A substation engineering complex scene reconstruction method based on three-dimensional Gauss
By using a scene reconstruction method based on 3D Gaussian, the problem of balancing model accuracy and rendering efficiency in complex substation engineering scenarios has been solved, achieving efficient and accurate reconstruction of complex scenes and improving the level of engineering digitization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID JIANGXI ELECTRIC POWER CO LTD
- Filing Date
- 2026-06-01
- Publication Date
- 2026-06-26
AI Technical Summary
Existing technologies struggle to balance model accuracy and rendering efficiency in the reconstruction of complex substation engineering scenarios, lack the ability to express dynamic scenes, and face difficulties in fusing multi-source data, thus limiting the improvement of engineering digitization.
A scene reconstruction method based on 3D Gaussian is adopted. The distribution of scene data points in 3D space is described by the 3D Gaussian distribution function. Combined with camera coordinate system analysis and fast re-iteration algorithm, scene optimization and reconstruction are performed. Object-level key points are extracted and scale calibration and normal vector angle constraint matching are performed to achieve multi-source data fusion.
It improves the reconstruction efficiency and accuracy of complex substation engineering scenarios, solves the reconstruction challenges of large-scale, multi-structure, and dynamically evolving scenarios, expands the research field of complex scenario optimization, and lays the foundation for the future development of related fields.
Smart Images
Figure CN122289573A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of three-dimensional scene reconstruction technology, specifically a method for reconstructing complex substation engineering scenes based on three-dimensional Gaussian. Background Technology
[0002] Current substation construction faces severe challenges due to complex geographical environments, diverse facility structures, and dynamically intertwined construction processes. Traditional modeling technologies, primarily relied upon in the industry, such as 2D drawings, BIM, and oblique photogrammetry, have significant shortcomings in addressing these complex scenarios: 2D drawings are not intuitive and easily lead to design conflicts; while BIM models are accurate, they struggle to represent large-scale real-world scenes; and oblique photogrammetry data is bloated and lacks interactivity. These technologies generally suffer from core defects such as the difficulty in balancing model accuracy and rendering efficiency, insufficient ability to express dynamic scenes, and challenges in integrating multi-source data, severely hindering the improvement of engineering digitization and refined management.
[0003] 3D Gaussian splatting, as an emerging 3D reconstruction technology, offers new hope for overcoming existing technological bottlenecks due to its fast training speed, high real-time rendering efficiency, and excellent visual fidelity. However, its application is currently mostly limited to small-scale object or indoor scene reconstruction. How to effectively utilize and adapt this technology in complex industrial scenarios such as substation engineering, which involve large scale, multiple structures, and dynamic evolution, remains an unexplored area and constitutes a pressing frontier technical challenge. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a method for reconstructing complex substation engineering scenarios based on three-dimensional Gaussian, aiming to solve the problems in the background technology.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for reconstructing complex substation engineering scenes based on three-dimensional Gaussian, comprising the following steps: Step S1: Scene optimization based on 3D Gaussian distribution: The distribution of scene data points in the substation engineering scene in 3D space is described by the 3D Gaussian distribution function, and the camera coordinate system analysis is introduced to optimize the scene representation. Step S2, Fast Iterative Scene Reconstruction: Based on the scene representation optimized in Step S1, fast iterative scene reconstruction is performed by integrating stochastic gradient descent, Gaussian function adaptive control, and fast rasterization algorithm. Step S3, Feature-based model coupling: Based on the rapid re-iteration scene reconstruction results, extract object-level key points including equipment and structures, and complete scale calibration; minimize the distance between key points by matching the scene data points in the rapid re-iteration scene reconstruction results with the normal vector angle constraints of the corresponding surfaces of the preset parameter CAD model; embed the equipment and structures after scale calibration and normal vector angle constraint matching into the rapid re-iteration scene reconstruction results to complete the fusion of multi-source data and form the final substation engineering scene.
[0006] Furthermore, the specific process of step S1 is as follows: Step S1.1: First, perform a three-dimensional Gaussian distribution analysis. Use the three-dimensional Gaussian distribution function to describe the distribution of scene data points in the three-dimensional space in the substation engineering scenario. The mean of the center point of the scene data point set represents the center position of the scene data point. The covariance matrix in the three-dimensional Gaussian distribution function is used to describe the degree and direction of change of scene data points in different dimensions. Step S1.2: Conduct 3D Gaussian sputtering analysis, treating each scene data point as a Gaussian point. Calculate the color value of each pixel by projecting the Gaussian point onto a 2D plane to generate a 2D image. During this process, camera coordinate system analysis is introduced: First, the projection result of the substation engineering scene onto the 2D plane is obtained through the camera, and the position and distribution of the scene data points are analyzed in the coordinate system to map the scene data points to the 2D plane projection result. Simultaneously, the scene data points in 3D space are matched with the corresponding 2D images. Step S1.3: After completing the 3D Gaussian sputtering analysis, optimize the scene representation: First, by introducing a 3D Gaussian distribution representation, the substation engineering scene is represented as a probability distribution; then, the diagonal elements of the 3D Gaussian covariance matrix are used to represent the variance of each coordinate axis of the 3D global coordinate system of the substation engineering scene, reflecting the extent of expansion of equipment and structures in the substation engineering scene. At the same time, the off-diagonal elements of the 3D Gaussian covariance matrix are used to represent the covariance of each coordinate axis of the 3D global coordinate system of the substation engineering scene.
[0007] Furthermore, the specific process of step S2 is as follows: Step S2.1: Based on the scene representation optimized in Step S1, create a set of Gaussian points to represent the free-view synthetic scene of the substation engineering project; set the initial covariance matrix of the created Gaussian points; based on the created Gaussian points and the corresponding initial covariance matrix, synthesize the simulated image of the substation engineering scene; define a combined loss function of pixel color loss and structural similarity loss; using the combined loss function as the evaluation criterion, compare the simulated image of the substation engineering scene with the real substation engineering scene image to quantify the difference between the two; introduce the stochastic gradient descent algorithm, with the goal of minimizing the combined loss function, iteratively solve to obtain the adjustment direction and magnitude parameters of the equipment / structure graphics; based on the optimization output of stochastic gradient descent, adjust the incorrectly located equipment / structure graphics in the projection process of Step S1.2; Step S2.2: Based on the analysis of the combined loss function, two types of regions to be optimized are identified from the scene representation after optimization and adjustment in Step S2.1: Area 1: Areas where geometric features are missing, including equipment connections and structural junctions; Region 2: The region covered by the Gaussian point set whose spatial range is larger than the preset value; A dual-region dynamic parameter adjustment strategy based on engineering features is implemented for the region to be optimized. Step S2.3: Based on the results of parameter adjustment in step S2.2, Gaussian rasterization technology is used to perform a three-dimensional Gaussian rasterization iterative reconstruction process.
[0008] Furthermore, the dual-region dynamic parameter adjustment strategy based on engineering features for the region to be optimized includes: For region 1, first obtain the device connection type information and preset the corresponding feature weight coefficients for different connection types; by calculating the curvature mutation value of Gaussian points at the connection point, construct the "curvature mutation value - encryption multiple" mapping relationship. When the curvature mutation value is ≥ the preset threshold, the number of Gaussian points per unit volume is encrypted to the first preset multiple of the original number. When the curvature mutation value is < the preset threshold, it is encrypted to the second preset multiple of the original number. For region 2, a density adjustment benchmark threshold is set. By calculating the distribution entropy of Gaussian points in region 2 in real time, the density of Gaussian points is increased when the distribution entropy is greater than or equal to the first density adjustment benchmark threshold; and the density of Gaussian points is decreased when the distribution entropy is less than the second density adjustment benchmark threshold.
[0009] Furthermore, the specific process of step S2.3 is as follows: Step S2.31: Filter and remove Gaussian points at extreme positions: Filter Gaussian points according to the field of view and the cyclic block situation, and at the same time use a guard band to remove Gaussian points at extreme positions; Step S2.32, Gaussian point instantiation and key assignment: For each overlapping cyclic block of Gaussian points, the corresponding Gaussian point is instantiated, and a key combining the view space depth and the block ID is assigned to each instance; Step S2.33, Gaussian sort: Use the radix sort algorithm to sort the corresponding Gaussian points according to the key; Step S2.34, Pixel rendering based on sorted Gaussian points: For the sorted Gaussian point set, traverse each Gaussian point in cyclic block and depth order, calculate its coverage and contribution weight in the two-dimensional image plane through the projection formula of the three-dimensional Gaussian distribution, combine the color information and covariance matrix of the Gaussian points, and accumulate the color value of each pixel by weighted fusion to generate a continuous and smooth intermediate image for scene reconstruction. Step S2.35, Reconstruction Result Verification and Optimization: Compare the intermediate scene reconstruction image generated in step S2.34 with the actual substation engineering scene image, and calculate the combined loss function value; if the loss function value is higher than the preset threshold, return to step S2.2 to readjust the Gaussian point parameters, and repeat steps S2.31-S2.34; if the combined loss function value is lower than or equal to the preset threshold, output the scene reconstruction result of this iteration, and complete the rapid re-iteration scene reconstruction.
[0010] Furthermore, the specific process of step S3 is as follows: Step S3.1, Initial orientation alignment of the preset parameter CAD model: Align the front of the preset parameter CAD model with the positive X-axis direction; align the top of the preset parameter CAD model with the positive Z-axis direction; Step S3.2: Extract object-level key points from the fast re-iteration scene reconstruction results in Step S2; Step S3.3: Complete scale calibration: Take the line segment where the intersection point between the planes of equipment and structures in the aligned preset parameter CAD model is located as the target line segment, and take the line segment where the intersection point between the planes of equipment and structures in the rapid re-iteration scene reconstruction result of step S2 as the source line segment; determine the scale of equipment and structures by calculating the Euclidean distance ratio between the target line segment and the source line segment. Step S3.4, Constraint matching between scene data points and preset parameter CAD model: Using the normal vector angle of the corresponding face of the scene data points in the fast re-iteration scene reconstruction result and the aligned preset parameter CAD model as constraints, optimize the normal vector angle of the corresponding face and the distance between object-level key points. Step S3.5, Model Fusion: The equipment and structures after scale calibration and normal vector angle constraint matching are embedded into the rapid re-iteration scene reconstruction results to complete the fusion of multi-source data and form the final substation engineering scene.
[0011] Furthermore, the specific process of step S3.2 is as follows: Step S3.21: Extract the intersection points between the planes of equipment and structures in the fast re-iteration scene reconstruction results of step S2; Step S3.22: Use a directional bounding box to divide the line segments where the intersection points between the equipment and structure planes are located; Step S3.23: Use the endpoints of the segmented line segments as object-level key points for equipment and structures.
[0012] An electronic device includes a processor, a memory, and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program code, and the processor is used to call the program code stored in the memory to execute a method for reconstructing complex substation engineering scenarios based on three-dimensional Gaussian.
[0013] A non-volatile computer storage medium storing computer-executable instructions that execute a method for reconstructing complex substation engineering scenarios based on three-dimensional Gaussian.
[0014] Compared with existing technologies, this invention offers the following advantages: By deeply analyzing methods for representing the data structure of complex substation engineering scene models, studying the three-dimensional Gaussian distribution function, and integrating camera coordinate system transformation analysis and scene optimization representation, this invention provides an efficient solution for improving the reconstruction efficiency of complex substation engineering scenes. Based on this approach, the distribution of data points in three-dimensional space is represented by a three-dimensional Gaussian distribution function, addressing issues such as occlusion and viewpoint changes in complex substation engineering scenes. This enables rapid and accurate structural reconstruction of complex scenes, providing a new perspective and method for scene reconstruction. This innovation not only expands the research field of optimizing complex substation engineering scenes but also lays a solid foundation for future development in related fields.
[0015] Given the constraints of both time and space costs, efficiently and cost-effectively reconstructing complex 3D scenes with multiple granularities remains a challenging research problem. This invention employs a multi-granularity 3D scene reconstruction method based on rapid re-iteration, separately reconstructing large scenes and optimizing components with lower reconstruction quality, significantly reducing the workload and complexity of optimizing the accuracy of complex scenes.
[0016] After segmenting the component to be optimized from the reconstructed scene and optimizing it using a fast re-iteration method, matching the point cloud of the scene with the target model is a very challenging problem due to the inconsistent scale of different data. This invention treats the reconstruction problem as a model fusion problem, extracts object-level key points and performs subsequent optimization to minimize the distance between corresponding key points, thereby accurately placing the target model at the target location. Attached Figure Description
[0017] Figure 1 This is a flowchart of the method of the present invention.
[0018] Figure 2 This is a diagram of the three-dimensional Gaussian projection process.
[0019] Figure 3 This is a flowchart of the adaptive Gaussian densification scheme of the present invention.
[0020] Figure 4 This is a flowchart of the 3D Gaussian rasterization re-iteration reconstruction process. Detailed Implementation
[0021] Example 1
[0022] like Figure 1 As shown, the present invention provides a technical solution: a method for reconstructing complex substation engineering scenes based on three-dimensional Gaussian, comprising the following steps: Step S1: Scene optimization based on 3D Gaussian distribution: The distribution of scene data points (equipment, structures, etc. in the substation engineering scene composed of dense scene data points) in 3D space is described by the 3D Gaussian distribution function. At the same time, camera coordinate system analysis is introduced to handle the problem of viewpoint changes and facility cross-occlusion under different observation angles, thereby optimizing the scene representation (making it adaptable to the characteristics of multiple structures and details of the project, effectively filtering noise interference in the substation engineering real scene data, and providing accurate basic scene data support for subsequent efficient reconstruction).
[0023] Step S2, Rapid Re-iteration Scene Reconstruction: Based on the scene representation optimized in Step S1, and considering the large-scale scene and dynamic construction requirements of substation engineering, a rapid re-iteration scene reconstruction is performed by integrating stochastic gradient descent, Gaussian function adaptive control, and fast rasterization algorithm (reducing the computational power and time consumption costs of large-scale scene reconstruction, and adapting to the timeliness and accuracy requirements of dynamic construction in engineering).
[0024] Step S3, Feature-based model coupling: Based on the rapid re-iteration scene reconstruction results, extract object-level key points including equipment and structures, and complete scale calibration; minimize the distance between key points by matching the scene data points in the rapid re-iteration scene reconstruction results with the normal vector angle constraints of the corresponding surfaces of the preset parameter CAD model; embed the equipment and structures after scale calibration and normal vector angle constraint matching into the rapid re-iteration scene reconstruction results to complete the fusion of multi-source data and form the final substation engineering scene.
[0025] The specific process of step S1 is as follows: Step S1.1: First, perform a three-dimensional Gaussian distribution analysis. Use the three-dimensional Gaussian distribution function to describe the distribution of scene data points constituting equipment and structures in the three-dimensional space of the substation engineering scenario, adapting to the characteristics of multi-structure layout in the project. The mean of the center point of the scene data point set represents the center position of the scene data point, and the covariance matrix is used to describe the degree and direction of change of the scene data points in different dimensions, which is represented as follows: ; In the formula, The probability density function representing a three-dimensional Gaussian distribution; A three-dimensional spatial coordinate vector representing a single scene data point; A vector representing the mean of the set of scene data points; Represent the covariance matrix; Indicates the transpose operation; It represents the base of the natural logarithm.
[0026] Step S1.2: As Figure 2 As shown, a 3D Gaussian sputtering analysis is conducted. Addressing the issues of multiple obstructions and varying viewing angles in substation engineering facilities, each scene data point is treated as a Gaussian point. The color value of each pixel is calculated by projecting the Gaussian points onto a 2D plane, generating a smooth and continuous 2D image (more accurately representing the surface shape of the original point cloud data). In this process, camera coordinate system analysis is introduced (introducing camera coordinate system analysis during the projection of the 3D Gaussian distribution onto 2D space improves the accuracy and stability of the engineering scene reconstruction). First, the projection results of the substation engineering scene onto the 2D plane are obtained through a camera. Then, the position and distribution of scene data points are analyzed in the coordinate system to map the scene data points to the 2D plane projection results. The projection process is represented as follows: ; ; In the formula, This represents the pixel coordinates of scene data points (Gaussian points) projected onto a two-dimensional plane. The corresponding pixel position in the horizontal direction (horizontal axis) of the two-dimensional plane. The pixel position in the vertical direction (vertical axis) of the corresponding two-dimensional plane; Indicates the camera's focal length; , , This represents the 3D coordinates of scene data points (Gaussian points) in the camera coordinate system. , These are the x and y coordinates of scene data points (Gaussian points) in the camera coordinate system plane (a plane perpendicular to the camera's optical axis). It is the distance (depth value) from the scene data point (Gaussian point) to the camera lens; Represents the coordinates of the principal point in a two-dimensional plane (also called the coordinates of the image center). Meanwhile, to eliminate interference from equipment cross-obstruction and different observation angles in substation engineering scenarios, scene data points in three-dimensional space are matched with corresponding two-dimensional images. This is achieved by describing the positional distribution of scene data points in the camera coordinate system. Specifically, after matching scene data points in three-dimensional space with corresponding two-dimensional images, the covariance matrix of the updated scene data points (Gaussian points) in the camera coordinate system is... Represented as: ; In the formula, This represents the Jacobian matrix of the projection transformation that supports the matching of "3D scene data points → 2D image"; express The transpose of the matrix; This represents the noise term corresponding to interference caused by equipment cross-blocking and different observation angles.
[0027] Step S1.3: After completing the 3D Gaussian sputtering analysis, scene representation optimization is performed: Addressing the difficulty of traditional point cloud methods in capturing the complex structural details of substation engineering projects, a 3D Gaussian distribution representation is first introduced to represent the substation engineering scene as a probability distribution, reducing uncertainty in the scene and helping to reduce noise in subsequent processing and analysis. Then, the diagonal elements of the 3D Gaussian covariance matrix (the three elements on the main diagonal of the 3×3 covariance matrix) represent the variance of each coordinate axis in the 3D global coordinate system of the substation engineering scene, reflecting the extent of expansion of equipment and structures within the scene. Simultaneously, the off-diagonal elements of the 3D Gaussian covariance matrix represent the covariance of each coordinate axis in the 3D global coordinate system, revealing its shape and structure, thereby quickly and accurately restoring the details and structure of complex substation engineering scenes. The 3D Gaussian covariance matrix... Represented as: ; In the formula, This represents a 3×3 orthogonal rotation matrix, used to represent the spatial rotation orientation of equipment and structures in a three-dimensional global coordinate system in a substation engineering scenario. Represents a 3×3 diagonal scaling matrix, used to describe the spatial expansion of equipment and structures in each coordinate axis direction; express The transpose of the matrix; express The transpose of .
[0028] The specific process of step S2 is as follows: Step S2.1: Based on the scene representation optimized in Step S1, create a dense set of Gaussian points to accurately represent the free-view synthetic scene of the substation engineering project (such as scenes at different stages of construction); set the initial covariance matrix (isotropic Gaussian matrix) of the created Gaussian points; based on the created Gaussian points and the corresponding initial covariance matrix, synthesize the simulated image of the substation engineering scene; define a combined loss function of pixel color loss and structural similarity loss; using the combined loss function as the evaluation criterion, compare the simulated image of the substation engineering scene with the real substation engineering scene image (on-site photos of the construction stage, actual images of equipment layout) to quantify the differences between the two; introduce the stochastic gradient descent algorithm, with the goal of minimizing the combined loss function, and iteratively solve to obtain the adjustment direction and magnitude parameters of the equipment / structure graphics; based on the results of the stochastic gradient descent optimization, adjust (move / destroy) the incorrectly located equipment / structure graphics in the projection process of Step S1.2, and complete the optimization and adjustment of the scene representation.
[0029] Among them, pixel color loss Represented as: ; In the formula, In the image representing a real substation engineering scene, the first... 1 pixel The corresponding actual color value; In the simulated image of a substation engineering scenario, the first... 1 pixel The corresponding actual color value; This represents the L1 norm.
[0030] Among them, structural similarity loss Represented as: ; In the formula, This represents a simulated image of a substation engineering scenario. Images representing real substation engineering scenarios; , These represent simulated images of substation engineering scenarios. Images of real substation engineering scenarios The average pixel brightness; , These represent simulated images of substation engineering scenarios. Images of real substation engineering scenarios The pixel brightness variance; Image representing a substation engineering scenario simulation Images of real substation engineering scenarios The pixel brightness covariance between them; , All of these represent constants for stable calculations (to avoid denominators being 0 or numerical fluctuations), and are usually small values set according to the dynamic range of the image.
[0031] Among them, the combination loss function Represented as: ; In the formula, This represents the weighting coefficient.
[0032] This invention adaptively controls the number and density of Gaussians per unit volume, transforming an initially sparse Gaussian set into a denser one, thereby representing the scene with optimized parameters. For regions lacking geometric features and large areas within a Gaussian-covered scene, Gaussians with an average view-space gradient exceeding a threshold are densified, such as... Figure 3 As shown.
[0033] Step S2.2: Based on the analysis of the combined loss function, two types of regions to be optimized are identified from the scene representation after optimization and adjustment in Step S2.1: Area 1: Areas where geometric features are missing at equipment connections and structural junctions (the original Gaussian representation is insufficient for these areas, resulting in missing details). Region 2: Large areas within the scene covered by the created Gaussian point set (the original Gaussian distribution in these areas is sparse, indicating insufficient accuracy); a large area is defined as: the portion of the scene covered by the Gaussian point set whose spatial range exceeds the preset range (such as large open areas of a construction site, the main area of large equipment, or areas with a large number of structures); large areas within the scene covered by the Gaussian point set are located using a combined loss function: because the original Gaussian distribution in large areas is small and has low density, the simulated image generated using these sparse Gaussians will show significant deviations in pixel color and structural details from the real image in this area—reflected in the loss function, the loss function value corresponding to this area will be significantly higher; For the region to be optimized, implement a dual-region dynamic parameter adjustment strategy based on engineering features: For region 1, first obtain the device connection type information and preset the corresponding feature weight coefficients for different connection types; by calculating the curvature mutation value of Gaussian points at the connection point, construct the mapping relationship of "curvature mutation value - encryption multiple". When the curvature mutation value is ≥ the preset threshold (e.g., 0.8), the number of Gaussian points per unit volume is encrypted to 3 times the original number. When the curvature mutation value is < the preset threshold, it is encrypted to 1.5 times the original number, so as to achieve accurate completion of the geometric feature missing area. For region 2, a density adjustment benchmark threshold is set. By calculating the distribution entropy of Gaussian points in region 2 in real time, when the distribution entropy is greater than or equal to the first density adjustment benchmark threshold of 1.2 (dense equipment area), the density of Gaussian points is increased by 2 times; when the distribution entropy is less than the second density adjustment benchmark threshold of 0.5 (open area), the density of Gaussian points is reduced by 30%, balancing reconstruction accuracy and computing cost.
[0034] Before performing the above parameter adjustments, a multi-constraint collaborative optimization selection is performed on the Gaussian points in the region to be optimized, specifically including: First, import the equipment ledger data of the substation project, perform semantic annotation on all Gaussian points, retain the core equipment or structure Gaussian points related to transformers, switchgear, and frameworks, and remove irrelevant background Gaussian points. Then, a neighborhood connectivity graph of Gaussian points is constructed, and the neighborhood search radius is set to R (where R = 5cm). Isolated Gaussian points with fewer than 5 points in the neighborhood are removed. The spacing between connected components is calculated. When the spacing is greater than the engineering accuracy threshold D (where D = 3cm), Gaussian points are densified in the spacing region to ensure the topological connectivity of the device structure.
[0035] A feedback optimization mechanism is initiated for the adjusted Gaussian point set: First, the combined loss function is recalculated for the adjusted Gaussian point set, and the gradient rate of change of the loss function is extracted. If the gradient rate of change is > 0.3, the current adjustment parameters are retained and the next iteration is initiated; if the gradient rate of change is < 0.1, the adjustment parameters are corrected in reverse. Then, the adjusted Gaussian point set (optimized scene) is projected to generate a two-dimensional image. Key engineering dimensions are extracted from the two-dimensional image and compared with the corresponding dimensions in the substation engineering design drawings. When the dimension deviation is > ±5cm, a secondary fine-tuning of the parameters is triggered to ensure that the scene reconstruction meets both visual fidelity and the actual engineering dimension requirements.
[0036] Step S2.3: Based on the results of parameter adjustment in Step S2.2, with the core objective of achieving rapid overall reconstruction and rapid sorting of large-scale substation engineering scenarios, Gaussian rasterization technology is used to perform a three-dimensional Gaussian rasterization iterative reconstruction process, such as... Figure 4 As shown; specifically: Step S2.31: Filter and remove Gaussian points in extreme positions: Filter Gaussian points according to the field of view and the cyclic block situation, and use a guardband to remove Gaussian points in extreme positions (mean is close to the near plane but far from the view). The definition of field of view: refers to the three-dimensional spatial range that the "camera" (or observation view) can cover when rendering / reconstructing a substation engineering scene (i.e., the visible area under the observation view); the definition of looping blocks: since substation engineering is a large-scale scene, the entire three-dimensional scene will be divided into multiple small blocks; the field of view and looping blocks are the current field of view range (observation view) plus the range of the scene block currently being looped to. Step S2.32, Gaussian point instantiation and key assignment: For each overlapping cyclic block of Gaussian points, the corresponding Gaussian point is instantiated, and a key combining the view space depth and the block ID is assigned to each instance; Step S2.33, Gaussian point sorting: The Radix Sort algorithm of a single fast GPU is used to sort the corresponding Gaussian points according to the key, ensuring that the Gaussian points in the same cyclic block are arranged in depth order to avoid pixel occlusion logic errors during subsequent rendering; Step S2.34, Pixel rendering based on sorted Gaussian points: For the sorted Gaussian point set, traverse each Gaussian point in cyclic block and depth order. Calculate its coverage and contribution weight on the two-dimensional image plane using the projection formula of the three-dimensional Gaussian distribution. Combine the color information and covariance matrix of the Gaussian points, and accumulate the color value of each pixel using a weighted fusion method to generate a continuous and smooth intermediate image for scene reconstruction. The pixel color accumulation formula is as follows: ; In the formula, Represents pixels The final color value; Indicates the covered pixels The number of Gaussian points; Indicates the first The contribution weight of each Gaussian point to the corresponding pixel (determined by the covariance matrix of the Gaussian point and its distance from the pixel). Indicates the first The color value of a Gaussian point; Step S2.35, Reconstruction Result Verification and Optimization: Compare the intermediate reconstruction image generated in step S2.34 with the actual substation engineering scene image, and calculate the combined loss function value; if the loss function value is higher than the preset threshold (e.g., 0.05), return to step S2.2 to readjust the Gaussian point parameters (encrypt the missing feature areas, optimize the density of large areas), and repeat steps S2.31-S2.34; if the combined loss function value is lower than or equal to the preset threshold, output the scene reconstruction result of this iteration (after optimization in the current iteration, a complete and parameterized set of three-dimensional Gaussian point models that can represent the entire three-dimensional scene of the substation engineering), and complete the rapid re-iteration scene reconstruction.
[0037] In the parameter adjustment stage of rapid re-iteration scene reconstruction, this invention innovatively introduces a dual-region dynamic parameter adjustment strategy based on engineering features, a multi-constraint collaborative Gaussian point optimization selection mechanism, and an adaptive feedback parameter iterative optimization mechanism. By integrating the attributes of substation engineering equipment, semantic information, and engineering accuracy requirements, it achieves precise control of Gaussian point distribution. This not only solves the problems of inaccurate geometric feature completion and waste of computing power in large areas in traditional methods, but also ensures that the reconstruction results are highly consistent with the actual structure and size of the engineering, further improving the efficiency and accuracy of complex scene reconstruction.
[0038] The specific process of step S3 is as follows: Step S3.1: Initial orientation alignment of the preset parameter CAD model (a parametric 3D model used for feature matching with data points and providing precise design parameters for parts): Align the front of the preset parameter CAD model with the positive X-axis direction; align the top of the preset parameter CAD model with the positive Z-axis direction.
[0039] Step S3.2: Extract object-level key points from the fast re-iteration scene reconstruction results in Step S2: Step S3.21: Extract the intersection points between the planes of equipment and structures in the fast re-iteration scene reconstruction results of step S2; Step S3.22: Use OBB (Oriented Bounding Box) to segment the line segments where the intersection points between the planes of equipment and structures are located; Step S3.23: Use the endpoints of the segmented line segments as object-level key points for "equipment and structure".
[0040] Step S3.3: Complete Scale Calibration: Use the line segment containing the intersection point between the equipment and structure planes in the aligned CAD model with preset parameters as the target line segment, and use the line segment containing the intersection point between the equipment and structure planes in the rapid re-iteration scene reconstruction result of Step S2 as the source line segment; determine the scale of the equipment and structure by calculating the ratio of the Euclidean distance (L2 norm) between the target line segment and the source line segment; expressed as: ; In the formula, , These represent the first and second endpoints of the target line segment, respectively. , These represent the first and second endpoints of the source line segment, respectively. The scale parameter represents the length ratio between the target line segment and the source line segment, and is used to determine the scale of equipment and structures. This represents the L2 norm.
[0041] Step S3.4, Constraint Matching of Scene Data Points and Preset Parameter CAD Model: Using the normal vector angles of the corresponding faces of the scene data points in the fast re-iteration scene reconstruction results and the aligned preset parameter CAD model as constraints, optimize the distance between the normal vector angles of the corresponding faces and the object-level key points; optimizing the distance between the normal vector angles of the corresponding faces and the object-level key points means minimizing the distance between the normal vector angles of the corresponding faces and the object-level key points; minimizing the normal vector angles of the corresponding faces is expressed as: ; In the formula, Represents a constrained minimization optimization operator; This refers to the adjustment parameters (such as attitude, position, etc.) of the preset parameter CAD model, used to make the CAD model match the actual scene; This indicates the number of corresponding faces between the fast re-iteration scene reconstruction result and the aligned CAD model with preset parameters; In the result of fast re-iteration scene reconstruction, the first... Normal vectors of each face; This indicates the alignment of the preset parameter CAD model with the first [parameter] in the fast re-iteration scene reconstruction result. The normal vector of the face corresponding to each face; This represents a vector angle calculation function used to solve for the angle between two three-dimensional normal vectors.
[0042] Step S3.5, Model Fusion: The equipment and structures after scale calibration and normal vector angle constraint matching are embedded into the rapid re-iteration scene reconstruction results to complete the fusion of multi-source data and form the final substation engineering scene.
[0043] A second embodiment of the present invention also provides an electronic device, including a processor, a memory, and a bus, wherein the processor and the memory are connected through the bus, wherein the memory is used to store a set of program code, and the processor is used to call the program code stored in the memory to execute a method for reconstructing complex substation engineering scenarios based on three-dimensional Gaussian.
[0044] A third embodiment of the present invention also provides a non-volatile computer storage medium storing computer-executable instructions that execute a method for reconstructing complex substation engineering scenarios based on three-dimensional Gaussian.
[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for reconstructing complex substation engineering scenes based on three-dimensional Gaussian, characterized in that, Includes the following steps: Step S1: Scene optimization based on 3D Gaussian distribution: The distribution of scene data points in the substation engineering scene in 3D space is described by the 3D Gaussian distribution function, and the camera coordinate system analysis is introduced to optimize the scene representation. Step S2, Fast Iterative Scene Reconstruction: Based on the scene representation optimized in Step S1, fast iterative scene reconstruction is performed by integrating stochastic gradient descent, Gaussian function adaptive control, and fast rasterization algorithm. Step S3, Feature-based model coupling: Based on the rapid re-iteration scene reconstruction results, extract object-level key points including equipment and structures, and complete scale calibration; minimize the distance between key points by matching the scene data points in the rapid re-iteration scene reconstruction results with the normal vector angle constraints of the corresponding surfaces of the preset parameter CAD model; embed the equipment and structures after scale calibration and normal vector angle constraint matching into the rapid re-iteration scene reconstruction results to complete the fusion of multi-source data and form the final substation engineering scene.
2. The method for reconstructing complex substation engineering scenes based on three-dimensional Gaussian as described in claim 1, characterized in that: The specific process of step S1 is as follows: Step S1.1: First, perform a three-dimensional Gaussian distribution analysis. Use the three-dimensional Gaussian distribution function to describe the distribution of scene data points in the three-dimensional space in the substation engineering scenario. The mean of the center point of the scene data point set represents the center position of the scene data point. The covariance matrix in the three-dimensional Gaussian distribution function is used to describe the degree and direction of change of scene data points in different dimensions. Step S1.2: Conduct 3D Gaussian sputtering analysis, treating each scene data point as a Gaussian point. Calculate the color value of each pixel by projecting the Gaussian point onto a 2D plane to generate a 2D image. During this process, camera coordinate system analysis is introduced: First, the projection result of the substation engineering scene onto the 2D plane is obtained through the camera, and the position and distribution of the scene data points are analyzed in the coordinate system to map the scene data points to the 2D plane projection result. Simultaneously, the scene data points in 3D space are matched with the corresponding 2D images. Step S1.3: After completing the 3D Gaussian sputtering analysis, optimize the scene representation: First, by introducing a 3D Gaussian distribution representation, the substation engineering scene is represented as a probability distribution; then, the diagonal elements of the 3D Gaussian covariance matrix are used to represent the variance of each coordinate axis of the 3D global coordinate system of the substation engineering scene, reflecting the extent of expansion of equipment and structures in the substation engineering scene. At the same time, the off-diagonal elements of the 3D Gaussian covariance matrix are used to represent the covariance of each coordinate axis of the 3D global coordinate system of the substation engineering scene.
3. The method for reconstructing complex substation engineering scenes based on three-dimensional Gaussian as described in claim 2, characterized in that: The specific process of step S2 is as follows: Step S2.1: Based on the scene representation optimized in step S1, create a set of Gaussian points to represent the free-view synthetic scene of the substation project; set the initial covariance matrix of the created Gaussian points; based on the created Gaussian points and the corresponding initial covariance matrix, synthesize the simulated image of the substation project scene. Define a combined loss function of pixel color loss and structural similarity loss; use the combined loss function as the evaluation criterion to compare the simulated image of the substation engineering scene with the real substation engineering scene image and quantify the difference between the two; introduce the stochastic gradient descent algorithm, with the goal of minimizing the combined loss function, and iteratively solve to obtain the adjustment direction and magnitude parameters of the equipment / structure graphics; Based on the results of the stochastic gradient descent optimization, the graphics of the equipment / structures that were incorrectly located during the projection process in step S1.2 are adjusted. Step S2.2: Based on the analysis of the combined loss function, two types of regions to be optimized are identified from the scene representation after optimization and adjustment in Step S2.1: Area 1: Areas where geometric features are missing, including equipment connections and structural junctions; Region 2: The region covered by the Gaussian point set whose spatial range is larger than the preset value; A dual-region dynamic parameter adjustment strategy based on engineering features is implemented for the region to be optimized. Step S2.3: Based on the results of parameter adjustment in step S2.2, Gaussian rasterization technology is used to perform a three-dimensional Gaussian rasterization iterative reconstruction process.
4. The method for reconstructing complex substation engineering scenes based on three-dimensional Gaussian as described in claim 3, characterized in that: The dual-region dynamic parameter adjustment strategy based on engineering features for the region to be optimized includes: For region 1, first obtain the device connection type information and preset the corresponding feature weight coefficients for different connection types; by calculating the curvature mutation value of Gaussian points at the connection point, construct the "curvature mutation value - encryption multiple" mapping relationship. When the curvature mutation value is ≥ the preset threshold, the number of Gaussian points per unit volume is encrypted to the first preset multiple of the original number. When the curvature mutation value is < the preset threshold, it is encrypted to the second preset multiple of the original number. For region 2, a density adjustment benchmark threshold is set. By calculating the distribution entropy of Gaussian points in region 2 in real time, the density of Gaussian points is increased when the distribution entropy is greater than or equal to the first density adjustment benchmark threshold; and the density of Gaussian points is decreased when the distribution entropy is less than the second density adjustment benchmark threshold.
5. The method for reconstructing complex substation engineering scenes based on three-dimensional Gaussian as described in claim 4, characterized in that: The specific process of step S2.3 is as follows: Step S2.31: Filter and remove Gaussian points at extreme positions: Filter Gaussian points according to the field of view and the cyclic block situation, and at the same time use a guard band to remove Gaussian points at extreme positions; Step S2.32, Gaussian point instantiation and key assignment: For each overlapping cyclic block of Gaussian points, the corresponding Gaussian point is instantiated, and a key combining the view space depth and the block ID is assigned to each instance; Step S2.33, Gaussian sort: Use the radix sort algorithm to sort the corresponding Gaussian points according to the key; Step S2.34, Pixel rendering based on sorted Gaussian points: For the sorted Gaussian point set, traverse each Gaussian point in cyclic block and depth order, calculate its coverage and contribution weight in the two-dimensional image plane through the projection formula of the three-dimensional Gaussian distribution, combine the color information and covariance matrix of the Gaussian points, and accumulate the color value of each pixel by weighted fusion to generate a continuous and smooth intermediate image for scene reconstruction. Step S2.35, Reconstruction Result Verification and Optimization: Compare the intermediate scene reconstruction image generated in step S2.34 with the actual substation engineering scene image, and calculate the combined loss function value; if the loss function value is higher than the preset threshold, return to step S2.2 to readjust the Gaussian point parameters, and repeat steps S2.31-S2.34; if the combined loss function value is lower than or equal to the preset threshold, output the scene reconstruction result of this iteration, and complete the rapid re-iteration scene reconstruction.
6. The method for reconstructing complex substation engineering scenes based on three-dimensional Gaussian as described in claim 5, characterized in that: The specific process of step S3 is as follows: Step S3.1, Preset Parameter CAD Model Initial Orientation Alignment: Set the preset parameters to align the front of the CAD model with the positive X-axis direction; The top of the CAD model is aligned with the positive Z-axis direction according to the preset parameters. Step S3.2: Extract object-level key points from the fast re-iteration scene reconstruction results in Step S2; Step S3.3: Complete scale calibration: Take the line segment where the intersection point between the planes of equipment and structures in the aligned preset parameter CAD model is located as the target line segment, and take the line segment where the intersection point between the planes of equipment and structures in the rapid re-iteration scene reconstruction result of step S2 as the source line segment. The dimensions of equipment and structures are determined by calculating the Euclidean distance ratio between the target line segment and the source line segment. Step S3.4, Constraint matching between scene data points and preset parameter CAD model: Using the normal vector angle of the corresponding face of the scene data points in the fast re-iteration scene reconstruction result and the aligned preset parameter CAD model as constraints, optimize the normal vector angle of the corresponding face and the distance between object-level key points. Step S3.5, Model Fusion: The equipment and structures after scale calibration and normal vector angle constraint matching are embedded into the rapid re-iteration scene reconstruction results to complete the fusion of multi-source data and form the final substation engineering scene.
7. The method for reconstructing complex substation engineering scenes based on three-dimensional Gaussian as described in claim 6, characterized in that: The specific process of step S3.2 is as follows: Step S3.21: Extract the intersection points between the planes of equipment and structures in the fast re-iteration scene reconstruction results of step S2; Step S3.22: Use a directional bounding box to divide the line segments where the intersection points between the equipment and structure planes are located; Step S3.23: Use the endpoints of the segmented line segments as object-level key points for equipment and structures.
8. An electronic device, characterized in that, The system includes a processor, a memory, and a bus. The processor and the memory are connected via the bus. The memory is used to store a set of program code, and the processor is used to call the program code stored in the memory to execute the method for reconstructing complex substation engineering scenarios based on three-dimensional Gaussian as described in any one of claims 1-7.
9. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer can execute instructions to perform the method for reconstructing complex substation engineering scenes based on three-dimensional Gaussian as described in any one of claims 1-7.