the geometric shape representations of the fragment nodes at both ends of the candidate topological connection. When generating the topological connection form, the surface primitive descriptions are read from the two geometric shape representations respectively, and the relative direction, boundary spacing, and spatial overlap state of the two local surfaces are determined based on the two surface primitive descriptions. When two surfaces described by the surface primitives have similar directions and their boundaries extend continuously, the topological connection is registered as a surface continuation relationship; when two surfaces described by the surface primitives intersect at their boundaries, the topological connection is registered as a boundary intersection relationship; when the boundary range of one surface primitive description is located within the boundary range of another surface primitive description, the topological connection is registered as a nesting relationship. The surface continuation relationship, the boundary intersection relationship, and the nesting relationship all belong to the topological connection forms and are used to describe the connection method of the candidate topological connection in geometric space.

[0040] The topology connection rules are configured before generating the topology consistency record. When configuring the topology connection rules, for each 3D object type, another 3D object type allowed to connect to that 3D object type is recorded, along with the allowed topology connection forms between the two 3D object types. For example, the allowed topology connection forms between wall-type 3D objects may include surface continuation relationships or boundary intersection relationships; the allowed topology connection forms between door-type 3D objects and wall-type 3D objects may include embedding relationships; and the allowed topology connection forms between window-type 3D objects and wall-type 3D objects may include embedding relationships. The 3D object types in the topology connection rules and the 3D object types in the object semantic category index use the same category number, allowing the object semantic representation to directly participate in the verification of the topology connection rules. When verifying the topology connection form using the object semantic representation and the topology connection rules, the candidate object semantic categories of the fragment nodes at both ends of the candidate topology connection are read, and the allowed topology connection forms allowed by the corresponding candidate object semantic category combinations are retrieved from the topology connection rules. The actually generated topology connection form is compared with the allowed topology connection form to generate the topology consistency record. The topology consistency record at least records the fragment nodes at both ends of the candidate topology connection, the corresponding candidate object semantic category, the actually generated topology connection form, the allowed topology connection form, and whether the topology connection rule is satisfied.

[0041] When a candidate topological connection corresponds to multiple combinations of candidate object semantic categories, topological rule verification results are generated for each combination of candidate object semantic categories, and then the topological consistency record is determined by combining the semantic credibility record of each candidate object semantic category. When the candidate object semantic category combination with a high degree of semantic credibility is consistent with the actually generated topological connection form, the topological consistency record corresponds to a lower uncertainty in the topological connection representation; when multiple candidate object semantic category combinations with similar degrees of semantic credibility correspond to different allowed topological connection forms, the topological consistency record corresponds to a higher uncertainty in the topological connection representation. The candidate topological connection, the topological connection form, and the topological consistency record are jointly registered to the corresponding topological connection representation position to form the topological connection representation. The candidate topological connection and the topological connection form are used to generate the semantic topological consistency record and the geometric topological consistency record, and the topological consistency record is used to determine the uncertainty in the topological connection representation. The uncertainty in the topological connection representation is also used to determine whether the topological connection representation is the representation to be corrected or the supporting representation in the collaborative reconstruction constraints. In a preferred embodiment, for candidate topological connections whose topological consistency records do not satisfy the topological connection rules, the candidate topological connections are not immediately deleted. Instead, the unsatisfied state and the corresponding topological connection form are retained in the topological connection representation. The retained candidate topological connections can be re-examined by the topological connection rules after the object semantic representation or the geometric shape representation has been corrected. This allows the topological connection representation to participate in the correction of the object semantic representation and the geometric shape representation, and also to accept the subsequent correction results of the object semantic representation and the geometric shape representation.

[0042] Optionally, generating collaborative reconstruction constraints based on the consistent relationship between the geometric shape representation, the object semantic representation, and the topological connectivity representation includes: A geometric semantic consistency record is generated based on the surface primitive description, the candidate object semantic category, and pre-configured geometric semantic correspondence rules. A semantic topological consistency record is generated based on the candidate object semantic category, the topological connection form, and the topological connection rules. A surface boundary is generated based on the surface primitive description, and the candidate topological connection is used to check whether the surface boundary of adjacent surface mesh fragments satisfies the topological connection form, generating a geometric topological consistency record. Based on the geometric semantic consistency record, the semantic topological consistency record, and the unsatisfied states in the geometric topological consistency record, a representation to be corrected is determined from the geometric shape representation, the object semantic representation, and the topological connection representation, and a support representation for correcting the representation to be corrected is determined from the geometric shape representation, the object semantic representation, and the topological connection representation. A correction pointing record is generated based on the representation to be corrected and the support representation, and the geometric semantic consistency record, the semantic topological consistency record, the geometric topological consistency record, and the correction pointing record are combined into the collaborative reconstruction constraint.

[0043] Specifically, in this application, the geometric semantic correspondence rule is used to record the allowed correspondence between the surface primitive type and the three-dimensional object type. When configuring the geometric semantic correspondence rule, each three-dimensional object type in the object semantic category index is associated with the surface primitive type that typically corresponds to that three-dimensional object type. For example, a wall-type three-dimensional object typically corresponds to a plane type, a column-type three-dimensional object can correspond to a cylindrical surface type or a combination of surfaces composed of multiple planes, and a plate-type three-dimensional object typically corresponds to a plane type. The geometric semantic correspondence rule can also record multiple surface primitive types that are allowed to be used for the same three-dimensional object type, to accommodate irregularly shaped components or three-dimensional objects composed of multiple local surfaces. For each fragment node, the surface primitive description is read from the geometric shape representation, the candidate object semantic category is read from the object semantic representation, and then the geometric semantic correspondence rule is queried based on the surface primitive type in the surface primitive description and the candidate object semantic category. When the surface primitive type and the candidate object semantic category satisfy the geometric semantic correspondence rule, a satisfied state is written into the geometric semantic consistency record; when the surface primitive type and the candidate object semantic category do not satisfy the geometric semantic correspondence rule, a not satisfied state is written into the geometric semantic consistency record. The geometric semantic consistency record is also associated with the geometric credibility record corresponding to the surface primitive description and the semantic credibility record corresponding to the candidate object semantic category, so as to subsequently determine whether the geometric shape representation or the object semantic representation should be corrected.

[0044] For each candidate topology connection, the candidate object semantic category of the fragment nodes at both ends of the candidate topology connection is read, and the corresponding topology connection form is read. Then, based on the candidate object semantic category combination, the topology connection rules are queried, and the queried allowed topology connection forms are compared with the actually generated topology connection forms to generate a semantic topology consistency record. When the candidate object semantic category combination and the actually generated topology connection form satisfy the topology connection rules, a satisfied state is written into the semantic topology consistency record; when the candidate object semantic category combination and the actually generated topology connection form do not satisfy the topology connection rules, a unsatisfied state is written into the semantic topology consistency record. The unsatisfied state in the semantic topology consistency record indicates that the current candidate object semantic category combination and the current topology connection form cannot be simultaneously satisfied. Subsequently, the correction direction of the object semantic representation or the topology connection representation is determined by combining the semantic reliability record and the topology consistency record.

[0045] The surface boundary is generated based on the surface primitive description and the boundary point data in the surface mesh fragments. For planar surface primitive descriptions, the boundary point data is projected onto the corresponding plane, and the surface boundary is formed according to the spatial adjacency order of the boundary point data. For cylindrical or conical surface primitive descriptions, the boundary point data is converted to surface coordinates based on the surface axis, and the surface boundary is formed according to the adjacency order in the surface coordinates. The surface boundary continues to maintain association with the original surface mesh fragments and their corresponding fragment nodes through the fragment node mapping record. When using the candidate topology connection to check whether the surface boundaries of adjacent surface mesh fragments satisfy the topology connection form, the surface boundaries at both ends of the candidate topology connection are read respectively, and the corresponding boundary check is performed according to the topology connection form. For the surface continuation relationship, the spacing and directional changes of the two surface boundaries in the extension direction are examined; for the boundary intersection relationship, it is examined whether the two surface boundaries form an intersection within a pre-configured allowable space; for the embedding relationship, it is examined whether one surface boundary is located within the space defined by the other surface boundary. The results of the boundary checks are written into the geometric topology consistency record. When the surface boundary and the topological connection form are compatible in geometric space, the satisfied state is written into the geometric topology consistency record; when the surface boundary formed by the surface primitive description and the current topological connection form cannot be simultaneously satisfied, the unsatisfied state is written into the geometric topology consistency record. The geometric topology consistency record is also associated with the geometrically reliable record corresponding to the surface primitive description and the topology consistency record corresponding to the candidate topological connection, so as to subsequently determine the correction direction of the geometric shape representation or the topological connection representation.

[0046] When determining the representation to be corrected, the unsatisfied states in the geometric semantic consistency record, the semantic topology consistency record, and the geometric topology consistency record are first read. Then, the two types of representations involved in each unsatisfied state are determined. Subsequently, the uncertainty of the two types of representations involved in the unsatisfied state is compared. The type of representation with higher uncertainty is determined as the representation to be corrected, and the type of representation with lower uncertainty is determined as the supporting representation. For example, when the geometric semantic consistency record is in an unsatisfied state, and the uncertainty of the geometric shape representation is higher than the uncertainty of the object semantic representation, the geometric shape representation is determined as the representation to be corrected, and the object semantic representation is determined as the supporting representation. When the same fragment node corresponds to multiple unsatisfied states simultaneously, the number of unsatisfied states corresponding to the geometric shape representation, the object semantic representation, and the topology connection representation, and the uncertainty of each unsatisfied state are counted respectively. The representation corresponding to multiple unsatisfied states and with higher uncertainty is preferentially determined as the representation to be corrected. The supporting representation is determined from another type of representation that has a consistency relationship with the representation to be corrected and is located in the corresponding consistency record. Therefore, both the characterization to be corrected and the supporting characterization have traceable and consistent records.

[0047] The correction pointing record is used to register the fragment nodes or candidate topological connections, the representation to be corrected, the support representation, and the corresponding consistency record involved in a correction. When the representation to be corrected is a geometric shape representation, the correction pointing record records the constraint direction formed by the object semantic representation or the topological connection representation on the surface primitive description; when the representation to be corrected is an object semantic representation, the correction pointing record records the constraint direction formed by the geometric shape representation or the topological connection representation on the candidate object semantic category; when the representation to be corrected is a topological connection representation, the correction pointing record records the constraint direction formed by the geometric shape representation or the object semantic representation on the candidate topological connection and the topological connection form. The geometric semantic consistency record, the semantic topological consistency record, the geometric topological consistency record, and the correction pointing record combine to form the collaborative reconstruction constraint. The collaborative reconstruction constraint records both the coordination state between the geometric shape representation, the object semantic representation, and the topological connection representation, and the correction direction corresponding to the unsatisfied state. Therefore, the collaborative reconstruction constraint can directly participate in the subsequent cyclic correction, rather than merely serving as evaluation information for the parametric 3D mesh model.

[0048] In a preferred embodiment, the collaborative reconstruction constraints may further include geometric deviation constraint components, object semantic constraint components, topological consistency constraint components, and cross-consistency constraint components. The geometric deviation constraint components are determined based on the geometric deviation record and the geometric reliability record, and are used to constrain the spatial deviation between the surface primitive description and the point data. The object semantic constraint components are determined based on the candidate object semantic category and the semantic reliability record, and are used to constrain the correspondence between the candidate object semantic category and the fragment semantic features. The topological consistency constraint components are determined based on the topological consistency record, and are used to constrain the candidate topological connections and the topological connection forms to satisfy the topological connection rules. The cross-consistency constraint components are determined based on the geometric semantic consistency record, the semantic topological consistency record, and the geometric topological consistency record, and are used to constrain the consistency between the geometric shape representation, the object semantic representation, and the topological connection representation. The geometric deviation constraint components, the object semantic constraint components, the topological consistency constraint components, and the cross-consistency constraint components can each be configured with constraint weights, and the constraint weights can be adjusted according to the uncertainty of the geometric shape representation, the object semantic representation, and the topological connection representation. In an exemplary embodiment, the initial constraint weights of the geometric deviation constraint component, the object semantic constraint component, the topological consistency constraint component, and the cross-consistency constraint component can be set to 1.0, 0.8, 1.2, and 0.6, respectively. When one of the geometric shape representation, the object semantic representation, or the topological connection representation is in a corresponding high uncertainty state, the constraint weight corresponding to the constraint component formed by that type of representation alone is reduced, and the constraint weight corresponding to the cross-consistency constraint component related to that type of representation is increased, so that the other two types of representations participate more in the subsequent correction of that type of representation.

[0049] Optionally, before performing cyclic correction on the geometric shape representation, the object semantic representation, and the topological connection representation using the collaborative reconstruction constraints, the method further includes: generating a geometrically uncertain record based on the geometric deviation record and the geometrically reliable record; generating a semantically uncertain record based on the semantically reliable record; and generating a topologically uncertain record based on the topological consistency record; arranging the geometric shape representation, the object semantic representation, and the topological connection representation according to the degree of uncertainty represented by each of the geometrically uncertain record, the semantically uncertain record, and the topologically uncertain record to generate an uncertain ranking record; determining the representation to be corrected and the supporting representation using the correction pointing record; adjusting the arrangement position of the representation to be corrected in the uncertain ranking record according to the arrangement position of the representation to be corrected and the supporting representation in the uncertain ranking record to obtain a corrected uncertain ranking record; and arranging the geometric shape representation, the object semantic representation, and the topological connection representation according to the corrected uncertain ranking record to obtain a correction order, so that the correction order controls the order in which the geometric shape representation, the object semantic representation, and the topological connection representation are cyclically corrected.

[0050] Specifically, in this application, the geometric uncertainty record, the semantic uncertainty record, and the topological uncertainty record are used to convert the sources of uncertainty in the geometric shape representation, the object semantic representation, and the topological connection representation into uncertainties that can be compared sequentially. Since the geometric deviation record represents the spatial deviation between the surface primitive description and the point data, and the geometric reliability record represents the geometric reliability of the surface primitive description relative to the 3D scanned point cloud, when generating the geometric uncertainty record, the overall deviation level, the proportion of inlier data, and the proportion of point data exceeding the fitting distance range are first read from the geometric deviation record. Then, the geometric reliability level and the geometrically high uncertainty state are read from the geometric reliability record. Subsequently, according to the correspondence that the higher the overall deviation level, the lower the proportion of inlier data, and the lower the geometric reliability level, the higher the uncertainty level of the geometric shape representation, the overall deviation level, the proportion of inlier data, the proportion of point data exceeding the fitting distance range, the geometric reliability level, and the geometrically high uncertainty state are converted into the geometric uncertainty record. The geometric uncertainty record is associated with the corresponding fragment node, surface primitive description, and geometric shape representation position, enabling subsequent processing to locate the geometric shape representation that needs correction based on the geometric uncertainty record. When generating the semantic uncertainty record, the category probability and semantic credibility of each candidate object semantic category are read from the semantic credibility record, and the category probability difference between the top-ranked candidate object semantic category and the remaining candidate object semantic categories is calculated. When multiple top-ranked candidate object semantic categories have similar category probabilities, the category probability difference is small, and the surface mesh fragment lacks a stable distinguishing basis among multiple 3D object types, thus recording a high degree of uncertainty in the semantic uncertainty record. When the top-ranked candidate object semantic category has a high degree of semantic credibility, and the category probability difference is large, recording a low degree of uncertainty in the semantic uncertainty record. When a high-uncertainty state has already been written into the semantic credibility record, the high-uncertainty state is synchronously written into the semantic uncertainty record. The semantically uncertain record is also associated with the corresponding fragment node, the candidate object semantic category, and the object semantic representation position, so that the arrangement order of the candidate object semantic categories can be corrected subsequently based on the semantically uncertain record. When generating the topology uncertain record, the candidate topology connection, the topology connection form, the allowed topology connection form, and the status of whether the topology connection rule is satisfied are read from the topology consistency record.When a candidate topological connection corresponds to only one topological connection form that is consistent with the semantic category of the candidate object, and the topological connection form satisfies the topological connection rules, a lower degree of uncertainty is recorded in the topological uncertainty record. When a candidate topological connection corresponds to multiple different topological connection forms, or when the topological connection form changes with the combination of semantic categories of the candidate objects, a higher degree of uncertainty is recorded in the topological uncertainty record. When the topological consistency record is in the unsatisfied state, the topological uncertainty record also records the combination of semantic categories of the candidate objects and the topological connection form that caused the unsatisfied state. The topological uncertainty record is associated with the corresponding candidate topological connection, the fragment node, and the topological connection representation position, so that the topological uncertainty record can directly participate in the correction of the topological connection representation, rather than just recording whether the candidate topological connection exists.

[0051] The geometrically uncertain records, semantically uncertain records, and topologically uncertain records all use the same uncertainty range. When transforming the sources of uncertainty in the geometric shape representation, the object semantic representation, and the topological connection representation, the geometric deviation scale, the semantic high uncertainty threshold, and the verification status of the topological connection rule are used as the transformation criteria, so that the geometrically uncertain records, semantically uncertain records, and topologically uncertain records can be arranged under the same comparison scale. As one implementation, the uncertainty level can be divided into multiple continuous intervals from low to high, or it can be represented by a numerical value from 0 to 1; where, close to 0 indicates that the corresponding geometric shape representation, object semantic representation, or topological connection representation is relatively stable, and close to 1 indicates that the corresponding geometric shape representation, object semantic representation, or topological connection representation needs to be corrected first. The specific numerical value of the uncertainty level is used to form an order comparison without changing the technical relationship between the geometrically uncertain records, semantically uncertain records, and topologically uncertain records. When generating the uncertain sorting record, for the same fragment node, the geometric uncertain record, the semantic uncertain record, and the topological uncertain record associated with the fragment node are arranged from high to low according to their degree of uncertainty, and the geometric shape representation, the object semantic representation, and the topological connection representation are respectively associated with their corresponding arrangement positions. The uncertain sorting record not only records the arrangement order of the geometric shape representation, the object semantic representation, and the topological connection representation, but also records the degree of uncertainty corresponding to each of the geometric shape representation, the object semantic representation, and the topological connection representation, as well as the source record that forms the degree of uncertainty. Therefore, the corresponding geometric deviation record, the geometrically reliable record, the semantically reliable record, or the topological consistency record can be retrieved from the uncertain sorting record.

[0052] The correction pointer record further defines the correction relationship between the representation to be corrected and the supporting representation. When the geometric shape representation, object semantic representation, or topological connection representation that is ranked first in the uncertain sorting record does not appear in the unsatisfied state recorded in the correction pointer record, the corresponding sorting position is retained, and when processing the current unsatisfied state, the representation to be corrected is determined from the geometric shape representation, object semantic representation, and topological connection representation involved in the correction pointer record. When the uncertainty of the supporting representation is lower than the uncertainty of the representation to be corrected, and the corresponding consistency record between the supporting representation and the other representations is in the satisfied state, the representation to be corrected is kept before the supporting representation, so that the representation to be corrected receives the correction provided by the supporting representation first; when the uncertainty of the supporting representation is higher than the uncertainty of the representation to be corrected, or the corresponding consistency record between the supporting representation and the other representations is in the unsatisfied state, the supporting representation is first sorted before the representation to be corrected, so that the supporting representation completes the correction before participating in the correction of the representation to be corrected, thereby obtaining the correction uncertain sorting record. In a preferred embodiment, the order of the correction uncertainty sorting records is adjusted according to the dependency relationship between the representation to be corrected and the supporting representation. When the geometric representation is determined to be the representation to be corrected, and both the object semantic representation and the topological connection representation are determined to be the supporting representation, the semantic topological consistency record between the object semantic representation and the topological connection representation is checked first. When the semantic topological consistency record is in the satisfied state, the geometric representation is arranged before the object semantic representation and the topological connection representation to correct the geometric representation using the object semantic representation and the topological connection representation in a coordinated state; when the semantic topological consistency record is in the unsatisfied state, the correction order between the object semantic representation and the topological connection representation is determined first according to the semantic uncertainty record and the topological uncertainty record, and then the geometric representation is arranged after the correction between the object semantic representation and the topological connection representation is completed. By adjusting the arrangement of the correction uncertainty sorting records, the situation where two incompatible supporting representations act on the representation to be corrected simultaneously can be reduced. The correction order is obtained by arranging the geometric representation, the object semantic representation, and the topological connection representation according to the correction uncertainty sorting records. The correction order is registered according to the fragment node or the candidate topology connection and associated with the corresponding correction pointer record.When performing the cyclic correction in the subsequent steps, the representations to be corrected that are ranked earlier in the correction order are processed first, and then the correction results of the representations to be corrected are written to the corresponding geometric shape representation position, the object semantic representation position, or the topological connection representation position, so that the representations to be corrected that are ranked later in the correction order can read the correction results that have already been written.

[0053] Optionally, the step of cyclically correcting the geometric representation, object semantic representation, and topological connection representation using the collaborative reconstruction constraints includes: determining the current representation to be corrected according to the correction order; determining the current supporting representation based on the correction pointing record; and determining the target consistency record corresponding to the current representation to be corrected from the collaborative reconstruction constraints; correcting the current representation to be corrected using the current supporting representation according to the target consistency record to obtain the current corrected representation; determining the current corrected representation as the update result of the corresponding representation according to the correction order; and repeating the aforementioned process to form a correction cycle to obtain updated geometric representation, updated object semantic representation, and updated topological connection representation; and using the updated geometric representation... The geometric semantic consistency record, semantic topology consistency record, and geometric topology consistency record are regenerated based on the updated object semantic representation, the updated topology connection representation, and the updated geometric shape representation to obtain a periodic consistency result. A representation evolution record is generated based on the changes in the updated geometric shape representation, the updated object semantic representation, and the updated topology connection representation relative to the correction cycle input. If the periodic consistency result does not meet the pre-configured collaborative convergence condition, the next correction cycle begins. If the periodic consistency result meets the collaborative convergence condition, the initial 3D mesh topology is updated with the updated geometric shape representation, the updated object semantic representation, and the updated topology connection representation to obtain the optimized 3D mesh topology.

[0054] Specifically, in this application, each correction cycle uses the geometric shape representation, object semantic representation, and topological connection representation in the correction cycle input as the correction basis. When the correction cycle is executed for the first time, the correction cycle input is the geometric shape representation, object semantic representation, and topological connection representation registered in the initial 3D mesh topology; when entering the next correction cycle, the updated geometric shape representation, updated object semantic representation, and updated topological connection representation formed in the previous correction cycle become the correction cycle input for the next correction cycle. When determining the current representation to be corrected according to the correction order, the geometric shape representation, object semantic representation, or topological connection representation that is ranked high in the correction order and has not yet been corrected in the current correction cycle is read, and then the current support representation is read according to the correction direction record associated with the current representation to be corrected. Subsequently, the geometric semantic consistency record, semantic topological consistency record, or geometric topological consistency record that causes the current representation to be corrected to need correction is read from the collaborative reconstruction constraints, and the read geometric semantic consistency record, semantic topological consistency record, or geometric topological consistency record is determined as the target consistency record. The target consistency record records the unmet correspondence between the current representation to be corrected and the current supporting representation, thus enabling the determination of the technical content that needs to be changed in this correction.

[0055] When the current representation to be corrected is the geometric shape representation, the current supporting representation is the object semantic representation, the topological connection representation, or a combination of the object semantic representation and the topological connection representation. When the target consistency record is the geometric semantic consistency record, based on the candidate object semantic category in the current supporting representation, candidate surface primitive descriptions that satisfy the geometric semantic correspondence rule with the candidate object semantic category are selected from the multiple retained candidate surface primitive descriptions, and the spatial fitting records corresponding to the selected candidate surface primitive descriptions are re-compared to determine the updated surface primitive description. When the target consistency record is the geometric topological consistency record, the surface parameters corresponding to the surface primitive description are adjusted according to the topological connection form in the current supporting representation, so that adjacent surface boundaries satisfy the surface continuity relationship, the boundary intersection relationship, or the embedding relationship at the boundary continuation position, to form the updated surface primitive description. The adjustment range of the surface parameters is constrained by the fitting distance range from the point data to the corresponding surface, so that the updated surface primitive description continues to maintain spatial correspondence with the 3D scanned point cloud. After determining the updated surface primitive description, the geometric deviation record and the geometric reliability record are regenerated based on the updated surface primitive description. The updated surface primitive description, the regenerated geometric deviation record, and the regenerated geometric reliability record are then combined to form the current correction representation. When the current representation to be corrected is the object semantic representation, the current supporting representation is the geometric shape representation, the topological connection representation, or a combination of the geometric shape representation and the topological connection representation. When the target consistency record is the geometric semantic consistency record, based on the surface primitive type corresponding to the surface primitive description, candidate object semantic categories that satisfy the geometric semantic correspondence rules are selected from the candidate object semantic categories, and the semantic reliability of the selected candidate object semantic categories in the semantic reliability record is increased. When the target consistency record is the semantic topological consistency record, the topological connection rules are queried based on the candidate topological connections and the topological connection forms. Candidate object semantic categories that can jointly satisfy the topological connection rules with the candidate object semantic categories of adjacent fragment nodes are retained, and the semantic reliability corresponding to candidate object semantic categories that are inconsistent with the topological connection forms is reduced. Subsequently, the candidate object semantic categories are rearranged according to the adjusted semantic credibility to form updated candidate object semantic categories, and the updated candidate object semantic categories and the regenerated semantic credibility records are combined into the current correction representation. When the current representation to be corrected is the topological connectivity representation, the current support representation is the geometric shape representation, the object semantic representation, or a combination of the geometric shape representation and the object semantic representation.When the target consistency record is a geometric topological consistency record, the spatial distance, directional changes, and spatial overlap between adjacent surface boundaries are used to re-determine whether to retain the candidate topological connections, and if so, to determine an updated topological connection form. When the target consistency record is a semantic topological consistency record, the topological connection rules are queried based on the candidate object semantic category combination, and an updated topological connection form that is compatible with the surface boundaries is determined from the allowed topological connection forms recorded in the topological connection rules. Subsequently, the topological connection rule verification is re-executed using the updated topological connection form to generate an updated topological consistency record, and the candidate topological connections, the updated topological connection forms, and the updated topological consistency records are combined into the current correction representation.

[0056] According to the correction order, the current correction representation is written to the corresponding geometric shape representation position, object semantic representation position, or topological connection representation position, and the current correction representation is determined as the update result of the current representation to be corrected. The current representation to be corrected, arranged in a subsequent position in the correction order, reads the already written current correction representation, so that the previous correction result directly participates in the subsequent correction process. After completing all corrections of the geometric shape representation, object semantic representation, and topological connection representation in the correction order, a correction cycle is formed, and the updated geometric shape representation, updated object semantic representation, and updated topological connection representation are obtained. The geometric semantic correspondence rule is re-executed using the updated geometric shape representation and the updated object semantic representation to regenerate the geometric semantic consistency record; the topological connection rule is re-executed using the updated object semantic representation and the updated topological connection representation to regenerate the semantic topological consistency record; the boundary check is re-executed using the surface boundary formed by the updated geometric shape representation and the candidate topological connection and the updated topological connection form in the updated topological connection representation to regenerate the geometric topological consistency record. The regenerated geometric semantic consistency record, the semantic topology consistency record, and the satisfied states, unsatisfied states, and corresponding uncertainties in the geometric topology consistency record are summarized into the periodic consistency result.

[0057] The representation evolution record is used to record the changes in the geometric shape representation, object semantic representation, and topological connection representation between adjacent correction cycles. When generating the representation evolution record, for each fragment node, the updated geometric shape representation is compared with the geometric shape representation in the correction cycle input, recording whether the surface primitive type has changed and the degree of change in the surface parameters; the updated object semantic representation is compared with the object semantic representation in the correction cycle input, recording whether the top-ranked candidate object semantic category has changed and the change in the semantic credibility corresponding to each candidate object semantic category; the updated topological connection representation is compared with the topological connection representation in the correction cycle input, recording whether the candidate topological connection has changed and whether the topological connection form has changed. These changes are associated according to the fragment node and the correction cycle to form the representation evolution record, which is subsequently used to identify the mesh region to be reconstructed. The cooperative convergence conditions are configured before performing the cyclic correction. The cooperative convergence conditions include consistency state conditions and representation change conditions. The consistency state condition is used to determine whether the number of unsatisfied states in the periodic consistency result has decreased to a pre-configured threshold for the number of unsatisfied states. The representation change condition is used to determine whether the changes in the updated geometric representation, the updated object semantic representation, and the updated topological connection representation relative to the correction cycle input have decreased to a pre-configured threshold for representation change. As one implementation, when the number of unsatisfied states no longer decreases in multiple consecutive correction cycles, and the degree of change in the updated geometric representation, the updated object semantic representation, and the updated topological connection representation is lower than the corresponding representation change threshold, the periodic consistency result satisfies the collaborative convergence condition. As another implementation, when the unsatisfied states in the geometric semantic consistency record, the semantic topological consistency record, and the geometric topological consistency record are all eliminated, the periodic consistency result satisfies the collaborative convergence condition. A maximum number of correction cycles (e.g., 20 correction cycles) can also be configured to limit the processing range of one cycle correction. When the periodic consistency result does not meet the collaborative convergence condition, and the current number of correction periods does not reach the maximum number of correction periods, the updated geometric shape representation, the updated object semantic representation, and the updated topological connection representation are used as the correction period input for the next correction period, and the correction order is regenerated based on the regenerated geometric uncertainty record, the semantic uncertainty record, the topological uncertainty record, and the correction pointing record.When the number of current correction cycles reaches the maximum number of correction cycles, the current cyclic correction ends, and the fragment nodes that do not meet the cooperative convergence condition and the candidate topological connections are written into the representation evolution record for subsequent identification of the mesh region to be reconstructed. When the cycle consistency result meets the cooperative convergence condition, the updated geometry representation, the updated object semantic representation, and the updated topological connection representation are written into the corresponding fragment nodes and the node adjacency record to update the initial 3D mesh topology and obtain the optimized 3D mesh topology.

[0058] Optionally, the step of cyclically correcting the geometric shape representation, the object semantic representation, and the topological connectivity representation using the collaborative reconstruction constraints further includes: generating a geometric shape representation admission record based on the geometric uncertainty record, generating an object semantic representation admission record based on the semantic uncertainty record, and generating a topological connectivity representation admission record based on the topological uncertainty record; determining the target representation according to the correction order, determining the target representation admission record corresponding to the target representation from the geometric shape representation admission record, the object semantic representation admission record, and the topological connectivity representation admission record, and determining the target based on the correction direction record. Supporting representation; when the target representation admission record does not meet the pre-configured constraint output conditions, the target representation is corrected using the target supporting representation to obtain a corrected representation, and the corrected representation is prohibited from being identified as the supporting representation; the geometrically uncertain record, the semantically uncertain record, or the topologically uncertain record corresponding to the target representation is regenerated based on the corrected representation to update the target representation admission record; when the updated target representation admission record meets the constraint output conditions, the corrected representation is identified as the supporting representation so that the corrected representation can participate in the subsequent generation of the collaborative reconstruction constraints.

[0059] Specifically, in this application, the geometric shape representation admission record, the object semantic representation admission record, and the topological connection representation admission record are used to control whether the corrected geometric shape representation, the object semantic representation, or the topological connection representation can continue to influence other representations as supporting representations. When generating the geometric shape representation admission record, the uncertainty level in the geometric uncertainty record, the geometric credibility level in the geometric credibility record, and the geometric semantic consistency record and geometric topological consistency record corresponding to the surface primitive description are read. Based on the reading results, the geometric shape representation admission record records whether it satisfies the constraint output condition. When generating the object semantic representation admission record, the uncertainty level in the semantic uncertainty record, the semantic credibility level in the semantic credibility record, and the geometric semantic consistency record and semantic topological consistency record corresponding to the candidate object semantic category are read. Based on the reading results, the object semantic representation admission record records whether it satisfies the constraint output condition. When generating the topology connectivity representation admission record, the uncertainty level, the topology consistency record, and the corresponding semantic topology consistency record and geometric topology consistency record in the topology uncertainty record are read. Based on the reading results, the topology connectivity representation admission record records whether the topology connectivity representation satisfies the constraint output conditions. The constraint output conditions are configured before generating the geometric shape representation admission record, the object semantic representation admission record, and the topology connectivity representation admission record. For the geometric shape representation, the constraint output conditions include that the uncertainty level of the geometric shape representation is lower than a pre-configured geometric admission threshold, the geometric confidence record is not in the geometric high uncertainty state, and the geometric semantic consistency record and the geometric topology consistency record associated with the geometric shape representation do not have a non-satisfied state with opposite correction directions. For the object semantic representation, the constraint output conditions include that the uncertainty level of the object semantic representation is lower than a pre-configured semantic admission threshold, the semantic confidence level corresponding to the first-ranked candidate object semantic category is not lower than the semantic admission threshold, and the geometric semantic consistency record and the semantic topology consistency record are not simultaneously in a non-satisfied state with opposite correction directions. For the topology connectivity representation, the constraint output conditions include that the uncertainty of the topology connectivity representation is lower than a pre-configured topology admission threshold, the topology consistency record satisfies the topology connectivity rule, and the topology connectivity form is consistent with the corresponding surface boundary. The geometric admission threshold, the semantic admission threshold, and the topology admission threshold all use the same uncertainty range as the geometrically uncertain record, the semantically uncertain record, and the topologically uncertain record, respectively.

[0060] After determining the target representation according to the correction order, based on whether the target representation belongs to the geometric representation, the object semantic representation, or the topological connection representation, the corresponding target representation admission record is read from the geometric representation admission record, the object semantic representation admission record, and the topological connection representation admission record. Then, the target support representation corresponding to the target representation is determined according to the correction direction record. The target support representation already satisfies its corresponding constraint output condition. When the correction direction record contains multiple selectable target support representations, the target support representation is selected from low to high uncertainty in the corresponding geometric representation admission record, the object semantic representation admission record, or the topological connection representation admission record. When the target representation admission record does not satisfy the constraint output condition, the target representation still accepts the correction provided by the target support representation, but the corrected representation is temporarily not used as the support representation for other target representations. For example, when the geometric shape representation is in a state of high geometric uncertainty, the surface primitive description is redefined using the object semantic representation and the topological connectivity representation. When the regenerated geometrically reliable record is still in a state of high geometric uncertainty, the corrected representation does not participate in the correction of other object semantic representations or topological connectivity representations, thereby reducing the continuous propagation of the same unstable corrected representation along the fragment adjacency relationship. Based on the corrected representation, the geometrically uncertain record, the semantically uncertain record, or the topologically uncertain record corresponding to the target representation is regenerated. When the target representation is the geometric shape representation, the geometric deviation record, the geometric reliability record, and the geometric uncertainty record are regenerated based on the corrected representation, and the geometric shape representation admission record is updated using the regenerated geometric uncertainty record. When the target representation is the object semantic representation, the semantic reliability record and the semantic uncertainty record are regenerated based on the corrected representation, and the object semantic representation admission record is updated using the regenerated semantic uncertainty record. When the target representation is the topological connectivity representation, the topological consistency record and the topological uncertainty record are regenerated based on the corrected representation, and the topological connectivity representation admission record is updated using the regenerated topological uncertainty record. The updated geometric shape representation admission record, the object semantic representation admission record, or the topological connectivity representation admission record is determined as the updated target representation admission record. When the updated target representation admission record satisfies the constraint output condition, the corrected representation is determined as the support representation, and the corrected representation is written into the geometric semantic consistency record, the semantic topological consistency record, or the geometric topological consistency record used by the subsequent collaborative reconstruction constraints.As one implementation, only the corrected representation that satisfies the constraint output condition in multiple consecutive correction cycles (e.g., two consecutive correction cycles) is determined as the supporting representation, thereby reducing the repeated changes in the corrected representation between satisfying and not satisfying the constraint output condition. The updated changes in the target representation admission record are also written into the representation evolution record so that the reconstructed mesh region that fails to satisfy the constraint output condition in multiple consecutive correction cycles can be identified subsequently based on the representation evolution record.

[0061] Optionally, identifying the mesh region to be reconstructed based on the consistency of the geometric shape representation, object semantic representation, and topological connection representation in the optimized 3D mesh topology includes: extracting the semantic category changes of the same surface mesh fragment in consecutive correction cycles from the representation evolution record to generate a semantic change trajectory, and extracting the changes of the periodic consistency results in consecutive correction cycles to generate a consistency change trajectory; obtaining pre-configured oscillation identification conditions, and determining oscillating fragment nodes based on the semantic change trajectory and the consistency change trajectory; determining associated fragment nodes adjacent to the oscillating fragment nodes using the node adjacency record, and determining the surface mesh fragments corresponding to the oscillating fragment nodes and the associated fragment nodes based on the fragment node mapping record; and determining the mesh region occupied by the oscillating fragment nodes and the associated fragment nodes in the optimized 3D mesh topology as the mesh region to be reconstructed, so that the local point cloud corresponding to the mesh region to be reconstructed is used to reconstruct the 3D mesh topology of the mesh region to be reconstructed.

[0062] Specifically, in this application, the semantic change trajectory is used to represent the change process of the semantic category of the candidate object of the same surface mesh fragment in consecutive correction cycles. When generating the semantic change trajectory, the fragment node corresponding to the surface mesh fragment is determined based on the fragment node mapping record. Then, the first-ranked candidate object semantic category, the arrangement position of each candidate object semantic category, and the corresponding semantic credibility are read from the representation evolution record according to the order of the correction cycles. The first-ranked candidate object semantic category, the arrangement position of each candidate object semantic category, and the corresponding semantic credibility are arranged according to the correction cycle to form the semantic change trajectory. When the first-ranked candidate object semantic category changes between two adjacent correction cycles, a change marker is added to the semantic change trajectory. This change marker is subsequently used to determine the number of semantic category changes. The consistency change trajectory is used to represent the change process of the periodic consistency result corresponding to the same fragment node and the fragment nodes adjacent to the fragment node in consecutive correction cycles. When generating the consistency change trajectory, the satisfied and unsatisfied states of the geometric semantic consistency record, the semantic topology consistency record, and the geometric topology consistency record are read from the periodic consistency results corresponding to each correction cycle. The number of consistency state changes for the same fragment node, from satisfied to unsatisfied or vice versa, is counted. Both the semantic change trajectory and the consistency change trajectory are associated with the same fragment node. The number of semantic category changes and the number of consistency state changes together characterize whether the surface mesh fragment changes repeatedly during the cyclic correction process. The oscillation identification conditions are configured before identifying the oscillating fragment node. The oscillation identification conditions include the number of observed correction cycles, the number of semantic category changes, and the number of consistency state changes. As one implementation, the semantic change trajectory and the consistency change trajectory (e.g., the most recent 5 correction cycles) are extracted from the most recent correction cycles, and the number of correction cycles is determined as the number of observation correction cycles. When the semantic category of the candidate object ranked first changes multiple times within the number of observation correction cycles (e.g., the semantic category changes no less than 3 times), or when the same geometric semantic consistency record, semantic topological consistency record, or geometric topological consistency record changes multiple times between the satisfied state and the unsatisfied state, the corresponding fragment node is determined as the oscillating fragment node. In a preferred embodiment, the oscillation identification condition is further combined with the geometric shape representation admission record, the object semantic representation admission record, or the topological connection representation admission record.When the geometric shape representation, object semantic representation, or topological connectivity representation corresponding to the fragment node fails to meet the constraint output condition in multiple consecutive correction cycles, and the corresponding semantic change trajectory or consistency change trajectory reaches the oscillation identification condition, the fragment node is identified as the oscillating fragment node. Thus, changes in the geometric shape representation, object semantic representation, or topological connectivity representation caused by a single normal correction can be distinguished from changes in the geometric shape representation, object semantic representation, or topological connectivity representation that remain unstable after multiple corrections.

[0063] The node adjacency record is used to read the fragment nodes directly connected to the oscillating fragment node, and these directly connected fragment nodes are identified as the associated fragment nodes. The associated fragment nodes are included in subsequent processing because the geometric semantic consistency record, semantic topological consistency record, and geometric topological consistency record corresponding to the oscillating fragment node are all formed by the relationship between adjacent surface mesh fragments. Reprocessing only the oscillating fragment node would retain the boundary point data that caused the unsatisfied state. Based on the fragment node mapping record, the surface mesh fragments corresponding to the oscillating fragment node and the associated fragment node are determined respectively. The spatial range occupied by the surface mesh fragments corresponding to the oscillating fragment node and the associated fragment node in the optimized 3D mesh topology is merged to determine the mesh region to be reconstructed. The mesh region to be reconstructed at least covers the surface mesh fragments corresponding to the oscillating fragment node, the surface mesh fragments corresponding to the associated fragment node, and the boundary connection positions between them. Subsequently, based on the position of the mesh region to be reconstructed in the unified three-dimensional coordinate space, the local point cloud is extracted from the three-dimensional scanned point cloud. This local point cloud contains not only the point data within the mesh region to be reconstructed but also the neighboring point data outside the boundary of the mesh region. The neighboring point data is used to restore the boundary continuity between the mesh region to be reconstructed and the surrounding optimized three-dimensional mesh topology when reconstructing the three-dimensional mesh topology of the mesh region to be reconstructed.

[0064] Optionally, the step of reconstructing the 3D mesh topology of the region to be reconstructed using the local point cloud to update the optimized 3D mesh topology and obtain a corrected 3D mesh topology includes: obtaining a local processing resolution smaller than the point cloud processing resolution; spatially regularizing the local point cloud according to the local processing resolution to obtain a locally regularized point cloud; constructing a local 3D mesh topology based on the spatial continuity of the locally regularized point cloud; and generating local geometric shape representations, local object semantic representations, and local topological connection representations based on the locally regularized point cloud and the local 3D mesh topology; and verifying the local geometric shape representations, local object semantic representations, and local topology according to the geometric semantic correspondence rules and the topological connection rules. The connection representation is used to obtain a local consistency result. The geometric shape representation, object semantic representation, and topological connection representation corresponding to the mesh region to be reconstructed are checked according to the geometric semantic correspondence rule and the topological connection rule to obtain the original region consistency result. When the consistency degree of the local consistency result representation is higher than the consistency degree of the original region consistency result representation, the local three-dimensional mesh topology and the local geometric shape representation, local object semantic representation, and local topological connection representation in it are used to replace the three-dimensional mesh topology, geometric shape representation, object semantic representation, and topological connection representation corresponding to the mesh region to be reconstructed in the optimized three-dimensional mesh topology to obtain the corrected three-dimensional mesh topology.

[0065] Specifically, in this application, the local processing resolution is used to improve the point data retention density in the grid region to be reconstructed. When configuring the local processing resolution, the point cloud processing resolution, the original point density of the local point cloud, and the minimum boundary scale to be recovered in the grid region to be reconstructed are read. The spatial scale that is smaller than the point cloud processing resolution but not smaller than the original point spacing determined based on the original point density of the local point cloud is determined as the local processing resolution. As one implementation, the local processing resolution can be configured to half of the point cloud processing resolution (for example, when the point cloud processing resolution is 0.02 meters, the local processing resolution is 0.01 meters) to retain more point data near the edges of door and window openings, wall and column intersections, and local occlusion edges. The three-dimensional space where the local point cloud is located is divided into local voxel units according to the local processing resolution. Then, locally regularized point data is generated based on the spatial position of the point data in each local voxel unit. Multiple locally regularized point data form the locally regularized point cloud. The locally regularized point cloud uses the same unified 3D coordinate space as the regularized 3D point cloud. Therefore, the subsequently generated local 3D mesh topology can correspond positionally with the optimized 3D mesh topology. Local point normal representations are generated based on the spatial positions of the locally regularized point data in the locally regularized point cloud. Then, local surface continuity representations are generated based on the spatial distance between adjacent locally regularized point data and the changes in the local point normal representations. Subsequently, local surface mesh fragments are generated based on the local surface continuity representations, and local fragment adjacency relationships are generated based on the boundary continuity between adjacent local surface mesh fragments. Local fragment nodes are formed using the local surface mesh fragments, and local node adjacency records are formed using the local fragment adjacency relationships, thereby constructing the local 3D mesh topology. The local point normal representation, the local surface continuity representation, the local surface mesh fragment, the local fragment adjacency relationship, the local fragment node, and the local node adjacency record are all registered to the local 3D mesh topology, so that the local 3D mesh topology participates in the generation of the local geometric shape representation, the local object semantic representation, and the local topology connection representation.

[0066] When generating the local geometric shape representation, the local object semantic representation, and the local topological connection representation based on the local regularized point cloud and the local 3D mesh topology, the same technical processing relationship is used as that for the geometric shape representation, the object semantic representation, and the topological connection representation. Specifically, local surface primitive candidate descriptions are generated based on the local regularized point data in the local surface mesh fragments, and the spatial fit test is performed on the local surface primitive candidate descriptions and the local regularized point data to determine the local surface primitive descriptions. Then, the local geometric shape representation is formed based on the local surface primitive descriptions. Local fragment semantic features are generated based on the local point normal representation, the local surface continuity representation, and the boundary morphology of the local surface mesh fragments. Then, the local object semantic representation is generated based on the local fragment semantic features and the object semantic category index. Local candidate topological connections are generated based on the local node adjacency records. Then, the local topological connection representation is generated by combining the local candidate topological connections, the local geometric shape representation, and the topological connection rules. The correspondence between the local geometric shape representation and the local object semantic representation is verified using the geometric semantic correspondence rules to generate a local geometric semantic consistency record. The correspondence between the local object semantic representation and the local topological connection representation is verified using the topological connection rules to generate a local semantic topological consistency record. The local surface boundary formed by the local geometric shape representation is used to verify the local topological connection form in the local topological connection representation to generate a local geometric topological consistency record. The local geometric semantic consistency record, the local semantic topological consistency record, and the local geometric topological consistency record are combined to form the local consistency result. The same geometric semantic correspondence rules are used to verify the original geometric shape representation and the object semantic representation in the mesh region to be reconstructed to generate an original region geometric semantic consistency record. The same topological connection rules are used to verify the original object semantic representation and the topological connection representation in the mesh region to be reconstructed to generate an original region semantic topological consistency record. The surface boundary formed by the original geometric shape representation in the mesh region to be reconstructed is used to verify the topological connection form in the original topological connection representation to generate an original region geometric topological consistency record. The original region geometric semantic consistency record, the original region semantic topology consistency record, and the original region geometric topology consistency record are combined to form the original region consistency result.

[0067] To enable comparison between the local consistency results and the original region consistency results, the same consistency degree calculation method is used for both. For the local consistency results, the local geometric semantic satisfaction ratio is determined based on the proportion of local geometric semantic consistency records in the satisfied state out of all local geometric semantic consistency records; the local semantic topology satisfaction ratio is determined based on the proportion of local semantic topology consistency records in the satisfied state out of all local semantic topology consistency records; and the local geometric topology satisfaction ratio is determined based on the proportion of local geometric topology consistency records in the satisfied state out of all local geometric topology consistency records. The local geometric semantic satisfaction ratio, the local semantic topology satisfaction ratio, and the local geometric topology satisfaction ratio are then aggregated according to pre-configured comparison weights to form the consistency degree represented by the local consistency results. Regarding the original region consistency results, the original region geometric semantic satisfaction ratio is determined based on the proportion of original region geometric semantic consistency records in the satisfied state to all original region geometric semantic consistency records. The original region semantic topology satisfaction ratio is determined based on the proportion of original region semantic topology consistency records in the satisfied state to all original region semantic topology consistency records. The original region geometric topology satisfaction ratio, the original region semantic topology satisfaction ratio, and the original region geometric topology satisfaction ratio are then aggregated using the same comparison weight to form the consistency degree represented by the original region consistency results. The comparison weights can utilize the constraint weights in the collaborative reconstruction constraints, ensuring that the comparison direction between the local consistency results and the original region consistency results remains consistent with the evaluation direction of the aforementioned cyclic correction.

[0068] When the consistency level of the local consistency result is higher than that of the original region consistency result, the boundary correspondence between the local 3D mesh topology and the optimized 3D mesh topology is first determined based on the unified 3D coordinate space. Then, based on the boundary correspondence, a new node adjacency record is established between the local fragment node and the fragment node located outside the mesh region to be reconstructed in the optimized 3D mesh topology. Subsequently, the 3D mesh topology corresponding to the mesh region to be reconstructed in the optimized 3D mesh topology is replaced with the local 3D mesh topology, and the corresponding geometric shape representation, object semantic representation, and topological connection representation are replaced with the local geometric shape representation, local object semantic representation, and local topological connection representation to obtain the corrected 3D mesh topology. After the replacement is completed, the fragment node mapping record is updated so that the newly generated local fragment node can return the point data in the local point cloud; at the same time, a replacement correspondence record is generated, and the correspondence between the replaced original fragment node and the local fragment node is recorded in the replacement correspondence record. The replacement correspondence record is used to compare the changes in the geometric shape representation, object semantic representation, and topological connection representation before and after local reconstruction. When the consistency level of the local consistency result is not higher than that of the original region consistency result, the original three-dimensional mesh topology, geometric shape representation, object semantic representation, and topological connection representation in the optimized three-dimensional mesh topology are retained, and the local consistency result and the original region consistency result are written into the representation evolution record for subsequent verification of the surface closure rule or the object embedding rule to continue locating the corresponding region.

[0069] Optionally, the step of generating a three-dimensional surface based on the geometric shape representation in the corrected three-dimensional mesh topology, and organizing the three-dimensional surface according to the object semantic representation and the topological connection representation to obtain a parametric three-dimensional mesh model includes: extracting surface shape parameters from the geometric shape representation in the corrected three-dimensional mesh topology, and generating the three-dimensional surface based on the surface shape parameters; determining the three-dimensional object type of each three-dimensional surface based on the object semantic representation in the corrected three-dimensional mesh topology, and combining the corresponding three-dimensional surfaces according to the same three-dimensional object type to generate three-dimensional object units; generating an object source record based on the position of the three-dimensional surface used to generate each three-dimensional object unit in the corrected three-dimensional mesh topology, so as to associate the three-dimensional object unit with the corrected three-dimensional mesh topology using the object source record; generating object topological connection relationships between the three-dimensional object units based on the topological connection representation in the corrected three-dimensional mesh topology, and connecting the corresponding three-dimensional object units using the object topological connection relationships to obtain the parametric three-dimensional mesh model.

[0070] Specifically, in this application, the surface shape parameters are extracted from the surface primitive description in the geometric shape representation. For the planar type surface primitive description, the surface shape parameters include the plane normal direction, the plane's position relative to the unified three-dimensional coordinate space, and the surface boundary; for the cylindrical type surface primitive description, the surface shape parameters include the axial direction, the position through which the axial direction passes, the radius, and the surface boundary; for the conical type surface primitive description, the surface shape parameters include the axial direction, vertex position, opening degree, and the surface boundary. A corresponding parametric surface is generated in the unified three-dimensional coordinate space based on the surface shape parameters, and the effective region of the parametric surface is truncated according to the surface boundary to obtain the three-dimensional surface. The parametric surface is used to form the three-dimensional surface within the range defined by the surface boundary, and the three-dimensional surface continues to participate in the generation of the three-dimensional object unit. Each three-dimensional surface continues to be associated with the geometric shape representation that generated the surface shape parameters and the corresponding fragment node. Based on the object semantic representation of the corresponding fragment node, the candidate object semantic category that ranks first and satisfies the constraint output condition is read, and the 3D object type corresponding to the candidate object semantic category is determined as the 3D object type of the 3D surface. When the candidate object semantic category that ranks first does not satisfy the constraint output condition, multiple candidate object semantic categories are retained, and the combination of the corresponding 3D surface into the 3D object unit is temporarily suspended until the subsequent verification of the surface closure rule, the verification of the object embedding rule, or the collaborative reconstruction constraint makes the object semantic representation satisfy the constraint output condition.

[0071] When combining the corresponding 3D surfaces according to the same 3D object type, it is also determined whether the 3D surfaces can form the same 3D object unit based on the node adjacency record and topological connection representation in the corrected 3D mesh topology. 3D surfaces with the same 3D object type, whose corresponding fragment nodes are adjacent to each other, and whose corresponding topological connection forms represent surface continuity or internal connections within the same 3D object, are combined into the same 3D object unit. 3D surfaces with the same 3D object type but without continuous object topological connections in the corrected 3D mesh topology form different 3D object units, reducing the possibility of combining spatially separated walls, columns, or slabs into the same 3D object unit. In existing building scanning scenarios, the 3D object unit can correspond to walls, columns, beams, slabs, doors, windows, or staircases. As one implementation method, according to a pre-configured Building Information Exchange (BIE) format, the 3D object units of the wall class can be mapped to wall entities in the BIE format, the 3D object units of the column class can be mapped to column entities in the BIE format, the 3D object units of the beam class can be mapped to beam entities in the BIE format, and the 3D object units of the door or window class can be mapped to opening objects and filling objects in the BIE format. The wall entities, column entities, beam entities, opening objects, and filling objects are used to organize the parametric 3D mesh model according to the BIE format without changing the source relationship between the 3D object units and the corrected 3D mesh topology.

[0072] When generating the object source record, all the three-dimensional surfaces constituting each three-dimensional object unit are extracted, and then the fragment nodes, surface mesh fragments, and point data indexes corresponding to each three-dimensional surface are read. The correspondence between the three-dimensional object unit, the three-dimensional surface, the fragment node, the surface mesh fragment, and the point data index is written into the object source record. The object source record can be used to reverse locate from the three-dimensional object unit to the fragment node in the corrected three-dimensional mesh topology, and can be further located to the point data in the three-dimensional scan point cloud through the fragment node mapping record. When generating the object topology connection relationship based on the topology connection characterization in the corrected three-dimensional mesh topology, the candidate topology connections and topology connection forms between the fragment nodes are first converted into the corresponding connection relationships between the three-dimensional object units. When multiple candidate topology connections connect the same pair of three-dimensional object units, the multiple candidate topology connections are merged into one object topology connection relationship, and the corresponding topology connection form and the boundary continuation range are recorded in the object topology connection relationship. When the topological connection is a nested relationship, the object topological connection relationship also records the 3D object unit as the embedded object and the 3D object unit as the receiving object. The corresponding 3D object units are connected using the object topological connection relationship, ensuring that each 3D object unit maintains a spatial organization consistent with the corrected 3D mesh topology in the parametric 3D mesh model. The parametric 3D mesh model includes at least the 3D surface, the 3D object unit, the object source record, and the object topological connection relationship. The object source record is subsequently used to return the corrected 3D mesh topology based on the surface closure verification result or the object nesting verification result. The object topological connection relationship is subsequently used to perform the verification corresponding to the surface closure rule and the verification corresponding to the object nesting rule.

[0073] Optionally, after obtaining the parameterized 3D mesh model, the method further includes: acquiring pre-configured surface closure rules and object embedding rules; tracing the surface connection order between adjacent 3D object units according to the object topology connection relationship in the parameterized 3D mesh model, generating surface boundary continuity records, verifying the surface boundary continuity records using the surface closure rules, and obtaining a surface closure verification result; determining the 3D object units with embedding relationships according to the 3D object type and the object topology connection relationship, generating object embedding relationships based on the 3D object units with embedding relationships, verifying the object embedding relationships using the object embedding rules, and obtaining an object embedding verification result; when the surface closure verification result does not satisfy the surface closure rule or the object embedding verification result does not satisfy the object embedding rule, determining the 3D object unit corresponding to the unsatisfied state, determining the position of the corresponding 3D object unit in the corrected 3D mesh topology according to the object source record, and determining the abnormal mesh region according to the position.

[0074] Specifically, in this application, the surface closure rule is used to check whether the enclosing boundary formed by multiple three-dimensional object units is continuous, and the object embedding rule is used to check whether two three-dimensional object units with the embedding relationship are compatible in geometric position and object topological connection relationship. When configuring the surface closure rule, the types of three-dimensional objects that can form the enclosing boundary, the allowed object topological connection relationships, the maximum boundary spacing between adjacent three-dimensional surfaces, and the maximum closure spacing between the first and last three-dimensional surfaces are recorded. The maximum boundary spacing is used to check whether adjacent three-dimensional surfaces are continuous, and the maximum closure spacing is used to check whether the end of the enclosing boundary returns to the starting position. When configuring the object embedding rule, the combinations of three-dimensional object types that are allowed to form the object embedding relationship are recorded, as well as the spatial inclusion conditions and boundary overlap conditions that the three-dimensional surface of the embedded object must satisfy relative to the three-dimensional surface of the receiving object. When tracing the surface connection order between adjacent 3D object units based on the object topology connection relationship, starting from a 3D object unit that participates in forming the enclosing boundary, the next 3D object unit connected to that 3D object unit is read, and the subsequent 3D object units are read according to the object topology connection relationship until the starting 3D object unit is returned or no subsequent 3D object unit exists. The 3D object units traversed during the tracing process, the corresponding 3D surfaces, the surface connection order between adjacent 3D surfaces, and the boundary spacing are written into the surface boundary continuation record. When using the surface closure rule to check the surface boundary continuation record, each item is checked to see if the boundary ends of adjacent 3D surfaces are continuating within the maximum boundary spacing, and it is also checked whether the end of the surface boundary continuation record returns to the starting 3D surface. When the boundary ends of all adjacent 3D surfaces satisfy the maximum boundary spacing, and the distance between the end of the continuous record of the surface boundary and the starting 3D surface satisfies the maximum closure spacing, the satisfied state is written into the surface closure verification result; when there are broken boundaries, duplicate connections, or the inability to return to the starting 3D surface, the unsatisfied state is written into the surface closure verification result, and the 3D object unit and adjacent 3D surfaces that caused the unsatisfied state are recorded. As one implementation, the maximum closure spacing can be configured according to the scanning resolution and the allowable deviation of the building boundary (e.g., 0.05 meters).

[0075] When determining the 3D object units with the embedding relationship based on the 3D object type and the object topology connection relationship, the embedding relationship in the object topology connection relationship is first filtered, and then the 3D object type of the 3D object units at both ends of the embedding relationship is read. 3D object units that satisfy the combination of 3D object types in the object embedding rules are registered as the object embedding relationship. Taking the 3D object units of the door class and the 3D object units of the wall class as examples, the 3D object units of the door class constitute the embedded object, and the 3D object units of the wall class constitute the receiving object; taking the 3D object units of the window class and the 3D object units of the wall class as examples, the 3D object units of the window class constitute the embedded object, and the 3D object units of the wall class constitute the receiving object. When verifying the object embedding relationship using the object embedding rules, the receiving boundary range is first generated based on the 3D surface of the receiving object, and then the 3D surface of the embedded object is projected onto the surface coordinates corresponding to the receiving object to generate the projection boundary. Subsequently, it is determined whether the projected boundary is within the range of the receiving boundary, and whether there is an object topological connection relationship between the embedded object and the receiving object that records the embedding relationship. When the projected boundary is within the range of the receiving boundary, and the object topological connection relationship records the embedding relationship, the satisfied state is written into the object embedding verification result; when the projected boundary exceeds the range of the receiving boundary, there is no embedding relationship between the embedded object and the receiving object, or the same embedded object is simultaneously associated with multiple mutually separate receiving objects, the unsatisfied state is written into the object embedding verification result. When the surface closure verification result or the object embedding verification result is in the unsatisfied state, the three-dimensional object unit that caused the unsatisfied state is read from the surface closure verification result or the object embedding verification result in the unsatisfied state. Subsequently, the three-dimensional surface constituting the three-dimensional object unit is determined according to the object source record, and the fragment node and the surface mesh fragment corresponding to the three-dimensional surface are further determined. The spatial range occupied by the corresponding fragment node in the corrected 3D mesh topology, and the spatial range occupied by adjacent fragment nodes related to the unsatisfied state, are merged to determine the abnormal mesh region. The abnormal mesh region is obtained by reverse positioning of the unsatisfied state in the surface closure verification result or the object embedding verification result, and therefore can continue to participate in the update of the collaborative reconstruction constraints.

[0076] Optionally, after determining the abnormal mesh region, the method further includes: determining abnormal representations to be corrected from the geometric representation, the object semantic representation, and the topological connection representation based on the surface closure verification result or the object embedding verification result corresponding to the abnormal mesh region; determining abnormal support representations for correcting the abnormal representations to be corrected from the geometric representation, the object semantic representation, and the topological connection representation; updating the cooperative reconstruction constraints based on the abnormal representations to be corrected and the abnormal support representations; using the updated cooperative reconstruction constraints to correct the geometric representation, the object semantic representation, and the topological connection representation in the abnormal mesh region to update the corrected 3D mesh topology; and regenerating the parametric 3D mesh model based on the updated corrected 3D mesh topology to obtain... The parameterized 3D mesh model is updated, and the checks corresponding to the surface closure rule and the object embedding rule are re-executed on the updated parameterized 3D mesh model to obtain the updated surface closure check result and the updated object embedding check result. When the updated surface closure check result does not meet the surface closure rule or the updated object embedding check result does not meet the object embedding rule, the abnormal mesh region is re-determined based on the updated surface closure check result or the updated object embedding check result, and the collaborative reconstruction constraints are updated again based on the re-determined abnormal mesh region. When the updated surface closure check result meets the surface closure rule and the updated object embedding check result meets the object embedding rule, the updated parameterized 3D mesh model is determined as the 3D mesh topology reconstruction result.

[0077] Specifically, in this application, when determining the abnormal characterization to be corrected based on the surface closure verification result, the boundary spacing of the three-dimensional object unit causing the unsatisfied state, the adjacent three-dimensional surfaces, and the corresponding object topological connection relationship are first read. When the surface boundary of the adjacent three-dimensional surfaces has a high geometric confidence level, but the object topological connection relationship fails to connect the adjacent three-dimensional object units, the topological connection characterization is determined as the abnormal characterization to be corrected, and the geometric shape characterization and the object semantic characterization are determined as the abnormal support characterization. When the object topological connection relationship has recorded the required connection relationship, but the boundary spacing of the adjacent three-dimensional surfaces exceeds the maximum boundary spacing in the surface closure rule, the corresponding geometric uncertainty record and the topological uncertainty record are compared to determine the geometric shape characterization or the topological connection characterization as the abnormal characterization to be corrected. When determining the abnormal characterization to be corrected based on the object embedding verification result, the embedded object, the receiving object, the corresponding object topological connection relationship, and the position of the projected boundary relative to the receiving boundary range are read. When the 3D object type of the embedded object and the 3D object type of the receiving object do not satisfy the object embedding rules, the corresponding object semantic representation is determined as the abnormal representation to be corrected, and the geometric shape representation and the topological connection representation are determined as the abnormal support representation. When the 3D object type satisfies the object embedding rules, but the projection boundary exceeds the receiving boundary range, the geometric shape representation corresponding to the embedded object or the receiving object is determined as the abnormal representation to be corrected based on the geometrically reliable record. When the projection boundary is within the receiving boundary range, but the embedding relationship between the embedded object and the receiving object is lacking, the topological connection representation is determined as the abnormal representation to be corrected.

[0078] When updating the collaborative reconstruction constraints based on the abnormal characterization to be corrected and the abnormal support characterization, the unsatisfied states in the surface closure verification result or the object embedding verification result are associated with the corresponding fragment nodes and the candidate topological connections, and a correction pointing record corresponding to the abnormal characterization to be corrected is added. For the abnormal mesh region caused by the break of the surface boundary, the constraint weight of the cross-consistency constraint component corresponding to the geometric topological consistency record is increased, and the constraint weight of the topological consistency constraint component corresponding to the topological consistency record is also increased; for the abnormal mesh region caused by the abnormal object embedding relationship, the constraint weight of the cross-consistency constraint component corresponding to the semantic topological consistency record and the geometric topological consistency record is increased. As one implementation, the constraint weight corresponding to the unsatisfied state can be increased (e.g., increased by 20%) based on the original constraint weights, while keeping the other constraint weights unchanged. When correcting the geometric characterization, object semantic characterization, and topological connection characterization in the abnormal mesh region using the updated collaborative reconstruction constraints, the cyclic correction is performed only on the abnormal mesh region and the fragment nodes directly adjacent to the abnormal mesh region. The abnormal representation to be corrected and the abnormal support representation are determined according to the updated correction pointing record. Then, the abnormal representation to be corrected is corrected according to the target consistency record corresponding to the abnormal representation to be corrected. After each correction, the geometric semantic consistency record, the semantic topology consistency record, and the geometric topology consistency record are regenerated until the periodic consistency result corresponding to the abnormal mesh region satisfies the cooperative convergence condition. Subsequently, the updated geometric shape representation, the object semantic representation, and the topological connection representation are written into the corrected 3D mesh topology to update the corrected 3D mesh topology.

[0079] When regenerating the parametric 3D mesh model based on the updated corrected 3D mesh topology, the surface shape parameters are re-extracted from the updated geometric shape representation to generate a new 3D surface; the 3D object type is re-determined based on the updated object semantic representation to recombine the 3D object units; and the object topological connection relationship is regenerated based on the updated topological connection representation. The regenerated 3D surface, 3D object units, object source record, and object topological connection relationship are combined into the updated parametric 3D mesh model. The surface closure rule verification is re-executed on the updated parametric 3D mesh model to obtain the updated surface closure verification result; the object embedding rule verification is re-executed on the updated parametric 3D mesh model to obtain the updated object embedding verification result. When the updated surface closure verification result or the updated object embedding verification result is still in the unsatisfied state, the corresponding 3D object unit is re-read based on the unsatisfied state in the updated surface closure verification result or the updated object embedding verification result, and the abnormal mesh region is re-determined based on the object source record. The redefined anomalous mesh region can be the same as the previously defined anomalous mesh region, or it can be expanded to include fragment nodes adjacent to the previously defined anomalous mesh region that participate in the unmet states. Based on the redefined anomalous mesh region, the anomalous characterization to be corrected and the anomalous support characterization are determined again, and the collaborative reconstruction constraints are updated again. Each time the collaborative reconstruction constraints are updated, the previous round of updated surface closure verification results and updated object embedding verification results are retained, so that it is possible to determine whether the number of unmet states has decreased. When the updated surface closure verification result satisfies the surface closure rule, and the updated object embedding verification result satisfies the object embedding rule, the updated parametric 3D mesh model is determined as the 3D mesh topology reconstruction result.

[0080] In one specific embodiment, a six-story reinforced concrete frame office building in North China is used as the reconstruction target. Each floor has a floor area of ​​approximately 2400 square meters, and the total building area is approximately 14400 square meters. A Leica RTC360 terrestrial laser scanner is used to collect 3D point clouds of the existing building surface. The scanning resolution is set to 6 mm corresponding to a scanning distance of 10 meters. Scanning stations are arranged at intervals of 8 to 10 meters, with a total of 168 scanning stations throughout the building. A Leica Cyclone REGISTER 360 is used to perform multi-station registration of the point data collected from different scanning stations, ensuring that the registration error of the overlapping area of ​​adjacent scanning stations is within 3 mm. The original 3D point cloud of the entire building includes approximately 3.8 billion point data points. Taking the standard office floor with 27 rooms on the third floor as an example, the 3D point cloud of this floor includes approximately 19 million point data points. First, a fixed number of nearest neighbor point data points are found for each point data point. To avoid the nearest neighbor number using the same letter as the subsequent semantic category index, this embodiment records the nearest neighbor number as... Wherein, the subscript "near" indicates nearest neighbor retrieval; in this embodiment, we take... Regarding the first For each data point, calculate its correlation with... The average neighborhood distance between the nearest neighbors ,in, This indicates the sequential index of point data or subsequent fragment nodes. Indicates the first The average neighborhood distance of each data point; then calculate the total... Mean distance and distance standard deviation ,in, The population mean representing the average neighborhood distance. Indicates the average neighborhood distance relative to The degree of dispersion. When satisfying

[0081] At that time, the first The data points were identified as outliers; thus, approximately 230,000 outliers were removed from this floor. Subsequently, voxel downsampling was performed at a point cloud processing resolution of 0.02 meters, preserving the centroid of each voxel, thereby normalizing the cleaned 3D point cloud of this floor to approximately 6.4 million points. The normal representation of each point was estimated based on the KD-tree nearest neighbor search results of 15 nearest neighbors. KD-tree is an abbreviation for k-dimensional tree, a multidimensional spatial tree-like index structure used to quickly retrieve neighboring point data according to the spatial location of the point; then, normal flipping is eliminated through consistency propagation of the normal representations of adjacent points. VCCS voxelized hyper-clustering was used to over-segment the normalized 3D point cloud. VCCS is an abbreviation for Voxel Cloud Connectivity Segmentation; the seed voxel side length was set to 0.15 meters, thereby generating 17,842 hyperpoints and approximately 52,600 hyperpoint adjacency edges. To avoid using the same letter for the super-vertex adjacency edge set and the joint energy function, this embodiment represents the super-vertex adjacency graph as follows:

[0082] in, This represents the super-adjacency graph corresponding to the initial 3D mesh topology described above. This represents the set of nodes consisting of all fragment nodes. This represents the set of edges consisting of all adjacent records of nodes. Indicates the relationship with the first Fragment nodes corresponding to each surface mesh fragment. Indicates the first The fragment node and the first There are node adjacency records between each fragment node. Indicates the relationship with the first The sequential index of another adjacent fragment node of each fragment node. This application uses the following at the graph structure initialization: Summary of the first The node-level geometric parameters of each fragment node are used in subsequent geometric fitting and joint optimization. This represents the same node-level geometric parameters; in this embodiment, the following will uniformly adopt... Indicates the first The surface primitive type and surface parameters in the geometric shape representation of each fragment node. Indicates the first The current semantic label in the object semantic representation of each fragment node. Indicates the first The fragment node and the first The topological edge variables between fragment nodes, and satisfy the following conditions:

[0083] in, This indicates that there is a physical connection between two fragment nodes. This indicates that there is no physical connection between the two fragment nodes. The initial 3D mesh topology is stored using a JSON data structure, where JSON is an abbreviation for JavaScript Object Notation. The code field `node_id` represents the unique node number of the fragment node, `centroid` represents the 3D spatial position of the centroid of the surface mesh fragment, `normal` represents the statistical direction of the point normal within the surface mesh fragment, `point_indices` represents the index of the original point data contained in the surface mesh fragment, `edge_id` represents the unique number of the node adjacency record, and `node_i` and `node_j` represent the first and second edges of the node adjacency record, respectively. The fragment node and the first Each fragment node, with `shared_boundary_length` representing the length of the shared boundary between two surface mesh fragments. The above code fields correlate the 3D scan point cloud, surface mesh fragments, fragment nodes, and node adjacency records, enabling the subsequent return of the corresponding original point data based on the fragment node mapping records.

[0084] Furthermore, for the 17842 surface mesh fragments, the RANSAC (Random Sample Consensus) method was used to attempt to fit three types of surface primitives: planes, cylinders, and cones. RANSAC is an abbreviation for Random Sample Consensus. The interior point distance threshold is denoted as... Among them, Greek letters This represents the distance threshold, with the subscript "geo" indicating that this threshold is used for geometric fitting. In this embodiment, it is set... The maximum number of random sampling iterations is set to 1000. After comparing the three types of candidate surface primitive descriptions for the same surface mesh fragment, the surface primitive type with the highest proportion of interior point data is selected as the first... Initial geometric parameters of each fragment node And calculate the root mean square of the fitting residuals. RMSE is an abbreviation for Root Mean Square Error, which is the root mean square error. This indicates that the root mean square error belongs to the th Each fragment node This corresponds to the overall degree of deviation in the aforementioned geometric deviation record. The geometric parameters... Recorded as dictionary fields, where `basis_type` represents the surface primitive type, with values ​​`plane`, `cylinder`, or `cone` representing a plane, cylinder, or cone surface, respectively; `params` represents the surface parameter sequence; and `rmse` represents... ;confidence indicates the degree of geometric confidence described later. Planar type params are recorded as ,in, , and These represent the components of the plane's normal direction along the three coordinate axes in a unified three-dimensional coordinate space. This represents the directed distance of the plane relative to the origin of the coordinate system; the `params` parameter for the cylindrical surface type records the three components of the axis direction, the three coordinate components of a point on the axis, and the radius, where the radius represents the typical distance from the point data on the cylindrical surface to the axis. The geometric reliability of a fragment node is denoted as . ,in, Indicates the degree of credibility, subscript Indicates the corresponding number Each fragment node, with the superscript "geo" indicating its geometric shape representation, is organized according to...

[0085] Confirmed. Among them, This represents an exponential mapping with the natural constant as the base. This indicates the geometric confidence bandwidth used when converting geometric deviations into geometric confidence levels. The subscript "geo" indicates that this bandwidth is used for geometric shape characterization. In this embodiment, it is set... ; Indicates the first The root mean square error of each fragment node This is used to control the rate at which geometric reliability decreases as geometric deviation increases. It will satisfy... Fragment node written with geometric uncertainty mask ,in, This represents a mask used to record highly uncertain states, with the superscript "geo" indicating geometric shape representation; this floor identifies a total of 1203 fragment nodes representing these geometrically highly uncertain states. For each of the curved mesh fragments, 32-dimensional fragment semantic features are extracted and then input into a graph attention classification structure consisting of two graph attention encoding layers, where GAT is an abbreviation for Graph Attention Network. To avoid using the same symbol for the semantic category index and the aforementioned nearest neighbor count, this embodiment uses... Represents a semantic category index, in This represents a set of eight semantic categories, corresponding to walls, columns, beams, slabs, doors, windows, stairs, and other objects, respectively. The semantic probability distribution of each fragment node is represented as follows:

[0086] Among them, bold Indicates the first The complete semantic probability distribution corresponding to each fragment node. Indicates the first The fragment node belongs to the first The probability of each semantic category, Furthermore, the sum of the probabilities of the eight categories is 1. The semantic credibility of each fragment node is denoted as . The superscript "sem" indicates the semantic representation of the object, and is arranged according to...

[0087] Sure; Indicates the index of all semantic categories Select the class with the highest probability. This will satisfy... Write the semantic uncertainty mask into the fragment node The superscript "sem" indicates the semantic representation of the object. This graph attention classification structure was pre-trained using 42 existing buildings and approximately 310 million labeled points, achieving a mean semantic confidence level of 0.89 in the third layer of data, and identifying 896 fragment nodes with high semantic uncertainty. Therefore, , and These correspond to the surface primitive description, the geometric deviation record, and the geometric reliability record, respectively. , and These correspond to the semantic category of the candidate object, the current semantic tag, and the semantically trusted record, respectively. and This is used to determine the adjustment direction of the geometrically uncertain record, the semantically uncertain record, and the subsequent constraint weights.

[0088] Furthermore, when generating the collaborative reconstruction constraints, the combination of geometric parameters of all fragment nodes is denoted as... The current semantic tag combination of all fragment nodes is denoted as... The combination of topological edge variables of all node adjacency records is denoted as And the joint energy function is denoted as ;in, This indicates the overall degree of inconsistency between the geometric shape representation, the object semantic representation, and the topological connectivity representation under the current values, and the aforementioned edge set. Different fonts are used to represent different concepts. The joint energy function is...

[0089] The four energy components are represented as follows:

[0090]

[0091]

[0092]

[0093] in, The constraint weights represent the geometric deviation constraint components. The constraint weights represent the semantic constraint components of an object. The constraint weights represent the components of the topological consistency constraint. The constraint weights represent the cross-consistency constraint components; the subscripts "geo", "sem", "topo", and "cross" represent geometric shape representation, object semantic representation, topological connection representation, and cross-representation consistency relationship, respectively. In this embodiment, the initial values ​​are 1.0, 0.8, 1.2, and 0.6, respectively. The summation of nodes in each energy component represents the index of all fragment nodes. By summing the edges sequentially, the set of opposite edges can be represented by the summation of the edges. All adjacent node pairs Add them up sequentially. Indicates the first The point data of each fragment node relative to geometric parameters The geometric primitives of the surface are fitted with a residual vector, where each element of the residual vector represents a directed spatial deviation of a point data point from the surface. This represents the sum of the squares of each element in the residual vector fitted by the geometric primitive. Indicates the first Each fragment node in the semantic probability distribution In relation to the current semantic tag The corresponding category probability, Represents the natural logarithm. This represents the semantic classification cost corresponding to the current semantic label. Indicates the first The fragment node and the first The topological consistency potential function among fragment nodes. The input consists of the current semantic labels of the two fragment nodes in sequence. and Topological edge variables and geometric parameters and When the above input does not satisfy the topology connection rules, Returns a larger topology consistency cost. Indicates the first Geometric semantic cross-consistency term of each fragment node, Greek letter Used to represent geometric parameters The corresponding surface primitive type and the current semantic label The cross-consistency cost between the corresponding 3D object types. The joint energy function is cyclically corrected using a block coordinate descent approach. Indicates the correction period index. , and They represent the first The combinations of geometric parameters, semantic labels, and topological edge variables in the input of each correction cycle. , and These represent the initial values ​​used in the first cyclic calibration. This indicates the maximum number of correction cycles; in this embodiment, it is taken as... The update of the geometric representation is expressed as:

[0094] in, Indicates all possible combinations of geometric parameters Select the combination of geometric parameters that minimizes the cost within the parentheses. Indicates the first The combination of updated geometric parameters used in each correction cycle Indicates the first In the correction cycle, the _th The semantic labels of each fragment node. Before updating the object semantic representation, the topological consistency potential function is... Convert to positive topology compatibility The wavy symbol indicates that the quantity is obtained from the topological consistency cost, avoiding directly taking the logarithm for non-positive values. The semantic probability distribution after fusing topological priors is represented as follows:

[0095] in, This represents the updated semantic probability distribution after incorporating topological priors. The apostrophe indicates that this probability distribution is derived from the original semantic probability distribution. Updated to obtain; This represents converting the combined scores of each semantic category into category probabilities that sum to 1; Indicates the relationship with the first Each fragment node has a set of adjacent fragment node indices for its adjacent records. Indicates the first The fragment node belongs to the first The adjacent fragment nodes of each fragment node; Indicates the first Topological edge variables in each correction period and These represent the geometric parameters of the two updated fragment nodes. Then, according to...

[0096] Determine to update semantic tags, where, Indicates updating the semantic probability distribution The Middle The probability of each semantic category, Indicates from the set of categories Selected The largest semantic category index, Indicates the first In the correction cycle, the _th The updated semantic labels of each fragment node. The updated representation of the topological connectivity is expressed as follows:

[0097] in, Represents the combination of all possible topological edge variables. Choose the combination that minimizes the cost of topological consistency. Indicates the first The combination of updated topological edge variables for each correction cycle. The joint energy of the previous correction cycle is denoted as... The joint energy of the current correction period is denoted as Where the subscript "prev" represents the previous correction period and the subscript "cur" represents the current correction period; the convergence threshold is denoted as... In this embodiment, When satisfied

[0098] At that time, the periodic consistency result satisfies the cooperative convergence condition. The final result is... , and These represent the updated geometric parameter combinations, updated semantic label combinations, and updated topological edge variable combinations, respectively, when the aforementioned collaborative convergence condition is met. The subscript "final" indicates the final value after iterative correction. The components of the joint energy in each round are written to iteration_log.csv, where CSV stands for Comma-Separated Values; iter_index represents the correction cycle index. E_geo represents the energy value of the geometric deviation constraint component, E_sem represents the energy value of the object semantic constraint component, E_topo represents the energy value of the topological consistency constraint component, E_cross represents the energy value of the cross consistency constraint component, and E_total represents the joint energy function. The current total energy value. After 19 correction cycles in the third layer, the cooperative convergence condition is met, the mean root mean square error of the geometric fit decreases from 0.024 m to 0.013 m, and the number of fragment nodes in the semantically high uncertainty state decreases from 896 to 211; the energy values ​​of each item in the iteration log and , , The changes together form the above-mentioned record of characterization and evolution.

[0099] After completing the global cyclic correction, the semantic tags of each fragment node in the last 5 correction cycles are counted. Number of changes; when When at least two flips occur, the corresponding fragment node is identified as the oscillating fragment node, and the one-hop neighbor node directly connected to the oscillating fragment node is identified as the associated fragment node. A total of 97 oscillating fragment nodes are identified in the third layer, mainly distributed at the junction of movable partition walls and fixed walls in two large conference rooms. Based on the fragment node mapping records, the original 3D scanned point cloud without voxel downsampling is returned. The local point cloud is re-extracted from the spatial range corresponding to the oscillating fragment node and the associated fragment node, restoring the point data density of the local point cloud to 500 to 1500 points per square meter. Then, geometric primitive fitting, object semantic classification, and geometric-semantic-topological co-correction are re-performed on the local point cloud, with the upper limit of the local correction cycle set to 10. The third layer, corresponding to a local region, converged after six calibration cycles. The local joint energy of 81 fragment nodes was lower than the original joint energy; therefore, the original results were replaced using the local geometric representation, the local object semantic representation, and the local topological connectivity representation. The remaining 16 fragment nodes retained their original results and were appended with confidence labels. Based on the final... , and When generating the parametric 3D mesh model, based on The final semantic tags represent how the three-dimensional object units corresponding to walls, columns, beams, slabs, doors, windows, and staircases are mapped to the IfcWall, IfcColumn, IfcBeam, IfcSlab, IfcDoor, IfcWindow, and IfcStair entities in the IFC 4.3 standard, where IFC is an abbreviation for Industry Foundation Classes, representing the Industrial Foundation Classes Building Information Interchange Standard. The final geometric parameters represent the parametric geometric representations of each 3D object element. Wall-type 3D object elements are generated by extruding along the wall axis using IfcExtrudedAreaSolid, where IfcExtrudedAreaSolid represents the parametric geometric representation of the solid formed according to the cross-sectional profile and extrusion direction. The extrusion thickness is taken from the statistical value of the normal thickness fitted from the planar primitives. The final topological edge variables represent IfcRelConnectsElements and IfcRelVoidsElement, where IfcRelConnectsElements represents the connection relationship between two 3D object units, and IfcRelVoidsElement represents the opening relationship of door / window openings relative to the host wall. The third layer ultimately generates 156 IfcWall instances, 32 IfcColumn instances, 28 IfcBeam instances, 1 IfcSlab instance, 31 IfcDoor instances, and 44 IfcWindow instances. For the 16 3D object units that still require review, PSet_ConfidenceFlag is written into the IFC attribute set, where PSet is an abbreviation for Property Set, and ConfidenceFlag represents the confidence level flag. The value of this attribute indicates the overall confidence level of the corresponding 3D object unit after collaborative correction. The PSet_ConfidenceFlag, together with the fragment nodes, surface mesh fragments, and point data indexes that constitute the 3D object unit, is written into the object source record, enabling the low-confidence 3D object unit in the parameterized 3D mesh model to return the corresponding corrected 3D mesh topology and 3D scan point cloud.

[0100] After generating the IFC model, the surface closure rule was checked on the 3D object units corresponding to the 27 rooms, with the room boundary closure error threshold set to 0.03 meters. The object embedding rule was also checked on the 31 door-type 3D object units and 44 window-type 3D object units, with the allowable deviation of door and window openings relative to the host wall bounding box set to 0.01 meters. The initial check found that the surface closure check results for two rooms were not satisfied, with an unclosed room ratio of 7.4%, higher than the pre-configured 5% re-trigger ratio threshold. Therefore, based on the object source record, the two rooms were located to the corresponding partition wall element nodes, and the topology consistency constraint component weights of the corresponding local areas were adjusted. The error was increased from 1.2 to 1.8, and then a local collaborative correction of up to 15 correction cycles was performed on the abnormal mesh area. This local collaborative correction actually converged after 9 correction cycles, reducing the boundary closure errors of the two rooms to 0.021 meters and 0.018 meters respectively, ensuring that the updated surface closure verification results met the surface closure rules. The updated object embedding verification results of all 31 door-type 3D object units and 44 window-type 3D object units also met the allowable deviation of 0.01 meters. After performing the above process on all 6 floors and 168 rooms, approximately 2860 3D object units were finally formed, of which 107 3D object units retained the PSet_ConfidenceFlag attribute, accounting for 3.7%. The closure rate of the room boundaries after the initial cycle correction was 96.4%, and after the second collaborative correction of the abnormal mesh area, the closure rate reached 100%. The object embedding verification pass rate of door and window openings and the host wall reached 100%. Therefore, the 3D mesh topology reconstruction result was... Corresponding updated geometric shape representation, The corresponding updated object semantic representation and The corresponding updated topology connection representations are jointly formed, and the consistency with the technology chain described in this application is maintained through local resampling, local three-dimensional mesh topology replacement, parametric three-dimensional mesh model generation, and model integrity recalibration.

[0101] like Figure 3The diagram illustrates a 3D mesh topology reconstruction device based on geometric-semantic-topology collaboration, as described in this application. This device is a virtual device configured by computer program instructions, comprising an initial topology construction module, a collaborative correction module, a local topology reconstruction module, and a mesh model generation module. The initial topology construction module acquires a 3D scanned point cloud, constructs an initial 3D mesh topology based on the spatial continuity of the point cloud, and generates geometric shape representations, object semantic representations, and topological connectivity representations based on the point cloud and the initial topology. The collaborative correction module generates collaborative reconstruction constraints based on the consistency relationships between the geometric shape representations, object semantic representations, and topological connectivity representations. It then uses these constraints to iteratively correct the geometric shape representations, object semantic representations, and topological connectivity representations, thereby updating the initial 3D mesh topology and obtaining an optimized 3D mesh topology. The local topology reconstruction module is used to identify the mesh region to be reconstructed based on the consistency of the geometric shape representation, object semantic representation, and topological connection representation in the optimized 3D mesh topology. It extracts the local point cloud corresponding to the mesh region to be reconstructed from the 3D scanned point cloud, and uses the local point cloud to reconstruct the 3D mesh topology of the mesh region to be reconstructed, thereby updating the optimized 3D mesh topology and obtaining a corrected 3D mesh topology. The mesh model generation module is used to generate a 3D surface based on the geometric shape representation in the corrected 3D mesh topology, and organize the 3D surface according to the object semantic representation and the topological connection representation to obtain a parametric 3D mesh model.

[0102] like Figure 4 As shown, an electronic device according to an embodiment of this application includes a processor, a memory, a communication interface, and a data bus. The processor, the memory, and the communication interface transmit data via the data bus. The processor executes the computer program stored in the memory to implement the solution described in the above-described method of this application.

[0103] In one embodiment, the electronic device can be a 3D modeling server, a cloud-based modeling device, a graphics workstation, or a computing device equipped with a 3D mesh topology reconstruction program. The electronic device receives the 3D scanned point cloud formed by a laser scanning device through the communication interface; the processor executes the computer program to generate the parametric 3D mesh model; and then outputs the parametric 3D mesh model through the communication interface.

[0104] like Figure 5The illustration shows a computer-readable storage medium according to an embodiment of this application. The computer-readable storage medium stores a computer program, which, when executed by the processor of an electronic device, causes the electronic device to perform the steps of the method described above. The computer-readable storage medium can be a non-transitory storage medium capable of storing a computer program and being read by an electronic device, including read-only memory, random access memory, solid-state memory, disk storage, optical disk storage, or a combination of the above storage media. The computer-readable storage medium can be disposed inside the electronic device or can be read by the electronic device via a communication connection.

[0105] like Figure 6 The diagram illustrates an automatic modeling system for scanning existing buildings into a Building Information Modeling (BIM) according to an embodiment of this application. The system includes a 3D scanning device, a 3D mesh topology reconstruction device, and a model application terminal. The 3D scanning device is communicatively connected to the 3D mesh topology reconstruction device, and the 3D mesh topology reconstruction device is communicatively connected to the model application terminal. The 3D scanning device scans the surface space of the existing building to generate a 3D scan point cloud, which is then sent to the 3D mesh topology reconstruction device. The 3D scanning device can be a ground laser scanning device or a mobile scanning device capable of generating 3D point data. The 3D scan point cloud generated by the 3D scanning device represents the surface space positions of walls, columns, beam bottoms, floor slabs, door and window openings, and component junction areas. The 3D mesh topology reconstruction device executes the steps of the method described above. The model application terminal receives the parametric 3D mesh model and displays, stores, or converts its format. The model application terminal can be a building information modeling workstation, a 3D model browsing terminal, or a model data management terminal. The model application terminal displays the spatial form of existing buildings according to the 3D surfaces, 3D object types, and topological connections in the parametric 3D mesh model, and can convert the parametric 3D mesh model into a data format that building information modeling editing software can read.

[0106] In one embodiment, the 3D scanning device is deployed in corridors, room entrances, and spatial corners within an existing building to collect point cloud data from multiple scanning stations. The 3D scanning device registers the point cloud data from the multiple scanning stations into a 3D scanned point cloud and sends the 3D scanned point cloud to the 3D mesh topology reconstruction device. The 3D mesh topology reconstruction device uses the 3D scanned point cloud to generate the parametric 3D mesh model, which is then displayed by the model application terminal as 3D objects corresponding to walls, columns, beams, slabs, doors, windows, or staircases, and the spatial connection relationships between these 3D objects are also displayed.

[0107] Figures 2-6 For an exemplary description, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. A three-dimensional mesh topology reconstruction method based on geometry-semantics-topology collaboration, characterized in that, include: A 3D scanned point cloud is acquired, an initial 3D mesh topology is constructed based on the spatial continuity of the 3D scanned point cloud, and geometric shape representation, object semantic representation, and topological connection representation are generated based on the 3D scanned point cloud and the initial 3D mesh topology. Based on the consistency relationship between the geometric shape representation, the object semantic representation, and the topological connection representation, a cooperative reconstruction constraint is generated. The cooperative reconstruction constraint is used to perform cyclic correction on the geometric shape representation, the object semantic representation, and the topological connection representation, thereby updating the initial 3D mesh topology and obtaining an optimized 3D mesh topology. The mesh region to be reconstructed is identified based on the consistency of the geometric shape representation, the object semantic representation, and the topological connection representation in the optimized 3D mesh topology. The local point cloud corresponding to the mesh region to be reconstructed is extracted from the 3D scan point cloud. The 3D mesh topology of the mesh region to be reconstructed is reconstructed using the local point cloud to update the optimized 3D mesh topology and obtain the corrected 3D mesh topology. A three-dimensional surface is generated based on the geometric shape representation in the corrected three-dimensional mesh topology. The three-dimensional surface is then organized according to the object semantic representation and the topological connection representation to obtain a parametric three-dimensional mesh model.

2. The method according to claim 1, characterized in that, The construction of the initial 3D mesh topology based on the spatial continuity of the 3D scanned point cloud includes: The neighborhood distribution test is performed on the point data in the three-dimensional scanned point cloud to generate the neighborhood distribution test result, and outlier point data is determined based on the neighborhood distribution test result; The outlier data is removed from the three-dimensional scanned point cloud to obtain a purified three-dimensional point cloud; Obtain the pre-configured point cloud processing resolution, and perform spatial regularization on the purified 3D point cloud according to the point cloud processing resolution to obtain a regularized 3D point cloud. Extract the spatial positions of adjacent points in the regular 3D point cloud, generate point normal representations based on the point spatial positions, and generate continuous surface representations based on the changes in the point normal representations of adjacent point data. Based on the spatial location of the point, the normal representation of the point, and the continuous representation of the surface, surface mesh fragments are generated. Fragment adjacency relationships are generated according to the boundary continuity between adjacent surface mesh fragments. The initial three-dimensional mesh topology is constructed using the surface mesh fragments and the fragment adjacency relationships.

3. The method according to claim 2, characterized in that, The generation of geometric shape representation, object semantic representation, and topological connection representation based on the 3D scanned point cloud and the initial 3D mesh topology includes: Fragment nodes are generated based on each surface mesh fragment, and a fragment node mapping record representing the correspondence between the surface mesh fragments and the fragment nodes is generated. The fragment node mapping record is used to associate the fragment adjacency relationship with the corresponding fragment node, thereby generating a node adjacency record; Register the point data, point spatial position, point normal representation, and surface continuity representation in the surface mesh fragment to the corresponding fragment node to generate a basic fragment description; In each fragment node, a geometric shape representation position, an object semantic representation position, and a topological connection representation position are configured. The fragment basic description is associated with the geometric shape representation position and the object semantic representation position, respectively. The fragment basic description and the node adjacency record are associated with the topological connection representation position, so as to generate the geometric shape representation, the object semantic representation, and the topological connection representation, respectively.

4. The method according to claim 3, characterized in that, Generating the geometric representation includes: For each of the aforementioned surface mesh fragments, a candidate description of a surface primitive corresponding to a pre-configured surface primitive type is generated using the point data and spatial position of the points in the surface mesh fragments. A spatial fit test is performed on each candidate description of the surface primitive and the point data to generate a spatial fit record. The surface primitive description is determined based on the spatial fit record, and a geometric deviation record is generated based on the spatial deviation between the surface primitive description and the point data. A geometrically reliable record is generated based on the geometric deviation record. The surface primitive description, the geometric deviation record, and the geometrically reliable record are registered to the corresponding geometric shape representation position to obtain the geometric shape representation. The surface primitives are used to describe the generation of the constraints involved in the collaborative reconstruction, and the geometric deviation record and the geometric reliability record are used to determine the degree of uncertainty of the geometric shape representation.

5. The method according to claim 4, characterized in that, Generating the semantic representation of the object includes: The point normal representation, the surface continuity representation, and the boundary morphology of the surface mesh fragments are used to generate fragment semantic features; Obtain the object semantic category index pre-configured according to the 3D object type, and filter candidate object semantic categories from the object semantic category index based on the fragment semantic features; By combining the geometric shape representation, the basic fragment description of the adjacent fragment nodes, and the fragment semantic features, a semantically reliable record of the semantic category of each candidate object is generated. The candidate object semantic category and the semantic trust record are registered to the corresponding object semantic representation position to obtain the object semantic representation; The semantic categories of the candidate objects are used to participate in the generation of the collaborative reconstruction constraints, and the degree of uncertainty of the semantic representation of the objects is determined by the semantic trust record.

6. The method according to claim 5, characterized in that, Generating the topology connectivity representation includes: Candidate topological connections between adjacent fragment nodes are generated based on the node adjacency records, and a topological connection form is generated by combining the candidate topological connections and the geometric shape representation. Obtain the topology connection rules pre-configured according to the allowed connection methods between 3D objects, use the object semantic representation and the topology connection rules to verify the topology connection form, and generate a topology consistency record; The candidate topology connection, the topology connection form, and the topology consistency record are registered to the corresponding topology connection representation position to obtain the topology connection representation; The candidate topological connections and the forms of the topological connections are used to participate in the generation of the collaborative reconstruction constraints, and the uncertainty of the topological connection representation is determined by the topological consistency record.

7. The method according to claim 6, characterized in that, The step of generating collaborative reconstruction constraints based on the consistent relationship between the geometric shape representation, the object semantic representation, and the topological connection representation includes: A geometric semantic consistency record is generated based on the surface primitive description, the candidate object semantic category, and the pre-configured geometric semantic correspondence rules; A semantic topology consistency record is generated based on the candidate object semantic category, the topology connection form, and the topology connection rule; A surface boundary is generated based on the surface primitive description. The candidate topological connection is used to check whether the surface boundaries of adjacent surface mesh fragments satisfy the topological connection form, and a geometric topological consistency record is generated. Based on the geometric semantic consistency record, the semantic topology consistency record, and the unsatisfied states in the geometric topology consistency record, a representation to be corrected is determined from the geometric shape representation, the object semantic representation, and the topological connection representation, and a supporting representation for correcting the representation to be corrected is determined from the geometric shape representation, the object semantic representation, and the topological connection representation; A correction pointing record is generated based on the characterization to be corrected and the supporting characterization. The geometric semantic consistency record, the semantic topology consistency record, the geometric topology consistency record, and the correction pointing record are combined into the collaborative reconstruction constraint.

8. The method according to claim 7, characterized in that, Before performing cyclic correction on the geometric shape representation, the object semantic representation, and the topological connectivity representation using the collaborative reconstruction constraints, the method further includes: A geometrically uncertain record is generated based on the geometric deviation record and the geometrically reliable record; a semantically uncertain record is generated based on the semantically reliable record; and a topologically uncertain record is generated based on the topological consistency record. Based on the degree of uncertainty represented by the geometrically uncertain record, the semantically uncertain record, and the topologically uncertain record, the geometric shape representation, the object semantic representation, and the topological connection representation are arranged to generate an uncertain sorting record; The character to be corrected and the supporting character are determined using the correction pointing record. Based on the arrangement positions of the character to be corrected and the supporting character in the uncertain sorting record, the arrangement positions of the character to be corrected in the uncertain sorting record are adjusted to obtain the corrected uncertain sorting record. The geometric shape representation, the object semantic representation, and the topological connection representation are arranged according to the correction uncertainty sorting record to obtain a correction order, so that the correction order controls the order in which the geometric shape representation, the object semantic representation, and the topological connection representation are cyclically corrected.

9. The method according to claim 8, characterized in that, The step of cyclically correcting the geometric shape representation, the object semantic representation, and the topological connectivity representation using the collaborative reconstruction constraints includes: The current character to be corrected is determined according to the correction order, the current support character is determined according to the correction direction record, and the target consistency record corresponding to the current character to be corrected is determined from the collaborative reconstruction constraints. The current supporting representation is used to correct the current representation to be corrected according to the target consistency record to obtain the current corrected representation. The current corrected representation is determined as the update result of the corresponding representation according to the correction order, and the above process is repeated to form a correction cycle to obtain the updated geometric shape representation, the updated object semantic representation, and the updated topological connection representation. The updated geometric shape representation, the updated object semantic representation, and the updated topological connection representation are used to regenerate the geometric semantic consistency record, the semantic topological consistency record, and the geometric topological consistency record to obtain a periodic consistency result. A representation evolution record is generated based on the changes of the updated geometric shape representation, the updated object semantic representation, and the updated topological connection representation relative to the correction periodic input. When the periodic consistency result does not meet the pre-configured collaborative convergence condition, the next correction cycle begins; when the periodic consistency result meets the collaborative convergence condition, the initial 3D mesh topology is updated with the updated geometry representation, the updated object semantic representation, and the updated topology connection representation to obtain the optimized 3D mesh topology.

10. The method according to claim 9, characterized in that, The method of cyclically correcting the geometric shape representation, the object semantic representation, and the topological connectivity representation using the collaborative reconstruction constraints further includes: A geometric shape representation admission record is generated based on the geometric uncertainty record; an object semantic representation admission record is generated based on the semantic uncertainty record; and a topological connection representation admission record is generated based on the topological uncertainty record. The target representation is determined according to the correction order. The target representation admission record corresponding to the target representation is determined from the geometric shape representation admission record, the object semantic representation admission record and the topological connection representation admission record. The target support representation is determined according to the correction pointing record. When the target representation admission record does not meet the pre-configured constraint output conditions, the target representation is corrected using the target support representation to obtain the corrected representation, and the corrected representation is prohibited from being identified as the support representation. Based on the corrected representation, the geometrically uncertain record, the semantically uncertain record, or the topologically uncertain record corresponding to the target representation is regenerated to update the target representation admission record; When the updated target representation admission record satisfies the constraint output condition, the corrected representation is determined as the supporting representation, so that the corrected representation participates in the generation of the subsequent collaborative reconstruction constraints.

11. The method according to claim 9, characterized in that, The step of identifying the mesh region to be reconstructed based on the consistency of the geometric shape representation, the object semantic representation, and the topological connectivity representation in the optimized 3D mesh topology includes: Extract the semantic category changes of the candidate object from the same surface mesh fragment in the continuous correction period from the representation evolution record to generate a semantic change trajectory, and extract the changes of the periodic consistency result in the continuous correction period to generate a consistency change trajectory. Obtain pre-configured oscillation identification conditions, and determine the oscillation fragment node based on the semantic change trajectory and the consistency change trajectory; The associated fragment nodes adjacent to the oscillating fragment node are determined using the node adjacency record, and the surface mesh fragments corresponding to the oscillating fragment node and the associated fragment node are determined based on the fragment node mapping record. The mesh regions occupied by the oscillating fragment nodes and the surface mesh fragments corresponding to the associated fragment nodes in the optimized 3D mesh topology are determined as the mesh regions to be reconstructed, so that the local point cloud corresponding to the mesh regions to be reconstructed is used to reconstruct the 3D mesh topology of the mesh regions to be reconstructed.

12. The method according to claim 11, characterized in that, The step of reconstructing the 3D mesh topology of the region to be reconstructed using the local point cloud to update the optimized 3D mesh topology and obtain the corrected 3D mesh topology includes: A local processing resolution smaller than the point cloud processing resolution is obtained, and the local point cloud is spatially regularized according to the local processing resolution to obtain a locally regularized point cloud. A local three-dimensional mesh topology is constructed based on the spatial continuity of the local regular point cloud, and local geometric shape representation, local object semantic representation, and local topological connection representation are generated based on the local regular point cloud and the local three-dimensional mesh topology. The local geometric shape representation, the local object semantic representation, and the local topological connection representation are examined according to the geometric semantic correspondence rules and the topological connection rules to obtain a local consistency result. The corresponding geometric shape representation, the object semantic representation, and the topological connection representation in the mesh region to be reconstructed are examined according to the geometric semantic correspondence rules and the topological connection rules to obtain a region consistency result. When the consistency of the local consistency result representation is higher than the consistency of the original region consistency result representation, the local three-dimensional mesh topology, the geometric representation, the object semantic representation, and the topological connection representation in the optimized three-dimensional mesh topology corresponding to the mesh region to be reconstructed are replaced by the local three-dimensional mesh topology and the local geometric shape representation, the local object semantic representation, and the local topological connection representation therein, to obtain the corrected three-dimensional mesh topology.

13. The method according to claim 1, characterized in that, The process of generating a three-dimensional surface based on the geometric shape representation in the corrected three-dimensional mesh topology, and organizing the three-dimensional surface according to the object semantic representation and the topological connectivity representation to obtain a parametric three-dimensional mesh model includes: Surface shape parameters are extracted from the geometric shape representation in the corrected 3D mesh topology, and the 3D surface is generated based on the surface shape parameters; The three-dimensional object type of each three-dimensional surface is determined based on the semantic representation of the object in the corrected three-dimensional mesh topology. The three-dimensional surfaces corresponding to the same three-dimensional object type are combined to generate three-dimensional object units. Based on the position of the three-dimensional surface used to generate each three-dimensional object unit in the corrected three-dimensional mesh topology, an object source record is generated to associate the three-dimensional object unit with the corrected three-dimensional mesh topology using the object source record; Based on the topological connection characterization in the corrected 3D mesh topology, the object topological connection relationship between the 3D object units is generated, and the corresponding 3D object units are connected using the object topological connection relationship to obtain the parameterized 3D mesh model.

14. The method according to claim 13, characterized in that, After obtaining the parameterized 3D mesh model, the process further includes: Obtain pre-configured surface closure rules and object embedding rules; Based on the object topological connection relationship in the parameterized 3D mesh model, the surface connection order between adjacent 3D object units is traced to generate a surface boundary continuity record. The surface closure rule is used to check the surface boundary continuity record to obtain the surface closure check result. Based on the three-dimensional object type and the object topology connection relationship, the three-dimensional object units with embedding relationship are determined, the object embedding relationship is generated based on the three-dimensional object units with embedding relationship, and the object embedding relationship is verified using the object embedding rules to obtain the object embedding verification result; When the surface closure verification result does not meet the surface closure rule or the object embedding verification result does not meet the object embedding rule, the three-dimensional object unit corresponding to the unmet state is determined, the position of the corresponding three-dimensional object unit in the corrected three-dimensional mesh topology is determined according to the object source record, and the abnormal mesh region is determined according to the position.

15. The method according to claim 14, characterized in that, After determining the abnormal grid region, the process also includes: Based on the surface closure verification result or the object embedding verification result corresponding to the abnormal mesh region, an abnormal representation to be corrected is determined from the geometric representation, the object semantic representation and the topological connection representation, and an abnormal support representation for correcting the abnormal representation to be corrected is determined from the geometric representation, the object semantic representation and the topological connection representation. The cooperative reconstruction constraints are updated based on the anomaly to be corrected characterization and the anomaly support characterization. The updated cooperative reconstruction constraints are then used to correct the geometric shape characterization, the object semantic characterization, and the topological connectivity characterization in the anomaly mesh region, so as to update the corrected 3D mesh topology. Based on the updated corrected 3D mesh topology, a parametric 3D mesh model is regenerated to obtain an updated parametric 3D mesh model. The surface closure rule check and the object embedding rule check are then re-executed on the updated parametric 3D mesh model to obtain the updated surface closure check result and the updated object embedding check result. When the updated surface closure verification result does not meet the surface closure rule or the updated object embedding verification result does not meet the object embedding rule, the abnormal mesh region is re-determined based on the updated surface closure verification result or the updated object embedding verification result, and the collaborative reconstruction constraints are updated again based on the re-determined abnormal mesh region; when the updated surface closure verification result meets the surface closure rule and the updated object embedding verification result meets the object embedding rule, the updated parameterized 3D mesh model is determined as the 3D mesh topology reconstruction result.

Images