Power transmission tower image-driven reconstruction and dynamic visualization method and system

Abstract

This invention belongs to the field of digitalization, 3D reconstruction, and dynamic visualization technology of transmission tower structures, and particularly relates to an image-driven reconstruction and dynamic visualization method and system for transmission towers. It can be used for digital twins of transmission tower structures and transmission tower-line systems, disaster response assessment, and interactive visualization. The method includes: multi-view image acquisition and camera prior establishment; differentiable rendering optimization based on anisotropic 3D Gaussian primitives; key area capacity allocation and adaptive refinement during the training period for structure-terrain collaboration; terrain extraction and digital elevation model construction after training; multi-layer boundary expansion and completion; physical-driven data acquisition; coordinate registration and Gaussian deformation mapping based on right-pole decomposition; and sparse-to-dense anisotropic interpolation and dynamic rendering based on component principal axis constraints.

Description

Technical Field

[0001] This invention belongs to the field of digital reconstruction, three-dimensional reconstruction and dynamic visualization technology of transmission tower structures. In particular, it relates to a method and system for image-driven reconstruction and dynamic visualization of transmission towers. It is a dynamic visualization method and system that realizes three-dimensional digital reconstruction of transmission towers, scene expansion and completion, and physical consistency coupling with finite element / monitoring displacement data based solely on multi-view images. It can be used for digital twins (i.e., digital mirrors corresponding to physical objects that can be updated with physical states), disaster response assessment and interactive visualization of transmission tower structures and transmission tower-line systems. Background Technology

[0002] Transmission lines are a crucial component of the power grid. Lattice-type transmission towers support and transmit force to conductors, ground wires, and insulator strings; their structural configuration and stress performance directly impact line operational safety. With the development of drone inspections and visual perception, image-based 3D modeling provides a new data source for the visualization, digital twinning, and intelligent operation and maintenance of transmission assets. In current engineering practices, 3D reconstruction often relies on LiDAR point clouds or parametric model libraries. However, in large-scale inspection and rapid modeling scenarios, LiDAR equipment is costly and task organization is complex, and parametric library methods are often limited by the completeness and robustness of the component library, making it difficult to simultaneously capture realistic appearance and complex details.

[0003] 3D modeling methods based on pure images mainly include: geometric reconstruction methods based on motion-reconstructed structures and multi-view stereo vision, view synthesis methods based on neural radiation fields, and real-time rendering modeling methods based on explicit particles or Gaussian units that have emerged in recent years. Among them, explicit 3D Gaussian representation is suitable for rapid visualization of engineering scenes due to its high rendering efficiency and interpretable geometric parameters. However, for targets such as transmission towers with "dense slender rods, complex node connections, sharp structural boundaries, and severe occlusion," existing image-driven explicit reconstruction still generally suffers from the following problems: (1) Blurring of slender components and sharp edges: When the model capacity (number of primitives) is limited or approximately uniformly distributed, slender components such as tower legs, diagonal braces, and web members are prone to excessive smoothing, breakage, ghosting, or blurred boundaries, resulting in unclear geometry and texture, which makes it difficult to meet the needs of engineering-level display and subsequent analysis.

[0004] (2) Insufficient coverage of the outer region of the view frustum: The image observation range is usually limited by the flight path and camera angle. The reconstruction results are often truncated at the boundary of the view frustum. The lack of observation of the outer region of the view frustum leads to discontinuity of the background such as terrain / roads / vegetation and the presence of holes, which affects the integrity of the scene and the immersive display.

[0005] (3) Unstable coupling in dynamic visualization: In engineering, it is often necessary to map the finite element analysis results or the displacement time history monitored on site to a three-dimensional model to visualize the structural response under conditions such as wind-induced, icing, and earthquake. Existing practices mostly use global rigid body transformation or simple neighborhood interpolation, which only updates the translation of points and makes it difficult to express local rotation and tension at the same time. Under explicit Gaussian element representation, it is also easy to cause problems such as covariance degradation, rotation and scale coupling, cross-component crosstalk, time series flickering and ghosting, thus affecting the physical consistency and visual stability of dynamic visualization.

[0006] Therefore, there is an urgent need for a new image-driven 3D reconstruction and dynamic visualization method for transmission tower structures: without relying on lidar and component libraries, it can prioritize the details of slender components and sharp edges under a fixed computational budget, and can perform consistent outward expansion and completion of the terrain background after training; at the same time, it can stably map the finite element / monitoring displacement to a 3D Gaussian representation at the "translation-rotation-stretching" level, suppress dynamic rendering artifacts, and achieve physically interpretable, continuous and stable response animation, providing support for digital twins of power facilities. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for image-driven reconstruction and dynamic visualization of transmission towers, in order to solve problems in the prior art such as difficulty in preserving the details of slender components, difficulty in consistently completing the regions outside the view frustum, and ghosting / blurring / instability caused by displacement field mapping.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for image-driven reconstruction and dynamic visualization of transmission towers includes the following steps: Step (1) Multi-view image acquisition and camera prior establishment: A multi-view image sequence containing the target transmission tower and its surrounding terrain environment is acquired; feature point extraction and multi-view matching are performed on the image sequence, and mismatched points are eliminated by combining random sampling consistency algorithm with epipolar geometric constraints; then bundle adjustment is performed to jointly optimize and restore the camera intrinsic and extrinsic pose parameters and the sparse 3D point cloud structure of the scene for each image; the feature point extraction, multi-view matching, random sampling consistency geometric verification and bundle adjustment can be completed using commonly used software, algorithms or equivalent implementations in this field, and the restored results are used as the initial space and view prior for subsequent 3D Gaussian scene reconstruction and rendering optimization of the transmission tower.

[0009] Step (2) Differentiable rendering optimization based on anisotropic 3D Gaussian primitives: Each 3D point in the sparse 3D point cloud recovered in step (1) is initialized as an anisotropic 3D Gaussian primitive, and the set of all 3D Gaussian primitives is used as the explicit representation of the target scene. The center spatial position, unit quaternion representing orientation, logarithmic scale parameter representing anisotropy, spherical harmonic coefficient representing appearance color, and opacity of each Gaussian primitive are independently parameterized. In the differentiable rendering optimization loop, the Gaussian primitives projected onto the 2D image plane are compared pixel-by-pixel with the actual acquired image to combine... The multi-view reconstruction error of the norm and structural similarity index is used as the target loss function. The parameters of the Gaussian meta-set are iteratively optimized using the backpropagation mechanism to obtain the initial three-dimensional Gaussian model.

[0010] Step (3) Training Period Structure-Terrain Coordination: Key Area Capacity Allocation and Adaptive Refinement Under a fixed or controlled primitive quantity budget, the importance criterion of "reconstruction error gradient + shape statistics" triggers directional splitting and cloning, prioritizing the allocation of modeling resources to slender components such as tower legs, crossarms, diagonal braces, and web members, as well as sharp boundary areas. At the same time, a more lenient retention and delayed pruning strategy is adopted for terrain primitives that are responsible for subsequent digital elevation model construction and scene expansion constraints near the ground surface, road boundaries, and view cone boundaries around the tower base. A region-differentiated pruning strategy based on screen-space effective opacity contribution is introduced: within a preset statistical window, the effective opacity contribution of each Gaussian primitive to pixel color synthesis during forward rendering at each training viewpoint is statistically analyzed, and the average value is calculated across all pixels and viewpoints to obtain the screen-space effective opacity contribution value of the Gaussian primitive; the three-dimensional bounding box (core bounding box) surrounding the main body of the transmission tower is set as the key region, and different pruning thresholds are applied to the inside and outside of the key region. When the contribution value is lower than the threshold corresponding to its region, pruning is performed on the Gaussian primitive; at the same time, a "pruning grace period" is set for newly generated Gaussian primitives to protect the details of newly split slender components from being prematurely removed, effectively suppressing the disorderly growth of redundant Gaussian primitives in the background and non-critical regions.

[0011] Step (4) Post-training terrain extraction and digital elevation model construction: After the differentiable rendering optimization described in step (2) and the adaptive refinement described in step (3) reach the convergence condition, the local geometric features of the Gaussian core points are extracted (including the local normal slope angle and local elevation undulation based on principal component analysis). The terrain Gaussian cores and non-terrain Gaussian cores are distinguished by a multi-feature joint threshold. An adaptive height consistency filter is applied to the XY plane grid to remove micro-terrain noise and outliers, and a clean terrain sample set is obtained. Based on the terrain sample set, a spatial variogram (i.e., a function used to describe the correlation between samples at different spatial locations) is fitted and a common Kriging interpolation algorithm (i.e., an interpolation method for unbiased optimal estimation of unknown locations based on spatial statistical correlation) is used to construct a continuous digital elevation model with physical spatial correlation, which is used as the absolute geometric elevation constraint for subsequent view frustum expansion and completion.

[0012] Step (5) Multi-layer boundary expansion and completion: The effective observation boundary of the current scene is identified on the planar occupied grid (i.e., the discrete grid used to indicate whether the planar grid cell has been occupied by the current scene) of the scene around the transmission tower. Empty grid cells outside the boundary are collected layer by layer from the inside out as candidate expansion cells, and spatial sparsity is controlled by minimum Chebyshev distance constraint and upper limit of single-layer candidate number. For each candidate cell, the best matching high-quality source cell is retrieved from the effective boundary area based on the cosine similarity of the two-dimensional spatial feature vector and the high consistency score combined with the digital elevation model. All Gaussian primitives in the source cell are cloned to the candidate cell, and in-plane rotation correction around the Z-axis, XY translation registration, and dynamic alignment of elevation consistency based on the offset of the digital elevation model are performed in sequence. This expansion process is iterated until no new cells are added or the global primitive growth limit is reached, generating an expanded scene model with a wider coverage and natural and continuous texture and geometry.

[0013] Step (6) Physical drive data acquisition: Obtain nodal displacement time history data (including responses to external loads such as wind vibration and earthquakes) obtained from finite element transient dynamic analysis or on-site sensor network monitoring, as input for subsequent coordinate registration and Gaussian deformation mapping.

[0014] Step (7) Coordinate registration and Gaussian deformation mapping based on right-pole decomposition (a matrix decomposition method that decomposes local deformation gradients into rotational and stretching components): Within the local neighborhood of discrete physical nodes, the Jacobian matrix of the local displacement gradient is estimated using a weighted least squares method with distance decay weights. And construct the local deformation gradient matrix. ,in, For nodes The local deformation gradient matrix at the location, It is a third-order identity matrix. For nodes The local displacement gradient Jacobian matrix at the given location. Perform right-polar decomposition on the local deformation gradient matrix. ;when Non-singular and At that time, the local deformation is explicitly and uniquely decoupled into an orthogonal rotation increment matrix. With symmetric positive definite stretching increment matrix .

[0015] Furthermore, the decoupled increments are mapped to the Gaussian meta-parameter space: translation increments are directly linearly accumulated; rotation increments are converted into unit quaternions and updated by multi-node stable averaging and interpolation in the Lie algebra tangent space (i.e., the locally linearized representation space of the unit quaternion space near the identity element); scale increments are updated in the logarithmic scale space after extracting the principal stretching components. Finally, the Gaussian covariance matrix is ​​directly reconstructed using the updated quaternions and the scale matrix to maintain the positive definiteness of the deformed covariance matrix and reduce the numerical truncation and accumulated errors caused by direct matrix multiplication.

[0016] Step (8) Sparse to dense anisotropic interpolation and dynamic rendering based on component principal axis constraints: Anisotropic interpolation refers to the interpolation method using different propagation bandwidths along different directions so that deformation is preferentially transmitted along the component principal axis direction.

[0017] To address the spatial sampling differences where physical nodes are sparse and Gaussian cells are dense, we utilize... Principal component analysis of nearest neighbor nodes estimates the local principal axis direction vectors of the physical nodes and the components in which the Gaussian elements are located, where, To determine the number of nearest neighbors, an anisotropic interpolation weight kernel with a large smoothing bandwidth along the component's axial direction and a small smoothing bandwidth laterally is constructed. A "direction consistency penalty term" (i.e., a penalty term that reduces the interpolation weight when the source node's principal axis direction is inconsistent with the predicted principal axis direction of the target position) is added to the kernel function. Through this anisotropic weight, translations, rotations, and scale increments at the sparse physical node level are continuously propagated to the dense Gaussian elements, obtaining a Gaussian element parameter set containing a complete time series. Finally, a response animation is generated based on Gaussian rasterization rendering, achieving high frame rate rendering and display of the structural physical response.

[0018] A transmission tower image-driven reconstruction and dynamic visualization system for implementing the above method includes: Data acquisition and initialization module: used to acquire multi-view image data of transmission lines and transmission towers, and extract sparse 3D structure point cloud and camera pose through the structure-reconstruction-motion method, and initialize 3D Gaussian model; Adaptive model training module: used to train and iterate the three-dimensional Gaussian model under a fixed budget constraint of Gaussian element quantity, and to construct a static three-dimensional Gaussian scene of the transmission tower structure and its surrounding environment through a structure-terrain collaborative adaptive model optimization strategy. Terrain Modeling and Boundary Expansion Module: This module is used to construct a continuous digital elevation model based on adaptive terrain extraction and ordinary kriging, and to perform scene completion outside the view frustum of the static 3D Gaussian scene using the digital elevation model as a constraint, combined with a multi-layer boundary expansion strategy, to obtain an expanded scene model. Physical drive data acquisition module: used to acquire displacement time history data of key nodes of the transmission tower; wherein, the displacement time history data is obtained by establishing a finite element model of the transmission tower and performing time-varying wind vibration response analysis, or by monitoring through on-site sensor network; Coordinate registration and deformation mapping solution module: It is used to unify the coordinate system of the physical model and the coordinate system of the three-dimensional Gaussian scene model by using iterative nearest point or other rigid registration algorithms, and to use the Gaussian deformation mapping method based on local right pole decomposition and component principal axis constraint anisotropic interpolation to map the discrete displacement time history data to the corresponding three-dimensional Gaussian element parameters in the extended scene model. Dynamic visualization rendering module: used to perform three-dimensional visualization rendering of the dynamic response and deformation process of the transmission tower structure based on the updated three-dimensional Gaussian primitive parameters.

[0019] Compared with the prior art, the present invention has the following advantages: (1) Improved modeling fidelity of slender components and achieved efficient and intensive use of computing resources. Without significantly increasing the overall model size, this invention employs a splitting and cloning mechanism jointly triggered by "reconstruction error gradient + geometric shape statistics," coupled with a region-differentiated pruning strategy based on opacity contribution, to preferentially allocate limited modeling resources to slender components and geometrically sharp edges of the tower structure. This not only significantly improves the clarity of 3D details and structural continuity of key load-bearing components such as tower legs and diagonal braces, but also effectively eliminates background stripes, edge burrs, and visual artifacts that are easily caused by traditional uniform sampling.

[0020] (2) This invention solves the problem of scene boundary truncation that often accompanies image-based modeling, thus enhancing the immersive experience of panoramic roaming. The invention employs a two-stage scene completion mechanism: "terrain adaptive extraction—construction of a conventional kriging digital elevation model—multi-layer boundary feature matching and expansion." In blind areas completely lacking original camera frustum observations, this algorithm, relying on the physical spatial correlation of the internal terrain and the texture features of high-quality source units, automatically generates an expanded scene with continuous geometric shape and appearance texture through rigorous rotation correction and elevation alignment. This significantly reduces the sense of boundary truncation and abrupt holes in large-scale engineering visualization.

[0021] (3) From a mathematical perspective, artifacts in dynamic rendering are significantly reduced, improving the consistency between visual response and physical laws. This invention introduces the right-pole decomposition theory from finite element analysis to handle the deformation update of Gaussian parameters of transmission tower structures. By strictly decoupling the mixed local deformations into rigid body rotation and pure tension components, and independently updating the Gaussian parameters in a more stable Lie algebra and logarithmic scale space, the numerical instability problem caused by traditional direct matrix derivation is avoided. Combined with the positive definiteness protection mechanism of "parameter update-covariance reconstruction", this method can still maintain high stability of the rendered image when facing large deformation and large torsion conditions of transmission towers, suppressing Gaussian element tearing, ghosting, and double image phenomena caused by rotation-scale coupling.

[0022] (4) It alleviates the problem of cross-component motion interference and achieves smooth transitions with kinematic compatibility at complex nodes. Addressing the characteristics of crisscrossing components and sparse physical nodes in transmission towers, this invention proposes an anisotropic interpolation propagation algorithm constrained by the component principal axes. By constructing a weight kernel with smooth axial transmission and rapid lateral decay, supplemented by a stringent directional consistency penalty, this method prioritizes the propagation of deformation increments along the physical axis of the corresponding member. This significantly reduces numerical crosstalk between adjacent anisotropic components, ensuring the motion continuity of each slender member and the physical compatibility of deformation at complex nodes such as connecting plates and flanges.

[0023] (5) It has good engineering application value and cross-scenario promotion potential. The overall process of this invention uses a low-cost UAV two-dimensional image array as the core driving data. The physical driving source is compatible with standard finite element output or conventional field sensors, without the need to forcibly bind expensive airborne lidar equipment or prefabricated standard component libraries. This workflow is highly adaptable to the current demand for rapid modeling of UAV autonomous inspection in the context of ubiquitous Internet of Things in the power grid, and can be extended to the construction of digital twin platforms for various slender pole structures such as communication base station towers, wind turbine transmission towers, large building scaffolding, and long-span truss bridges. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating an embodiment of the transmission tower image-driven reconstruction and dynamic visualization method provided by the present invention. Figure 2 This is a schematic diagram of the logical architecture of image-driven reconstruction and dynamic visualization of transmission towers provided by the present invention. It shows that after multi-view images are sparsely reconstructed by motion recovery structure and initialized with three-dimensional Gaussian splashing, they enter a differentiable rendering loop optimization, and the process of physically consistent dynamic rendering is driven by the finite element time history displacement field through right pole decomposition decoupling and principal axis constrained anisotropic interpolation. Figure 3 This is a schematic diagram of an embodiment of the transmission tower image-driven reconstruction and dynamic visualization system provided by the present invention. Detailed Implementation

[0025] The specific embodiments of the present invention are further described below with reference to the accompanying drawings and technical solutions. The image feature extraction, multi-view matching, rigid registration, finite element solution, and graphics rendering processes, unless otherwise specified, can be accomplished using commonly used software, algorithms, or equivalent implementations in the field.

[0026] In modern power transmission systems, transmission towers, as the main load-bearing structures supporting conductors and equipment, directly affect the reliability of power grid operation due to their mechanical properties and structural condition. However, lattice-type transmission tower structures contain numerous slender components and sharp edges, and traditional image-driven 3D reconstruction methods often result in problems such as blurring of slender components, insufficient coverage of occluded areas, and discontinuities. Furthermore, stably mapping the time-varying displacement field of the structure to a 3D reconstruction model for dynamic visualization faces serious visual artifacts such as ghosting, double images, and tearing caused by rotation and scaling coupling.

[0027] See Figure 1 and Figure 2 To address the technical problems of low accuracy in the reconstruction of transmission tower structures, difficulty in expanding the scene, and distortion of dynamic response mapping in existing technologies, this invention provides a method and system for transmission tower image-driven reconstruction and dynamic visualization. The method uses multi-view images as the input data source and employs three-dimensional Gaussian splashing (a modeling method that uses three-dimensional Gaussian primitives as explicit scene representation and performs real-time rendering) as the means of explicit scene representation and rendering. Figure 1 Show the overall process, Figure 2 This embodiment illustrates the logical architecture from multi-view images, motion-reconstructed sparse reconstruction of structures and 3D Gaussian splash initialization, to differentiable rendering loop optimization, and finally to physically consistent dynamic rendering driven by finite element time-history displacement fields. The method includes steps 101 to 108, which respectively correspond to... Figure 1 The steps in the process shown and Figure 2 The logical architecture shown includes multi-view image input, motion-reconstructed sparse reconstruction of structures, 3D Gaussian splash initialization, differentiable rendering loop optimization, finite element time-history displacement field input, right-pole decomposition decoupling, and principal axis constrained anisotropic interpolation dynamic rendering. The specific steps of each step are as follows: Step 101: Obtain multi-view images containing the transmission tower structure and surrounding terrain environment, and perform pose estimation and sparse 3D reconstruction using the structure-in-motion algorithm to obtain the initial camera pose and sparse point cloud.

[0028] In this embodiment of the invention, a drone equipped with a high-resolution camera performs a surround flight to capture images of ultra-high voltage or extra-high voltage transmission towers and their corridors, acquiring a multi-view image sequence. An incremental structure-of-motion (SOMO) algorithm is used to process the multi-view image sequence. The specific process includes: first, extracting scale-invariant features from the images and performing multi-view feature matching; then, applying a random sampling consensus algorithm combined with an essential matrix or homography matrix for geometric verification, eliminating mismatches, and constructing a common view. Based on the initial solution provided by the two-view geometry, pose initialization and triangulation are performed, new images are gradually added, and local and global bundle adjustments are performed alternately. During the global bundle adjustment stage, the camera's intrinsic parameters (focal length, principal point, radial and tangential distortion coefficients) and 3D point coordinates are jointly optimized to reduce reprojection errors to the sub-pixel level. Finally, the model is aligned to the engineering coordinate system using a small number of control points or similarity transformations, outputting a sparse 3D point cloud structure and calibrated camera pose for subsequent modeling, serving as a geometric prior for subsequent 3D Gaussian scene model rendering optimization. The above SOMO process can be implemented using common methods in the field or other equivalent systems.

[0029] Step 102: Based on the initial camera pose and sparse 3D point cloud, initialize a 3D Gaussian set and establish a differentiable rendering optimization loop based on image fidelity.

[0030] As a preferred embodiment, initializing the 3D Gaussian set and establishing a differentiable rendering optimization loop includes: Each 3D point in the sparse 3D point cloud is initialized as an anisotropic 3D Gaussian primitive, and the scene vector is represented as follows: ,in, Let be the scene vector, representing the set of all Gaussian meta-parameters.

[0031] in, Indicates the first The center coordinates of each Gaussian element; Indicates the first The unit quaternion rotation parameter of Gaussian elements; Opacity; These are the spherical harmonic coefficients; For Gaussian meta-index, Let be the total number of Gaussian elements in the scene. To maintain the positive definiteness of the scaling parameter and improve numerical stability, a logarithmic scaling parameter is used. ,in, Indicates the first A logarithmic scaling vector of Gaussian elements along the three principal axes. The corresponding scale vector is represented by the diagonal scaling matrix, which restores the original value to its original form. , and These represent element-wise logarithmic operations and element-wise exponential operations, respectively. This represents an operator for constructing a diagonal matrix from vectors.

[0032] The covariance matrix of Gaussian elements is calculated based on standard three-dimensional Gaussian splash factorization: in, For the first The covariance matrix of Gaussian elements, in the formula Indicates by parameters The corresponding rotation matrix obtained after transformation, This represents the transpose of the rotation matrix. For the reason The diagonal scaling matrix is ​​recovered; This indicates that the covariance matrix is ​​a positive definite matrix.

[0033] In the differentiable rendering optimization loop, the forward propagation process performs camera projection, elliptical weighted average Gaussian splashing, and opacity-based color synthesis. The 3D Gaussian primitives are projected onto the 2D image plane to obtain the rendered image. ,in, For the parameter set The corresponding rendered image, The set of all Gaussian parameters vectorized in order. This represents a vectorization operator.

[0034] During backpropagation, the rendered image is calculated. Compared with the observed real image The reconstruction loss function between them is used to calculate the gradient and update the Gaussian parameters: In the formula, To reconstruct the loss function, For a set of pixels, | | For pixel collections The total number of pixels in To render an image in pixels Pixel value at that location, For real images in pixels Pixel value at that location, These are actual observation images; for L 1-norm; It is a structural similarity index; and These are preset weighting coefficients. L The 1 loss is used to ensure the accuracy of low-frequency colors, while the structural similarity index loss emphasizes the consistency of image structure and high-frequency details.

[0035] Step 103: During the iterative training process of the differentiable rendering optimization loop, an adaptive model optimization strategy of structure-terrain collaboration is applied. Based on shape statistics and reconstruction gradient, the splitting, cloning and pruning operations of Gaussian units are triggered to optimize the modeling quality of key regions.

[0036] This invention addresses the issue of numerous slender components (such as tower legs, cross braces, and crossarm web members) in lattice-type transmission towers. Traditional uniform sampling easily leads to edge blurring and loss of detail in these components during rendering. Furthermore, limited video memory resources necessitate strict control over model size. Therefore, an importance-guided adaptive splitting and density control strategy is introduced. Additionally, a more lenient retention and delayed pruning strategy is employed for terrain Gaussian elements near the tower base, road boundaries, and view frustum boundaries, which are responsible for subsequent digital elevation model construction and scene expansion constraints.

[0037] As a preferred embodiment, the adaptive model optimization strategy for applying structure-terrain collaboration includes an importance-guided adaptive splitting and density control strategy: The adaptive splitting strategy is as follows: Calculate the first The aspect ratio of each high-speed element and volume These serve as representations of local geometric elongation and local volume size, respectively. Calculate the reconstruction loss function relative to the Gaussian metacenter. Position gradient magnitude ,in, The reconstruction loss function represents the first... The gradient magnitude at the center of each Gaussian element. The reconstruction loss function is expressed as follows: gradient, express Norm; Define Gaussian split trigger indicator function : In the formula, For the first The split trigger indicator value of each Gaussian unit. and These represent the global statistical quantile functions for aspect ratio and volume, respectively. and The set percentile level; For indicator functions; For cloning gradient threshold; symbol and These represent logical OR and logical AND, respectively. When At that time, a directional splitting or cloning operation is performed on the Gaussian unit.

[0038] Furthermore, in order to control the overall size of the model, an adaptive density control strategy is implemented, as follows: Let... For critical regions (such as the 3D bounding box of a transmission tower), in the iteration steps The number of Gaussian elements in this region at that time was Apply a capacity limit constraint: In the formula, Representing the iterative steps Key areas The number of Gaussian elements within; Represents the critical region at the initial time. The number of Gaussian elements within; This is the capacity control coefficient for critical areas, used to limit the maximum growth multiple of the number of Gaussian elements within critical areas.

[0039] Meanwhile, based on the effective opacity contribution of screen space Pruning operations are performed, among which... Indicates the first Each Gaussian element in a pixel The effective opacity contribution to pixel color composition. This represents the average effective opacity contribution of the Gaussian pixel across all pixels. Represents a set of pixels The total number of pixels. Different pruning thresholds are applied inside and outside the key area: In the formula, For key areas Internal pruning threshold For key areas External pruning threshold, and This is to preserve more detail in key areas.

[0040] This region-aware capacity control strategy can prioritize the allocation of limited expressive power to slender components of transmission towers and discontinuous areas in the surrounding terrain without increasing the total number of global primitives, effectively suppressing artifacts such as background stripes and ghosting. Furthermore, a "pruning grace period" is set for newly generated Gaussian primitives, which are temporarily excluded from the pruning process within a preset number of iterations to prevent the premature removal of details from newly split slender components.

[0041] Step 104: After the differentiable rendering optimization described in step 102 and the adaptive model optimization strategy described in step 103 reach the convergence condition, adaptive terrain extraction is performed on Gaussian elements based on local geometric features, and a continuous digital elevation model of the terrain is constructed using ordinary kriging interpolation.

[0042] In the transmission tower scenario embodiment of the present invention, the original camera's field of view often lacks observation data, and simple extrapolation can lead to severe geometric and textural inconsistencies. To achieve consistent scene completion, it is first necessary to isolate the terrain from the complex background.

[0043] As a preferred embodiment, the construction of the continuous digital elevation model of the terrain includes: For each Gaussian core center In its Nearby Neighborhood Within, local normal vectors are calculated using principal component analysis. And then calculate the local slope angle. and elevation fluctuations : In the formula, For the first A high-ranking official nearest neighbor set The number of nearest neighbors. The local normal vector is obtained through principal component analysis. The global vertical coordinate axis vector. For the neighboring region The elevation of each element Representing vectors Norm, This indicates the conversion of radian angles to degrees.

[0044] Then, a joint threshold is established to extract the initial terrain candidate set. : In the formula, This is the initial terrain candidate set obtained based on the joint threshold; This is the local slope angle threshold; Local elevation fluctuation threshold; symbol This indicates a logical AND operation.

[0045] Calculating the adaptive robust threshold using the median absolute deviation and To resist local noise interference: In the formula, and These are the empirical quantiles for slope angle and elevation undulation, respectively; and These are the corresponding empirical quantile functions; This is the amplification factor for the absolute deviation of the median; Represents the set of slope angles for all samples; Represents the set of elevation fluctuations for all samples; This indicates the absolute deviation from the median.

[0046] To suppress minor terrain undulations and residual noise, step sizes are used on the XY plane. Divide into grids, where, This defines the grid step size for the XY plane grid during the terrain extraction stage. The mean elevation of the initial terrain points is calculated within each grid cell. and standard deviation The final terrain points are retained using a dynamic tolerance range: In the formula, Indicates the first Line 1 The mean elevation of the terrain samples within the column grid cell. This represents the standard deviation of the elevation of the terrain samples within the grid cell. and These represent the adaptive tolerance coefficients for the lower and upper bounds, respectively.

[0047] Subsequently, the ordinary kriging algorithm was used to process the cleaned terrain point samples. Processing and reconstructing the continuous digital elevation model surface. ,in, This is a cleaned terrain point sample set. At any predicted location... Its elevation is estimated by a weighted sum of sample elevations: ,in, This represents the number of terrain sample points. To reconstruct the digital elevation model, For position The estimated elevation of the location, These are the planar coordinates of the terrain sample points. To correspond to the elevation, The location to be predicted. The weights are Kriging weights and satisfy the unbiased constraint. This digital elevation model not only fills in the gaps between scattered terrain points around the transmission towers, but also provides global elevation consistency constraints for subsequent boundary expansion.

[0048] Step 105: Based on the continuous digital elevation model, by constructing a joint matching mechanism of feature similarity and elevation consistency, multi-layer boundary expansion is performed on the region outside the view frustum to obtain a static three-dimensional scene model after view frustum completion.

[0049] As a preferred embodiment, the multi-layer boundary extension includes: In the XY plane, with step size Construct an occupying grid, in which, This defines the grid step size for the XY plane occupied grid during the boundary expansion phase. Record the current occupied boundary edge of the non-terrain region, and collect the blank grids adjacent to the occupied boundary as the first... Candidate unit set corresponding to the outer expansion iteration of the wheel boundary ,in, Number the iterative layers for boundary expansion; Applying minimum Chebyshev spacing constraints and maximum candidate number constraints in a single layer to Perform sparsification: In the formula, and For candidate unit set Any two candidate units in the, Denotes the Chebyshev norm. The minimum Chebyshev spacing threshold, This represents the upper limit of the number of candidate units in a single layer.

[0050] Extract high-quality source unit sets near the occupied boundaries. High-quality source cells refer to occupied cells located near occupied boundaries, containing a sufficient number of Gaussian cells, with stable local geometry and texture features, and not contaminated by obvious holes or anomalous noise. For each candidate cell... Based on spatial feature similarity Elevation Consistency under Digital Elevation Model Constraints Select the best matching source unit : The final score is based on the overall score. The highest-ranking pair is determined; among them, Candidate unit and source unit Spatial feature similarity score between them To score for elevation consistency, The normalized spatial feature similarity score, The combined score is based on the similarity of spatial features and the consistency of elevation. and These are the aggregate space feature vectors of the candidate unit and the source unit, respectively. To prevent tiny positive numbers with a denominator of zero, and These represent the elevation values ​​of the digital elevation model at the candidate and source elements, respectively. These are elevation consistency control parameters.

[0051] All Gaussian elements within the matching source element are cloned to the candidate element. Texture orientation misalignment is eliminated by rotation alignment around the Z-axis, and elevation compensation is calculated using a digital elevation model. The center coordinates of the cloned Gaussian elements are then determined. The updated formula is as follows: In the formula, The coordinates are the original center coordinates of the Gaussian cell to be cloned within the source cell. The updated Gaussian metacenter coordinates, Rotation angle around the Z-axis The corresponding yaw angle correction matrix, and The centroids in the XY plane of the source element and the candidate element are respectively. and These represent the elevation values ​​of the digital elevation model at the candidate and source elements, respectively. This is a global vertical unit vector. This operation iterates outward layer by layer until no new units are added or the global expansion limit is reached. This method achieves seamless expansion of the scene surrounding the transmission tower under a fixed primitive budget, while maintaining visual compatibility of texture and height.

[0052] Step 106: Obtain physical drive displacement time history data. In one embodiment, the time history displacement field of key nodes of the transmission tower can be obtained by constructing a transmission tower-line coupled finite element model, synthesizing transient wind fields and calculating wind loads, and performing geometric nonlinear transient analysis. In another embodiment, displacement time history data obtained from on-site sensor network monitoring can also be used directly.

[0053] In this embodiment of the invention, finite element models of beam and cable elements of the transmission tower are established using finite element software. According to specifications, the harmonic superposition method (a method of superimposing sinusoidal components of different frequencies to synthesize a time-history signal) and fast Fourier transform are used to synthesize a transient wind field that satisfies the target average wind profile and fluctuating wind spectrum (such as the Davenport spectrum, i.e., the fluctuating wind power spectrum model commonly used in engineering wind-induced vibration analysis). In the formula, Indicates height The average wind speed at that location Indicates the height above the ground. The average wind speed at a height of 10 meters. This is the terrain roughness index.

[0054] Davenport power spectral density function of pulsating wind By combining the spatial coherence function (a function characterizing the correlation of wind speed fluctuations at different spatial locations, considering exponential decay constants in the downwind, crosswind, and vertical directions) generated by the dual-index frequency method (i.e., a method of constructing multi-point correlated stochastic wind fields by introducing dual frequency indices), multi-point correlated gust time histories are generated. Represents frequency The power spectral density of the pulsating wind at that location, For frequency variables.

[0055] The wind speed time history is converted into equivalent nodal wind loads using dynamic wind pressure theory: In the formula, Indicates the first Each loading point at time The equivalent nodal wind load, This indicates that the loading point is at time [time]. Total wind speed, air density, Indicates the relationship with the first The component group to which each loading point belongs The corresponding body type coefficient, This indicates the corresponding equivalent windward width. This indicates the corresponding representative length. This indicates the corresponding floor height magnification factor. Indicates the first The layer index corresponding to each loading point Indicates the first The component group index to which each loading point belongs is then determined. Subsequently, a transient dynamic analysis considering large deformation geometric nonlinearity is performed, outputting the three-component time-history displacements of representative nodes. ,in, , and They represent , and Displacement time history in the direction. In this embodiment, the process of acquiring physical driving data is illustrated using wind-induced vibration as an example. In other embodiments, the above-mentioned node-level displacement time history data can also be directly acquired by the field sensor network and used as the input for step 107 after time synchronization and coordinate unification are completed. The acquisition method of node response under other external load excitations such as earthquakes is similar.

[0056] Step 107: Using iterative nearest point or other rigid registration algorithms, the coordinate system of the physical model and the coordinate system of the 3D Gaussian scene model are uniformly and accurately registered to establish the initial spatial mapping relationship between discrete physical nodes and Gaussian primitives; after completing the coordinate registration, the local deformation is decoupled into rotation increment and stretching increment by estimating the local displacement gradient and using the right pole decomposition strategy.

[0057] Existing dynamic mapping techniques, if they only transfer displacement through rigid body translation, will cause severe distortion and tearing of the model during rotational deformation; if they directly interpolate the noisy deformation gradient matrix and update the covariance in the numerical implementation, it is easy to cause scale imbalance, numerical instability and visual artifacts.

[0058] As a preferred embodiment, at the nodes of the finite element model First-order topological neighborhood Internally, a weighted least squares method with Gaussian decay weights is used, combined with... Regularization, fitting the local displacement Jacobian matrix : In the formula, For nodes The local displacement Jacobian matrix at that point. For the solution to be found Matrix variables, express The set of real matrices This represents the value of the independent variable that minimizes the objective function. For nodes The first-order topological neighborhood, For nodes With nodes Distance decay weight between them and They are nodes and nodes The initial coordinate vector, and They are nodes and nodes The displacement vector, for Regularization coefficient, express Norm, This represents the Frobenius norm, which is the square root of the sum of the squares of the elements of the matrix.

[0059] Constructing the local deformation gradient matrix ,in, For nodes The local deformation gradient matrix at the location, It is a third-order identity matrix. For nodes The local displacement gradient Jacobian matrix at the location. To explicitly decouple rigid body rotation from local tension, for... Perform right-pole decomposition: when Non-singular ( When ), the decomposition is unique, where, Represents determinant operations; For nodes The local deformation gradient matrix at the location, It is a third-order identity matrix. It is a rotation increment matrix, belonging to a special three-dimensional orthogonal group. , The stretching increment matrix belongs to the set of three-dimensional symmetric positive definite matrices. .

[0060] Let stretching increment matrix The main stretch amount is ,in, For nodes In the The main tensile amount in the main direction, ;make This indicates the node formed by the three main tensile amounts. The scale increment vector is given. To improve numerical stability, it is transformed into a logarithmically scaled increment space for representation: For the rotation matrix extracted by polar decomposition Especially at the connection points of component nodes, local noise interference may occur. This embodiment further introduces a Lie algebra averaging scheme based on edge vector alignment for smoothing: for each structural edge... The edge vector pairs before and after the transformation are extracted and mapped to the shortest unit quaternion. After averaging in the tangent space of the Lie algebra, the result is obtained through exponential mapping. Convert back to quaternion increment This step, while maintaining the decoupling characteristics of polar decomposition, enhances the continuity and physical reliability of node rotation. Specifically, For nodes Rotational increment quaternions, This represents the indices of the two adjacent nodes corresponding to the structural edge. This represents the exponential mapping from the space of Lie algebra tangents to the space of the unit quaternions. This represents the logarithmic mapping from the space of unit quaternions to the space of tangents of Lie algebras.

[0061] Step 108: Construct an anisotropic interpolation weight kernel based on the principal axis constraints of the components, and interpolate the translation increment, rotation increment and stretch increment from the finite element nodes to all three-dimensional Gaussian element parameter spaces to achieve physically consistent dynamic visualization rendering.

[0062] Since the finite element displacement field is defined only at sparse nodes, it must be interpolated onto millions of dense Gaussian elements. Traditional isotropic interpolation can lead to severe motion crosstalk between adjacent but unconnected components (such as at cross braces).

[0063] As a preferred embodiment, the anisotropic interpolation of the component principal axis constraint includes: For the source finite element node, utilize its Principal component analysis of nearest neighbor nodes estimates the local principal axis direction vector of the component in which it is located. ,in, The number of nearest neighbors. For the source node The local principal axis direction of the corresponding component. For the location of the target Gaussian element that needs interpolation. First, estimate its predicted spindle direction. Then calculate the relative distance vector from the source node to the target, and decompose it into axial components parallel to the principal axis of the component. and the transverse component perpendicular to the principal axis of the component .

[0064] Construct an anisotropic interpolation weight kernel with direction consistency constraints : In the formula, For the source node For the target location Anisotropic interpolation weights, The position of the target Gaussian element. Let be the axial component of the relative distance vector from the source node to the target position along the principal axis of the component. For the corresponding horizontal components, and Control the influence bandwidth along the axial and transverse directions of the component separately (typically set) ), The coefficient for the directional consistency penalty term. The estimated principal axis direction for the target location. This represents the local principal axis direction of the component corresponding to the source node. This represents the vector dot product.

[0065] Normalize the anisotropic interpolation weights to obtain ,in, The normalized interpolation weights, To the target location Center, search radius The set of source nodes participating in interpolation within the range, For the search radius, For target location The sum of the unnormalized weights of all source nodes in the neighborhood at that location; then the translation displacement of the finite element nodes. Lie algebra rotation increment and logarithmic scaling increment Perform weighted aggregation: Finally, the increment obtained from the aggregation is directly written back into the parameters of the target Gaussian unit: In the formula, and The first The center coordinates of the target Gaussian element before and after update. The translation increment obtained by interpolation. and These are the unit quaternion orientation parameters before and after the update, where... Compared with the unit quaternion stored in step 102 These represent similar orientation parameters, differing only in their index and notation; The rotation increment quaternion obtained by interpolation. This represents quaternion composition operations. This is the scale increment vector obtained through interpolation. Represents a unit quaternion Logarithmic mapping in the tangent space of Lie algebras Represents the scale increment vector The logarithmic representation of , This represents the scale increment vector obtained through interpolation. The logarithmic representation of , and Let represent the logarithmic scaling parameters of the Gaussian elements before and after the update, respectively. For the target Gaussian meta-index.

[0066] After the update, the covariance matrix of Gaussian elements will be based on the new quaternions. and new logarithmic scaling parameters Refactor and rebuild: .in, For the first The covariance matrix updated by Gaussian elements for each objective. For the reason The corresponding rotation matrix, For the reason The recovered diagonal scaling matrix For the first The covariance matrix before updating the target Gaussian elements. Compared to directly deriving the covariance through the locally deformed gradient matrix in numerical implementation (theoretically, it can be written as...), this is different. ,in, Let be the notation of the deformed gradient matrix of the local neighborhood of the corresponding target Gaussian element, and when (The theoretical form also maintains a positive definite covariance matrix when it is non-singular). This "write parameters → reconstruct" approach is more conducive to reducing the accumulation of numerical rounding errors and scaling imbalance.

[0067] Based on the dynamically updated 3D Gaussian parameter set, the elliptic weighted average Gaussian splashing algorithm can generate continuous bending and twisting deformation animations from any viewpoint, providing support for disaster prevention and mitigation and full life-cycle digital twin management of transmission towers.

[0068] like Figure 3 As shown, a corresponding system implementation is provided based on the above method implementation. This invention provides an image-based three-dimensional digital reconstruction and physical consistency dynamic visualization system for transmission tower structures. The system includes: a data acquisition and initialization module, an adaptive model training module, a terrain modeling and boundary expansion module, a physical driving data acquisition module, a coordinate registration and deformation mapping solution module, and a dynamic visualization rendering module. The data acquisition and initialization module is used to acquire multi-view image data of transmission lines and transmission towers, and extract sparse three-dimensional structure point cloud and camera pose through the motion reconstruction structure method, and initialize a three-dimensional Gaussian model. The adaptive model training module is used to train and iterate the three-dimensional Gaussian model under a fixed Gaussian element quantity budget constraint through a structure-terrain collaborative adaptive model optimization strategy, thereby constructing a static three-dimensional Gaussian scene of the transmission tower structure and its surrounding environment. The terrain modeling and boundary expansion module is used to construct a continuous digital elevation model based on adaptive terrain extraction and ordinary kriging method of the site surrounding the transmission tower, and to perform scene completion outside the view frustum of the static three-dimensional Gaussian scene by combining the digital elevation model as a constraint and a multi-layer boundary expansion strategy to obtain an expanded scene model. The physical drive data acquisition module is used to acquire displacement time history data of key nodes of the transmission tower; wherein, the displacement time history data is obtained by establishing a finite element model of the transmission tower and performing time-varying wind vibration response analysis, or by monitoring through a field sensor network, and is used as input for coordinate registration and Gaussian deformation mapping. The coordinate registration and deformation mapping solution module is used to unify the physical model coordinate system and the three-dimensional Gaussian scene model coordinate system by using iterative nearest point or other rigid registration algorithms, and to use a Gaussian deformation mapping method based on local right pole decomposition and component principal axis constraint anisotropic interpolation to map the discrete displacement time history data to the corresponding three-dimensional Gaussian element parameters in the extended scene model. The dynamic visualization rendering module is used to perform three-dimensional visualization rendering of the dynamic response and deformation process of the transmission tower structure based on the mapped and updated three-dimensional Gaussian primitive parameters.

Claims

1. A method for transmission tower image-driven reconstruction and dynamic visualization, characterized in that, The steps are as follows: Step 101: Obtain multi-view images containing the transmission tower structure and surrounding terrain environment, and perform pose estimation and sparse 3D reconstruction using the structure-in-motion algorithm to obtain the initial camera pose and sparse point cloud. Step 102: Based on the initial camera pose and sparse 3D point cloud, initialize the 3D Gaussian set and establish a differentiable rendering optimization loop based on image fidelity; Step 103: During the iterative training process of the differentiable rendering optimization loop, an adaptive model optimization strategy of structure-terrain collaboration is applied to optimize the modeling quality of key regions by triggering the splitting, cloning, and pruning operations of Gaussian units based on shape statistics and reconstruction gradients. Step 104: After the differentiable rendering optimization described in step 102 and the adaptive model optimization strategy described in step 103 reach the convergence condition, adaptive terrain extraction is performed on Gaussian elements based on local geometric features, and a continuous digital elevation model of the terrain is constructed using ordinary kriging interpolation. Step 105: Based on the continuous digital elevation model, by constructing a joint matching mechanism of feature similarity and elevation consistency, multi-layer boundary expansion is performed on the region outside the view frustum to obtain a static three-dimensional scene model after view frustum completion. Step 106: Obtain physical drive displacement time history data; Step 107: Using iterative nearest point or other rigid registration algorithms, the coordinate system of the physical model and the coordinate system of the 3D Gaussian scene model are uniformly and accurately registered to establish the initial spatial mapping relationship between discrete physical nodes and Gaussian primitives; after completing the coordinate registration, the local deformation is decoupled into rotation increment and stretching increment by estimating the local displacement gradient and using the right pole decomposition strategy. Step 108: Construct an anisotropic interpolation weight kernel based on the principal axis constraints of the components, and interpolate the translation increment, rotation increment and stretch increment from the finite element nodes to all three-dimensional Gaussian element parameter spaces to achieve physically consistent dynamic visualization rendering.

2. The method of transmission tower image-driven reconstruction and dynamic visualization of claim 1, wherein, In step 102, the 3D Gaussian metaset is initialized and a differentiable rendering optimization loop is established, including: Each 3D point in the sparse 3D point cloud is initialized as an anisotropic 3D Gaussian primitive, and the scene vector is represented as follows: ,in, The scene vector represents the set of all Gaussian meta-parameters; in, Indicates the first The center coordinates of each Gaussian element; Indicates the first The unit quaternion rotation parameter of Gaussian elements; Opacity; These are the spherical harmonic coefficients; For Gaussian meta-index, Let be the total number of Gaussian elements in the scene; to maintain the positive definiteness of the scaling parameter and improve numerical stability, a logarithmic scaling parameter is used. ,in, Indicates the first A logarithmic scaling vector of Gaussian elements along the three principal axes. The corresponding scale vector is represented by the diagonal scaling matrix, which restores the original value to its original form. , and These represent element-wise logarithmic operations and element-wise exponential operations, respectively. This represents an operator for constructing a diagonal matrix from vectors; The covariance matrix of Gaussian elements is calculated based on standard three-dimensional Gaussian splash factorization: in, For the first The covariance matrix of Gaussian elements, in the formula Indicates by parameters The corresponding rotation matrix obtained after transformation, This represents the transpose of the rotation matrix. For the reason The diagonal scaling matrix is ​​recovered; This indicates that the covariance matrix is ​​a positive definite matrix; In the differentiable rendering optimization loop, the forward propagation process performs camera projection, elliptical weighted average Gaussian splashing, and opacity-based color synthesis; the three-dimensional Gaussian primitives are projected onto the two-dimensional image plane to obtain the rendered image. ,in, For the parameter set The corresponding rendered image, The set of all Gaussian parameters vectorized in order. Represents vectorization operators; During backpropagation, the rendered image is calculated. Compared with the observed real image Reconstruction loss function between Calculate the gradient and update the Gaussian parameters.

3. The transmission tower image-driven reconstruction and dynamic visualization method according to claim 1, characterized in that, In step 103, the adaptive model optimization strategy for applying structure-terrain collaboration includes an importance-guided adaptive splitting and density control strategy: The adaptive splitting strategy is as follows: Calculate the first The aspect ratio of each high-speed element and volume These serve as representations of local geometric elongation and local volume size, respectively. Calculate the reconstruction loss function relative to the Gaussian metacenter. Position gradient magnitude ,in, The reconstruction loss function represents the first... The gradient magnitude at the center of each Gaussian element. The reconstruction loss function is expressed as follows: gradient, express Norm; Define Gaussian split trigger indicator function : In the formula, For the first The split trigger indicator value of each Gaussian unit. and These represent the global statistical quantile functions for aspect ratio and volume, respectively. and The set percentile level; For indicator functions; For cloning gradient threshold; symbol and They represent logical OR and logical AND respectively; when At that time, perform a directed splitting or cloning operation on the Gaussian unit; The adaptive density control strategy is as follows: Let For the key region, in the iteration steps The number of Gaussian elements in this region at that time was Apply a capacity limit constraint: In the formula, Representing the iterative steps Key areas The number of Gaussian elements within; Represents the critical region at the initial time. The number of Gaussian elements within; This refers to the capacity control coefficient for key areas. Meanwhile, based on the effective opacity contribution of screen space Pruning operations are performed, among which... Indicates the first Each Gaussian element in a pixel The effective opacity contribution to pixel color composition. This represents the average effective opacity contribution of the Gaussian pixel across all pixels. Represents a set of pixels The total number of pixels; different pruning thresholds are applied inside and outside the key area: In the formula, For key areas Internal pruning threshold For key areas External pruning threshold, and .

4. The transmission tower image-driven reconstruction and dynamic visualization method according to claim 1, characterized in that, In step 104, the construction of the continuous digital elevation model of the terrain includes: For each Gaussian core center In its Nearby Neighborhood Within, local normal vectors are calculated using principal component analysis. And then calculate the local slope angle. and elevation fluctuations : In the formula, For the first A high-ranking official nearest neighbor set The number of nearest neighbors. The local normal vector is obtained through principal component analysis. The global vertical coordinate axis vector. For the neighboring region The elevation of each element Representing vectors Norm, This indicates the conversion of radian angles to degrees. Then, a joint threshold is established to extract the initial terrain candidate set. : In the formula, This is the initial terrain candidate set obtained based on the joint threshold; This is the local slope angle threshold; Local elevation fluctuation threshold; symbol This indicates a logical AND operation. Calculating the adaptive robust threshold using the median absolute deviation and To resist local noise interference: In the formula, and These are the empirical quantiles for slope angle and elevation undulation, respectively; and These are the corresponding empirical quantile functions; This is the amplification factor for the absolute deviation of the median; Represents the set of slope angles for all samples; Represents the set of elevation fluctuations for all samples; Indicates the absolute deviation of the median; To suppress minor terrain undulations and residual noise, step sizes are used on the XY plane. Divide into grids, where, The grid step size of the XY plane grid is set for the terrain extraction stage; the mean elevation of the initial terrain points is calculated within each grid cell. and standard deviation The final terrain points are retained using a dynamic tolerance range: In the formula, Indicates the first Line 1 The mean elevation of the terrain samples within the column grid cell. This represents the standard deviation of the elevation of the terrain samples within the grid cell. and These represent the adaptive tolerance coefficients for the lower and upper bounds, respectively; Subsequently, the ordinary kriging algorithm was used to process the cleaned terrain point samples. Processing and reconstructing the continuous digital elevation model surface. ,in, This is a cleaned set of terrain point samples; at any predicted location Its elevation is estimated by a weighted sum of sample elevations: ,in, This represents the number of terrain sample points. To reconstruct the digital elevation model, For position The estimated elevation of the location, These are the planar coordinates of the terrain sample points. To correspond to the elevation, The location to be predicted. The weights are Kriging weights and satisfy the unbiased constraint. .

5. The transmission tower image-driven reconstruction and dynamic visualization method according to claim 1, characterized in that, In step 105, the multi-layer boundary expansion includes: In the XY plane, with step size Construct an occupying grid, in which, The grid step size for the XY plane occupied by the grid during the boundary expansion phase; record the occupied boundary edge of the current non-terrain area, and collect the blank grids adjacent to the occupied boundary as the first... Candidate unit set corresponding to the outer expansion iteration of the wheel boundary ,in, Number the iterative layers for boundary expansion; Applying minimum Chebyshev spacing constraints and maximum candidate number constraints in a single layer to Perform sparsification: In the formula, and For candidate unit set Any two candidate units in the, Denotes the Chebyshev norm. The minimum Chebyshev spacing threshold, This represents the upper limit of the number of candidate units in a single layer; Extract high-quality source unit sets near the occupied boundaries. High-quality source cells refer to occupied cells located near occupied boundaries, containing a sufficient number of Gaussian cells, with stable local geometry and texture features, and not contaminated by obvious holes or anomalous noise; for each candidate cell Based on spatial feature similarity Elevation Consistency under Digital Elevation Model Constraints Select the best matching source unit : The final score is based on the overall score. The highest-ranking pair is determined; among them, Candidate unit and source unit Spatial feature similarity score between them To score for elevation consistency, The normalized spatial feature similarity score, The combined score is based on the similarity of spatial features and the consistency of elevation. and These are the aggregate space feature vectors of the candidate unit and the source unit, respectively. To prevent tiny positive numbers with a denominator of zero, and These represent the elevation values ​​of the digital elevation model at the candidate and source elements, respectively. These are elevation consistency control parameters; All Gaussian elements within the matching source element are cloned to the candidate element; texture orientation misalignment is eliminated by rotation alignment around the Z-axis, and elevation compensation is calculated using a digital elevation model; the center coordinates of the cloned Gaussian elements are then determined. The updated formula is as follows: In the formula, The coordinates are the original center coordinates of the Gaussian cell to be cloned within the source cell. The updated Gaussian metacenter coordinates, Rotation angle around the Z-axis The corresponding yaw angle correction matrix, and The centroids in the XY plane of the source element and the candidate element are respectively. and These represent the elevation values ​​of the digital elevation model at the candidate and source elements, respectively. It is the global vertical unit vector.

6. The transmission tower image-driven reconstruction and dynamic visualization method according to claim 1, characterized in that, Step 107 is as follows: At the nodes of the finite element model First-order topological neighborhood Internally, a weighted least squares method with Gaussian decay weights is used, combined with... Regularization, fitting the local displacement Jacobian matrix : In the formula, For nodes The local displacement Jacobian matrix at that point. For the solution to be found Matrix variables, express The set of real matrices This represents the value of the independent variable that minimizes the objective function. For nodes The first-order topological neighborhood, For nodes With nodes Distance decay weight between them and They are nodes and nodes The initial coordinate vector, and They are nodes and nodes The displacement vector, for Regularization coefficient, express Norm, This represents the Frobenius norm, which is the square root of the sum of the squares of the elements of the matrix; Constructing the local deformation gradient matrix ,in, For nodes The local deformation gradient matrix at the location, It is a third-order identity matrix. For nodes The Jacobian matrix of the local displacement gradient at the location; in order to explicitly decouple rigid body rotation and local tension, for Perform right-pole decomposition: in, It is a rotation increment matrix. It is a special orthogonal group in three dimensions. The stretching increment matrix belongs to the set of three-dimensional symmetric positive definite matrices. ; Let stretching increment matrix The main stretch amount is ,in, For nodes In the The main tensile amount in the main direction, ;make This indicates the node formed by the three main tensile amounts. The scale increment vector is used; to improve numerical stability, it is transformed into a logarithmically scaled increment space for representation. 。 7. The transmission tower image-driven reconstruction and dynamic visualization method according to claim 1, characterized in that, In step 108, the anisotropic interpolation of the principal axis constraint of the component includes: For the source finite element node, utilize its Principal component analysis of nearest neighbor nodes estimates the local principal axis direction vector of the component in which it is located. ,in, The number of nearest neighbors. For the source node The local principal axis direction of the corresponding component; the location of the target Gaussian element to be interpolated. First, estimate its predicted spindle direction. Then calculate the relative distance vector from the source node to the target, and decompose it into axial components parallel to the principal axis of the component. and the transverse component perpendicular to the principal axis of the component ; Construct an anisotropic interpolation weight kernel with direction consistency constraints : In the formula, For the source node For the target location Anisotropic interpolation weights, The position of the target Gaussian element. Let be the axial component of the relative distance vector from the source node to the target position along the principal axis of the component. For the corresponding horizontal components, and The influence bandwidth along the axial and transverse directions of the component is controlled separately. The coefficient for the directional consistency penalty term. The estimated principal axis direction for the target location. This represents the local principal axis direction of the component corresponding to the source node. Represents the dot product of vectors; Normalize the anisotropic interpolation weights to obtain ,in, The normalized interpolation weights, To the target location Center, search radius The set of source nodes participating in interpolation within the range, For the search radius, For target location The sum of the unnormalized weights of all source nodes in the neighborhood at that location; then the translation displacement of the finite element nodes. Lie algebra rotation increment and logarithmic scaling increment Perform weighted aggregation: Finally, the increment obtained from the aggregation is directly written back into the parameters of the target Gaussian unit: In the formula, and The first The center coordinates of the target Gaussian element before and after update. The translation increment obtained by interpolation. and These are the unit quaternion orientation parameters before and after the update; The rotation increment quaternion obtained by interpolation. This represents quaternion composition operations. This is the scale increment vector obtained through interpolation. Represents a unit quaternion Logarithmic mapping in the tangent space of Lie algebras Represents the scale increment vector The logarithmic representation of , This represents the scale increment vector obtained through interpolation. The logarithmic representation of , and Let represent the logarithmic scaling parameters of the Gaussian elements before and after the update, respectively. For the target Gaussian meta-index; After the update, the covariance matrix of Gaussian elements will be based on the new quaternions. and new logarithmic scaling parameters Refactor and rebuild: ;in, For the first The covariance matrix updated by Gaussian elements for each objective. For the reason The corresponding rotation matrix, For the reason The recovered diagonal scaling matrix For the first The covariance matrix of each objective Gaussian element before update.

8. A dynamic visualization system for three-dimensional digital reconstruction and physical consistency of transmission tower structures based on images, used to implement the method described in any one of claims 1-7, characterized in that, The system includes: a data acquisition and initialization module, an adaptive model training module, a terrain modeling and boundary expansion module, a physical data acquisition module, a coordinate registration and deformation mapping solution module, and a dynamic visualization rendering module.

9. The image-based three-dimensional digital reconstruction and physical consistency dynamic visualization system for transmission tower structures according to claim 8, characterized in that, The specific modules are as follows: The data acquisition and initialization module is used to acquire multi-view image data of transmission lines and transmission towers, and extract sparse three-dimensional structure point cloud and camera pose through the motion reconstruction structure method, and initialize a three-dimensional Gaussian model. The adaptive model training module is used to train and iterate the three-dimensional Gaussian model under a fixed Gaussian element quantity budget constraint through a structure-terrain collaborative adaptive model optimization strategy, thereby constructing a static three-dimensional Gaussian scene of the transmission tower structure and its surrounding environment. The terrain modeling and boundary expansion module is used to construct a continuous digital elevation model based on adaptive terrain extraction and ordinary kriging method of the site surrounding the transmission tower, and to perform scene completion outside the view frustum of the static three-dimensional Gaussian scene by combining the digital elevation model as a constraint and a multi-layer boundary expansion strategy to obtain an expanded scene model. The physical drive data acquisition module is used to acquire displacement time history data of key nodes of the transmission tower; wherein, the displacement time history data is obtained by establishing a finite element model of the transmission tower and performing time-varying wind vibration response analysis, or by monitoring through a field sensor network, and is used as input for coordinate registration and Gaussian deformation mapping. The coordinate registration and deformation mapping solution module is used to unify the physical model coordinate system and the three-dimensional Gaussian scene model coordinate system by using iterative nearest point or other rigid registration algorithms, and to use a Gaussian deformation mapping method based on local right pole decomposition and component principal axis constraint anisotropic interpolation to map the discrete displacement time history data to the corresponding three-dimensional Gaussian element parameters in the extended scene model. The dynamic visualization rendering module is used to perform three-dimensional visualization rendering of the dynamic response and deformation process of the transmission tower structure based on the mapped and updated three-dimensional Gaussian primitive parameters.

Images