Power distribution network scene three-dimensional reconstruction method based on 3DGAS

Abstract

The invention discloses a power distribution network scene three-dimensional reconstruction method based on 3DGAS, and particularly relates to the technical field of three-dimensional scene reconstruction, and the method comprises the following steps: obtaining a multi-view power distribution network scene image through preset sampling equipment, so as to completely cover the equipment and the surrounding environment; inputting the image into a computer vision system, generating sparse point clouds by adopting a matching and reconstruction algorithm, defining discrete Gaussian point clouds based on the sparse point clouds, and simulating three-dimensional radiation field characteristics; changing rendering of color values under different visual angles is achieved through spherical harmonic coefficients, traditional projection errors and point cloud distribution characteristics are evaluated, and a projection strategy is dynamically selected to generate a two-dimensional image; performing depth sorting on the two-dimensional Gaussian distribution, obtaining a two-dimensional image and point cloud data through rendering, and finally forming a three-dimensional Gaussian scene model; according to the method, complex manual intervention and trial and error processes in a traditional method are reduced, projection strategies such as unified weight or ellipse weighted average can be intelligently selected, errors are effectively reduced, and the reconstruction precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional scene reconstruction, and more specifically, to a three-dimensional reconstruction method for a distribution network scene based on 3DGAS. Background Art

[0002] The distribution network is an important part of the power system, and its facilities are widely distributed and complex, including key equipment such as transformers, switchgear, cable trays and transmission lines. Traditional distribution network scene monitoring and planning mainly rely on two-dimensional images and manual on-site surveys, which have the following limitations: Limitations of two-dimensional data: Traditional plane images and engineering drawings are difficult to fully reflect the spatial layout and structural details in the distribution network scene. The manual method is inefficient: On-site surveys require a lot of manpower, and there are errors in the accurate measurement of complex environments. Poor dynamic adaptability: It is impossible to dynamically adjust the modeling strategy according to the scene characteristics to meet the needs of different scene characteristics. Therefore, the present invention proposes a three-dimensional reconstruction method for distribution network scenes based on 3DGAS, in order to solve the above problems. Summary of the invention

[0003] To achieve the above object, the present invention provides the following technical solutions:

[0004] The 3DGAS-based 3D reconstruction method for distribution network scenes includes the following steps:

[0005] Multiple images of the distribution network scene are obtained through a preset sampling device, and the multiple images cover multiple perspectives of the distribution network equipment and its surrounding environment;

[0006] Multiple images are input into a computer vision system, and a sparse point cloud is generated using an image matching and reconstruction algorithm. Based on the sparse point cloud, the radiation field in the three-dimensional scene is defined as a discrete Gaussian point cloud.

[0007] The spherical harmonic coefficients are used to simulate the changes in color values ​​under different viewing angles, perform viewing angle-dependent color rendering, evaluate the errors of traditional projection methods in point cloud data mapping, and evaluate the spatial distribution characteristics of point cloud data. Based on the evaluation results, a corresponding strategy is selected to project the three-dimensional Gaussian point cloud onto a two-dimensional plane to generate a two-dimensional image.

[0008] Countless two-dimensional Gaussian distributions are deeply sorted, and a clear three-dimensional Gaussian scene model is formed through the rendered two-dimensional image and processed point cloud data.

[0009] In a preferred embodiment, multiple images are input into a computer vision system, and a sparse point cloud is generated using an image matching and reconstruction algorithm, which means:

[0010] In multiple images, the correspondence is found by matching the feature points in each image, and the space reconstruction under different perspectives is realized based on these correspondences. The camera is calibrated, and the position of each 3D point is determined by triangulation through the matched feature points and the internal and external parameters of the calibrated camera to form a sparse point cloud.

[0011] In a preferred embodiment, when constructing a Gaussian point cloud, the point cloud is divided into multiple clusters by the clustering method K-means. The points in each cluster collectively represent an "area" in the scene. Each cluster is modeled with a Gaussian distribution to obtain a parameter combination for each Gaussian point cloud.

[0012] In a preferred embodiment, the parameter combination of each Gaussian point cloud includes the following parameters: position, covariance, color, and opacity.

[0013] In a preferred embodiment, evaluating the error of the traditional projection method in point cloud data mapping refers to:

[0014] Suppose there is an original data set X, which contains N samples, each sample has M features, that is, the dimension of the data set is N×M. Suppose the projected data set is the data set Y obtained after dimensionality reduction, and its dimension is N×K, where K is the dimension of the projected data.

[0015] The projection matrix P is a matrix that maps the original data to a low-dimensional space. Its dimension is M×K, that is, the transformation matrix from the original data space to the low-dimensional space. The projected data Y can be obtained by the following formula: Y=XP, where P is the projection matrix, X is the original data matrix, and Y is the projected data.

[0016] By calculating the reconstruction error, we can measure the difference between the projected data Y and the original data X when reconstructed back to the original space through inverse projection:

[0017] E=‖XX′‖ F ; X′ is the reconstructed data, E is the reconstruction error, ‖·‖ F is the Frobenius norm, which is used to measure the overall error of the matrix;

[0018] The projection error index can be calculated by the following formula:

[0019] ‖X‖ F is the Frobenius norm of the original data, PEI is the projection error index, which is used to reflect the error degree of the traditional projection method in the point cloud data mapping, ‖X‖ F Defined as:

[0020] ‖X‖ FIt is obtained by calculating the square root of the sum of the squares of the elements of the original data matrix. i and j are the position indexes of the data in the matrix.

[0021] In a preferred embodiment, evaluating the spatial distribution characteristics of point cloud data refers to:

[0022] The point cloud data is divided into several volume units of fixed size, the points in each unit are counted, and the point density of the unit is calculated. The density mean and standard deviation of all units are calculated to evaluate the density inhomogeneity of the point cloud.

[0023] In a preferred embodiment, selecting a corresponding strategy according to the evaluation result refers to:

[0024] The projection error index, density mean, and density standard deviation are input into the pre-trained machine learning model together, and the output result is 1 or 0. When the output result is 1, the strategy selected is the projection strategy of uniform weight, and when the output result is 0, the strategy selected is the projection strategy of elliptical weighted average.

[0025] In a preferred embodiment, the pre-trained machine learning model is a convolutional neural network model.

[0026] Technical effects and advantages of the present invention:

[0027] The present invention can comprehensively capture the geometric features of the distribution network scene through multi-view image acquisition and sparse point cloud generation technology, and utilizes the Gaussian point cloud modeling method to convert the sparse point cloud into a structured three-dimensional data model, thereby improving the spatial description capability of the point cloud.

[0028] The present invention can intelligently select projection strategies (such as uniform weight or elliptical weighted average) according to the lighting conditions, equipment density and distribution characteristics in the distribution network scene, effectively reducing errors and improving reconstruction accuracy. In the face of changes in scene complexity, this method ensures the stability and consistency of model reconstruction by dynamically adjusting weights and parameters.

[0029] The present invention uses spherical harmonic coefficients to achieve perspective-dependent color rendering, accurately simulates illumination changes and texture features under different perspectives, and enhances the realism and detail performance of the three-dimensional scene model by introducing the transparency and color parameters of the Gaussian point cloud. The K-means clustering algorithm is used to cluster the point cloud data, and combined with the Gaussian distribution modeling technology, it can quickly extract the key areas in the scene and efficiently calculate the parameters.

[0030] With the help of pre-trained convolutional neural network models, the distribution characteristics of point cloud data are intelligently analyzed, reducing the complex manual intervention and trial-and-error process in traditional methods. This method is not only suitable for the reconstruction of distribution network scenes, but can also be widely used in other scenes such as urban planning, digital twins and virtual reality, providing high-precision 3D modeling solutions for different fields. The automated 3D reconstruction process reduces the manpower and time costs of traditional on-site mapping. Through accurate scene modeling, the operation and maintenance and monitoring efficiency of equipment is improved, and the operation and maintenance risks of the power system are reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to facilitate understanding by those skilled in the art, the present invention is further described below in conjunction with the accompanying drawings;

[0032] Figure 1 This is a schematic diagram of the 3DGAS-based three-dimensional reconstruction method for distribution network scenes in the present invention. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] Reference Figure 1 The following embodiments are obtained:

[0035] Embodiment 1:

[0036] In the process of surveying and managing the distribution network, traditional two-dimensional or low-quality three-dimensional reconstruction technology often fails to truly and accurately display the structure of distribution lines, equipment and the surrounding environment. By adopting three-dimensional Gaussian point cloud technology, higher quality and more realistic reconstruction effects can be provided in terms of details and perspective changes, thus helping managers make more accurate decisions in remote monitoring, fault detection, line optimization and other aspects.

[0037] Traditional 3D reconstruction methods (such as reconstruction based on depth maps or stereo vision) often face problems such as low resolution, insufficient accuracy, and large amount of calculation when dealing with complex scenes. 3D Gaussian point cloud technology can not only efficiently describe the position, shape, color and other information of objects in space by representing the scene as a discrete Gaussian point cloud, but also accurately describe the details of the object through parameters such as the covariance and opacity of the Gaussian points. Especially in complex environments, it can better reconstruct distribution network equipment and its surrounding environment.

[0038] The present invention adopts differentiable volume rendering technology and simulates color changes under different viewing angles through spherical harmonic coefficients, which can effectively improve the accuracy and quality of the rendering effect. At the same time, when necessary, elliptical weighted averaging technology is used to achieve efficient projection from three-dimensional to two-dimensional, so that more point cloud data can be processed during image rendering, reducing the reliance on complexity and calculation in traditional rendering methods, thereby improving the efficiency of the system.

[0039] In the 3D reconstruction of distribution networks, especially when generating point clouds using images collected by drones and other means, the reconstruction effect of small objects is often not ideal. Technology based on 3D Gaussian point clouds helps improve the reconstruction effect of small objects through depth sorting and transparency blending, thereby improving the accuracy and visualization of the entire distribution network scene.

[0040] The survey of distribution networks usually requires efficient and accurate acquisition of on-site three-dimensional information, especially in complex environments where the relationship between distribution network equipment and the surrounding environment needs to be clearly visible. Generating sparse point clouds from multiple images acquired by drones and then performing three-dimensional reconstruction using Gaussian point cloud technology can significantly improve the quality of the reconstruction effect, meet the high-precision requirements in the actual survey process, and help engineers monitor, maintain and plan the distribution network more efficiently.

[0041] With the development of smart grid, automation equipment and Internet of Things technology, the management of distribution network will gradually turn to a more intelligent direction. 3D Gaussian point cloud technology provides a new technical means for 3D reconstruction of distribution network scenes, which will help the digital and intelligent management of future distribution network systems, making the status monitoring, fault diagnosis, emergency handling, etc. of distribution networks more automated and intelligent.

[0042] This paper proposes a 3D reconstruction method for distribution network scenes based on 3D Gaussian point cloud technology, which aims to improve the accuracy, efficiency and rendering quality of 3D reconstruction, solve the limitations of traditional 3D reconstruction technology in distribution network applications, meet actual survey and management needs, and improve the visualization and intelligence level of distribution networks. This technical solution can not only provide clearer and more accurate 3D scenes, but also effectively cope with the challenges of reconstructing small objects, thus providing strong support for the management, maintenance and development of distribution networks.

[0043] Therefore, a 3DGAS-based 3D reconstruction method for distribution network scenes is proposed, which includes the following steps:

[0044] Multiple images of the distribution network scene are obtained through preset sampling devices. The multiple images cover multiple perspectives of the distribution network equipment and its surrounding environment. This step is the data collection stage. Images of the distribution network equipment and its surrounding environment are taken from multiple perspectives to obtain as much perspective information as possible. Images from multiple perspectives help improve the accuracy and completeness of 3D reconstruction because images from different perspectives contain different geometric and texture information of the scene. This ensures that richer and more accurate 3D data can be obtained in the subsequent image matching and reconstruction process.

[0045] Multiple images are input into the computer vision system, and sparse point clouds are generated using image matching and reconstruction algorithms. Based on the sparse point clouds, the radiation field in the three-dimensional scene is defined as a discrete Gaussian point cloud; this stage is the core of three-dimensional reconstruction. The computer vision system extracts feature points in the image through image matching algorithms (such as feature point matching) and generates sparse point clouds through triangulation techniques. Sparse point clouds are composed of discrete points in three-dimensional space, which represent the locations of important feature points in the scene. Although these point clouds are sparse, they are the basis for building a complete three-dimensional model. Further processing of the point cloud converts the sparse point cloud into a "Gaussian point cloud". Each point is regarded as a Gaussian distribution to simulate the spatial characteristics and radiation characteristics of the point cloud (such as lighting, color, etc.). Gaussian point clouds approximate the shape of the surface of an object or scene through mathematical models. These Gaussian distributions can more accurately describe the properties of each point in the scene, such as lighting and color, which helps to improve the realism and visual effects of the image, especially when rendering.

[0046] The spherical harmonic coefficients are used to simulate the changes in color values ​​under different viewing angles, perform viewing angle-dependent color rendering, evaluate the errors of traditional projection methods in point cloud data mapping, and evaluate the spatial distribution characteristics of point cloud data. According to the evaluation results, the corresponding strategy is selected to project the three-dimensional Gaussian point cloud onto a two-dimensional plane to generate a two-dimensional image; the spherical harmonic function is a mathematical tool used to describe the illumination distribution on a spherical surface. Here, the spherical harmonic coefficients are used to simulate the changes in color and illumination of the point cloud surface under different viewing angles. Specifically, the spherical harmonic coefficients can be used to describe the impact of the lighting environment on each point in the scene, and the color and brightness will also change as the viewing angle changes. In this way, the illumination changes and color rendering in the real world can be simulated more realistically. Traditional projection methods (such as plane projection) may introduce errors in the three-dimensional to two-dimensional mapping process. These errors are mainly reflected in the distortion of geometric shapes and the inaccuracy of color rendering. This step ensures that the projection method used can minimize these errors by evaluating the projection errors. In addition, the spatial distribution characteristics of the point cloud data, such as the density of the point cloud and the uniformity of distribution, are also evaluated to better select a suitable projection strategy. According to the results of the previous evaluation, a suitable projection method is selected to convert the three-dimensional Gaussian point cloud into a two-dimensional image. The goal of this stage is to minimize the geometric distortion caused by projection while preserving the lighting and texture information of the 3D scene to adapt to different viewpoints and scene characteristics.

[0047] Countless two-dimensional Gaussian distributions are deeply sorted, and a clear three-dimensional Gaussian scene model is formed by rendering the obtained two-dimensional images and the processed point cloud data. In this stage, the Gaussian point cloud after two-dimensional projection is deeply sorted, that is, the display order of each point in the final image is determined for correct lighting and color rendering. The purpose of this step is to synthesize all two-dimensional images into a clear and realistic three-dimensional Gaussian scene model through depth sorting and color rendering. Depth sorting is an important technology for solving occlusion problems. It ensures that objects in the scene are correctly displayed according to their positions in three-dimensional space, avoiding visual incoherence or confusion.

[0048] Inputting multiple images into a computer vision system and using image matching and reconstruction algorithms to generate a sparse point cloud means:

[0049] In multiple images, the correspondence is found by matching the feature points in each image, and the space reconstruction under different perspectives is realized based on these correspondences. The camera is calibrated, and the position of each 3D point is determined by triangulation through the matched feature points and the internal and external parameters of the calibrated camera to form a sparse point cloud.

[0050] Specifically: First, the multiple collected images need to be preprocessed to ensure the quality of the data and the feasibility of subsequent processing. Preprocessing includes image denoising, distortion correction, color correction, etc. Then, key feature points (such as corners, edges or other significant features) are extracted from the image for subsequent matching and reconstruction. Common feature extraction methods include: SIFT (Scale Invariant Feature Transform): Scale and rotation invariant feature description of key points in the image. SURF (Speeded Up Robust Features): Compared with SI FT, the SURF algorithm is faster to calculate and suitable for real-time applications. ORB: An efficient binary descriptor with fast calculation speed and suitable for large-scale matching.

[0051] In multiple images, correspondences are found by matching feature points in each image, and spatial reconstruction under different perspectives is achieved based on these correspondences. The goal of image matching is to find correspondences between points of the same object in different images. Since there may be false matches in feature matching (such as noise, background, etc.), geometric verification is needed to eliminate false matches. The most commonly used geometric verification method is the RANSAC (Random Sample Consensus) algorithm. This method estimates the projection matrix by randomly selecting point pairs, and then verifies which matching points conform to the model.

[0052] Through the matching image feature points, stereo vision or multi-view geometry methods are used to perform 3D reconstruction and generate a sparse point cloud. Camera calibration is a necessary step in the reconstruction process, the purpose of which is to determine the internal and external parameters of the camera (the camera's intrinsic matrix and extrinsic matrix). Assume that the camera's intrinsic matrix is ​​U and the extrinsic matrix is ​​[R|T], where R is the rotation matrix and T is the translation vector. The projection model of the camera is:

[0053] (X, Y, Z) are the coordinates of a point in three-dimensional space, (x', y') are the coordinates of a point in a two-dimensional image, and s is the scale factor (usually 1). Through camera calibration, the internal and external parameters of the camera are obtained, and then the points on the image can be matched with the points in three-dimensional space.

[0054] The position of the 3D point is estimated by using the matching feature points and the internal and external parameters of the camera using triangulation. Suppose there are two images, where the matching points of the first image are (x1, y1) and the matching points of the second image are (x2, y2). The internal and external parameters of the camera are U1, [R1|T1] and U2, [R2|T2] respectively. According to the geometric principle of triangulation, the 3D point P (X, Y, Z) satisfies the following projection relationship:

[0055] λ1 and λ2 are both scale factors, which can be solved by the least squares method. The existing technology will not be described here. In this way, the three-dimensional point P (X, Y, Z) can be calculated from the two-dimensional image points under multiple perspectives to form a sparse point cloud. The generated point cloud is usually sparse and needs to be further improved through optimization algorithms, such as Bundle Adjustment: This is the most commonly used optimization algorithm, which optimizes the camera parameters and the position of the point cloud by minimizing the reprojection error. The objective function is:

[0056] x ij is the projection of the i-th feature point in the j-th image, π is the projection function, and in order to improve the accuracy of feature matching, an adaptive feature matching method can be introduced, such as RANSAC (random sampling consensus algorithm): the RANSAC algorithm is applied in the feature matching process to remove false matches and improve the accuracy of the triangulation results.

[0057] When constructing a Gaussian point cloud, the point cloud is divided into multiple clusters using the K-means clustering method. The points in each cluster collectively represent a "region" in the scene. Each cluster is modeled using a Gaussian distribution to obtain a parameter combination for each Gaussian point cloud. Specifically:

[0058] Assume that when constructing the Gaussian point cloud, a sparse point cloud set is obtained Where each point p i It is obtained from a 3D scene, a total of N points are obtained, and the point cloud is divided into multiple clusters through clustering methods (such as K-means). The points in each cluster are relatively close, representing an "area" in the scene. Each cluster can be modeled with a Gaussian distribution to obtain the parameters of each Gaussian point cloud.

[0059] Position p i : Calculate the centroid or mean of each cluster as the center position of the Gaussian point cloud:

[0060] C i is the i-th cluster, |C i | is the number of points in the i-th cluster, p j is the position of the jth point in the cluster, p i is the center position of the ith cluster.

[0061] Covariance matrix∑ i : Use the distribution of points within the cluster to calculate the covariance matrix to describe the scalability of the cluster point cloud:

[0062] This covariance matrix reflects the distribution of points within the cluster and determines the size and shape of the Gaussian point cloud.

[0063] Color ci : For each Gaussian point cloud, the color average can be calculated: c j Yes j color.

[0064] Opacity i : Opacity can be obtained by counting the point cloud density of each cluster.

[0065] The parameter combination of each Gaussian point cloud includes the following parameters: position, covariance, color and opacity. All Gaussian point clouds can be synthesized according to their parameters (position, covariance, color, opacity) to obtain an approximate representation of a complete 3D scene.

[0066] Using spherical harmonic coefficients to simulate the change of color values ​​at different viewing angles, and performing perspective-dependent color rendering means:

[0067] Spherical harmonic basis functions are orthogonal functions that describe any direction (longitude, latitude) on the sphere, and can usually be expressed by the following formula:

[0068] l is the order of the spherical harmonics, which determines the complexity and frequency of the function. m is the frequency of the azimuth angle, which satisfies the range of [-l, l]. θ and φ are the polar angle and azimuth angle in the spherical coordinate system, respectively. is the adjoint Legendre polynomial associated with the spherical harmonics, and the spherical harmonic coefficients c l , m is the projection of the basis function on a specific spherical data. For a lighting or color value function I(θ,φ) on a sphere, the spherical harmonic coefficients can be calculated by the following formula:

[0069]

[0070] In the context of color rendering, spherical harmonic coefficients are often used to represent the distribution of color on a sphere. For example, assuming that the change in color with viewing angle can be represented by directions on a sphere, it can be represented by expansion:

[0071] C(θ,φ) is the distribution of color on the sphere, c l,m are the spherical harmonic coefficients, and L is the maximum order (which also determines the accuracy of the spherical harmonic expansion).

[0072] View-dependent color rendering, especially lighting modeling in computer graphics, relies on the expansion of spherical harmonic coefficients to handle color changes at different view angles. The key steps include:

[0073] Calculation of spherical harmonic coefficients: In order to simulate the color changes under different viewing angles, the spherical harmonic coefficients under different viewing angles are first calculated through spherical harmonic expansion. This process is usually completed in the preprocessing stage. For example, for a texture image or an ambient light map (such as an HDR image), its spherical harmonic coefficients can be extracted to represent the color and light distribution.

[0074] Calculation of viewing angle and color adjustment: During the rendering process, the first thing to do is to determine the viewer's viewing angle (i.e. θ vicw ,φ vicw ). Then adjust the spherical harmonic coefficients according to the viewing angle and lighting conditions. By selecting the appropriate order L, you can decide how many frequency components to retain to adjust the color details. Usually, the lower-order spherical harmonic coefficients capture large-scale lighting information, while higher-order coefficients can represent more detailed local changes.

[0075] Reconstruction of spherical harmonic basis functions: According to the spherical harmonic coefficients adjusted by the viewing angle, spherical harmonic basis functions are used for reconstruction to obtain a new color distribution:

[0076] The pixels in the scene are then rendered using this distribution. Calculating the spherical harmonic coefficients requires integrating the entire sphere, which is a large computational overhead in real-time rendering. By pre-calculating radiation transfer, the lighting can be pre-calculated in the scene, and the ambient lighting information can be stored using the spherical harmonic coefficients, allowing for quick query of lighting from different perspectives during rendering. This method improves the efficiency of lighting calculations in real-time rendering.

[0077] Evaluating the error of traditional projection methods in point cloud data mapping refers to:

[0078] Suppose there is an original data set X, which contains N samples, each sample has M features, that is, the dimension of the data set is N×M. Suppose the projected data set is the data set Y obtained after dimensionality reduction, and its dimension is N×K, where K is the dimension of the projected data.

[0079] The projection matrix P is a matrix that maps the original data to a low-dimensional space. Its dimension is M×K, that is, the transformation matrix from the original data space to the low-dimensional space. The projected data Y can be obtained by the following formula: Y=XP, where P is the projection matrix, X is the original data matrix, and Y is the projected data.

[0080] By calculating the reconstruction error, we can measure the difference between the projected data Y and the original data X when reconstructed back to the original space through inverse projection:

[0081] E=‖XX′‖ F ; X′ is the reconstructed data, E is the reconstruction error, ‖·‖ Fis the Frobenius norm, which is used to measure the overall error of the matrix; the reconstruction error reflects the error between the projection and the reconstruction.

[0082] The projection error index can be calculated by the following formula:

[0083] ‖X‖ F is the Frobenius norm of the original data, PEI is the projection error index, which is used to reflect the error degree of the traditional projection method in the point cloud data mapping, ‖X‖ F Defined as:

[0084] ‖X‖ F It is obtained by calculating the square root of the sum of the squares of the elements of the original data matrix. i and j are the position indexes of the data in the matrix.

[0085] The smaller the PEI value, the smaller the difference between the projected data and the original data, and the smaller the error in the projection process. Ideally, if the projection is perfect, then X' = X, PE I = 0. When the PEI value is large, it means that the projection error is large and a lot of correct information is lost.

[0086] By calculating the change in point density in different regions or volume units, the distribution characteristics of the point cloud can be evaluated. Evaluating the spatial distribution characteristics of point cloud data refers to:

[0087] The point cloud data is divided into several fixed-size volume units, the points in each unit are counted, and the point density of the unit is calculated. The density mean and standard deviation of all units are calculated to evaluate the density inhomogeneity of the point cloud. When reconstructing three-dimensional terrain or buildings, the distribution characteristics of point cloud data directly affect the accuracy and detail of the model.

[0088] Choosing the corresponding strategy based on the evaluation results means:

[0089] The projection error index, density mean, and density standard deviation are input into the pre-trained machine learning model. The output result is 1 or 0. When the output result is 1, the projection strategy with uniform weight is selected. When the output result is 0, the projection strategy with elliptical weighted average is selected. If the point cloud data is unevenly distributed in space, elliptical weighted average may be a better choice because it can dynamically adjust the weight according to the spatial position of each point. If the error of the traditional projection method is small and the applicability is strong, the traditional method, i.e., the projection strategy with uniform weight, may be selected. When the data is relatively uniform or the error is small, the projection strategy with uniform weight can be selected, such as parallel projection. Parallel projection refers to a projection method in which the projection lines are parallel. It includes different types such as orthographic projection, axonometric projection, and oblique projection. In parallel projection, the projection lines are parallel, and the distance of the object will not affect its projection size, similar to orthographic projection. Parallel projection will not have a perspective effect either. Explanation of uniform weight: The lengths of different projection lines in parallel projection remain unchanged, and will not shrink or enlarge due to the distance of the object. The relative position and shape of the object will not change due to the perspective effect. Similar to orthographic projection, parallel projection can also be considered uniformly weighted, especially when orthographic projection is a special case of parallel projection.

[0090] The pre-trained machine learning model is a convolutional neural network model. By inputting features such as projection error index, density mean, and density standard deviation into the pre-trained machine learning model, the system can evaluate the quality and spatial distribution characteristics of point cloud data. Then, the result (1 or 0) output by the model determines which projection strategy to choose later. The core significance of this mechanism is:

[0091] Intelligent decision-making: Traditional projection methods or weighted average methods have their own advantages and disadvantages, while machine learning models can integrate these features and dynamically make reasonable decisions based on the specific circumstances of the data.

[0092] Reduced human intervention: Machine learning models can automatically select the most appropriate strategy without the need for manual adjustment, which can improve processing efficiency and reduce human errors.

[0093] When the model output is 1, a uniform weight projection strategy is selected. This method assumes that all points are equally important in space, that is, each point has an equal weight in the projection. This strategy is usually applicable when:

[0094] When the projection error is small and evenly distributed: When the point cloud data is relatively evenly distributed, or the projection error is small, a uniform weight projection strategy may be simpler and more effective, without considering a complex weighting mechanism.

[0095] The calculation is simple and efficient: The unified weight method has a small amount of calculation and is suitable for large-scale point cloud data processing. It has advantages in scenarios that do not require overly fine adjustments.

[0096] Ellipse weighted average projection strategy: When the model output is 0, the ellipse weighted average projection strategy is selected. This method is based on the spatial distribution of point cloud data. Each point is weighted by the weight distribution of the ellipse shape, and its weight in the projection can be dynamically adjusted according to the spatial position of each point. Its advantages are:

[0097] Processing unevenly distributed point clouds: If the point cloud data is unevenly distributed in space, some areas may have higher point density. Elliptical weighted averaging can adjust the weights so that the points in these dense areas are processed more accurately during the projection process.

[0098] Reducing projection errors: By considering the differences in spatial distribution, elliptical weighted averaging can effectively reduce projection errors, especially when point cloud data are locally dense or sparse.

[0099] Strong adaptability: This method is highly adaptable and can flexibly adjust the weights according to the different characteristics of the point cloud to achieve better projection effects.

[0100] The significance of strategy selection: Data-driven optimization: By automatically selecting the appropriate projection strategy through the machine learning model, when processing complex or changing point cloud data, the optimal strategy can be selected based on the characteristics of the data itself to achieve the purpose of minimizing errors or maximizing computational efficiency.

[0101] Improve processing accuracy: The elliptical weighted average projection strategy is particularly suitable for point cloud data with uneven spatial distribution. It can more accurately handle the projection problems of local dense areas or sparse areas, thereby improving the overall projection accuracy.

[0102] Improve computational efficiency and stability: By automatically selecting appropriate strategies in different scenarios, we can avoid overly complex calculations while ensuring accuracy, thereby improving the stability and efficiency of the system.

[0103] The specific elliptical weighted average projection strategy is:

[0104] Traditional point cloud projection techniques, such as orthographic projection, usually simply project 3D point cloud data onto a 2D plane. This method does not take into account the density distribution characteristics of point cloud data in 3D space, which may lead to the loss of some important information or unsatisfactory projection effects. Especially for Gaussian point clouds (i.e., point cloud distribution follows Gaussian distribution), traditional methods may not be able to effectively capture the spatial relationship and density changes of point clouds.

[0105] By introducing weighting factors, the ellipse weighted averaging technology can take into account the distribution characteristics of the point cloud in three-dimensional space during the projection process, and simulate the geometric transformation in the actual scene through the shape of the ellipse, thereby achieving a more accurate two-dimensional projection effect.

[0106] Suppose there is a set of 3D point cloud data N represents the total amount of data, i represents the index, where each point P i =(x i ,y i ,z i ) is a point in three-dimensional space. These points usually follow a Gaussian distribution, and the center (mean) of the distribution is known. These points need to be projected onto a two-dimensional plane.

[0107] Define ellipse weighting factor: The ellipse weighting factor adjusts the weight of each point according to the position and distribution of the point cloud in space. Define an ellipse weighting function W(P i ) for each point P i =(x i ,y i ,z i ) is assigned a weight.

[0108] The calculation formula of the ellipse weighting factor is as follows

[0109] Among them, σ is the parameter that controls the shape of the ellipse, which controls the range and degree of weighting. The function is based on the point P i The coordinates of the point are calculated by taking into account the distance of the point from the origin, thereby assigning different weights to the points in space. In addition, the axis of the ellipse can be further adjusted, for example, considering the length of the ellipse in different directions, to better simulate the actual point cloud distribution characteristics.

[0110] Weighted average projection to a 2D plane: Use an elliptical weighting factor to perform a weighted average on the 3D point cloud and calculate the weighted 2D projection. Assuming that you want to project the 3D point cloud to the xy plane, you can use the following weighted average formula:

[0111] are all weighted projection coordinates, and these weighted averages are the two-dimensional projections of the three-dimensional point cloud after elliptical weighting. The projection is projected onto a pixel grid to generate the final two-dimensional image. In this process, an interpolation algorithm can be used to fill in the blank areas to ensure the integrity of the image. For example, bilinear interpolation or Gaussian interpolation can be used to smooth the projection results to ensure that the value of each pixel in the image represents the weighted information of the surrounding point cloud. It should be noted that the parameter σ of the ellipse weighting factor can be adaptively adjusted according to the density distribution of the point cloud to further ensure the projection effect.

[0112] Countless two-dimensional Gaussian distributions are deeply sorted, and a clear three-dimensional Gaussian scene model is formed through the rendered two-dimensional image and the processed point cloud data, and a mean vector and a covariance matrix are selected for each two-dimensional Gaussian distribution. These parameters determine the center and shape of each Gaussian distribution.

[0113] Sampling to generate data: Sample multiple data points (e.g. 1000 points) from each two-dimensional Gaussian distribution. These points will be distributed in two-dimensional space, and the value of each point corresponds to the probability density of the point under the Gaussian distribution.

[0114] Calculate the depth of each point: Assuming that the center of these Gaussian distributions has a certain height or depth value, a depth value can be assigned to each point according to preset rules (such as distance, intensity, etc.). For example, the closer the Gaussian distribution is to a certain observation point, the smaller its depth value, or it can be sorted according to the probability density.

[0115] Perform depth sorting: Sort all Gaussian distributed sample points by depth from near to far. This can be achieved by calculating the depth value of each point and sorting them by depth value. At this time, points with high depth values ​​should be in the front, and points with low depth values ​​should be in the back.

[0116] Select a viewing angle: Set a viewing angle (such as from directly above or from the side) and determine the height and position of the viewpoint.

[0117] Projecting depth-sorted points to a two-dimensional plane: Based on the viewing angle and depth information, the points in the three-dimensional space are projected to a two-dimensional plane. At this time, points with greater depth are projected to the foreground of the image, and points with smaller depth are projected to the background.

[0118] Rendered image: Render the projected point cloud using rendering methods in computer graphics (such as lighting, shadows, etc.). Depth changes are expressed through color, size, transparency, etc. Usually, points far away from the observation point will appear transparent or blurred, while points close to the observation point will be more clearly visible.

[0119] For example, during the rendering process, the contribution of each Gaussian point cloud to the viewpoint is considered. Through the interaction between light and each Gaussian point cloud, the color, transparency, and scalability of the Gaussian point cloud in space are considered. Assume that a ray starts from the viewpoint and passes through multiple Gaussian point clouds in the scene. Each Gaussian point cloud contributes a color and opacity. The final color C final is the weighted average of the color and transparency of each Gaussian point cloud. i The light whose color contribution is YS i It can be expressed as:

[0120] YS i =α i ·c i ; i represents the index of the Gaussian point cloud, and the final color of the light is obtained by accumulating the color and transparency of all Gaussian point clouds:

[0121] This formula describes the gradual mixing process of color and transparency, where the previous Gaussian point cloud affects the rendering effect of the subsequent point cloud.

[0122] Generate a 3D Gaussian scene model: Point cloud data processing: All 2D Gaussian distribution sample points are sorted and rendered into a point cloud data set according to depth. The point cloud contains the spatial position and depth information of each point. Construct a 3D model: Use 3D reconstruction technology (such as point cloud reconstruction, surface fitting, etc.) to generate a 3D model based on the processed point cloud data. For example, a complete 3D Gaussian scene can be formed using 3D mesh reconstruction technology or a surface reconstruction algorithm based on depth information.

[0123] Optimize model: Perform preset optimization on the generated 3D model to remove noise, fill holes, etc., to ensure the clarity and coherence of the scene.

[0124] Visualization: Finally, the 3D Gaussian scene model is visualized and displayed using 3D rendering technology. Through operations such as rotation and scaling, users can view the Gaussian scene model in all directions.

[0125] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0126] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0127] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0128] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0129] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A 3D reconstruction method for distribution network scenes based on 3DGAS, characterized in that: The following steps are involved: Multiple images of the distribution network scene are obtained through a preset sampling device, and the multiple images cover multiple perspectives of the distribution network equipment and its surrounding environment; Input multiple images into a computer vision system, use image matching and reconstruction algorithms to generate sparse point clouds, and define the radiation field in the three-dimensional scene as a discrete Gaussian point cloud based on the sparse point cloud; The spherical harmonic coefficients are used to simulate the changes in color values ​​under different viewing angles, perform viewing angle-dependent color rendering, evaluate the errors of traditional projection methods in point cloud data mapping, and evaluate the spatial distribution characteristics of point cloud data. Based on the evaluation results, a corresponding strategy is selected to project the three-dimensional Gaussian point cloud onto a two-dimensional plane to generate a two-dimensional image. Countless two-dimensional Gaussian distributions are deeply sorted, and a clear three-dimensional Gaussian scene model is formed through the rendered two-dimensional image and processed point cloud data.

2. The 3DGAS-based 3D reconstruction method for distribution network scenes according to claim 1 is characterized in that: Inputting multiple images into a computer vision system and using image matching and reconstruction algorithms to generate a sparse point cloud means: In multiple images, the correspondence is found by matching the feature points in each image, and the space reconstruction under different perspectives is realized based on these correspondences. The camera is calibrated, and the position of each 3D point is determined by triangulation through the matched feature points and the internal and external parameters of the calibrated camera to form a sparse point cloud.

3. The 3DGAS-based 3D reconstruction method for distribution network scenes according to claim 2 is characterized in that: When constructing a Gaussian point cloud, the point cloud is divided into multiple clusters using the clustering method K-means. The points in each cluster collectively represent an "area" in the scene. Each cluster is modeled with a Gaussian distribution to obtain a parameter combination for each Gaussian point cloud.

4. The 3DGAS-based 3D reconstruction method for distribution network scenes according to claim 3 is characterized in that: The parameter combination of each Gaussian point cloud includes the following parameters: position, covariance, color, and opacity.

5. The 3DGAS-based 3D reconstruction method for distribution network scenes according to claim 4 is characterized in that: Evaluating the error of traditional projection methods in point cloud data mapping refers to: Suppose there is an original data set X, which contains N samples, each sample has M features, that is, the dimension of the data set is N×M. Suppose the projected data set is the data set Y obtained after dimensionality reduction, and its dimension is N×K, where K is the dimension of the projected data. The projection matrix P is a matrix that maps the original data to a low-dimensional space. Its dimension is M×K, that is, the transformation matrix from the original data space to the low-dimensional space. The projected data Y can be obtained by the following formula: Y=XP, where P is the projection matrix, X is the original data matrix, and Y is the projected data. By calculating the reconstruction error, we can measure the difference between the projected data Y and the original data X when reconstructed back to the original space through inverse projection: E=‖XX ′ ‖ F ;X ′ is the reconstructed data, E is the reconstruction error, ‖·‖ F is the Frobenius norm, which is used to measure the overall error of the matrix; The projection error index can be calculated by the following formula: ‖X‖ F is the Frobenius norm of the original data, PEI is the projection error index, which is used to reflect the error degree of the traditional projection method in the point cloud data mapping, ‖X‖ F Defined as: ‖X‖ F It is obtained by calculating the square root of the sum of the squares of the elements of the original data matrix. i and j are the position indexes of the data in the matrix.

6. The 3DGAS-based 3D reconstruction method for distribution network scenes according to claim 5 is characterized in that: Evaluating the spatial distribution characteristics of point cloud data refers to: The point cloud data is divided into several volume units of fixed size, the points in each unit are counted, and the point density of the unit is calculated. The density mean and standard deviation of all units are calculated to evaluate the density inhomogeneity of the point cloud.

7. The 3DGAS-based 3D reconstruction method for distribution network scenes according to claim 6 is characterized in that: Choosing the corresponding strategy based on the evaluation results means: The projection error index, density mean, and density standard deviation are input into the pre-trained machine learning model together, and the output result is 1 or 0. When the output result is 1, the strategy selected is the projection strategy of uniform weight, and when the output result is 0, the strategy selected is the projection strategy of elliptical weighted average.

8. The 3DGAS-based 3D reconstruction method for distribution network scenes according to claim 7 is characterized in that: The pre-trained machine learning model is a convolutional neural network model.

Images