Spatial data governance method and system based on digital twinning

Abstract

The invention discloses a spatial data management method and system based on digital twinning, particularly relates to the field of spatial information processing, and is used for solving the problems that a digital twinning three-dimensional model is difficult to locate defect causes in multi-source observation updating and repair is non-traceable. A target three-dimensional model is divided into surface patch units and a contribution index table is established by constructing an alignment result packet unified observation fragment set and a projection relation, a residual image is generated based on multi-view texture alignment to position a defect candidate area, a boundary projection dislocation pixel potential and a dynamic shielding coverage ratio are calculated to obtain a defect attribution result, and a defect attribution result is obtained. And performing local grid reconstruction or candidate fragment screening according to a defect attribution result, recalculating local textures to generate a repair model, updating an empirical distribution library in combination with projection re-inspection and boundary continuity inspection, and outputting a tracing record, so that integrated closed-loop treatment of defect positioning, attribution, repair and re-inspection is realized, and the consistency and credibility of the model are improved.

Description

Technical Field

[0001] This invention relates to the field of spatial information processing, and more specifically, to a spatial data governance method and system based on digital twins. Background Technology

[0002] In spatial data governance, spatial data often includes not only traditional vector layers and attribute information, but also 3D scenes and model results obtained from the processing of observation data such as imagery and point clouds. Furthermore, it requires continuous integration of new data, updates, maintenance, unified release, and sharing services during long-term operation. Existing technologies have explored data aggregation, database construction, and model maintenance for digital twin cities. For example, the "A Resource Data Database Construction and Management Method for Digital Twin Cities" (application number 202111653057.8) emphasizes the integration and association of resource data, while the "Digital Twin Model Update and Maintenance Method and System Based on Multi-Source Multimodal Data" (application number 202111618605.3) focuses on a multi-source observation-driven model update and maintenance process. These efforts provide fundamental support for the construction and operation of the digital twin platform.

[0003] However, in practical applications, the existing digital twin 3D model update and maintenance chain still has the following shortcomings: when local anomalies or quality defects occur in the model results, it is usually only possible to trace back to the production batch or processing flow corresponding to the results. It is difficult to further locate which source observation segments contributed to the abnormal area, what the corresponding projection relationship and processing caliber are, which makes it difficult to explain the cause of defects, difficult to accurately converge the repair scope, and difficult to review and audit the repair results, thus affecting the consistency and credibility of the continuous model update.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of existing technologies, embodiments of the present invention provide a spatial data governance method and system based on digital twins. This method constructs an alignment result package to unify the observation fragment set and projection relationship, divides the target 3D model into patch units and establishes a contribution index table, generates residual maps based on multi-view texture alignment to locate candidate defect areas, calculates the boundary projection misalignment pixel potential and dynamic occlusion coverage ratio to obtain defect attribution results, performs local mesh reconstruction or candidate fragment screening based on the defect attribution results, recalculates local textures to generate and repair the model, and updates the experience distribution library and outputs traceability records by combining projection re-inspection and boundary continuity verification. This achieves integrated closed-loop governance of defect location, attribution, repair, and re-inspection, improving model consistency and reliability, thereby solving the problems mentioned in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A spatial data governance method based on digital twins includes the following steps: S1: Obtain the source observation dataset and the target 3D model, perform coordinate registration and time alignment, and generate an alignment result package. The alignment result package contains the set of observation segments and projection relationships. S2: The target 3D model is divided into patch units. Based on depth rendering, the view range determination and occlusion determination are completed to obtain a set of candidate fragments of patch units and establish a contribution index table. S3: Contribution index table drives back projection consistency calculation, forms residual map and locates defect candidate area; calculates boundary projection misalignment pixel potential and dynamic occlusion coverage ratio for each defect candidate area, calculates defect attribution posterior confidence based on empirical distribution library, and outputs defect attribution results and candidate segment set. S4: Defect attribution results drive the repair path. Geometric misalignment defects trigger local mesh reconstruction and recalculation of local textures. Transient texture defects trigger candidate fragment screening and recalculation of local textures, generate repair models and update the corresponding records in the contribution index table. S5: Perform projection backtesting on the repaired model, recalculate the residual map and evaluate boundary continuity. For failed items, return to step S3 and update the evidence distribution in the empirical distribution library. For successful items, output the repaired model and retrospective records.

[0007] In a preferred embodiment, the source observation dataset is divided into observation segment sets according to the continuity of acquisition. Feature points are extracted from the target 3D model and the source observation dataset to establish a matching pair set, and an initial pose solution set is obtained through consistency verification. The initial pose solution set is obtained through point-to-plane consistency iteration. Event anchor point sequences are extracted from the observation segment sets and matched with the projectible anchor point sequences of the target 3D model to form a temporal mapping relationship. The observation segment set, pose solution set, temporal mapping relationship, and projection relationship are encapsulated into an alignment result package. Among them, the event anchor point sequence refers to the key moment sequence representing the appearance or structural changes of the same spatial region extracted according to the acquisition time sequence of the observation segments; the projectible anchor point sequence refers to the sequence of geometric feature points in the target 3D model that can be projected onto the image plane and stably identified under multiple views, formed according to the model version time label; the temporal mapping relationship is obtained by order-preserving matching; order-preserving matching means that the matching relationship between the event anchor sequence and the projectible anchor sequence maintains the consistency of the temporal order.

[0008] In a preferred embodiment, the target 3D model is divided into facet units based on the topological relationship of triangular facets. The facet units satisfy the continuous constraint of the difference between the normal angle and the curvature indicator. A set of surface sampling points is generated on the surface of the facet unit. The projection relationship maps the set of surface sampling points to the image plane of the set of observation segments. The pixels are rasterized to form the projection coverage area.

[0009] In a preferred embodiment, the set of observed segments performs depth rendering within the effective area of ​​the viewport to generate the foremost surface depth. The patch unit calculates the surface depth of the patch unit pixel by pixel in the projected coverage area and compares it with the foremost surface depth to obtain the visible area. The projected coverage area and the visible area are statistically analyzed to form the visible coverage ratio. The set of candidate segments for patch units is obtained by filtering the visible coverage ratio. The contribution index table records the set of candidate segments for patch units, the projected coverage area, and the visible area.

[0010] In a preferred embodiment, the contribution index table reads the candidate fragment set of patch units, the projection coverage area, and the visible area. The projection relationship extracts the candidate fragment texture fragment within the visible area. The candidate fragment texture fragment is mapped to the patch parameter domain to form an aligned texture. The residual map is generated by the median statistics of the pairwise differences of the aligned texture. The defect candidate area is determined by the preset thresholded connected region.

[0011] In a preferred embodiment, the boundary projection misalignment pixel potential is obtained by establishing a pixel distance field within the boundary band and combining it with edge direction consistency screening. The dynamic occlusion coverage ratio is obtained by calculating the optical flow return offset and appearance mutation within the intersection of the projection coverage area and the visible area and generating a dynamic confidence map. The empirical distribution library queries the conditional probability density of two types of defects and combines it with the prior probability to form the defect attribution posterior confidence. The defect attribution result is determined based on the defect attribution posterior confidence. Among them, the optical flow return offset refers to the positional offset length generated when a pixel is mapped back to its original position by forward optical flow and then by backward optical flow, which is used to characterize the consistency of pixel motion estimation. The boundary projection misalignment pixel potential refers to the statistical quantity obtained by statistically aggregating the distances from the image edge pixels after edge direction consistency screening to the model projection boundary within the boundary band corresponding to the model projection boundary, which is used to characterize the degree of misalignment between the model projection boundary and the real edge of the image.

[0012] In a preferred embodiment, the defect candidate region is back-mapped onto the surface of the target 3D model within the intersection of the projected coverage area and the visible area, forming a set of local triangular facets and expanding to obtain a set of local vertices and a set of boundary vertices. The candidate fragment set calculates the texture consistency deviation and transient contamination intensity within the defect candidate region. The defect attribution result controls the screening criteria for the repair evidence set. The transient contamination intensity refers to a statistical measure obtained by statistically analyzing the corresponding pixel values ​​of the dynamic confidence map within the defect candidate region, used to characterize the degree of influence of occlusion, reflection, or other short-term changes on texture stability.

[0013] In a preferred embodiment, when the defect attribution result is labeled as a geometric misalignment defect, local vertex set performs local mesh reconstruction while the boundary vertex set remains in its original position. When the defect attribution result is labeled as a transient texture defect, local vertex set remains unchanged while candidate fragment filtering is performed. Local texture recalculation uses fragment assignment and seam regularization to generate a repair model, updates the candidate fragment set of patch elements in the contribution index table, and recalculates the projected coverage area and visible area during geometric changes. Here, fragment assignment and seam regularization refer to determining the texture source of each candidate fragment position during local texture recalculation and performing smooth transition processing on the boundary areas of different texture sources to reduce discontinuities at fragment switching boundaries.

[0014] In a preferred embodiment, the alignment result package provides projection relationships, the contribution index table provides a set of candidate fragments of patch units, projection coverage areas, and visible areas, the repair model generates a back-check rendering map within the intersection of the projection coverage area and the visible area and calculates a back-check residual map with the candidate fragment images, the back-check defect candidate area and area ratio complete the fitting judgment, the geometric boundary continuity index and texture boundary alignment continuity index are calculated at the boundary vertex set to complete the boundary judgment, the failed items generate labels based on the back-check evidence and update the experience distribution library, and the passed items output traceability records.

[0015] A spatial data governance system based on digital twins includes: Alignment Package Module: Acquires the source observation dataset and the target 3D model, performs coordinate registration and time alignment, and generates an alignment result package. The alignment result package contains the set of observation segments and projection relationships. Patch Index Module: The target 3D model is divided into patch units. Based on depth rendering, view range determination and occlusion determination are completed to obtain a set of candidate fragments of patch units and establish a contribution index table; The residual attribution module: The contribution index table drives the back projection consistency calculation, forms the residual map and locates the defect candidate area; for each defect candidate area, it calculates the boundary projection misalignment pixel potential and dynamic occlusion coverage ratio, calculates the defect attribution posterior confidence based on the empirical distribution library, and outputs the defect attribution results and candidate segment set. Local Repair Module: Defect attribution results drive the repair path. Geometric misalignment defects trigger local mesh reconstruction and recalculation of local textures. Transient texture defects trigger candidate fragment filtering and recalculation of local textures, generate repair models and update the corresponding records in the contribution index table. Backtesting and tracing module: The repair model performs a projection backtest, recalculates the residual map and evaluates the boundary continuity. For failed items, it returns to step S3 and updates the evidence distribution in the empirical distribution library. For successful items, it outputs the repair model and tracing records.

[0016] The technical effects and advantages of the spatial data governance method and system based on digital twins of this invention are as follows: First, the alignment result package fixes the observation segment set and projection relationship into a unified standard. The contribution index table implements the candidate segment set of patch unit, projection coverage area, and visible area into the visible range. The residual map and defect candidate area express the abnormal location in the patch parameter domain, reducing misjudgment caused by occlusion holes and viewpoint differences. This transforms spatial data governance from global investigation to precise positioning at the patch unit level, and the repair range converges and is traceable.

[0017] Secondly, the boundary projection misalignment pixel potential and dynamic occlusion coverage ratio form complementary evidence. The defect attribution results distinguish between geometric misalignment defects and transient texture defects. Local mesh reconstruction and candidate fragment screening correspond to structural repair and texture purification, respectively. Local texture recalculation and seam regularization unify the fragment assignment caliber. The repair model is stable on both the contour and texture sides, avoiding contour offset residue caused by texture replacement alone, and also avoiding the introduction of polluted textures by geometric reconstruction.

[0018] Third, the projection re-inspection places the repair model back into the intersection of the projection coverage area and the visible area of ​​the observed image. The re-inspection residual map, the re-inspection defect candidate area, the area ratio and boundary continuity test constitute a consistent acceptance link. For items that fail, the experience distribution library is updated according to the re-inspection evidence and the defect attribution results are recalculated. For items that pass, the traceability record is output, so that the evidence is consistent in the same area unit in multiple treatments, and the problem reproduction, repair review and responsibility audit are more direct. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a spatial data governance method based on digital twins according to the present invention.

[0020] Figure 2 This is a schematic diagram of the structure of a spatial data governance system based on digital twins according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1: Figure 1 This invention presents a spatial data governance method based on digital twins, comprising: S1: Obtain the source observation dataset and the target 3D model, perform coordinate registration and time alignment, and generate an alignment result package. The alignment result package contains the set of observation segments and projection relationships.

[0023] S2: The target 3D model is divided into patch units. Based on depth rendering, the view range determination and occlusion determination are completed to obtain a set of candidate fragments of patch units and establish a contribution index table.

[0024] S3: The contribution index table drives the back projection consistency calculation, forming a residual map and locating the defect candidate area; for each defect candidate area, the boundary projection misalignment pixel potential and dynamic occlusion coverage ratio are calculated, the defect attribution posterior confidence is calculated based on the empirical distribution library, and the defect attribution results and candidate fragment set are output.

[0025] S4: Defect attribution results drive the repair path. Geometric misalignment defects trigger local mesh reconstruction and recalculation of local textures. Transient texture defects trigger candidate fragment screening and recalculation of local textures, generating a repair model and updating the corresponding record in the contribution index table.

[0026] S5: Perform projection backtesting on the repaired model, recalculate the residual map and evaluate boundary continuity. For failed items, return to step S3 and update the evidence distribution in the empirical distribution library. For successful items, output the repaired model and retrospective records.

[0027] The source observation dataset comes from image sequences or point cloud sequences. The batch size and time stamp of the acquisition result in the same spatial object appearing from different perspectives and with different poses in different observation segments. The target 3D model serves as the basis for the digital twin. It is necessary to accurately map the set of observation segments onto the model coordinate system in order to establish projection comparison, field of view determination, and occlusion determination on the same spatial reference. However, the source observation dataset and the target 3D model are often in different coordinate frames and different time axes. The sensor intrinsic and extrinsic parameters will also deviate with the acquisition conditions. Direct projection is prone to positional drift and temporal mismatch, affecting the reliability of defect localization and attribution.

[0028] Construction of the S101 observation fragment set.

[0029] Multi-source acquisition data typically spans time periods and routes. Directly using the entire image sequence or point cloud sequence for registration can easily introduce acquisition discontinuities and pose abrupt changes into the same solution process, thus spreading local errors into overall errors. The source observation dataset is first organized into a frame sequence according to the acquisition time order. Each frame in the sequence retains the acquisition time label and sensor intrinsic parameter information. The observation segment set is formed through a two-condition segmentation. The first condition uses the difference between the acquisition time labels of two adjacent frames; a segment boundary is established when the difference exceeds a preset continuity limit. The second condition uses the change in the angle of attitude rotation between two adjacent frames. The angle change is obtained by multiplying adjacent rotation matrices, taking the trace value, and converting it into an angle. A segment boundary is established when the angle change exceeds a preset rotation continuity limit. After segmentation, the observation segment set is obtained. Each observation segment contains a continuous frame range, a time label sequence, and an intrinsic parameter sequence. The observation segment set naturally isolates acquisition discontinuities, ensuring that subsequent pose calculations are performed within continuous acquisition segments, reducing misregistration caused by cross-segment jumps.

[0030] S102 coarse registration obtains the initial pose solution set.

[0031] When the target 3D model and the source observation dataset are in different coordinate references, direct iteration in fine registration is prone to getting stuck in local convergence. The initial pose solution set needs to fall within an iteratively correctable neighborhood. A stable set of geometric feature points is extracted from the target 3D model, with priority given to corner points and curvature abrupt changes near boundary polylines. Corresponding spatial or visual features are extracted from the source observation dataset within the observation segment set. Spatial features are preferentially derived from curvature extrema of the point cloud, while visual features are preferentially derived from scale-invariant keypoints in the image. A set of matching pairs is established based on descriptor similarity between the two sides' features. Consistency verification is then performed on the matching pairs. Consistency verification uses random sampling to generate candidate rigid body transformations. After transforming the source observation feature points to the target 3D model coordinate system, the distance from the transformed point to the matching target point is calculated. Matching pairs with distances less than a preset consistency threshold are considered passed. Repeated sampling is then performed, and the candidate rigid body transformation with the highest number of passed pairs is selected. Based on the passed set of matching pairs, the coarse registration solution adopts the principle of minimizing the sum of absolute distances, with the objective of minimizing the sum of the absolute values ​​of the distances between the transformed point and the matching target point. The coarse registration outputs an initial pose solution set, which gives the initial rigid body pose for each observation segment in the observation segment set. Mismatches are removed during the consistency verification stage, and the initial pose solution set has stable and verifiable properties.

[0032] S103 fine-registers and refines the pose solution set and stabilizes the projection relationship.

[0033] Coarse registration results often still contain local drift, especially in building facades and areas with repetitive textures, where errors manifest as misaligned projection boundaries. Fine registration aims to compress these errors to a usable, patch-level range. Fine registration is performed independently on each observation segment, with inputs being the initial pose solution set and the set of sampled points on the target 3D model surface. Fine registration employs a point-to-plane consistency iteration. The target 3D model provides surface normals at each nearest point. After the source observation points are projected onto the target coordinate system through pose transformation, the nearest point is found, and its normal is taken. The point-to-plane residual is defined as the inner product of the normal and the point difference vector, with the residual dimension maintained as length. Pose update aims to minimize the sum of the absolute values ​​of the residuals. The iterative process involves alternating nearest-point re-pairing and pose re-estimation until the residual variation stabilizes. Fine registration outputs a refined pose solution set, which satisfies the consistency requirement for repeatable projection on each observation segment. The projection relationship no longer depends on accidental matching points, significantly improving the stability of the projection coverage area.

[0034] S104 time alignment establishes a time mapping relationship.

[0035] The digital twin platform undergoes version updates. When observation segments are collected across time periods, short-term occlusion and seasonal differences can alter texture appearance. Spatial alignment alone cannot distinguish between version differences and transient differences; therefore, the temporal mapping relationship needs to unify the expression of the observation timeline and the model version timeline. The source observation dataset extracts event anchor sequence sequences from the observation segment set. Event anchors are defined as moments in the same spatial region where the appearance or structure undergoes repeatable changes. Event anchors are determined by two types of evidence: the first is abrupt changes in global similarity between adjacent frames, and the second is abrupt changes in the number of local feature matches within the same region. A projectable anchor sequence is constructed for the corresponding region of the target 3D model. Projectable anchors are defined as geometric anchors that can be projected onto the image plane and stably appear in multiple viewpoints, carrying a model version time label. The temporal mapping relationship employs order-preserving matching, which requires that the matching between the event anchor sequence and the projectable anchor sequence cannot be reversed or many-to-one.

[0036] The first type of evidence uses global similarity mutation data. This data is calculated from adjacent frame image pairs and includes the similarity value, similarity difference, similarity difference sign, and adjacent frame image pair index. The similarity value reflects the overall consistency of the entire frame image in terms of grayscale and edge structure; the similarity difference reflects the degree to which adjacent frames change from continuous to significant variation; the similarity difference sign is used to distinguish the direction of mutation; and the adjacent frame image pair index is used to map the mutation location back to the acquisition time stamp.

[0037] The second type of evidence uses local feature match quantity mutation data. This data is obtained by statistically analyzing local feature matches of the same region across adjacent frame image pairs. The data includes region identifier, region pixel range, number of valid feature points within the region, number of valid match pairs within the region, difference in the number of valid match pairs, number of matches that pass consistency, difference in the number of matches that pass consistency, and adjacent frame image pair indices. The region identifier and region pixel range ensure a fixed statistical range. The number of valid feature points reflects the sufficiency of regional texture information, the number of valid match pairs reflects the trackability of the region across frames, and the number of matches that pass consistency reflects the stability and reliability of the matches. The two differences are used to identify significant changes in the number of matches within a short period, thus supporting the determination of the event anchor point.

[0038] The matching cost consists of two parts. The first part measures the overall offset consistency between the event anchor time stamp and the projectible anchor time stamp, which is determined by the median of the time difference between matching pairs. The second part measures the consistency between the time intervals of adjacent event anchors and the time intervals of adjacent projectible anchors. The time mapping relationship is obtained by solving the minimum cost path through dynamic programming. The time mapping relationship has sequential monotonicity, avoiding version confusion caused by time mapping jumps.

[0039] S105 generates the alignment result package and specifies the projection relationship content items.

[0040] The pose solution sets obtained in steps S2 and S3 need to be included in the subsequent depth rendering and contribution index table construction in a unified format. The temporal mapping relationship obtained in step S4 needs to be encapsulated together with the pose solution set to avoid inconsistencies caused by the separation of projection and temporal relationships. The alignment result package consists of four types of content items: the first type is the set of observation segments; the second type is the pose solution set, which provides rotation and translation for each observation segment; the third type is the temporal mapping relationship; and the fourth type is the projection relationship. The projection relationship is expressed using the camera imaging model. From the 3D points in the target 3D model coordinate system, the coordinate transformation is completed through the pose solution set, and then the pixel mapping is completed by the intrinsic parameters corresponding to the observation segments. The pixel coordinates are normalized in a homogeneous form to obtain 2D coordinates. After the alignment result package is generated, the set of observation segments and the pose solution set are consistent under the same coordinate reference, and the temporal mapping relationship and the projection relationship are consistent within the same encapsulation. Subsequent steps can directly perform depth rendering and occlusion determination based on the projection relationship, avoiding different interpretations of objects with the same name in different stages.

[0041] After coordinate registration and temporal alignment are completed, the alignment result package stably carries the set of observation segments, the pose solution of each observation segment, and the projection relationship. Coarse registration provides reliable initial values, fine registration converges the projection error to a stable range, and event anchor point alignment maps the time axis of the observation segments to the time axis of the model version, ensuring that the same spatial region has a consistent projection aperture in different observation segments. The resulting alignment result package can support depth rendering to obtain visibility masks, support the contribution index table to establish projection coverage areas and occlusion states, provide a recalcible basis for the candidate segment set of patch units, and establish various judgments of spatial data governance on a unified spatial and temporal reference.

[0042] The source observation dataset and the target 3D model are aligned in step S1 to form an alignment result package. The alignment result package contains a set of observation segments and a projection relation. The projection relation maps the surface of the target 3D model onto the image plane of the set of observation segments. However, the projection relation itself cannot directly provide a stable visible region. It is necessary to decompose the target 3D model into patch units and confirm the projection coverage area and occlusion state of each patch unit in the set of observation segments in order to associate the same patch unit with a set of candidate patch units with consistent observation conditions.

[0043] S201 patch unit division.

[0044] When the granularity of triangular facets is too fine, occlusion determination is easily affected by local noise. Facet units need to maintain a consistent shape within geometric continuity. The target 3D model first reads the topological relationships of the triangular facets, forming a set of adjacent triangular facet pairs. For each triangular facet, a unit normal is calculated, which is obtained by normalizing the cross product of the two edge vectors formed by the coordinates of the three vertices of the triangular facet. The included angle of normals between adjacent triangular facet pairs is calculated, which is obtained by taking the inverse cosine of the dot product of the two unit normals. For each triangular facet, a curvature indicator is calculated, which is implemented using dihedral angle variation, obtained by statistically analyzing the included angle of normals between the triangular facet and its adjacent triangular facets along their adjacent edges.

[0045] The aggregation process starts from any triangular facet and merges adjacent triangular facets that satisfy the condition that the included normal angle does not exceed a preset upper limit and the difference in curvature indices does not exceed a preset upper limit into the same aggregation result, until no adjacent triangular facets satisfy the conditions are found. The aggregation result forms a facet unit. After the facet unit set is formed, the facet unit boundary is naturally given by the aggregation termination position. The facet unit maintains continuity in local surface morphology, and the occlusion determination stage is no longer amplified by the occasional errors of individual triangular facets.

[0046] S202 projection coverage area generation.

[0047] View scope determination and occlusion determination require the projection coverage area of ​​the patch unit at the pixel level; otherwise, the pixel count will lack a unified boundary. The alignment result package provides the set of observation segments and projection relationships. Each observation segment contains the pose solution and imaging intrinsic parameters. Each patch unit generates a set of surface sampling points on its surface. The set of surface sampling points consists of the centroid of the triangular patch and the midpoint of the edge of the triangular patch, and the sampling point density is adaptively increased according to the area of ​​the triangular patch to ensure that the projection boundary of the slender triangular patch is not missing. The projection relationship maps the set of surface sampling points point by point to the image plane of the observation segment, obtaining the set of projected pixel coordinates. The set of projected pixel coordinates is pixel rasterized to form the projection coverage area, which is expressed in the form of a set of pixels. Once the projection coverage area is clear, view scope determination and occlusion determination only occur within the projection coverage area, and the pixel statistics remain consistent.

[0048] S203 Viewpoint Determination and Depth Rendering Generate Occlusion Reference.

[0049] When the projected coverage area includes both external and back-projection mappings, occlusion determination may result in errors such as mistaking unimageable areas for occlusion areas. Each observation segment transforms the target 3D model to the camera coordinate system under the observation segment pose. During the transformation process, rotation and translation are performed on each vertex of the target 3D model. View area determination is performed pixel-by-pixel on the projected coverage area. The determination criteria include that the pixel coordinates fall within the imaging plane range and the corresponding 3D point depth is positive. Pixels that meet the criteria form the effective view area. Depth rendering is performed within the effective view area. Depth rendering uses a rasterization method to traverse the triangular facets of the target 3D model. After projecting the triangular facets onto the image plane, a pixel coverage set is generated. For each covered pixel, the depth value of the triangular facet in the camera coordinate system is recorded. Only the pixel with the smallest depth value is retained to obtain the foreground surface depth. The foreground surface depth serves as a unified benchmark for occlusion determination, and occlusion determination no longer depends on the sparsity of sampling points within the facet unit.

[0050] S204 occlusion determination obtains the visible area and calculates the visible coverage ratio.

[0051] When the occlusion status of the same patch unit differs significantly across different observation segments, texture segment consistency comparison may mistakenly identify occlusion edges as defects. Within the effective viewing area of ​​each observation segment, the surface depth of the patch unit is calculated pixel-by-pixel for the projected coverage area of ​​the patch unit. The surface depth of the patch unit is obtained by interpolating the depth of the corresponding triangular patch at the pixel position. Occlusion determination compares the surface depth of the patch unit with the foremost surface depth. If the absolute value of the difference does not exceed a preset depth consistency tolerance, the pixel is marked as visible; otherwise, it is marked as occluded. The visible area consists of visible pixels, and the occlusion status consists of occluded pixels. The visible coverage ratio is obtained through pixel counting. First, the number of pixels within the effective viewing area of ​​the projected coverage area is counted, then the number of visible pixels within the same area is counted. The ratio of the number of visible pixels to the number of projected coverage pixels is defined as the visible coverage ratio. After the visible coverage ratio is formed, the visibility of the same patch unit across different observation segments has a comparable and unified indicator, and the influence of occlusion edges on the statistical results is limited to pixel-level determination.

[0052] In one implementation, the visible coverage ratio is denoted as The calculation method is as follows: within the effective viewing area, the number of pixels corresponding to the projected coverage area is denoted as... The number of visible pixels that satisfy the condition that the absolute value of the difference between the surface depth of a patch unit and the depth of the foremost surface does not exceed the depth consistency tolerance is denoted as . ,but .when When the observed fragment is not included in the candidate fragment set of the patch unit, the observed fragment will not be included in the candidate fragment set.

[0053] The candidate fragment set of S205 patch unit was determined and a contribution index table was established.

[0054] When the candidate fragment set of a patch unit lacks quantization filtering, grazing fragments and heavily occluded fragments will enter the candidate range, and the residual map will be perturbed by unstable observation conditions. Each patch unit traverses the observation fragment set and selects observation fragments with a visible coverage ratio not lower than a preset lower limit to form the patch unit candidate fragment set. The contribution index table uses the patch unit identifier as the index key to store the patch unit candidate fragment set, and records the projection coverage area and visible area for each observation fragment. The projection coverage area is used to limit the range of texture fragment extraction, and the visible area is used to exclude occluded pixels. After the contribution index table is generated, the correspondence between patch units and the observation fragment set has a stable boundary. During the residual map generation stage, texture fragments can be extracted within a consistent projection coverage area, and consistency comparison is completed within the visible area. Occluded pixels are no longer included in the texture difference statistics.

[0055] Patch unit partitioning transforms the target 3D model into geometrically continuous analysis units. Projection coverage region generation maps patch units to the pixel range of the observation fragment set. View scope determination and depth rendering provide a unified foreground surface depth benchmark. Occlusion determination fixes the visible area and occlusion state at the pixel level. The patch unit candidate fragment set is obtained by filtering based on the visible coverage ratio. The contribution index table solidifies the correspondence between patch units and the patch unit candidate fragment set, as well as the projection coverage region and the visible region. In step S3, when generating the residual map using the contribution index table, the texture fragment extraction range and the occlusion culling range remain consistent, and the defect candidate region localization aperture remains consistent.

[0056] In step S1, the source observation dataset and the target 3D model are aligned to form an alignment result package, which provides the set of observation segments and projection relationships. In step S2, the target 3D model generates a contribution index table, which provides the set of candidate segments for patch units, the projected coverage area, and the visible area. When digital twin spatial data governance enters the quality assessment stage, it is necessary to distinguish between texture differences and geometric misalignments, converge defect locations to defect candidate areas, and converge defect sources to the set of candidate segments. However, multi-view observations are significantly affected by occlusion and lighting changes. Step S3 uses a unified patch parameter domain expression and a dual-evidence attribution mechanism to complete the defect attribution determination.

[0057] S301 texture fragment extraction and patch parameter domain alignment.

[0058] Since patch units differ in pixel resolution and viewpoint scale across different candidate fragments, directly comparing textures on the pixel plane would treat scale differences as defects. The contribution index table reads the set of candidate fragments for each patch unit, along with the projected coverage area and visible area for each fragment. Projection relationships map the surface points of the patch units onto the image plane of the candidate fragments, extracting pixels only within the visible area to form texture fragments. A patch parameter domain mesh is established for each patch unit, covering its surface and maintaining topological consistency with the triangular patches. Each parameter position in the patch parameter domain mesh is used to find its corresponding pixel position on the image plane of the candidate fragment through projection relationships. Bilinear interpolation is then performed on the texture fragments to obtain the patch parameter domain texture value. After alignment, each candidate fragment texture fragment is transformed into the same patch parameter domain texture representation, and texture difference comparisons are no longer affected by pixel scale differences.

[0059] S302 Residual Map Generation and Defect Candidate Region Location.

[0060] When inconsistencies occur in local areas of multi-view textures, single-view differences are easily skewed by noise and local exposure. The residual map uses robust multi-view statistics to concentrate differences into true anomaly areas. Aligned texture values ​​of the candidate fragment set of patch units are collected at each parameter position in the patch parameter domain. The absolute differences between each pair are calculated, and the median value is taken to form the pixel value of the residual map. The calculation range of the residual map is limited to the effective parameter domain after the visible area is mapped to the patch parameter domain; parameter positions corresponding to occlusion holes are not included in the statistics. Within the effective parameter domain of the residual map, the residual values ​​are sorted, and the residual value corresponding to the preset quantile position is used as a threshold. Parameter positions in the residual map that are not lower than the threshold are used to extract defect candidate regions through four-neighbor connected components. Connectivity determination uses the four-neighbor domain of the patch parameter domain, and the boundary contour and parameter domain area information are preserved when outputting the defect candidate region. After the residual map and defect candidate regions are formed, the texture anomaly positions are fixed to the local connected regions of the patch parameter domain, and occluded pixels will not widen the defect candidate regions.

[0061] In one implementation, let the number of candidate fragments for the patch unit be... Patch parameter domain position The alignment texture value is denoted as ,in The residual plot value is defined as follows: .

[0062] The residual plot threshold is given using a quantile approach: taking values ​​within the effective parameter domain. quantiles As a threshold, where Preset quantile positions; satisfy The parameter positions are used to extract the defect candidate region through connected component extraction.

[0063] S303 boundary projection misaligned pixel potential calculation.

[0064] Geometric misalignment often manifests as displacement between the model projection boundary and the real edge of the image. The boundary projection misalignment pixel potential transforms the displacement scale into pixel-level evidence, facilitating complementarity with dynamic occlusion coverage ratio. For each candidate segment, the projection relationship maps the polygonal line of the patch unit boundary onto the image plane to form the model projection boundary. The model projection boundary extends to both sides to form a boundary band. Within the boundary band, each pixel calculates the shortest distance to the nearest model projection boundary pixel and records the sign and direction of the distance, forming a pixel distance field. Image edge pixels are extracted from the candidate segment image plane, confining them within the boundary band, while occluded pixels are removed from the visible area. To suppress edge interference within the facade texture, the image edge pixels are further filtered for edge direction consistency. The edge direction is determined by the local gradient direction, and pixels whose edge direction and the normal of the model projection boundary exceed the direction tolerance are excluded from the statistics. The absolute distance values ​​of the pixel distance field are read from the filtered image edge pixels, and quantile convergence is used to obtain the candidate segment misalignment potential. The median value of multiple candidate segment misalignment potentials is then taken to form the boundary projection misalignment pixel potential. After the potential for misaligned pixels in boundary projection is formed, the evidence of geometric misalignment mainly comes from stable edges near the model contour, while the influence of internal texture edges is suppressed.

[0065] In one implementation, for each candidate segment, the set of image edge pixels after edge direction consistency filtering is denoted as . Pixel distance field in pixels The signed distance at point is denoted as Candidate fragment misalignment potentials are aggregated according to quantile points. ,in The boundary projection misalignment pixel potential at the predefined quantile point is defined as the misalignment potential of all candidate segments for that patch unit. The median value is used to suppress the bias caused by abnormal edges in single segments.

[0066] S304 Dynamic Shading Coverage Ratio Calculation.

[0067] Transient occlusion and reflective changes exhibit temporal inconsistencies. Dynamic occlusion coverage ratio transforms these temporal inconsistencies into proportional evidence, enabling the separation of transient texture anomalies from geometric misalignments. The candidate fragment set for patch units is sorted by acquisition time label. Temporally adjacent candidate fragment pairs are selected, and forward and backward optical flow are calculated within the intersection of the projected coverage area and the visible area. Optical flow is obtained using pyramid-based encrypted matching, with matching costs constrained by a combination of brightness consistency and gradient consistency. The optical flow return test performs forward and backward mapping on each pixel, using the pixel offset length returning to its original position as the motion consistency residual. Pixels mapped outside the image range are directly discarded. The appearance abrupt change test compares the absolute value of the brightness difference at the corresponding optical flow position, while also incorporating the absolute value of the gradient magnitude difference to avoid false triggering by pure exposure changes. The motion consistency residual and appearance abrupt change residual are normalized using quantile stretching, with the stretching interval determined by the intermediate quantiles and extreme tails truncated. The normalization results remain within the same proportional range. The dynamic confidence map is calculated at the pixel level by multiplying the normalized residuals of two classes. Only pixels that simultaneously satisfy both motion inconsistency and abrupt appearance changes exhibit high confidence. Defect candidate regions are mapped back to the image plane from the patch parameter domain. The proportion of pixels in the set covered by the defect candidate region that exceed the dynamic confidence map threshold is statistically analyzed, forming the dynamic occlusion coverage ratio. After the dynamic occlusion coverage ratio is formed, transient change evidence is supported by temporal consistency, and the influence of gradual illumination changes and static texture differences is weakened.

[0068] In one implementation, for temporally adjacent candidate segment pairs, the forward optical flow is denoted as... Backward optical flow is denoted as Pixels Optical flow return offset is defined as The residual of appearance mutation is defined as follows: .right and Quantitative stretching and normalization were performed separately to obtain and Dynamic confidence graph is defined as The set of pixels after the defect candidate region is mapped to the pixel domain is denoted as... Dynamic occlusion coverage ratio is defined as ,in The threshold for the dynamic confidence graph.

[0069] S305 calculates and outputs the posterior confidence level of defect attribution and the defect attribution results.

[0070] Defect attribution determination requires mapping the boundary projection misalignment pixel potential and dynamic occlusion coverage ratio to a unified probability scale. An empirical distribution library provides statistical descriptions of evidence for two types of defects, ensuring interpretable evidence for attribution determination. The empirical distribution library contains samples of two types: geometric misalignment defects and transient texture defects, derived from historical repair records and manual verification results. The library establishes conditional probability densities for the boundary projection misalignment pixel potential and dynamic occlusion coverage ratio, respectively. These conditional probability densities are implemented using binned histograms with a smoothing kernel. Bin boundaries are determined by quantiles, and a triangular kernel is used for smoothing to avoid gaps with zero probability. The library also records the prior probabilities of both types of defects, derived from sample proportions and smoothed using Laplace's algorithm. For each defect candidate region, the boundary projection misalignment pixel potential and dynamic occlusion coverage ratio are read. The conditional probability density values ​​for both types of defects are retrieved from the empirical distribution library and multiplied by the prior probabilities to obtain the posterior probabilities. The difference between the two posterior probabilities forms the posterior confidence score for defect attribution. The posterior confidence sign of defect attribution determines the defect attribution result. The defect attribution result is either a geometric misalignment defect or a transient texture defect, and the candidate fragment set follows the same criteria as the patch element candidate fragment set. The output includes a residual map, defect candidate region, defect attribution posterior confidence, defect attribution result, and candidate fragment set. After the defect attribution determination is completed, the defect candidate region is clearly defined, the source of evidence is clear, and local repair can choose local mesh reconstruction and local texture recalculation based on the defect attribution result, or choose candidate fragment rejection and local texture recalculation.

[0071] In one implementation, the boundary projection misalignment pixel potential is denoted as... The dynamic occlusion coverage ratio is denoted as The empirical distribution library provides the conditional probability density for both types of defects. and And the prior probabilities of geometric misalignment defects and transient texture defects. The two categories have the following characteristics. Under the condition, the posterior probability of the two types of defects is as follows , Calculate and normalize; the posterior confidence score for defect attribution is defined as follows: Its symbol or size is used to determine the defect attribution result.

[0072] The residual map solidifies multi-view texture differences into the patch parameter domain, while the defect candidate region compresses the problem scope into locally connected regions. The boundary projection misalignment pixel potential transforms the displacement between the model projection boundary and the image's true edge into pixel evidence, and the dynamic occlusion coverage ratio transforms temporally inconsistent occlusion and reflection into proportional evidence. The empirical distribution library maps both types of evidence to the posterior probability space, ensuring that the defect attribution posterior confidence and defect attribution results have a stable caliber. The candidate fragment set remains consistent with the contribution index table, and step S4 can complete differentiated local repair under the same evidence system.

[0073] Step S3 has already output the residual map, defect candidate regions, defect attribution results, and candidate fragment set. The contribution index table has saved the candidate fragment set of patch units, the projected coverage area, and the visible area. When digital twin spatial data governance enters the repair stage, the repair scope needs to be locked on the local surface of the target 3D model corresponding to the defect candidate region, and the repair evidence needs to be locked on the candidate fragment set to avoid mistaking occlusion changes for geometric misalignment and contour misalignment for texture contamination. However, defect candidate regions come from differences in observation from multiple perspectives, and the repair action must use a unified standard to organize the local mesh and local texture. Step S4 generates the repair model accordingly and updates the contribution index table synchronously.

[0074] The candidate area for S401 defects is localized and a local repair area is formed.

[0075] When the defect candidate region is located in the patch parameter domain, direct repair in the patch parameter domain is insufficient to constrain the continuity of the true boundary of the target 3D model. The local repair area needs to fall on the surface of the target 3D model and have a clear boundary. The contribution index table reads the projection coverage area and visible area corresponding to the patch unit identifier. The defect candidate region performs a reverse mapping within the intersection of the projection coverage area and the visible area. The reverse mapping converts the patch parameter domain position of the defect candidate region into a point on the surface of the target 3D model. The surface points of the target 3D model are then mapped to the triangular patch index to obtain a local triangular patch set. The local triangular patch set is expanded to form a buffer loop. The expansion of the buffer loop adopts the grid geodesic distance rule. The grid geodesic distance is defined as the shortest path length of the accumulated edge length along the grid edge. The expansion range is controlled by a preset threshold and kept fixed. All vertices inside the buffer loop form a local vertex set, and the outermost vertices of the buffer loop form a boundary vertex set. The boundary vertex set retains its original position and does not participate in the displacement. After the local repair area is formed, local mesh reconstruction and local texture recalculation are limited to the local vertex set and the local triangular patch set. The repair boundary is naturally constrained by the boundary vertex set, and the connection between the repaired model and the original model is smoother.

[0076] The S402 candidate fragment set was reliably screened and a repair evidence set was generated.

[0077] When the candidate fragment set includes occluded and reflective fragments, backprojection of densely matched points will generate erroneous geometric points, and texture recalculation will bring contaminated textures into the repair results. The repair evidence set needs to be filtered within the defect candidate region. The contribution index table reads the candidate fragment set of patch units and reads the alignment texture and dynamic confidence map corresponding to each candidate fragment. The defect candidate region limits the statistical range. First, a reference alignment texture is constructed. The reference alignment texture gathers the alignment texture values ​​of the candidate fragment set at each position in the defect candidate region, and the median statistics are used to obtain the reference value. The reference alignment texture maintains the same dimension as the alignment texture. Then, the texture consistency deviation is calculated. For each candidate fragment, the absolute difference between the candidate fragment alignment texture and the reference alignment texture is calculated point by point in the defect candidate region, and then the median statistics are performed on the absolute difference. The texture consistency deviation is used to characterize the degree of consistency between the candidate fragment texture and the majority of observations. The transient contamination intensity is then calculated. A dynamic confidence map is generated within the pixel set after the defect candidate region is mapped to the pixel domain. When summarizing the dynamic confidence map from adjacent fragment pairs to a single fragment, a pixel-level maximum rule is used. Before calculation, pixels exceeding the optical flow mapping boundary and pixels outside the visible area are removed. The transient contamination intensity is calculated using the median of the pixel dynamic confidence values. When the defect attribution result is labeled as a geometric misalignment defect, the repair evidence set selects candidate fragments with texture consistency deviation not exceeding the texture consistency deviation quantile threshold and transient contamination intensity not exceeding the transient contamination intensity quantile threshold, avoiding contaminated textures and unstable observations from entering the reconstruction. When the defect attribution result is labeled as a transient texture defect, the repair evidence set selects candidate fragments with transient contamination intensity not exceeding the transient contamination intensity quantile threshold. The condition of texture consistency deviation not exceeding the texture consistency deviation quantile threshold is used to remove obviously abnormal fragments. After the repair evidence set is formed, the observation points used for local mesh reconstruction are more reliable, the texture sources used for local texture recalculation are cleaner, and the defect attribution results are consistently implemented in the repair process.

[0078] S403 geometric misalignment defects perform local mesh reconstruction while maintaining boundary continuity.

[0079] Geometric misalignment defects reflect offsets in contour and structural position. Simple texture replacement cannot eliminate projected boundary misalignment. Local mesh reconstruction requires adjusting the local surface corresponding to the defect candidate area to a consistent observation position while keeping the boundary vertex set unchanged to avoid seam steps. When the defect attribution result is labeled as a geometric misalignment defect, the repair evidence set generates a local geometric point set. The local geometric point set is derived from the backprojection of dense matching points from the candidate fragment set. The backprojection is only performed within the intersection of the projected coverage area and the visible area, and only matching points whose line of sight falls into the defect candidate area pixel set are retained to avoid interference from neighborhood structures in the defect candidate area. The surface normal is estimated for each local geometric point, and the normal is obtained by fitting the principal direction of the local geometric point's neighborhood. Local mesh reconstruction consists of two alternating actions: The first action maintains the relative shape of the mesh, calculating the displacement of each vertex in the local vertex set. The displacement difference between the two ends of each mesh edge in the local vertex set is consistent with the relative displacement difference between the two ends of the same mesh edge in the original model, ensuring that local deformation is mainly manifested as overall translation rather than abrupt changes in edge length. The cost is calculated by accumulating absolute distances to ensure that a small number of outliers do not dominate the overall displacement. The second action conforms to the observation plane, compressing the signed distance from the vertex to the tangent plane of the local geometric point to a tolerance range. Compression uses truncation correction along the normal direction, with each correction not exceeding a preset upper limit. The correction order is executed outward from the center of the defect candidate area, ensuring priority conformation to defect areas. The boundary vertex set remains unchanged throughout the process. These two actions are repeated until the vertex correction amplitude enters a stable range. After local mesh reconstruction, the local surface position corresponding to the defect candidate area closely matches the observed shape, the boundary vertex set ensures geometric continuity with the original model, and structural repair is achieved by correcting model contour misalignment.

[0080] S404 local texture recalculation and seam regularization.

[0081] Local mesh reconstruction alters the surface geometric mapping, candidate fragment elimination changes the texture source set, and local texture recalculation requires the formation of stable fragment assignments within the defect candidate region and the suppression of jagged edges at fragment switching boundaries. Local texture recalculation is performed in the patch parameter domain, with the defect candidate region limiting the recalculation range. The texture source set is determined by the defect attribution results; geometric misalignment defects use the geometric misalignment repair evidence set, and transient texture defects use the transient texture repair evidence set. Fragment assignment is performed at each location in the patch parameter domain, employing a deterministic dual-criteria sorting method: the first criterion uses the dynamic confidence map value, selecting the candidate fragment with the smallest dynamic confidence map value; the second criterion uses texture consistency deviation, selecting the candidate fragment with the smallest texture consistency deviation when the first criterion is tied. When the two criteria are tied, the candidate fragment with the earliest acquisition time stamp is selected to ensure unique assignment results. After fragment assignment is determined, the texture value at the corresponding location is read from the aligned texture of the assigned candidate fragment and written into the local texture. Seam regularization is performed within a thin band in the patch parameter domain. This thin band is formed by extending the fragment assignment change boundary to both sides, and the expansion width is controlled and kept fixed by a preset threshold. Inside the thin band, fragment assignments are replaced with the majority assignments in the neighborhood. The neighborhood uses the four-neighborhood of the patch parameter domain. Stability constraints are set during the replacement process; replacement is only performed when the majority assignment ratio exceeds the preset threshold to avoid excessive smoothing of detail textures. After regularization, local textures are regenerated. After the local texture recalculation, contamination textures in the defect candidate area are replaced, and the fragment switching boundary converges from a fragmented shape to a continuous shape, resulting in a more coherent texture in the repaired model.

[0082] S405 generates a patch model and updates the contribution index table synchronously.

[0083] The contribution index table carries the set of candidate fragments for facet units, the projected coverage area, and the visible area. When the repair model changes local geometry or local texture, the contribution index table needs to reflect the changes synchronously to avoid aperture drift when the same facet unit re-enters the residual map calculation. When the defect attribution result is marked as a geometric misalignment defect, the repair model replaces the vertex coordinates of the local vertex set with the local mesh reconstruction result and replaces the texture within the defect candidate area with the normalized local texture. When the defect attribution result is marked as a transient texture defect, the repair model keeps the vertex coordinates of the local vertex set unchanged and only replaces the texture within the defect candidate area with the normalized local texture. The contribution index table locates the record item by facet unit identifier, updates the set of candidate fragments for facet units to the repair evidence set, and records the list of candidate fragment sets participating in the repair and the repair action type in the record item. When local mesh reconstruction occurs, the contribution index table recalculates the projected coverage area and the visible area for the local triangular facet set, recalculates the depth rendering aperture following step S2 and limits it to the local range, and replaces the original recorded projected coverage area and visible area. After the contribution index table is updated, the patching model is consistent with the contribution index table, the source of patching evidence is clear, and the boundaries of the patching area are well-defined.

[0084] The local repair scope places the defect candidate area onto the surface of the target 3D model. The repair evidence set filters the candidate fragment set to stable observation. Geometric misalignment defects are repaired by local mesh reconstruction to correct structural offsets. Transient texture defects are removed by candidate fragment elimination and local texture recalculation to clear contaminated textures. Seam regularization stabilizes the fragment switching boundary into a continuous form. The contribution index table is updated synchronously with the candidate fragment set of patch units and refreshes the projection coverage area and visible area during geometric updates. The repaired model has an auditable source and scope.

[0085] In step S4, the repair model forms and updates the contribution index table, which stores the candidate fragment set of patch units, the projected coverage area, and the visible area. During the quality verification stage, the repair model needs to be reprojected onto the observed fragment set to verify its fit in real observations, while also verifying the geometric and texture continuity near the repair boundary vertex set. Due to occlusion and viewpoint differences in the observed fragment set, step S5 limits all statistics to a stable visible range, uses consistent statistical criteria to form a judgment conclusion, and transforms failed items into valid samples from the empirical distribution library.

[0086] S501 back-check range locking and input sorting.

[0087] Projection re-inspection requires excluding pixels outside the viewport and occluded pixels; otherwise, the re-inspection residual map will be mixed with differences from unobservable areas. The alignment result package reads the projection relationship, and the contribution index table reads the candidate fragment set of patch units, the projection coverage area, and the visible area. For each patch unit, the intersection of the projection coverage area and the visible area is taken on each candidate fragment to form the re-inspection calculation range. The repair model performs depth rendering and texture rendering from the candidate fragment's viewpoint to obtain the re-inspection rendering map. The re-inspection rendering map and the candidate fragment image are aligned pixel-by-pixel within the re-inspection calculation range; the re-inspection calculation does not include pixels outside the re-inspection calculation range. After the range is locked, the re-inspection evidence only reflects the repair quality of the observable area; occlusion changes will not interfere with the judgment.

[0088] S502 projection back inspection residual map generation.

[0089] Individual candidate segments are prone to local noise, and the re-examination residual map needs to incorporate evidence from the candidate segment set of patch units. For each candidate segment, the absolute difference between the observed image pixels and the re-examination rendered image pixels is calculated pixel-by-pixel within the re-examination calculation range, forming a candidate segment pixel residual map. After aligning all candidate segment pixel residual maps to the same pixel coordinates, the patch unit performs median statistics on the same pixel location to generate the re-examination residual map. Median statistics suppress single-frame exposure abrupt changes and local reflections, resulting in more concentrated abnormal areas in the re-examination residual map and more stable extraction of defect candidate areas.

[0090] S503 Defect candidate area location and consistency determination during re-inspection.

[0091] The re-inspection process requires both the location and scale of the anomaly to ensure that the pass / fail conclusion falls within a verifiable range. A quantile threshold is calculated within the valid pixel set of the re-inspection residual map. Pixels in the re-inspection residual map that are not below the quantile threshold form a binary anomaly map, which undergoes connected component extraction to form candidate regions for re-inspection defects. Re-inspection consistency is determined using area proportion, which is obtained by dividing the number of candidate region pixels by the number of valid pixels. The area proportion is compared with an area proportion threshold to obtain the re-inspection fit conclusion. When the area proportion falls within the area proportion threshold, the concentrated area of ​​the re-inspection residual is within an acceptable range, and the evidence of texture fit is sufficient.

[0092] S504 Boundary Continuity Check.

[0093] Geometric polylines and texture breaks are most likely to occur at the repair boundary. Boundary continuity testing requires measuring both geometric boundary continuity and texture boundary alignment separately. Geometric boundary continuity testing uses the boundary edge set corresponding to the boundary vertex set as the scope. Normals are calculated for the triangular facets on both sides of the boundary edge, and the included angle of these normals is used as the boundary fold angle. High-quantile convergence is performed on the fold angles of the boundary edge set to obtain the geometric boundary continuity index. Texture boundary alignment continuity testing uses the boundary band near the model projection boundary as the scope. A pixel distance field is established for each candidate segment, recording the shortest distance from the pixel to the model projection boundary. Image edge pixels are extracted from the candidate segments and restricted to the boundary band and the re-inspection calculation range. Edge direction consistency filtering is then performed on the image edge pixels. After filtering, the absolute distance of the pixel distance field is read and high-quantile convergence is performed to form the texture boundary misalignment amount of the candidate segment. Median statistics are performed on the texture boundary misalignment amount of the candidate segment by the facet unit to obtain the texture boundary alignment continuity index. Both indices are compared with preset conditions. When both fold angle evidence and boundary misalignment evidence simultaneously meet the preset conditions, the repaired boundary remains continuous on both the geometric and texture sides.

[0094] S505 updates the empirical distribution library of generated and failed items by determining the criteria.

[0095] The judgment requires traceable rules; otherwise, updating the experience distribution library will introduce unexplainable samples. The judgment is constrained by three pieces of evidence: the area proportion falls within the area proportion threshold, the geometric boundary continuity index falls within the geometric boundary continuity index threshold, and the texture boundary alignment continuity index falls within the texture boundary alignment continuity index threshold. A passing result is recorded when all three conditions are met. For failing items, a re-examination evidence rule is used to generate labels. When the geometric boundary continuity index exceeds the geometric boundary continuity index threshold, a geometric misalignment defect label is recorded. When the texture boundary alignment continuity index exceeds the texture boundary alignment continuity index threshold but the geometric boundary continuity index falls within the geometric boundary continuity index threshold, a transient texture defect label is recorded. When both types exceed their respective thresholds, a composite label is recorded, and the label information for both types is retained. The experience distribution library update uses the boundary projection misalignment pixel potential and dynamic occlusion coverage ratio output in step S3 as sample features. The sample features are placed within a preset binning interval, the binning count is incremented, and then the binning count is smoothed and normalized to generate the updated evidence distribution. After the update is completed, the defect attribution results for the failed items are recalculated based on the updated experience distribution library, and the defect attribution results form a consistent closure with the re-inspection labels.

[0096] S506 Traceability Record Generation and Output.

[0097] The traceability record needs to cover the spatial range, evidence source, and judgment conclusion. Missing traceability records will prevent the repair model from being audited and reviewed. Traceability records are generated by item, recording the patch unit identifier, the spatial range of the defect candidate area, the spatial range of the re-inspection defect candidate area, the defect attribution result, the list of patch unit candidate fragment sets, the repair action type, the judgment conclusion of the re-inspection residual map, the area ratio conclusion, the geometric boundary continuity index conclusion, the texture boundary alignment continuity index conclusion, and the experience distribution library version identifier. Failed items also generate traceability records and record the tags and the updated binning position of the experience distribution library. The traceability records are output together with the repair model, allowing reviewers to recalculate the re-inspection evidence and locate abnormal areas based on the traceability records. The repair model output has a verifiable path.

[0098] In one implementation, the experience distribution library includes at least: a library version identifier, a sample type label, sample values ​​of boundary projection misalignment pixel potentials and their binning indexes, sample values ​​of dynamic occlusion coverage ratios and their binning indexes, corresponding patch unit identifiers, corresponding candidate fragment set list indexes, and entry time and source. The traceability records are associated with the experience distribution library through the library version identifier and the patch unit identifier, enabling the attribution and repair evidence for the same patch unit to be recalculated and auditable.

[0099] Step S5 limits the projected backtesting of the repair model to the intersection of the projected coverage area and the visible area. The backtesting residual map uses median statistics of multiple candidate segments to suppress occasional noise. The backtesting defect candidate area and area ratio form fitting evidence. The geometric boundary continuity index and the texture boundary alignment continuity index together form boundary quality evidence, and the judgment is based on the rule that the three pieces of evidence are consistently satisfied. Items that fail are labeled based on the backtesting evidence. The experience distribution library is back-flowed by bin count and smoothed and normalized. The defect attribution results are recalculated after the update. The traceability records cover the spatial range, evidence source, and judgment conclusion. The output of the repair model has audit and review capabilities.

[0100] Example 2: Figure 2 This invention discloses a spatial data governance system based on digital twins, comprising: Alignment Package Module: Acquires the source observation dataset and the target 3D model, performs coordinate registration and time alignment, and generates an alignment result package. The alignment result package contains the set of observation segments and projection relationships.

[0101] Patch Index Module: The target 3D model is divided into patch units. Based on depth rendering, the view range determination and occlusion determination are completed to obtain a set of candidate fragments of patch units and establish a contribution index table.

[0102] The residual attribution module: It contributes to the index table to drive back-projection consistency calculation, form a residual map and locate defect candidate areas; it calculates the boundary projection misalignment pixel potential and dynamic occlusion coverage ratio for each defect candidate area, calculates the defect attribution posterior confidence based on the empirical distribution library, and outputs the defect attribution results and candidate fragment set.

[0103] Local Repair Module: Defect attribution results drive the repair path. Geometric misalignment defects trigger local mesh reconstruction and recalculation of local textures. Transient texture defects trigger candidate fragment filtering and recalculation of local textures, generate repair models and update the corresponding records in the contribution index table.

[0104] Backtesting and tracing module: The repair model performs a projection backtest, recalculates the residual map and evaluates the boundary continuity. For failed items, it returns to step S3 and updates the evidence distribution in the empirical distribution library. For successful items, it outputs the repair model and tracing records.

[0105] Specifically, the above description is only a preferred embodiment of this application and is not intended to limit this application.

[0106] The preset thresholds include at least: upper limit for continuous acquisition, upper limit for continuous rotation, consistency limit, depth consistency tolerance, quantile position corresponding to the residual map threshold, directional tolerance for edge direction consistency screening, quantile point used for misalignment potential convergence, dynamic confidence map threshold, lower limit for visible coverage ratio, area ratio threshold, geometric boundary continuity index threshold, and texture boundary alignment continuity index threshold. These thresholds can be determined through offline calibration: on samples with known true values ​​or manual verification, the empirical distribution of the corresponding indexes in the "pass / fail" samples is statistically analyzed, and threshold points that satisfy preset constraints on false alarm rate and false negative rate are selected; alternatively, initial values ​​can be given based on a proportional relationship established between sensor resolution, noise level, and sampling period, and then corrected through rolling correction using running data.

[0107] In an optional implementation, the rotational continuity upper limit can be taken as an angle range consistent with the upper limit of platform attitude jitter; the depth consistency tolerance can be set to the same order of magnitude as the depth quantization error and reprojection error; the residual map threshold quantile position and the misalignment potential quantile position can be taken as mid-to-high quantiles to suppress a small number of abnormal pixels; the dynamic confidence map threshold can be taken as the threshold point that maximizes the distinction between "transient texture-type defects / non-transient" samples in the defect candidate area. The above values ​​are optional examples and do not limit the scope of protection.

[0108] In the description of this specification, references to terms such as "an embodiment," "example," and "specific example" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0109] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A spatial data governance method based on digital twins, characterized in that, Including the following steps: S1: Obtain the source observation dataset and the target 3D model, perform coordinate registration and time alignment, and generate an alignment result package. The alignment result package contains the set of observation segments and projection relationships. S2: The target 3D model is divided into patch units. Based on depth rendering, the view range determination and occlusion determination are completed to obtain a set of candidate fragments of patch units and establish a contribution index table. S3: Contribution index table drives back projection consistency calculation, forms residual map and locates defect candidate area; calculates boundary projection misalignment pixel potential and dynamic occlusion coverage ratio for each defect candidate area, calculates defect attribution posterior confidence based on empirical distribution library, and outputs defect attribution results and candidate segment set. S4: Defect attribution results drive the repair path. Geometric misalignment defects trigger local mesh reconstruction and recalculation of local textures. Transient texture defects trigger candidate fragment screening and recalculation of local textures, generate repair models and update the corresponding records in the contribution index table. S5: Perform projection backtesting on the repaired model, recalculate the residual map and evaluate boundary continuity. For failed items, return to step S3 and update the evidence distribution in the empirical distribution library. For successful items, output the repaired model and retrospective records.

2. The spatial data governance method based on digital twins according to claim 1, characterized in that, Step S1 includes: The source observation dataset is divided into observation segment sets according to the continuity of acquisition. Feature points are extracted from the target 3D model and the source observation dataset to establish a matching pair set, and the initial pose solution set is obtained through consistency verification. The initial pose solution set is obtained through point-to-plane consistency iteration. Event anchor point sequences are extracted from the observation segment sets and matched with the projectible anchor point sequences of the target 3D model to form a temporal mapping relationship. The observation segment set, pose solution set, temporal mapping relationship, and projection relationship are encapsulated into an alignment result package. Among them, the event anchor point sequence refers to the key moment sequence representing the appearance or structural changes of the same spatial region extracted according to the acquisition time sequence of the observation segments; the projectible anchor point sequence refers to the sequence of geometric feature points in the target 3D model that can be projected onto the image plane and stably identified under multiple views, formed according to the model version time label; the temporal mapping relationship is obtained by order-preserving matching. Order-preserving matching means that the matching relationship between the event anchor sequence and the projectible anchor sequence maintains the consistency of the temporal order.

3. The spatial data governance method based on digital twins according to claim 2, characterized in that, Step S2 includes: The target 3D model is divided into facet units based on the topological relationship of triangular facets. The facet units satisfy the continuous constraint of the difference between the normal angle and the curvature indicator. A set of surface sampling points is generated on the surface of the facet unit. The projection relationship maps the set of surface sampling points to the image plane of the set of observation segments. Pixel rasterization forms the projection coverage area.

4. The spatial data governance method based on digital twins according to claim 3, characterized in that, Step S2 also includes: The set of observed fragments performs depth rendering within the effective area of ​​the viewport to generate the foreground surface depth. The patch unit calculates the surface depth of the patch unit pixel by pixel in the projection coverage area and compares it with the foreground surface depth to obtain the visible area. The projection coverage area and the visible area are statistically analyzed to form the visible coverage ratio. The set of candidate fragments of the patch unit is obtained by filtering the visible coverage ratio. The contribution index table records the set of candidate fragments of the patch unit, the projection coverage area, and the visible area.

5. A spatial data governance method based on digital twins according to claim 4, characterized in that, Step S3 includes: The contribution index table reads the candidate fragment set of the patch unit, the projection coverage area, and the visible area. The projection relationship extracts the candidate fragment texture fragment within the visible area. The candidate fragment texture fragment is mapped to the patch parameter domain to form an aligned texture. The residual map is generated by the median statistics of the pairwise differences of the aligned texture. The defect candidate area is determined by the preset thresholded connected region.

6. A spatial data governance method based on digital twins according to claim 5, characterized in that, Step S3 also includes: The boundary projection misalignment pixel potential is obtained by establishing a pixel distance field within the boundary band and combining it with edge direction consistency screening. The dynamic occlusion coverage ratio is obtained by calculating the optical flow return offset and appearance mutation within the intersection of the projection coverage area and the visible area and generating a dynamic confidence map. The empirical distribution library queries the conditional probability density of two types of defects and combines it with the prior probability to form the defect attribution posterior confidence. The defect attribution result is determined based on the defect attribution posterior confidence. Among them, the optical flow return offset refers to the positional offset length generated when a pixel is mapped back to its original position by forward optical flow and then by backward optical flow, which is used to characterize the consistency of pixel motion estimation. The boundary projection misalignment pixel potential is a statistical quantity used to characterize the degree of misalignment between the model projection boundary and the real edge of the image after statistically aggregating the distances from the image edge pixels after edge direction consistency screening to the model projection boundary within the boundary band corresponding to the model projection boundary.

7. A spatial data governance method based on digital twins according to claim 6, characterized in that, Step S4 includes: The defect candidate region is back-mapped onto the surface of the target 3D model within the intersection of the projected coverage area and the visible area, forming a set of local triangular facets and expanding to obtain a set of local vertices and a set of boundary vertices. The candidate fragment set calculates the texture consistency deviation and transient contamination intensity within the defect candidate region. The defect attribution result controls the screening criteria for the repair evidence set. Among them, the transient contamination intensity refers to the statistical quantity obtained by statistically analyzing the corresponding pixel values ​​of the dynamic confidence map within the defect candidate region, which is used to characterize the degree of influence of occlusion, reflection or other short-term changes on texture stability.

8. A spatial data governance method based on digital twins according to claim 7, characterized in that, Step S4 also includes: When the defect attribution result is labeled as a geometric misalignment defect, local vertex set is reconstructed locally while boundary vertex set remains in its original position. When the defect attribution result is labeled as a transient texture defect, local vertex set remains unchanged while candidate fragments are filtered out. Local texture recalculation uses fragment assignment and seam regularization to generate a repair model, updates the candidate fragment set of patch elements in the contribution index table, and recalculates the projected coverage area and visible area during geometric changes. Among these, fragment assignment and seam regularization refer to determining the texture source of each candidate fragment position during local texture recalculation and performing smooth transition processing on the boundary areas of different texture sources to reduce discontinuities at fragment switching boundaries.

9. A spatial data governance method based on digital twins according to claim 8, characterized in that, Step S5 includes: The alignment result package provides projection relationships, and the contribution index table provides a set of candidate fragments for patch units, projection coverage areas, and visible areas. The repair model generates a back-inspection rendering map within the intersection of the projection coverage area and the visible area and calculates a back-inspection residual map with the candidate fragment images. The back-inspection defect candidate area and area ratio complete the fitting judgment. The geometric boundary continuity index and texture boundary alignment continuity index are calculated at the boundary vertex set to complete the boundary judgment. For items that fail, labels are generated based on the back-inspection evidence and the experience distribution library is updated. For items that pass, traceability records are output.

10. A spatial data governance system based on digital twins, used to implement the spatial data governance method based on digital twins as described in any one of claims 1-9, characterized in that, include: Alignment Package Module: Acquires the source observation dataset and the target 3D model, performs coordinate registration and time alignment, and generates an alignment result package. The alignment result package contains the set of observation segments and projection relationships. Patch Index Module: The target 3D model is divided into patch units. Based on depth rendering, view range determination and occlusion determination are completed to obtain a set of candidate fragments of patch units and establish a contribution index table; Residual attribution module: The contribution index table drives back-projection consistency calculation, forms a residual map, and locates candidate defect regions; For each defect candidate region, calculate the boundary projection misalignment pixel potential and dynamic occlusion coverage ratio, calculate the defect attribution posterior confidence based on the empirical distribution library, and output the defect attribution results and candidate fragment set. Local Repair Module: Defect attribution results drive the repair path. Geometric misalignment defects trigger local mesh reconstruction and recalculation of local textures. Transient texture defects trigger candidate fragment filtering and recalculation of local textures, generate repair models and update the corresponding records in the contribution index table. Backtesting and tracing module: The repair model performs a projection backtest, recalculates the residual map and evaluates the boundary continuity. For failed items, it returns to step S3 and updates the evidence distribution in the empirical distribution library. For successful items, it outputs the repair model and tracing record.

Images