Real-scene three-dimensional dynamic change modeling method based on digital twinning

Abstract

The invention discloses a live-action three-dimensional dynamic change modeling method based on digital twinning. The method comprises the following steps: collecting multi-source data; carrying out space registration and time synchronization processing, and outputting a fusion data set; carrying out three-dimensional modeling, optimizing model parameters, and generating a live-action three-dimensional model; virtual-real mapping is carried out, and generated virtual space data and state parameters thereof are stored; performing differential analysis to generate a change detection result data set; performing dynamic change modeling to generate an updated virtual model; executing geometric accuracy optimization and semantic consistency check, and outputting an optimized virtual model; bidirectional data flow association and dynamic updating are executed, and the live-action three-dimensional dynamic change modeling process is completed. The real scene dynamic updating is realized by adopting the improved Gaussian sputtering modeling, and the method has the advantages of high precision, strong real-time performance and virtual-real synchronization.

Description

Technical Field

[0001] This invention relates to the field of real-scene 3D modeling technology, and in particular to a real-scene 3D dynamic change modeling method based on digital twins. Background Technology

[0002] Currently, real-scene 3D modeling technology has been widely used in fields such as urban modeling, engineering monitoring, and virtual simulation. Common methods include geometric reconstruction based on laser point clouds, texture mapping based on images, and semantic recognition based on deep learning. While these methods can generate high-precision static 3D models, they generally lack the ability to manage the temporal sequence of multi-source heterogeneous data, making it difficult to dynamically update the model when scene structure, lighting conditions, or object attributes change. In addition, existing digital twin systems mostly rely on manual triggering or periodic refresh methods to maintain the virtual model, resulting in delayed updates and high costs.

[0003] Current technologies lack a comprehensive method capable of automatically sensing changes in physical scenes, updating modeling parameters in real time, and maintaining semantic consistency between virtual and real models within a digital twin framework. Particularly in multi-temporal point cloud fusion, dynamic change detection, and bidirectional feedback between the virtual and real worlds, a complete algorithmic system has not yet been established. This results in asynchrony between the virtual model and the real physical environment in spatial, semantic, and temporal dimensions, limiting the application effectiveness and intelligence level of digital twin technology in complex dynamic scenarios.

[0004] Therefore, how to provide a method for real-world 3D dynamic change modeling based on digital twins is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a real-scene 3D dynamic change modeling method based on digital twins. This invention uses improved Gaussian sputtering modeling to achieve dynamic updates of the real scene, which has the advantages of high precision, strong real-time performance and virtual-real synchronization.

[0006] A method for real-scene 3D dynamic change modeling based on digital twin according to an embodiment of the present invention includes the following steps:

[0007] Multi-source data acquisition is performed on the target physical scene to obtain multi-source data including spatial coordinates, texture features and timestamps;

[0008] Perform spatial registration and temporal synchronization processing on multi-source data, and output a fused dataset after registration and synchronization processing;

[0009] Based on the fused dataset, an improved Gaussian sputtering neural field model is used to perform 3D modeling of the scene, and the model parameters are optimized to generate a realistic 3D model.

[0010] A virtual-real mapping is performed based on a digital twin mapping framework, and the generated virtual space data and its state parameters are stored.

[0011] By calling virtual space data and fused datasets, differential analysis is performed to identify newly added, deleted, and deformed regions, and a change detection result dataset is generated.

[0012] Dynamic change modeling and update mechanism are implemented to incrementally optimize local Gaussian distribution parameters and generate updated virtual model;

[0013] The updated virtual model is invoked, geometric accuracy optimization and semantic consistency verification are performed, the geometric error of the virtual model is adjusted and the semantic labels are corrected, and the optimized virtual model is output.

[0014] The optimized virtual model is rendered and dynamically displayed in real time, and bidirectional data flow is associated and dynamically updated to complete the real-scene 3D dynamic change construction process in the digital twin system.

[0015] Optionally, the generation of the fused dataset includes:

[0016] Spatial registration is performed on multi-source data to extract three-dimensional feature points. Based on the iterative nearest point registration algorithm of feature point matching, the multi-source data is aligned to a unified coordinate reference system. A spatial transformation matrix composed of rotation matrix and translation vector is established, and the multi-source data after preliminary registration is output.

[0017] Perform geometric optimization processing on the initially registered multi-source data to output geometrically optimized multi-source data;

[0018] The geometrically optimized multi-source data is time-synchronized to establish a unified time reference. Offset correction and time interpolation compensation are performed on each sampling time to output the time-synchronized multi-source data.

[0019] Multi-source data that has undergone spatial registration and time synchronization are uniformly encoded to generate a fused dataset. Outliers and noisy data are statistically removed and spatially interpolated to complete the dataset, which is then output as a fused dataset after spatial registration and time synchronization.

[0020] Optionally, the generation of the real-scene 3D model includes:

[0021] Using the fused dataset as input data, spatial coordinates, texture color values, illumination reflection parameters and time index information are extracted from the fused dataset to construct a three-dimensional spatial point set corresponding to the target physical scene;

[0022] Establish a Gaussian distribution parameter set for each point in the three-dimensional spatial point set;

[0023] Input the Gaussian distribution parameter set into the improved Gaussian sputtering neural field model to generate preliminary rendering results;

[0024] Using the preliminary rendering results and the original observations of the corresponding target physical scene in the fused dataset as reference inputs, pixel-by-pixel difference calculations are performed and a rendering error function is constructed.

[0025] An iterative optimization algorithm based on error gradient descent is adopted, with the rendering error function as the optimization objective. The Gaussian distribution parameter set is updated step by step. When the descent rate of the rendering error function is lower than the preset threshold or the error change is less than the stable value in multiple consecutive iterations, the iteration is stopped and the optimized Gaussian distribution parameter set is output.

[0026] The optimized Gaussian distribution parameter set is re-input into the improved Gaussian sputtering neural field model, and ray integral rendering is performed to generate a realistic 3D model corresponding to the target physical scene.

[0027] Optionally, the generation of the virtual space data and its state parameters includes:

[0028] Based on the real-world 3D model, it is instantiated as a 3D spatial object in the digital twin mapping framework. A unique identifier number is assigned to each spatial object, and a coordinate index table corresponding to the target physical scene is established to form a set of 3D spatial objects.

[0029] Taking a set of three-dimensional spatial objects as input, the spatial coordinates, geometric boundaries and texture features of each spatial object are extracted to generate a set of spatial feature descriptions. Based on this set, a spatial index structure is constructed to realize the geometric mapping alignment between the virtual space and the target physical scene, and output spatial index mapping data.

[0030] Semantic annotation and semantic synchronization are performed based on spatial index mapping data. Based on the semantic synchronization mechanism, the consistency of semantic labels under different time indices is maintained, and a semantic synchronization result set is output.

[0031] Using the semantic synchronization result set as input, the real-time physical attributes of the target physical scene are extracted from the fused dataset, and an attribute binding result set is generated by using an attribute binding algorithm based on the feature mapping matrix.

[0032] A virtual-real mapping data structure is constructed based on the attribute binding result set, the spatial feature description set, and the semantic synchronization result set, and the virtual-real mapping structured data is output.

[0033] A set of state parameters is generated for each virtual space object, forming a traceable temporal state parameter set;

[0034] The virtual-real mapping data structure and the set of state parameters are stored as a virtual space data file based on the time-series state parameter set.

[0035] Optionally, the generation of the change detection result dataset includes:

[0036] By calling virtual space data and fused datasets collected at different times, and based on the spatial index table and semantic label table in the virtual-real mapping structured data, spatial registration and temporal index matching are performed on the multi-temporal data to generate a multi-temporal fused data sequence.

[0037] Based on the multi-temporal fusion data sequence, spatial coordinates, reflection intensity, surface normal vector and semantic label features of point cloud data under each time index are extracted, a temporal feature set is established, and the correspondence of each feature in virtual space data is recorded by a unified node identifier.

[0038] Based on the time series feature set, the time series change detection mechanism is invoked to perform joint difference analysis on multi-source feature data under adjacent time indices and output a time series difference result set;

[0039] The multi-temporal point cloud comparison algorithm is invoked to perform spatial overlap calculation, point density difference analysis and semantic consistency verification on the corresponding node data of the virtual space data and the target physical scene. By fusing the weighted results of geometric distance difference, semantic consistency difference and attribute vector difference, the changed region is identified and the changed region result set is output.

[0040] Perform change type classification and spatial clustering processing, and update the time index and state parameters of the corresponding nodes in the virtual-real mapping structured data to generate a change detection result dataset.

[0041] Optionally, the generation of the updated virtual model includes:

[0042] The change region information in the change detection result dataset is analyzed, and the spatial location, semantic labels, and attribute parameters of the corresponding virtual model nodes are extracted.

[0043] The parameter update mechanism of the improved Gaussian sputtering neural field model is invoked based on the parameter set of the changed region to determine the set of local Gaussian distribution parameters that need to be updated, and the corresponding Gaussian distribution nodes are located in the virtual model according to the spatial boundary of the changed region to generate the parameter set to be optimized.

[0044] Based on the set of parameters to be optimized, an incremental optimization process is performed on the improved Gaussian sputtering neural field model to generate the optimized Gaussian distribution parameter set.

[0045] The optimized Gaussian distribution parameter set is written into the corresponding local Gaussian nodes of the virtual model to generate the updated virtual model.

[0046] Optionally, the generation of the optimized virtual model includes:

[0047] The updated virtual model is invoked, and the spatial coordinates, surface normal vectors, and semantic label information of each node in the virtual model are analyzed to establish the correspondence between the spatial features and semantic features of the nodes, forming a model feature index set.

[0048] Using the model feature index set as input, geometric accuracy optimization is performed. Local error analysis is conducted on the spatial geometry of the virtual model, the node geometric offset is calculated, and the geometric change corresponding to the change detection result is used as a constraint. The node spatial position and surface normal vector are adjusted by iteratively minimizing the geometric error function, and the geometrically optimized virtual model is output.

[0049] Perform a semantic consistency check process based on the geometrically optimized virtual model, and output a semantically corrected virtual model;

[0050] The semantically corrected virtual model is used as the result of the fusion of geometric optimization and semantic correction to generate an optimized virtual model.

[0051] Optionally, the real-scene 3D dynamic change construction includes:

[0052] The optimized virtual model generated based on the digital twin data framework is loaded into the rendering module, and the spatial structure, semantic labels and attribute parameters of the virtual model are loaded into the rendering module to generate rendering input data.

[0053] Perform real-time rendering, output the rendering results of the virtual model, and synchronously receive the perception feedback information of the target physical scene to generate virtual-real synchronization data containing time index and spatial coordinates.

[0054] Based on the virtual-real synchronous data, a two-way data flow association is established between the virtual model and the target physical scene. The rendering results and perception feedback information are matched in time and space to generate virtual-real difference data.

[0055] Perform adaptive adjustment of virtual-real difference based on virtual-real difference data, and output the virtual-real difference correction result;

[0056] Update the spatial structure, semantic labels, and attribute parameters of the virtual model to generate an updated virtual model;

[0057] The updated virtual model is written into the digital twin data framework and real-time rendering is re-executed. The virtual-real consistency index is calculated based on the virtual-real difference correction results, thus completing the dynamic change construction process of the real-world 3D scene under the digital twin data framework.

[0058] The beneficial effects of this invention are:

[0059] This invention proposes a digital twin-based method for dynamic change modeling of real-world 3D models. Addressing the problems of static 3D models, delayed updates, semantic inconsistencies, and asynchrony between virtual and real models in existing technologies, this invention constructs a complete dynamic modeling closed loop from data acquisition and model optimization to virtual-real feedback. By introducing an improved Gaussian sputtering neural field model, it achieves a joint representation of the physical scene across multiple dimensions of space, time, and illumination. It can adaptively adjust the Gaussian distribution parameters based on the temporal changes of multi-source data, enabling the real-world 3D model to possess temporal continuity and dynamic response capabilities. Compared to traditional static modeling methods, this invention can simultaneously process the temporal characteristics and physical properties of the scene during the modeling stage, providing structured support for subsequent dynamic change detection and virtual model updates.

[0060] This invention establishes a digital twin mapping framework, achieving a one-to-one correspondence between real-world spatial data and virtual spatial data in terms of geometric structure, semantic labels, and physical attributes. Utilizing attribute binding algorithms and semantic synchronization mechanisms, the virtual model can automatically associate semantic and attribute changes in the physical scene, thus maintaining consistency and integrity between virtual and real data in dynamic scenarios. Combining a temporal change detection mechanism and a multi-temporal point cloud comparison algorithm, this method can identify newly added, deleted, and deformed regions among multi-source data at different times, constructing a change detection result dataset. Based on Gaussian parameter temporal interpolation, local redistribution adjustment, and covariance adaptive correction, incremental updates of the virtual model are achieved, continuously optimizing the model in terms of spatial accuracy and semantic consistency.

[0061] Through a bidirectional data flow association mechanism and an adaptive adjustment strategy for virtual-real differences, this invention further realizes real-time interaction and feedback correction between the virtual model and the target physical scene. When differences arise between the virtual model and the real scene in spatial location, color brightness, and semantic labels, the system can automatically calculate the difference value and perform weighted correction to ensure that the virtual and real models are synchronized and consistent at the rendering, display, and perception levels. Ultimately, the digital twin model generated based on this method can continuously self-learn, self-update, and adaptively adjust, realizing the transformation from static mapping to dynamic intelligent evolution. Compared with existing technologies, this invention significantly improves the real-time performance, accuracy, and semantic consistency of digital twin models, enabling them to operate stably in complex dynamic environments and providing efficient and reliable virtual-real fusion support for fields such as smart cities, intelligent manufacturing, and digital factories. Attached Figure Description

[0062] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0063] Figure 1 This is a schematic diagram of the overall process of a real-scene 3D dynamic change modeling method based on digital twin proposed in this invention;

[0064] Figure 2 This is a flowchart of the improved Gaussian sputtering neural field modeling and rendering error minimization optimization method based on digital twin for real-scene 3D dynamic change modeling proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figure 1 and Figure 2 A method for real-world 3D dynamic change modeling based on digital twins includes the following steps:

[0067] Multi-source data acquisition is performed on the target physical scene to obtain multi-source data including spatial coordinates, texture features and timestamps;

[0068] Perform spatial registration and temporal synchronization processing on multi-source data, and output a fused dataset after registration and synchronization processing;

[0069] Based on the fused dataset, an improved Gaussian sputtering neural field model is used to perform 3D modeling of the scene, and the model parameters are optimized to generate a realistic 3D model.

[0070] A virtual-real mapping is performed based on a digital twin mapping framework, and the generated virtual space data and its state parameters are stored.

[0071] By calling virtual space data and fused datasets, differential analysis is performed to identify newly added, deleted, and deformed regions, and a change detection result dataset is generated.

[0072] Dynamic change modeling and update mechanism are implemented to incrementally optimize local Gaussian distribution parameters and generate updated virtual model;

[0073] The updated virtual model is invoked, geometric accuracy optimization and semantic consistency verification are performed, the geometric error of the virtual model is adjusted and the semantic labels are corrected, and the optimized virtual model is output.

[0074] The optimized virtual model is rendered and dynamically displayed in real time, and bidirectional data flow is associated and dynamically updated to complete the real-scene 3D dynamic change construction process in the digital twin system.

[0075] In this embodiment, the generation of the fused dataset includes:

[0076] Spatial registration processing is performed on multi-source data. Based on the spatial positioning information and attitude parameters of each data source, three-dimensional feature points are extracted. The iterative nearest point registration algorithm based on feature point matching aligns the multi-source data to a unified coordinate reference system. A spatial transformation matrix composed of rotation matrix and translation vector is established to describe the spatial position and orientation relationship of each data source in the unified coordinate reference system. The preliminarily registered multi-source data is then output.

[0077] After preliminary registration, the multi-source data is subjected to geometric optimization processing. The global geometric error is corrected by minimizing the sum of squared registration residuals, and the spatial deviation is constrained by the local weighted smoothing function. The geometrically optimized multi-source data is then output.

[0078] The geometrically optimized multi-source data is time-synchronized to establish a unified time reference. Offset correction and time interpolation compensation are performed on each sampling time so that the multi-source data corresponds to the same physical time under the unified time reference, and the time-synchronized multi-source data is output.

[0079] Multi-source data that has undergone spatial registration and time synchronization are uniformly encoded to generate a fused dataset. The fused dataset includes spatial coordinates, texture color values, illumination reflection parameters, and time index information. Outliers and noise data are statistically removed and spatially interpolated to complete the data. The fused dataset after spatial registration and time synchronization is then output.

[0080] In this embodiment, the generation of the real-scene 3D model includes:

[0081] Using the fused dataset as input data, spatial coordinates, texture color values, illumination reflection parameters and time index information are extracted from the fused dataset to construct a three-dimensional spatial point set corresponding to the target physical scene;

[0082] A Gaussian distribution parameter set is established for each spatial point in the three-dimensional spatial point set. The Gaussian distribution parameter set includes spatial location, color features, opacity and shape covariance parameters, which serve as inputs to the improved Gaussian sputtering neural field model.

[0083] The Gaussian distribution parameter set is input into the improved Gaussian sputtering neural field model to generate preliminary rendering results. The improved Gaussian sputtering neural field model adds a time-aware mechanism, a density adaptive mechanism, and an illumination consistency constraint mechanism to the traditional Gaussian sputtering neural field model. These mechanisms are used to represent the three-dimensional spatial structure, characterize the time-varying features, spatial density distribution changes, and illumination reflection changes.

[0084] The time-aware mechanism establishes a multi-temporal Gaussian distribution sequence by fusing time index parameters in the dataset. It serializes and associates spatial coordinates, color features and brightness information collected at different times, thereby achieving a continuous representation of the target physical scene in the time dimension and keeping the parameter changes of each Gaussian point consistent across multiple temporal phases.

[0085] The density adaptive mechanism is executed on the multi-temporal Gaussian distribution sequence generated by the time-aware mechanism. Specifically, it includes four steps: neighborhood density calculation, brightness gradient calculation, dynamic correction coefficient calculation, and covariance smoothing update. Neighborhood density calculation is used to determine the local density of each spatial point; brightness gradient calculation is used to evaluate the amplitude of illumination change; dynamic correction coefficient is determined jointly based on the normalized value of neighborhood density and the rate of change of brightness gradient; covariance smoothing update achieves adaptive adjustment of shape covariance by applying the dynamic correction coefficient to the main diagonal elements of the covariance matrix in the Gaussian distribution parameter set, thereby reducing the distribution range in dense point cloud regions and expanding the distribution range in sparse regions, and performing smoothing update under time index to ensure temporal continuity.

[0086] The illumination consistency constraint mechanism takes the multi-temporal Gaussian distribution sequence output by the time-aware mechanism as input and adopts a color constraint algorithm based on reflection intensity deviation compensation. It performs weighted correction on the color features under multi-temporal and multi-view sampling, and adjusts the reflection brightness weight according to the dynamic correction coefficient calculated by the density adaptive mechanism, so that the rendering output maintains color consistency and physical rationality under illumination change conditions, and outputs the preliminary rendering result.

[0087] Using the preliminary rendering results and the original observations of the corresponding target physical scene in the fused dataset as reference inputs, pixel-by-pixel difference calculations are performed and a rendering error function is constructed. The rendering error function specifically includes: comparing the color values, brightness values, and opacity values ​​in the preliminary rendering results with the original observations pixel by pixel, calculating the squares of the three types of differences and summing them with weights, forming a rendering error function with the weighted sum of squares as the core indicator, which is used to measure the deviation between the improved Gaussian sputtering neural field model and the real observations in terms of color reproduction, brightness consistency, and transparency matching.

[0088] An iterative optimization algorithm based on error gradient descent is adopted, with the rendering error function as the optimization objective. The Gaussian distribution parameter set is updated step by step. Specifically, it includes: calculating the direction of change of the partial derivative of each parameter in the rendering error function with respect to the error, adjusting the parameter values ​​according to the direction of the error gradient, so that the rendering error is gradually reduced, and making the rendering error function converge along the optimal direction in the multidimensional parameter space during the iterative optimization process. When the descent rate of the rendering error function is lower than a preset threshold or the error change is less than the stable value in multiple consecutive iterations, the iteration is stopped, and the optimized Gaussian distribution parameter set is output.

[0089] The optimized Gaussian distribution parameter set is re-input into the improved Gaussian sputtering neural field model, and ray integral rendering is performed to generate a realistic 3D model corresponding to the target physical scene. The realistic 3D model maintains a dynamic correspondence with the target physical scene in terms of spatial structure, texture color and lighting characteristics.

[0090] In this embodiment, the generation of the virtual space data and its state parameters includes:

[0091] Based on the real-world 3D model, it is instantiated as a 3D spatial object in the digital twin mapping framework. A unique identifier number is assigned to each spatial object, and a coordinate index table corresponding to the target physical scene is established to form a set of 3D spatial objects.

[0092] Taking a set of three-dimensional spatial objects as input, the spatial coordinates, geometric boundaries and texture features of each spatial object are extracted to generate a set of spatial feature descriptions. Based on this set, a spatial index structure is constructed to realize the geometric mapping alignment between the virtual space and the target physical scene, and output spatial index mapping data.

[0093] Semantic annotation and semantic synchronization processing are performed based on spatial index mapping data. Semantic annotation uses scene recognition algorithms based on geometric and texture features to perform semantic classification and matching of key entities in the target physical scene with three-dimensional spatial objects. Based on the semantic synchronization mechanism, the consistency of semantic labels under different time indices is maintained, and a semantic synchronization result set is output.

[0094] The semantic synchronization mechanism includes: establishing a triplet structure of "spatial node identifier - semantic category - temporal index" to perform temporal comparison, conflict detection and consistency correction on the semantic labels in the multi-temporal fusion dataset, so that the semantic labels remain continuous and consistent in the time series, and avoid semantic drift or duplicate classification problems in the process of multi-temporal data fusion.

[0095] Using the semantic synchronization result set as input, real-time physical attributes of the target physical scene are extracted from the fused dataset, including geometric state parameters, material reflection coefficient, illumination intensity and environmental attribute information. An attribute binding algorithm based on feature mapping matrix is ​​used to generate an attribute binding result set.

[0096] The generation of the attribute binding result set includes: extracting and normalizing features from semantic tags, spatial coordinates, and geometric state parameters of virtual space objects and real-time observed attributes from the target physical scene to generate a set of feature vectors; constructing a feature mapping matrix with the set of virtual space objects and the set of physical scene objects as input; each element of the feature mapping matrix is ​​used to characterize the degree of matching between virtual objects and physical objects in terms of spatial location, semantic category, and physical attribute dimensions; calculating the dynamic weight coefficient of each attribute based on the attribute change rate, observation confidence, and historical stability; weighting and fusing the feature mapping matrix and the dynamic weight coefficient; calculating the weighted matching degree of multiple attribute dimensions with the similarity value as the base factor and the weight coefficient as the adjustment factor; and normalizing the result by the sum of the weight coefficients to obtain the comprehensive attribute binding value; and performing a weighted smooth update of the binding results between the previous time and the current time in the time-series dimension to maintain the continuity and stability of attribute changes; and outputting an attribute binding result set containing virtual object identifier, physical object identifier, comprehensive attribute binding value, dynamic weight coefficient, and time index.

[0097] A virtual-real mapping data structure is constructed based on the attribute binding result set, the spatial feature description set, and the semantic synchronization result set. The virtual-real mapping data structure consists of a spatial index table, a semantic label table, and an attribute parameter table. The spatial index table is generated based on the spatial feature description set, and the semantic label table and the attribute parameter table use a unified node identifier as the primary key to realize the fusion of multi-dimensional information of spatial geometry, semantic category, and physical attributes, and output virtual-real mapping structured data.

[0098] The generation of the virtual-real mapping structured data includes three parts: a spatial index table, a semantic tag table, and an attribute parameter table. Using a unified node identifier as the primary key, the virtual space object is associated with the spatial location, semantic category, and physical attributes of the target physical scene. The spatial index table records the three-dimensional coordinates and geometric boundaries, the semantic tag table records the object category and temporal index, and the attribute parameter table records the geometric state, material, lighting, and environmental attributes. The three tables correspond one-to-one through the node identifier to form a dynamically updatable data structure. When the target physical scene changes, the system automatically updates the semantic tags and attribute parameters corresponding to the nodes based on the time index, so as to realize the synchronous adjustment and consistent maintenance of the virtual space data.

[0099] A set of state parameters is generated for each virtual space object. The set of state parameters includes spatial location parameters, semantic category parameters, lighting and material parameters, and dynamic change parameters. The parameters are continuously updated according to the time index to form a traceable temporal state parameter set.

[0100] The virtual-real mapping data structure and the set of state parameters are stored as a virtual space data file based on the time-series state parameter set.

[0101] In this embodiment, the generation of the change detection result dataset includes:

[0102] By calling virtual space data and fused datasets collected at different times, and based on the spatial index table and semantic label table in the virtual-real mapping structured data, spatial registration and temporal index matching are performed on the multi-temporal data to generate a multi-temporal fused data sequence.

[0103] Based on the multi-temporal fusion data sequence, spatial coordinates, reflection intensity, surface normal vector and semantic label features of point cloud data under each time index are extracted, a temporal feature set is established, and the correspondence of each feature in virtual space data is recorded by a unified node identifier.

[0104] Based on the time series feature set, the time series change detection mechanism is invoked to perform joint difference analysis on multi-source feature data under adjacent time indices and output a time series difference result set;

[0105] The generation of the temporal difference result set includes: matching nodes under adjacent time indices in the temporal feature set, establishing node correspondence based on unified node identifiers, calculating the changes of nodes in spatial, semantic, and attribute dimensions to obtain geometric difference components, semantic difference components, and attribute difference components, assigning temporal sensitivity coefficients to each difference component, multiplying each difference component by its corresponding temporal sensitivity coefficient and summing the results, and then normalizing by dividing by the sum of all temporal sensitivity coefficients to obtain a comprehensive change index. The comprehensive change index is used to characterize the overall changes of nodes in the three dimensions of spatial structure, semantic category, and physical attributes, and constitutes the main content of the temporal difference result set.

[0106] The multi-temporal point cloud comparison algorithm is invoked to perform spatial overlap calculation, point density difference analysis and semantic consistency verification on the corresponding node data of the virtual space data and the target physical scene. By fusing the weighted results of geometric distance difference, semantic consistency difference and attribute vector difference, the changed region is identified and the changed region result set is output.

[0107] The system performs change type classification and spatial clustering processing, dividing the changed regions into newly added regions, deleted regions, and deformed regions based on the change magnitude and semantic category. It also updates the time index and state parameters of the corresponding nodes in the virtual-real mapping structured data, generating a change detection result dataset. The change detection result dataset includes the spatial location of the changed region, change category, change magnitude, semantic change label, and corresponding time index.

[0108] In this embodiment, the generation of the updated virtual model includes:

[0109] The change region information in the change detection result dataset is analyzed, and the spatial location, semantic labels, and attribute parameters of the corresponding virtual model nodes are extracted.

[0110] The parameter update mechanism of the improved Gaussian sputtering neural field model is invoked based on the parameter set of the changed region to determine the set of local Gaussian distribution parameters that need to be updated, and the corresponding Gaussian distribution nodes are located in the virtual model according to the spatial boundary of the changed region to generate the parameter set to be optimized.

[0111] Based on the set of parameters to be optimized, an incremental optimization process is performed on the improved Gaussian sputtering neural field model to generate the optimized Gaussian distribution parameter set.

[0112] The incremental optimization includes: performing time interpolation updates in the time dimension, calculating the time interpolation position of the Gaussian distribution center based on the spatial position changes of nodes under adjacent time indices and the comprehensive change index, and outputting a time interpolation parameter set; performing local redistribution adjustments in the spatial dimension, calculating density adjustment coefficients based on the spatial density difference of newly added or deleted regions in the change detection results, dynamically adding or deleting and correcting the shape of local Gaussian distribution nodes, and outputting a local redistribution parameter set; and performing adaptive correction of the covariance matrix in the feature dimension, using the local redistribution parameter set as input, calculating covariance adjustment coefficients based on the uncertainty and attribute fluctuation level of nodes, dynamically adjusting the main diagonal elements of the covariance matrix, and optimizing the set by minimizing the negative log-likelihood of the rendering error function, and outputting the optimized Gaussian distribution parameter set.

[0113] The optimized Gaussian distribution parameter set is written into the local Gaussian node corresponding to the virtual model to generate an updated virtual model. The updated virtual model maintains dynamic consistency with the target physical scene in terms of spatial structure, semantic labels, and physical attributes.

[0114] In this embodiment, the generation of the optimized virtual model includes:

[0115] The updated virtual model is invoked, and the spatial coordinates, surface normal vectors, and semantic label information of each node in the virtual model are analyzed to establish the correspondence between the spatial features and semantic features of the nodes, forming a model feature index set.

[0116] Using the model feature index set as input, geometric accuracy optimization is performed. Local error analysis is conducted on the spatial geometry of the virtual model, the node geometric offset is calculated, and the geometric change corresponding to the change detection result is used as a constraint. The node spatial position and surface normal vector are adjusted by iteratively minimizing the geometric error function, and the geometrically optimized virtual model is output.

[0117] Perform a semantic consistency check process based on the geometrically optimized virtual model, and output a semantically corrected virtual model;

[0118] The generation of the virtual model includes: taking the geometrically optimized virtual model as input, extracting the semantic labels, semantic confidence scores, and spatial coordinate information of each node; establishing a semantic association graph between nodes based on the semantic topological relationships within the model; calculating the semantic similarity between each node and its spatial neighboring nodes, wherein the semantic similarity is determined by a weighted average of semantic label consistency and spatial proximity; calculating the semantic confidence difference of each node as the absolute difference between the semantic confidence score of that node and the average confidence score of its neighbors, and weighting the semantic confidence difference and semantic similarity proportionally to obtain a node semantic consistency score; and calculating the semantic consistency score based on the overall semantic score of the model. The variance of the data and the fluctuation range of the confidence level are used to calculate a dynamic threshold. When the semantic consistency score of a node is lower than this dynamic threshold, the node is determined to be semantically inconsistent. For semantically inconsistent nodes, the node with the highest semantic similarity and the highest confidence level is selected in its spatial neighborhood. The semantic label of the node with the highest confidence level is replaced with the current node's label, and the confidence level is updated to a weighted average. To maintain the spatial smoothness of the semantic distribution, a confidence decay averaging is performed on the neighboring nodes of the replaced node, so that the semantic confidence level transitions continuously in space. The above process is repeated until the average change of the overall semantic consistency score is less than the preset threshold, and the semantically corrected virtual model is output.

[0119] The semantically corrected virtual model is used as the result of the fusion of geometric optimization and semantic correction to generate an optimized virtual model. The optimized virtual model maintains high consistency with the target physical scene in terms of both geometric accuracy and semantic consistency.

[0120] In this embodiment, the real-scene 3D dynamic change construction includes:

[0121] The optimized virtual model generated based on the digital twin data framework is loaded into the rendering module, and the spatial structure, semantic labels and attribute parameters of the virtual model are loaded into the rendering module to generate rendering input data.

[0122] Perform real-time rendering, output the rendering results of the virtual model, and synchronously receive the perception feedback information of the target physical scene to generate virtual-real synchronization data containing time index and spatial coordinates.

[0123] Based on the virtual-real synchronous data, a two-way data flow association is established between the virtual model and the target physical scene. The rendering results and perception feedback information are matched in time and space to generate virtual-real difference data.

[0124] The bidirectional data stream association includes: using time indexing to realize the time correspondence between virtual model rendering frames and perception feedback information acquisition frames; using spatial coordinates to realize spatial matching between virtual nodes and physical scene observation points; establishing a mapping relationship table between the two to realize bidirectional update and transmission of rendering results and perception data; and outputting the matched virtual-real difference data.

[0125] Perform adaptive adjustment of virtual-real difference based on virtual-real difference data, and output the virtual-real difference correction result;

[0126] The adaptive adjustment of virtual-real differences includes: calculating the differences between the virtual model and the target physical scene in terms of spatial location, color distribution, and semantic labels. The three types of differences are defined as spatial difference, color difference, and semantic difference, respectively. Spatial difference is obtained by averaging the square root of the sum of the squares of the distance differences between virtual nodes and physical nodes in three directions. Color difference is obtained by averaging the absolute values ​​of the brightness differences of corresponding pixels in the virtual and real images. Semantic difference is obtained by comparing whether the semantic labels of nodes are consistent and calculating the average of the semantic confidence difference. During the adjustment process, spatial differences are corrected by translating virtual nodes along the difference direction. Color differences are corrected by adjusting the light intensity, reflectivity, and ambient brightness parameters of the virtual model to achieve color consistency. Semantic differences are corrected by calling the semantic label correction mechanism to replace the semantic labels of inconsistent nodes and update their confidence. The three types of difference results are weighted and averaged according to their importance. When the weighted average is less than the preset deviation range, the virtual-real difference adjustment is considered complete and the virtual-real difference correction result is output. Otherwise, the above adjustment process is automatically repeated until the virtual-real difference stabilizes within an acceptable range.

[0127] Update the spatial structure, semantic labels, and attribute parameters of the virtual model to generate an updated virtual model;

[0128] The updated virtual model is written into the digital twin data framework and real-time rendering is re-executed. The virtual-real consistency index is calculated based on the virtual-real difference correction results. The virtual-real consistency index is equal to the sum of the average value and fluctuation amplitude of the spatial residual, color residual and semantic residual of each node. When the virtual-real consistency index is lower than the set threshold, it is determined that virtual-real synchronization is completed. Otherwise, the current correction result is used as the new input to re-execute the virtual-real difference adaptive adjustment, so as to realize the automatic perception of the structural changes of the target physical scene and the continuous dynamic update of the real scene 3D spatial data, and complete the real scene 3D dynamic change construction process under the digital twin data framework.

[0129] Example 1:

[0130] To verify the feasibility of this invention in practice, it was applied to the construction and dynamic updating of a 3D digital twin in a central urban area. This area is densely built, with high traffic volume, and its surface structure and man-made facilities undergo significant dynamic changes within a short period, such as the addition of construction areas, alterations to road signs, the removal and installation of billboards, and periodic changes in green coverage. While traditional 3D modeling methods can generate high-precision static models, they often lag behind the real-world environment in the face of such frequent changes, leading to significant discrepancies between the virtual model and the physical environment. This not only affects the accurate decision-making of the urban management system but also reduces the practicality of the digital twin platform in monitoring, evaluation, and simulation analysis. The method proposed in this invention, by introducing an improved Gaussian sputtering neural field modeling algorithm, a digital twin mapping framework, and a temporal change detection mechanism, effectively achieves automatic synchronous updates between the virtual and real models, fundamentally solving the problem that traditional static models cannot reflect real-time changes in the scene.

[0131] In this scenario, the invention first collects data from multiple sources, including LiDAR point clouds, UAV aerial images, and multi-view ground photographs. All data are spatially registered and temporally synchronized within a unified coordinate system to ensure spatial alignment accuracy and temporal consistency across different data sources. The resulting multi-source dataset, after fusion processing, is input into an improved Gaussian sputtering neural field model. This model represents the spatial location, color features, opacity, and shape parameters of each point in the physical scene using a Gaussian distribution. A time-aware mechanism establishes a multi-temporal Gaussian distribution sequence, enabling the system to capture subtle differences in building appearance over time. A density adaptive mechanism further dynamically adjusts the distribution range between dense and sparse areas of the point cloud, maintaining overall structural stability while ensuring local detail accuracy. An illumination consistency constraint mechanism maintains color and brightness consistency under different shooting times, lighting conditions, and viewing angles, ensuring the final generated 3D model achieves a high level of both visual effect and physical consistency.

[0132] In practical applications, this invention maps the generated real-world 3D model onto a digital twin data framework. Through spatial indexing, semantic annotation, and attribute binding, a virtual-real correspondence is established. The system identifies entities such as buildings, roads, and transportation facilities based on their spatial location and semantic features, forming a dynamically trackable virtual spatial dataset. When the urban environment changes in subsequent time periods, such as the appearance of a new building or road widening and reconstruction, this invention uses the virtual spatial data and a new fused dataset to perform temporal change detection, automatically identifying newly added, deleted, and deformed areas, and generating a change detection result dataset. This dataset is input into the dynamic change modeling module. The system performs incremental optimization of Gaussian parameters based on the spatial characteristics of the changed areas, performing temporal interpolation, local redistribution, and covariance adaptive correction on the Gaussian distribution nodes of the corresponding areas, enabling the virtual model to automatically update its local structure. The updated model maintains consistency with the real environment in terms of spatial structure, semantic labels, and physical attributes, achieving real-time adaptive adjustment without manual intervention.

[0133] The advantage of this invention lies in its complete virtual-real bidirectional feedback mechanism. When the system renders the updated virtual model in real time within the digital twin framework, the perception module simultaneously receives observation feedback data from on-site sensors. By comparing the rendering results with actual observations, it automatically performs adaptive adjustments to address virtual-real differences. This process includes coordinate correction for spatial position deviations, brightness parameter adjustment for lighting and color differences, and dynamic replacement of inconsistent semantic labels, thereby gradually bringing the virtual model closer to a realistic state in terms of geometric accuracy, color reproduction, and semantic consistency. Once the difference correction converges to a preset range, the system rewrites the updated virtual model into the digital twin framework, achieving automatic perception and continuous updating of changes in the target physical scene structure.

[0134] Throughout the implementation process, the dynamic change modeling method constructed by this invention not only improved the real-time performance and accuracy of the digital twin model but also significantly reduced the workload of manual updates and modeling. The virtual model can automatically evolve over time, reflecting the changing state of the physical scene in real time, truly achieving virtual-real synchronization. Application results show that this method can operate stably in complex urban environments, exhibiting strong robustness to multiple factors such as lighting, viewing angle, and occlusion, and can support long-term continuous monitoring and dynamic management. Thus, the implementation of this invention effectively overcomes the limitations of traditional 3D modeling methods in adapting to dynamic environments, establishing a digital twin 3D modeling system with self-sensing, self-updating, and self-optimizing capabilities, providing reliable technical support and theoretical foundation for urban operation management, intelligent building monitoring, infrastructure operation and maintenance, and the construction of digital twin platforms.

[0135] Table 1 Performance Comparison Analysis of Improved Method and Traditional Method

[0136] Comparison Projects Traditional static modeling methods Multi-temporal point cloud fusion method Improved Gaussian sputtering modeling method The dynamic change modeling method of this invention Increase (%) Model update delay (seconds) 125 82 41 12 85.6 Geometric accuracy error (cm) 5.2 3.7 2.4 1.1 78.8 Semantic classification accuracy (%) 86.5 89.3 92.1 96.7 11.8 Light consistency index (%) 79.2 84.6 90.5 95.8 20.9 Virtual-to-real synchronization delay (seconds) 8.6 6.2 3.4 1.3 84.9 Model reconstruction frame rate (frames per second) 12 18 25 32 166.7

[0137] As shown in Table 1, the proposed digital twin-based real-scene 3D dynamic change modeling method significantly outperforms traditional methods in all key indicators. Regarding model update latency, this method reduces the latency from the traditional 125 seconds to 12 seconds, an improvement of 85.6%. This is mainly due to the incremental update mechanism of the improved Gaussian sputtering neural field model, which can reconstruct parameters only for the relevant region after detecting local changes, thus avoiding the time wastage caused by global recalculation. Geometric accuracy error is reduced from 5.2 cm to 1.1 cm, an improvement of nearly 79%, indicating that the time-aware and density-adaptive mechanisms effectively enhance the model's ability to recover details at the spatial structure level. Semantic classification accuracy is improved to 96.7%, an improvement of approximately 12% compared to traditional methods, indicating that the semantic synchronization and semantic consistency verification mechanisms significantly enhance the virtual-real semantic matching degree. The illumination consistency index is improved to 95.8%, proving that the illumination constraint mechanism can maintain color stability under different times and viewpoints. Furthermore, the proposed method excels in virtual-real synchronization latency and reconstruction frame rate, reducing latency to 1.3 seconds and increasing the frame rate to 32 frames per second, achieving near real-time dynamic rendering updates. Comprehensive analysis shows that by introducing a three-dimensional control mechanism of time, density, and illumination into the Gaussian sputtering neural field model, and combining it with a virtual-real bidirectional feedback adjustment and attribute dynamic binding algorithm, this invention effectively improves the real-time performance, accuracy, and semantic stability of model reconstruction, providing a reliable technical foundation for digital twin systems in complex dynamic scenarios.

[0138] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for modeling dynamic changes in three-dimensional real-world scenes based on digital twins, characterized in that, Includes the following steps: Multi-source data acquisition is performed on the target physical scene to obtain multi-source data including spatial coordinates, texture features and timestamps; Perform spatial registration and temporal synchronization processing on multi-source data, and output a fused dataset after registration and synchronization processing; Based on the fused dataset, an improved Gaussian sputtering neural field model is used to perform 3D modeling of the scene, and the model parameters are optimized to generate a realistic 3D model. A virtual-real mapping is performed based on a digital twin mapping framework, and the generated virtual space data and its state parameters are stored. By calling virtual space data and fused datasets, differential analysis is performed to identify newly added, deleted, and deformed regions, and a change detection result dataset is generated. Dynamic change modeling and update mechanism are implemented to incrementally optimize local Gaussian distribution parameters and generate updated virtual model; The updated virtual model is invoked, geometric accuracy optimization and semantic consistency verification are performed, the geometric error of the virtual model is adjusted and the semantic labels are corrected, and the optimized virtual model is output. The optimized virtual model is rendered and dynamically displayed in real time, and bidirectional data flow is associated and dynamically updated to complete the real-scene 3D dynamic change construction process in the digital twin system.

2. The method for real-scene 3D dynamic change modeling based on digital twins according to claim 1, characterized in that, The generation of the fused dataset includes: Spatial registration is performed on multi-source data to extract three-dimensional feature points. Based on the iterative nearest point registration algorithm of feature point matching, the multi-source data is aligned to a unified coordinate reference system. A spatial transformation matrix composed of rotation matrix and translation vector is established, and the multi-source data after preliminary registration is output. Perform geometric optimization processing on the initially registered multi-source data to output geometrically optimized multi-source data; The geometrically optimized multi-source data is time-synchronized to establish a unified time reference. Offset correction and time interpolation compensation are performed on each sampling time to output the time-synchronized multi-source data. Multi-source data that has undergone spatial registration and time synchronization are uniformly encoded to generate a fused dataset. Outliers and noisy data are statistically removed and spatially interpolated to complete the dataset, which is then output as a fused dataset after spatial registration and time synchronization.

3. The method for real-scene 3D dynamic change modeling based on digital twins according to claim 1, characterized in that, The generation of the real-scene 3D model includes: Using the fused dataset as input data, spatial coordinates, texture color values, illumination reflection parameters and time index information are extracted from the fused dataset to construct a three-dimensional spatial point set corresponding to the target physical scene; Establish a Gaussian distribution parameter set for each point in the three-dimensional spatial point set; Input the Gaussian distribution parameter set into the improved Gaussian sputtering neural field model to generate preliminary rendering results; Using the preliminary rendering results and the original observations of the corresponding target physical scene in the fused dataset as reference inputs, pixel-by-pixel difference calculations are performed and a rendering error function is constructed. An iterative optimization algorithm based on error gradient descent is adopted, with the rendering error function as the optimization objective. The Gaussian distribution parameter set is updated step by step. When the descent rate of the rendering error function is lower than the preset threshold or the error change is less than the stable value in multiple consecutive iterations, the iteration is stopped and the optimized Gaussian distribution parameter set is output. The optimized Gaussian distribution parameter set is re-input into the improved Gaussian sputtering neural field model, and ray integral rendering is performed to generate a realistic 3D model corresponding to the target physical scene.

4. The method for real-scene 3D dynamic change modeling based on digital twins according to claim 1, characterized in that, The generation of the virtual space data and its state parameters includes: Based on the real-world 3D model, it is instantiated as a 3D spatial object in the digital twin mapping framework. A unique identifier number is assigned to each spatial object, and a coordinate index table corresponding to the target physical scene is established to form a set of 3D spatial objects. Taking a set of three-dimensional spatial objects as input, the spatial coordinates, geometric boundaries and texture features of each spatial object are extracted to generate a set of spatial feature descriptions. Based on this set, a spatial index structure is constructed to realize the geometric mapping alignment between the virtual space and the target physical scene, and output spatial index mapping data. Semantic annotation and semantic synchronization are performed based on spatial index mapping data. Based on the semantic synchronization mechanism, the consistency of semantic labels under different time indices is maintained, and a semantic synchronization result set is output. Using the semantic synchronization result set as input, the real-time physical attributes of the target physical scene are extracted from the fused dataset, and an attribute binding result set is generated by using an attribute binding algorithm based on the feature mapping matrix. A virtual-real mapping data structure is constructed based on the attribute binding result set, the spatial feature description set, and the semantic synchronization result set, and the virtual-real mapping structured data is output. A set of state parameters is generated for each virtual space object, forming a traceable temporal state parameter set; The virtual-real mapping data structure and the set of state parameters are stored as a virtual space data file based on the time-series state parameter set.

5. The method for real-scene 3D dynamic change modeling based on digital twins according to claim 1, characterized in that, The generation of the change detection result dataset includes: By calling virtual space data and fused datasets collected at different times, and based on the spatial index table and semantic label table in the virtual-real mapping structured data, spatial registration and temporal index matching are performed on the multi-temporal data to generate a multi-temporal fused data sequence. Based on the multi-temporal fusion data sequence, spatial coordinates, reflection intensity, surface normal vector and semantic label features of point cloud data under each time index are extracted, a temporal feature set is established, and the correspondence of each feature in virtual space data is recorded by a unified node identifier. Based on the time series feature set, the time series change detection mechanism is invoked to perform joint difference analysis on multi-source feature data under adjacent time indices and output a time series difference result set; The multi-temporal point cloud comparison algorithm is invoked to perform spatial overlap calculation, point density difference analysis and semantic consistency verification on the corresponding node data of the virtual space data and the target physical scene. By fusing the weighted results of geometric distance difference, semantic consistency difference and attribute vector difference, the changed region is identified and the changed region result set is output. Perform change type classification and spatial clustering processing, and update the time index and state parameters of the corresponding nodes in the virtual-real mapping structured data to generate a change detection result dataset.

6. The method for real-scene 3D dynamic change modeling based on digital twins according to claim 1, characterized in that, The generation of the updated virtual model includes: The change region information in the change detection result dataset is analyzed, and the spatial location, semantic labels, and attribute parameters of the corresponding virtual model nodes are extracted. The parameter update mechanism of the improved Gaussian sputtering neural field model is invoked based on the parameter set of the changed region to determine the set of local Gaussian distribution parameters that need to be updated, and the corresponding Gaussian distribution nodes are located in the virtual model according to the spatial boundary of the changed region to generate the parameter set to be optimized. Based on the set of parameters to be optimized, an incremental optimization process is performed on the improved Gaussian sputtering neural field model to generate the optimized Gaussian distribution parameter set. The optimized Gaussian distribution parameter set is written into the corresponding local Gaussian nodes of the virtual model to generate the updated virtual model.

7. The method for real-scene 3D dynamic change modeling based on digital twins according to claim 1, characterized in that, The generation of the optimized virtual model includes: The updated virtual model is invoked, and the spatial coordinates, surface normal vectors, and semantic label information of each node in the virtual model are analyzed to establish the correspondence between the spatial features and semantic features of the nodes, forming a model feature index set. Using the model feature index set as input, geometric accuracy optimization is performed. Local error analysis is conducted on the spatial geometry of the virtual model, the node geometric offset is calculated, and the geometric change corresponding to the change detection result is used as a constraint. The node spatial position and surface normal vector are adjusted by iteratively minimizing the geometric error function, and the geometrically optimized virtual model is output. Perform a semantic consistency check process based on the geometrically optimized virtual model, and output a semantically corrected virtual model; The semantically corrected virtual model is used as the result of the fusion of geometric optimization and semantic correction to generate an optimized virtual model.

8. The method for real-scene 3D dynamic change modeling based on digital twins according to claim 1, characterized in that, The real-scene 3D dynamic change construction includes: The optimized virtual model generated based on the digital twin data framework is loaded into the rendering module, and the spatial structure, semantic labels and attribute parameters of the virtual model are loaded into the rendering module to generate rendering input data. Perform real-time rendering, output the rendering results of the virtual model, and synchronously receive the perception feedback information of the target physical scene to generate virtual-real synchronization data containing time index and spatial coordinates. Based on the virtual-real synchronous data, a two-way data flow association is established between the virtual model and the target physical scene. The rendering results and perception feedback information are matched in time and space to generate virtual-real difference data. Perform adaptive adjustment of virtual-real difference based on virtual-real difference data, and output the virtual-real difference correction result; Update the spatial structure, semantic labels, and attribute parameters of the virtual model to generate an updated virtual model; The updated virtual model is written into the digital twin data framework and real-time rendering is re-executed. The virtual-real consistency index is calculated based on the virtual-real difference correction results, thus completing the dynamic change construction process of the real-world 3D scene under the digital twin data framework.

Images