A High-Fidelity 3D Gaussian Reconstruction Method for Concrete Dam Scene Images Captured by UAVs
By employing a planar Gaussian splashing algorithm and a multi-stage training process, combined with RTK localization and SfM to generate sparse point clouds, and optimizing Gaussian point cloud densification and pruning, the problems of holes and local structural weakening in model reconstruction in concrete dam scenarios were solved, resulting in a high-fidelity 3D Gaussian model.
Patent Information
- Application Number
- CN202611310219.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-27
- Publication Date
- 2026-11-17
AI Technical Summary
Existing 3D reconstruction methods struggle to accurately reconstruct weakly textured areas in concrete dam scenes, resulting in holes and blurred local texture rendering in the model. Small-sized structures cannot be reconstructed, and local structures are weakened by the overall scene during full-scene training.
A planar Gaussian splashing algorithm combined with a multi-stage training process is adopted. Sparse point clouds are generated through RTK localization and SfM. A Gaussian metamorphic detection and rendering contribution-observability joint pruning mechanism is introduced to optimize Gaussian point cloud densification and pruning, generating a 3D Gaussian model with clear scene details and accurate geometric structure.
It realizes the generation of 3D Gaussian models with rich scene details and accurate geometric structure in complex water conservancy environments, solves the problem of difficulty in taking into account the overall structure, local details and geometric expression in large scene reconstruction, and reduces the impact of redundant Gaussian units and abnormal stretching.
Smart Images

Figure CN123118904A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of 3D reconstruction and deep learning technology, specifically to a method for high-fidelity 3D Gaussian reconstruction of UAV aerial images of concrete dam scenes using an improved planar Gaussian splashing algorithm and a multi-stage training process. Background Technology
[0002] Concrete dams play a crucial role in water storage and flood discharge. They are typically large in size, diverse in structural form, and complex in spatial relationships. While drone images can capture local details during dam analysis and inspection, they struggle to provide a coherent view of the dam's overall condition and the spatial relationships between different structures. 3D reconstruction methods can be used to create structural models with realistic textures and provide 3D geometric information. However, conventional photogrammetric 3D reconstruction methods often struggle to accurately reconstruct weakly textured areas such as dam gates and spillways, leading to voids in the model. Therefore, it is necessary to introduce deep learning-based 3D reconstruction methods.
[0003] The 3D Gaussian splashing method generates a high-quality 3D Gaussian model by establishing and optimizing a 3D Gaussian field. However, relying solely on photometric loss during reconstruction makes it difficult to establish stable constraints on the scene surface, easily leading to inconsistencies between the Gaussian primitive distribution and the real geometric surface, thus affecting the geometric accuracy of the 3D model. The planar Gaussian splashing algorithm improves geometric accuracy by planarizing Gaussian primitives and introducing geometric constraints to make them closer to the real scene surface. However, the concrete dam scene is large, and drone photography collects redundant information about the surrounding environment. If full-scene training is performed directly, the reconstruction will prioritize fitting the overall outline of the dam scene and the background area, while weakening small-sized feature areas such as joints, railings, and structural edges. This results in insufficient local surface constraints, causing problems such as blurred local texture rendering and the inability to reconstruct small-sized structures in the output concrete dam model.
[0004] To this end, this invention builds a multi-stage training process based on the planar Gaussian splashing algorithm. During the training process, the Gaussian point cloud densification and pruning logic are optimized to generate a three-dimensional Gaussian model with clear scene details and accurate geometric structure, providing a more reliable model foundation for the field of digital twins in water conservancy engineering. Summary of the Invention
[0005] The purpose of this invention is to propose a high-fidelity 3D Gaussian reconstruction method for UAV aerial images of concrete dam scenes, which can generate 3D Gaussian models with rich scene details and accurate geometric structure in complex hydraulic environments.
[0006] To achieve the above objectives, the technical solution of the present invention is: a high-fidelity 3D Gaussian reconstruction method for UAV aerial images of concrete dam scenes, comprising the following steps:
[0007] Step 1: Using a drone equipped with Real-Time Kinematic (RTK) positioning, multi-view aerial photography of the concrete dam scene is conducted to acquire aerial images covering the entire dam structure and its local components, along with corresponding location information. Combining the RTK positioning information, the Structure from Motion (SfM) method is used to generate sparse point clouds and camera parameters with a realistic scene scale, completing the image data preprocessing.
[0008] Step 2: Based on the aerial images, camera parameters, and sparse point clouds obtained in Step 1, a coarse training of the entire concrete dam scene is performed using the Planar-based Gaussian Splatting Reconstruction algorithm. A morphology detection mechanism is added to the original training process to determine the Gaussian morphology and restrict abnormally stretched Gaussian morphologies from further densifying, thereby generating a coarsely trained 3D Gaussian model covering the entire concrete dam scene.
[0009] Step 3: Based on the 3D Gaussian model generated in Step 2, project all Gaussian primitive center points onto a horizontal plane, statistically analyze their distribution, and divide the concrete dam scene into multiple blocks along the direction perpendicular to the river flow. Determine the boundary positions of each block based on the relatively balanced number of Gaussian primitives contained in each block. Combining the camera center position and the visibility of image rendering perturbations, determine the visibility of each aerial image relative to each block, assign images with high visibility to the corresponding blocks, and generate the image set used for training each block.
[0010] Step 4: Using the coarse-trained 3D Gaussian model obtained in Step 2 as the unified initial model for each block, and combining it with the block image set obtained in Step 3, perform fine training on the local scene of the concrete dam for each block, so that the visible Gaussian elements in the training image corresponding to the current block are optimized. During the training process, the Gaussian element morphology detection mechanism is also used, and rendering contribution-observability joint pruning is further introduced to remove low contribution, weak observation and redundant Gaussian elements, generating multiple refined 3D Gaussian models corresponding to different blocks.
[0011] Step 5: Read the refined 3D Gaussian model output from each block in Step 4, perform boundary trimming based on the block boundary information in Step 3, retain the Gaussian elements and their complete parameters within the corresponding block range in each model, merge the retained Gaussian elements in each block under a unified global coordinate system to form a 3D Gaussian expression covering the entire scene of the concrete dam, and output a high-fidelity refined 3D Gaussian model of the entire scene of the concrete dam with the scale of the real scene.
[0012] Furthermore, the specific steps for image data acquisition and preprocessing in step 1 are as follows:
[0013] 1-1 Multi-view image data of concrete dam scene was collected by drone with RTK module. During shooting, the drone flew at a certain height above the dam top along a preset flight path and captured the overall scene image of the dam vertically downward. Images of the dam surface, spillway and other structures were collected at different heights along the elevation direction. The flight height and shooting angle were then adjusted to capture close-up shots of the gate piers, gates and other structures. The RTK position information corresponding to each image was recorded during the shooting process, and the image shooting distance and overlap rate were kept consistent.
[0014] 1-2 Calculate the mean values of longitude, latitude, and geodetic height corresponding to the aerial image in 1-1, respectively, and use them as the reference origin of the local spatial rectangular coordinate system. Convert the location information of the aerial image and the reference origin from WGS84 geodetic coordinates to WGS84 geocentric-ground-fixed coordinates. The conversion formula is as follows:
[0015] (1)
[0016] (2)
[0017] (3)
[0018] (4)
[0019] In the formula, N represents the radius of curvature of the ramusoidal circle, a represents the semi-major axis of the reference ellipsoid (for the WGS84 ellipsoid, a is 6378137.0 m), e represents the first eccentricity, L represents the geodetic longitude, B represents the geodetic latitude, H represents the geodetic height, and X, Y, and Z represent the WGS84 geocentric coordinates.
[0020] Subtract the geocentric-geocentric coordinates (X0, Y0, Z0) of the reference origin from the geocentric-geocentric coordinates (X, Y, Z) of each image. Then, construct a rotation matrix based on the longitude L0 and latitude B0 of the reference point to perform a rotation transformation, further converting the geocentric-geocentric coordinate expression to the Northeast Eon (ENU) local spatial rectangular coordinate expression. The formula is as follows:
[0021] (5)
[0022] (6)
[0023] 1-3 Based on the aerial images acquired in 1-1, the SfM method is used to establish matching relationships between images. Based on these matching relationships, the camera parameters and sparse 3D points corresponding to each image are initially estimated. Simultaneously, the local spatial Cartesian coordinates obtained in 1-2 are used as a priori camera center position. A bundle adjustment method incorporating the prior error of the camera center position is used to jointly optimize the camera parameters and the sparse 3D point positions, resulting in a sparse 3D point cloud and camera parameters with a real-world scene scale. The objective function for bundle adjustment optimization with the camera center position prior constraint is as follows:
[0024] (7)
[0025] In the formula K i R is the camera intrinsic parameter corresponding to the i-th image. i Let t be the rotation matrix in the camera pose corresponding to the i-th image. i Let X be the translation vector. j For the j-th sparse 3D point, x ij Let j be the two-dimensional observation coordinates of the j-th 3D point in the i-th image. Let be the set of observable 3D points in the i-th image. C is the projection function of a 3D point onto the image plane. i The camera center is calculated from the camera pose. The prior position of the camera center after conversion from RTK latitude, longitude, and altitude to the ENU coordinate system. Let be the covariance matrix of the prior location of the i-th image. The prior error weights for the camera center position.
[0026] Furthermore, the specific steps for the full-scene coarse training of the planar Gaussian splashing method in step 2 are as follows:
[0027] 2-1 Based on the sparse point cloud in step 1, an initial three-dimensional Gaussian representation is constructed for each three-dimensional point, including the Gaussian center position, covariance matrix, color parameters and opacity, to form an initial three-dimensional Gaussian field.
[0028] 2-2 Select an image as the current training viewpoint, project all Gaussian primitives onto the image plane, determine the set of Gaussian primitives corresponding to each pixel, and use a tile-based differentiable Gaussian splash renderer to perform block traversal and visibility sorting of the planar Gaussian primitives in screen space. Then, accumulate the color contribution of each Gaussian primitive layer by layer using the alpha mixing method to generate the RGB image of the current viewpoint. During the rendering process, perform consistent weighted fusion of the normal vectors of each planar Gaussian and their signed planar distances to the camera, outputting a normal map and a planar distance map. Calculate pixel depth through ray and planar geometric relationships to achieve unbiased depth recovery, ensuring that the depth is strictly constrained to the planar geometric structure, and output an unbiased depth map.
[0029] 2-3 A scale compression loss constraint is constructed to minimize the scale along the principal axis of the Gaussian primitives, causing them to gradually become planar during optimization. A single-view local geometric consistency constraint is constructed based on the unbiased depth map. 3D points are recovered through backprojection of neighboring pixels and fitted to a local plane, with consistency constraints applied to the rendering normals. A multi-view geometric consistency constraint is constructed, achieving cross-view pixel mapping through camera homography and calculating the projection error to constrain geometric consistency. Further calculations of multi-view photometric consistency are performed, and patch alignment and normalized cross-correlation metrics ensure cross-view appearance consistency. The total loss function is constructed as follows:
[0030] (8)
[0031] In the formula, L total Let L be the total loss function. rgb For image reconstruction loss, L scale For the scale compression loss, L normal For single-view normal consistency loss, L geo For multi-view geometric consistency loss, L ncc λs, λn, λg, and λp represent the loss weights for each of the multiple-view NCC photometric consistency loss terms.
[0032] Backpropagation based on the total loss updates the position, opacity, covariance matrix, and color features of each Gaussian element, making the rendered image closer to the real aerial image and the Gaussian elements closer to the real scene surface.
[0033] 2-4 Calculate the gradient and absolute gradient of each Gaussian primitive on the two-dimensional image plane to determine whether there is a large rendering error in the region where the Gaussian primitive is located. At the same time, retain the magnitude information of the error change to avoid the gradients in different directions canceling each other out. When the gradient or absolute gradient of a Gaussian primitive is greater than a preset threshold, the Gaussian primitive is regarded as a candidate for densification.
[0034] 2-5 Before performing Gaussian point cloud densification, a Gaussian morphology detection mechanism is introduced to read the scale of candidate Gaussian morphologies in the three principal axes. Since the planar Gaussian splashing method compresses a certain principal axis size to achieve Gaussian morphology planarization, the ratio of the maximum scale value to the intermediate scale value is used to determine whether there is abnormal stretching of the Gaussian morphology. If the ratio is greater than a preset threshold, it is determined to be an extremely abnormal Gaussian morphology and is no longer allowed to participate in cloning or splitting.
[0035] 2-6 For the candidate Gaussian elements that pass the morphological detection, if the maximum scale value is not greater than the preset scale threshold, the Gaussian element is cloned to increase the number of local Gaussian elements; if the maximum scale value is greater than the preset scale threshold, the Gaussian element is split to generate multiple new Gaussian elements with smaller scales in the neighborhood of the original Gaussian element.
[0036] 2-7 Pruning is performed on Gaussian primitives based on their opacity, 2D projection radius, 3D scale, and number of multi-view projection observations during training. Gaussian primitives with excessively low opacity, excessively large projection range, excessively large 3D scale, or insufficient multi-view observations are deleted based on threshold criteria.
[0037] 2-8 Repeat the image selection, differentiable rendering, loss calculation, backpropagation, densification and pruning process until the preset number of iterations is reached, then the training ends and the coarse-trained 3D Gaussian model of the entire concrete dam scene is saved.
[0038] Furthermore, the specific steps of the full-scene coarse-training 3D Gaussian model block partitioning method in step 3 are as follows:
[0039] 3-1 Read the center point position of Gaussian elements in the coarse-trained model of the whole scene, and calculate the camera center position by combining the camera parameters obtained in 1-3, and determine the coverage area of the target scene in the horizontal plane.
[0040] 3-2 Project the center points of all Gaussian elements onto a horizontal plane, with the direction parallel to the river flow as the y-axis and the direction perpendicular to the river flow as the x-axis, so as to divide the concrete dam into blocks according to the dam section. According to the preset number of blocks, count the distribution of Gaussian elements, determine the boundary of each block according to the proportion of the number, and form a spatial block between adjacent boundaries to obtain a scene block with a relatively balanced number of Gaussian elements.
[0041] 3-3 Determine the basic attribution relationship of the image based on the camera center position corresponding to each training image. When the camera center projection position of the training image is located within the two-dimensional plane boundary of a certain spatial block, the training image is assigned to that spatial block.
[0042] 3-4 Based on the rendering results of the coarse-trained model across the entire scene at various training perspectives, calculate the visible content score for each spatial block of the training image. The formula is as follows:
[0043] (9)
[0044] In the formula, Let I represent the visibility contribution of the c-th camera to the (j,k)-th block; c This represents the original rendered image; This indicates that block B will be... j,kThe image obtained by re-rendering after setting the transparency of the inner Gaussian meta-transparency to zero.
[0045] When the visible content score is greater than a preset threshold, it is determined that the training image has a valid visible relationship with the spatial block, and the image is assigned to the training image set of the block.
[0046] 3-5 Merge the set of location-attributed images obtained in step 3-3 based on the camera center position with the set of content-visible images obtained in step 3-4 based on the visibility of image rendering perturbation, and remove duplicate images to obtain the final training image set corresponding to each spatial block.
[0047] Furthermore, the specific steps of the local scene fine-tuning in step 4 are as follows:
[0048] 4-1 Load the full-scene coarse-trained Gaussian model obtained in step 2 as the unified initial model for each scene block. Only read the image set corresponding to each block to participate in the block training, so as to realize the local refinement and optimization of the visible Gaussian elements in the block.
[0049] 4-2 During the local scene fine training process, the training total loss function in 2-3 is still used. The image reconstruction loss is included in the calculation of each iteration, the single-view normal consistency loss is calculated once every 2 iterations, and the multi-view geometric consistency loss and multi-view NCC photometric consistency loss are calculated once every 4 iterations, so as to reduce the memory usage and computational overhead of the fine training stage.
[0050] 4-3 Before the Gaussian morphology is increased, the Gaussian morphology is also checked to prevent the spread of Gaussian morphology with extreme abnormalities during block fine training.
[0051] 4-4 During Gaussian pruning, a joint pruning mechanism of rendering contribution and observability is adopted. This mechanism calculates the retention score based on the rendering contribution and observability of the Gaussian gram in the current block of the training image, as shown in the following formula:
[0052] (10)
[0053] (11)
[0054] In the formula, S i Let C be the overall score of the i-th Gaussian element. i The score contributes to the rendering of the i-th Gaussian pixel; the larger the value, the more important the Gaussian pixel is to the rendering result. i The observability score represents how many views in the training view set actually observe a Gaussian element, w. c and w o α represents the weight; K represents the number of perspectives participating in the final contribution statistics; α represents the weight.i,k T represents the effective opacity of the i-th Gaussian element at the k-th viewpoint. i,k It represents the cumulative transmission weight of the i-th Gaussian element at the k-th viewpoint, which is used to characterize the effective visibility of the Gaussian element after it is blocked by the preceding Gaussian element along the current line of sight.
[0055] By calculating the corresponding retention score for each Gaussian primitive and sorting them in descending order, a certain proportion of low-score Gaussian primitives are deleted, thereby reducing artifacts, floating points, and point cloud redundancy in local areas.
[0056] 4-5. The Gaussian point cloud parameters are updated using a learning rate lower than that used in the coarse training phase, which reduces the parameter fluctuations in the position, scale, and color representation of Gaussian points during iterative optimization. The number of iterations is increased while the frequency of Gaussian point cloud densification pruning is slowed down, so as to achieve more stable refinement and optimization of the position, color, opacity, scale, and rotation parameters of existing Gaussian points.
[0057] 4-6 Perform detailed training on each block separately to obtain multiple refined Gaussian models corresponding to each block.
[0058] Furthermore, the specific steps for merging local scene blocks and outputting a refined 3D Gaussian model of the entire concrete dam scene in step 5 are as follows:
[0059] 5-1 Read the spatial boundary information of each scene block generated in step 3 and the information of each local scene fine-trained 3D Gaussian model obtained in step 4, and extract the center position coordinates of each Gaussian element.
[0060] 5-2 For any scene block, only the Gaussian elements whose center point falls within the horizontal boundary of the block are retained. The elevation direction is not included in the clipping to avoid the accidental deletion of effective Gaussian elements due to changes in dam height, slope height difference, or differences in aerial photography perspective.
[0061] 5-3 A unified boundary assignment rule is used to handle the common boundaries of adjacent scene blocks. A closed boundary judgment is used for the lower boundary of the block, and an open boundary judgment is used for the upper boundary of the block, so that Gaussian elements located near the common boundary belong to only one scene block, thereby reducing duplicate retention and boundary overlap during block merging.
[0062] 5-4 Perform a field consistency check on each cropped local scene 3D Gaussian model to confirm that it contains all necessary fields such as position parameters, color spherical harmonic feature parameters, opacity parameters, scale parameters, and rotation parameters. For blocks with complete fields and consistent format, merge their Gaussian metadata according to the spatial block order. For blocks with missing fields or inconsistent format, record the exception information and prompt for reprocessing.
[0063] 5-5 The merged Gaussian meta-set is output as a PLY format file to obtain a refined 3D Gaussian model of the entire concrete dam scene. Since the model is trained and merged based on sparse point clouds and camera parameters at the real scene scale, the output model has a scale representation consistent with the actual scene.
[0064] Compared with existing technologies, the beneficial effects of this invention are as follows: Considering that the aerial photography scene of a concrete dam is large, the local structure is easily weakened by the overall scene when directly performing full-scene 3D Gaussian training, and redundant Gaussian primitives and abnormally stretched Gaussian primitives are easily generated during the training process. This invention adopts a planar Gaussian splashing algorithm combined with a multi-stage training process, and introduces a Gaussian primitive morphology detection mechanism and a rendering contribution-observability joint pruning mechanism. This solves the problem that it is difficult to take into account the overall structure, local details and geometric expression in large-scene reconstruction, as well as the problem that abnormal Gaussian primitives and redundant Gaussian primitives affect the reconstruction quality. Attached Figure Description
[0065] Figure 1 This is a flowchart illustrating a high-fidelity 3D Gaussian reconstruction method for UAV aerial images of concrete dam scenes in this embodiment.
[0066] Figure 2 This is a schematic diagram of the training process based on the planar Gaussian splashing method in this embodiment;
[0067] Figure 3 This is a schematic diagram of the full-scene block division process in this embodiment;
[0068] Figure 4 This is a schematic diagram of the detailed training process for a local scene in this embodiment;
[0069] Figure 5 This is a schematic diagram comparing the aerial image and the RGB image of the model trained by this method in this embodiment;
[0070] Figure 6 This is the normal map of the 3D Gaussian model trained in this embodiment. Detailed Implementation
[0071] The present invention will be further illustrated below with reference to specific embodiments. These embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.
[0072] Figure 1 The diagram shows the overall technical roadmap for a high-fidelity 3D Gaussian reconstruction method for concrete dam scenes captured by UAV aerial photography. Its specific implementation process includes:
[0073] Step 1: Acquire images of the concrete dam and its corresponding location information, and perform preprocessing.
[0074] Specifically, step 1 includes:
[0075] 1-1 Aerial image data of a concrete dam scene was collected from multiple perspectives using a drone equipped with an RTK module. During the acquisition process, the drone first flew along a preset flight path within a preset height range above the dam crest, capturing vertically downward aerial images of the dam crest, dam shoulders, and other overall dam scenes. Subsequently, horizontal images of the dam surface were collected at different heights along the vertical or diagonal flight path, focusing on the dam facade, spillway, and guide wall.
[0076] Further detailed shooting was conducted on specific structural elements such as sluice gates, gate towers, maintenance platforms, and flood discharge outlets. The drone's flight altitude, shooting distance, and gimbal angle were adjusted for close-up reshoots. During image acquisition, the forward and lateral overlap rates between adjacent images were controlled within preset ranges, and the latitude, longitude, and altitude information in the WGS84 coordinate system corresponding to each aerial image were recorded.
[0077] 1-2 Calculate the average longitude, latitude, and geodetic height of the location information corresponding to the aerial image in 1-1, and use this as a reference point near the center of the aerial camera's trajectory. This reference point is then used as the origin of the ENU local spatial rectangular coordinate system. Convert the image's location information and the reference point from WGS84 geodetic coordinates to WGS84 geocentric-ground-fixed coordinates using the following conversion formula:
[0078] (1)
[0079] (2)
[0080] (3)
[0081] (4)
[0082] In the formula, N represents the radius of curvature of the ramusoidal circle, a represents the semi-major axis of the reference ellipsoid (for the WGS84 ellipsoid, a is 6378137.0 m), e represents the first eccentricity, L represents the geodetic longitude, B represents the geodetic latitude, H represents the geodetic height, and X, Y, and Z represent the WGS84 geocentric-geofixed coordinates.
[0083] Subtract the geocentric and geofixed coordinates of the reference point from the geocentric and geofixed coordinates of each image. Then, construct a rotation matrix based on the longitude and latitude of the reference point to perform a rotation transformation, further converting the geocentric and geofixed coordinate representation to ENU local space rectangular coordinate representation. The formula is as follows:
[0084] (5)
[0085] (6)
[0086] 1-3 Extract feature points from the aerial images acquired in 1-1, match feature points in different images to obtain the correspondence between corresponding feature points between images, and eliminate incorrect matches by combining geometric consistency verification; select initial image pairs with a large number of corresponding feature points and good geometric consistency, estimate the relative camera pose of the initial image pairs through two-view geometry, and calculate the three-dimensional spatial position corresponding to the corresponding feature points using linear triangulation method to generate initial sparse three-dimensional points.
[0087] Subsequently, new images are added one by one. Based on the correspondence between the generated 3D points and the 2D feature points of the new images, the camera pose of the new images is solved using a PnP-based absolute pose estimation method. The sparse 3D point cloud is then further expanded using triangulation. Finally, the local spatial Cartesian coordinates obtained in steps 1-2 are used as the prior for the camera center position of the corresponding image. A bundle adjustment optimization objective is introduced to jointly optimize the camera intrinsic parameters, camera extrinsic parameters, and 3D point positions, ultimately obtaining a sparse 3D point cloud and camera parameters with a real-world scene scale. The bundle adjustment optimization objective function constrained by the camera center position prior is as follows:
[0088] (7)
[0089] In the formula K i R is the camera intrinsic parameter corresponding to the i-th image. i Let t be the rotation matrix in the camera pose corresponding to the i-th image. i Let X be the translation vector. j For the j-th sparse 3D point, x ij Let j be the two-dimensional observation coordinates of the j-th 3D point in the i-th image. Let be the set of observable 3D points in the i-th image. C is the projection function of a 3D point onto the image plane. i The camera center is calculated from the camera pose. The prior position of the camera center after conversion from RTK latitude, longitude, and altitude to the ENU coordinate system. Let be the covariance matrix of the prior location of the i-th image. The prior error weights for the camera center position.
[0090] Step 2: Perform coarse training across the entire scene using the planar Gaussian splashing method, see... Figure 2 .
[0091] Specifically, step 2 includes:
[0092] 2-1 Based on the sparse 3D point cloud from step 1, the coordinates of the points are used as the Gaussian center positions; the initial principal axis scale is estimated based on the distance between local neighborhood points in the point cloud, and all axes are initially set to the same scale to construct a covariance matrix, which controls the size and orientation of the Gaussian elements; the color information corresponding to the points is obtained through multi-view image projection sampling to obtain the initial color representation; a uniformly small initial opacity parameter is set. Finally, the sparse point cloud is converted into an initial 3D Gaussian field composed of multiple 3D Gaussian elements.
[0093] 2-2 An image is selected as the current training viewpoint, and all Gaussian primitives are projected onto the image plane to form a two-dimensional Gaussian distribution in screen space. The Gaussian set corresponding to each pixel is determined. A tile-based differentiable Gaussian splash renderer is used to perform block traversal and visibility sorting of the planar Gaussians in screen space. The color contribution of each Gaussian is accumulated layer by layer using the alpha mixing method to generate the RGB image of the current viewpoint. Simultaneously, during the same rendering process, the normal vectors of each planar Gaussian and their signed planar distances to the camera are consistently weighted and fused, outputting a normal map and a planar distance map. Pixel depth is calculated using ray and planar geometric relationships to obtain an unbiased depth map.
[0094] 2-3 Construct a scale compression loss to achieve Gaussian planarization by compressing the minimum principal axis scale of Gaussian cells; construct a single-view local geometric consistency constraint based on unbiased depth results, recover 3D points by backprojecting neighboring pixels and fitting a local plane, and constrain consistency with the rendering normal; construct a multi-view geometric consistency constraint, achieve cross-view pixel mapping through the homography relationship between cameras, and calculate the front and back projection errors to constrain geometric consistency.
[0095] Based on this, multi-view photometric consistency is further calculated, and cross-view appearance consistency is ensured through patch alignment and normalized cross-correlation metric. A total loss function is constructed based on the above loss, and backpropagation is performed based on the total loss to update the position, opacity, covariance matrix, and color features of each Gaussian element, making the rendered image closer to a realistic aerial image and making the Gaussian local plane closer to the surface of the real scene.
[0096] The total loss function is constructed as follows:
[0097] (8)
[0098] In the formula, L total Let L be the total loss function. rgb For image reconstruction loss, L scale For the scale compression loss, L normal For single-view normal consistency loss, L geo For multi-view geometric consistency loss, L nccλs, λn, λg, and λp represent the loss weights for each of the multiple-view NCC photometric consistency loss terms.
[0099] 2-4 Calculate the gradient and absolute gradient of each Gaussian primitive on the two-dimensional image plane. The gradient is the norm of the derivative of the loss function with respect to the position of the Gaussian on the image plane, and the absolute gradient is the sum of the absolute values of the gradients at all pixels affected by the Gaussian primitive. When the image spatial gradient or absolute gradient of the region corresponding to a Gaussian primitive is greater than a preset threshold, the Gaussian primitive is regarded as a densification candidate.
[0100] 2-5 Before performing Gaussian point cloud densification, a Gaussian morphology detection mechanism is introduced to read the three principal axis scale parameters of candidate Gaussian morphologies and sort them by size to obtain the minimum scale s. min Intermediate scale s mid and maximum scale s max , where s min ≤s mid ≤s max The degree of elongation within a plane is determined by the ratio of the largest scale to the intermediate scale.
[0101] (9)
[0102] In the formula, R represents the in-plane dimensional elongation of the candidate Gaussian element, and ε is a minimal constant to prevent division by zero.
[0103] By setting a threshold, it is determined whether Gaussian units are overstretched. If the ratio is greater than the preset threshold, it is determined to be an extreme abnormal Gaussian and is no longer allowed to participate in cloning or splitting.
[0104] 2-6 For the Gaussian candidate that passes the morphological detection, if the maximum scale value is not greater than the preset scale threshold, the Gaussian cell is cloned to increase the number of local Gaussians; if the maximum scale value is greater than the preset scale threshold, the Gaussian cell is split to generate multiple new Gaussian cells with smaller scales in the neighborhood of the original Gaussian cell.
[0105] 2-7 Pruning operations are performed on Gaussian primitives based on their opacity, 2D projection radius, 3D scene scale, and number of multi-view projection observations during training. Gaussian primitives with excessively low opacity, excessively large screen projection area, excessively large world space scale, and insufficient number of multi-view observations are deleted based on threshold criteria.
[0106] 2-8 Repeat the image selection, differentiable rendering, loss calculation, backpropagation optimization, morphological awareness densification constraint, Gaussian unit densification, and pruning process until the preset number of coarse training iterations of 30,000 is reached. After multiple rounds of optimization, the model forms a relatively stable color representation and overall geometric structure, resulting in a full-scene coarse-trained 3D Gaussian model.
[0107] Step 3, the method for coarsely training the 3D Gaussian model across the entire scene and dividing it into blocks, see... Figure 3 .
[0108] Specifically, step 3 includes:
[0109] 3-1 Calculate the camera center position corresponding to each training image using the camera parameters obtained in step 1. Determine the horizontal coverage of the target scene based on the distribution range of the Gaussian metacenter position and the camera center position in the horizontal plane in the full-scene coarse training model in step 2.
[0110] 3-2 Project all Gaussian element center points onto a horizontal plane, using the direction parallel to the river flow as the y-axis and the direction perpendicular to the river flow as the x-axis. Divide the concrete dam into blocks according to dam segments. In this embodiment, the preset number of blocks is 5×1. That is, on the x-axis, the Gaussian element center coordinates are sorted from smallest to largest, and the boundaries of 5 blocks are determined sequentially according to the proportion of Gaussian elements. No further subdivision is performed on the y-axis. This approach maintains the structural integrity of the dam segment in each block while also considering training efficiency. This results in narrower blocks for densely populated dam details and wider blocks for sparsely populated water surfaces, banks, or distant areas, leading to multiple local scene blocks with relatively balanced Gaussian element numbers and clear spatial boundaries.
[0111] 3-3 For any training image, read the projected coordinates of its camera center in the horizontal plane. When the projected coordinates fall within the two-dimensional boundary of a spatial block, assign the training image to that spatial block.
[0112] 3-4 Calculate the visible content score of each spatial block in the training image based on the rendering results of the full-scene coarse-trained model in step 2. For the (j, k)th spatial block, first determine the set of Gaussian elements within the two-dimensional boundary of the block. Then, render the original image from the c-th camera viewpoint, and then re-render after setting the opacity of the Gaussian elements within the block to zero to obtain the perturbed rendered image.
[0113] The visible content score is calculated based on the structural similarity difference between two rendered images. When the visible content score is greater than a preset threshold, it indicates that the rendering result of the current viewpoint has changed significantly after removing the Gaussian elements of that block. It is determined that the training image can effectively observe the block, and it is then added to the training image set for that spatial block.
[0114] 3-5 Merge the set of images with camera center attribution obtained in 3-3 with the set of rendered visible images obtained in 3-4 to obtain the final set of training images corresponding to each spatial block. Each block ultimately contains both images taken from its vicinity and images from which the camera center is not located but whose content can still be clearly observed.
[0115] Step 4, detailed training in local scenarios, see Figure 4 .
[0116] Specifically, step 4 includes:
[0117] 4-1 The coarse-trained Gaussian point cloud from step 2 is used as the unified initial Gaussian point cloud for each spatial block, allowing each block to inherit the overall geometric structure and color representation of the entire scene. For any spatial block, the relevant training image set corresponding to that block is read to participate in the fine-tuning of the local scene of the current block.
[0118] 4-2 During the block refinement training process, the total loss function established in the coarse training is continued to guide the optimization, but geometric constraint calculation is performed at a low frequency according to the preset interval. The image reconstruction loss participates in the calculation of each iteration, the single-view normal consistency loss is calculated once every 2 iterations, and the multi-view geometric consistency loss and multi-view NCC photometric consistency loss are calculated once every 4 iterations, so as to reduce the memory usage and computational overhead of the refinement training stage.
[0119] 4-3 Before each Gaussian element densification, Gaussian element morphology detection is also performed on candidate Gaussian elements to suppress the continued diffusion of abnormally long strip-shaped Gaussian elements near the dam surface edge, joints, etc., and reduce edge trailing and cross-structure coverage.
[0120] 4-4 During Gaussian pruning, a joint pruning mechanism of rendering contribution and observability is adopted for block-based fine training. This mechanism calculates the retention score based on the actual rendering contribution and observability of the Gaussian units in the current block training image, sorts the Gaussian units in descending order according to the retention score, and deletes a certain proportion of low-ranked Gaussian units to reduce artifacts, floating points, and point cloud redundancy in local regions.
[0121] The rendering contribution-observability joint pruning mechanism calculates the retention score based on the rendering contribution and observability of Gaussian elements in the current training image block. The calculation formula is as follows:
[0122] (10)
[0123] (11)
[0124] In the formula, S i Let C be the overall score of the i-th Gaussian element. i The score contributes to the rendering of the i-th Gaussian pixel; the larger the value, the more important the Gaussian pixel is to the rendering result. i The observability score represents how many views in the training view set actually observe a Gaussian element, w. c and w o α represents the weight; K represents the number of perspectives participating in the final contribution statistics; α represents the weight. i,kT represents the effective opacity of the i-th Gaussian element at the k-th viewpoint. i,k It represents the cumulative transmission weight of the i-th Gaussian element at the k-th viewpoint, which is used to characterize the effective visibility of the Gaussian element after it is blocked by the preceding Gaussian element along the current line of sight.
[0125] 4-5 Use a learning rate lower than that in the coarse training phase to update the Gaussian point cloud parameters, thereby reducing the parameter fluctuations in the position, scale, and color representation of the Gaussian point cloud during iterative optimization; increase the number of iterations while slowing down the frequency of Gaussian point cloud densification pruning.
[0126] 4-6 After each block reaches the preset number of training iterations, save the refined 3D Gaussian model corresponding to that block.
[0127] Step 5: Merge local scene blocks and output a refined 3D Gaussian model of the entire concrete dam scene.
[0128] Specifically, step 5 includes:
[0129] 5-1 Read the boundary information of each scene block generated in step 3 and the Gaussian element information of each local scene fine-trained 3D Gaussian model obtained in step 4, and extract the center point coordinates of each Gaussian element.
[0130] 5-2 Perform boundary clipping on the Gaussian point cloud output for each block. For any Gaussian element in the current block that satisfies the following formula, retain that Gaussian element:
[0131] (12)
[0132] (13)
[0133] In the formula and This indicates the boundary of the block along the x-axis. and This indicates the boundary of the block along the y-axis.
[0134] If the horizontal coordinates of a Gaussian element do not meet the above conditions, the Gaussian element is deleted from the current block result. The pruning is performed only in the horizontal plane, with the height coordinate z... i It does not participate in the cutting process.
[0135] 5-3 For Gaussian elements at the common boundary of adjacent blocks, a method of determining whether the lower boundary is closed and the upper boundary is open is adopted for preservation, i.e.:
[0136] (15)
[0137] This method ensures that Gaussian elements located near the common boundary belong to only one block, reducing duplicate points when merging adjacent blocks.
[0138] 5-4 Perform a field consistency check on each clipped local scene 3D Gaussian model to confirm that it contains all the necessary Gaussian metadata fields. For blocks with complete fields and consistent format, merge their Gaussian metadata according to the spatial block order. For blocks with missing fields or inconsistent format, record the exception information and prompt for reprocessing.
[0139] 5-5 The merged Gaussian meta-set is output as a PLY format file to obtain a refined 3D Gaussian model of the entire concrete dam scene. The model is trained based on sparse point clouds and camera parameters with a real scene scale, and block clipping and merging are completed in the same local spatial Cartesian coordinate system. Therefore, the output model has a scale representation consistent with the actual concrete dam scene.
[0140] like Figure 5 As shown, the left side is an aerial image, and the right side is the RGB image of the model trained by this method. The comparison reveals that, under the same camera viewpoint, the rendered image better reproduces the overall appearance and positional relationships of the main structures such as the dam, piers, and powerhouse. Components in the foreground, such as railings, ladders, and water level gauges, also have relatively clear outlines. The overall color, light and shadow distribution, and main texture of the concrete surface are basically consistent with the aerial image. Overall, this method can reconstruct the visible appearance of the main body of the concrete dam and its auxiliary components quite well.
[0141] like Figure 6 As shown in the figure, the 3D Gaussian model normal map trained in this embodiment uses color coding to represent the surface normal direction of different regions. The color changes reflect the local surface orientation estimated by the Gaussian model at the corresponding viewpoint. As can be seen from the figure, the colors of large structural areas such as the dam surface, gate piers, and powerhouse walls are relatively uniform, indicating that the local surface orientation changes in these areas are small. There are more obvious color changes at the edges of gate piers, dam surface transitions, railings, and equipment, corresponding to the geometric transitions and occlusion relationships between actual components. This normal map shows that the model can stably represent the local orientation and structural transitions of the main surface of the concrete dam.
[0142] In summary, the present invention proposes a high-fidelity 3D Gaussian reconstruction method for UAV aerial images of concrete dam scenes. This method uses RTK UAV to collect images of concrete dams and their corresponding location information, and combines the RTK location information with the SfM method for coordinate transformation and scale constraint sparse reconstruction to obtain sparse point clouds and camera parameters with real scale reference.
[0143] A coarse-trained 3D Gaussian model with a certain degree of color representation and relatively accurate geometric structure is output by performing full-scene coarse-training through an improved planar Gaussian splashing method. After Gaussian morphology detection, the further densification and diffusion of extreme abnormal Gaussian morphologies are suppressed.
[0144] Local scene blocks are divided based on the distribution of Gaussian elements. Images are allocated to the blocks according to camera position and rendering visibility to obtain a set of images with high visibility correlation to the corresponding blocks.
[0145] An improved planar Gaussian splashing method is used for fine training of local scenes. It also undergoes Gaussian metamorphic detection and introduces rendering contribution-observability joint pruning to reduce redundant Gaussians.
[0146] Finally, the results of each block are cropped and merged to obtain a refined 3D Gaussian model of the entire concrete dam scene with more accurate spatial location, more stable structural relationships, and clearer local details.
[0147] This invention is applicable to achieving high-fidelity 3D reconstruction of concrete dams under complex environmental interference, improving the expression of dam detail textures and the accuracy of geometric structure reconstruction, and has certain universality and operability.
[0148] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be modified within the scope of the concept described herein by means of the above teachings or the technology or knowledge in related fields.
Claims
1. A high-fidelity 3D Gaussian reconstruction method for UAV aerial images of concrete dam scenes, comprising the following steps: Step 1: Use a drone with real-time dynamic differential positioning (RTK) to take multi-view photos of the concrete dam scene, and collect aerial images covering the entire dam body and local structures, as well as the corresponding location information; combine the RTK positioning information, and use the structure recovery of motion (SfM) method to generate sparse point clouds and camera parameters with real scene scale to complete image data preprocessing. Step 2: Based on the aerial images, camera parameters and sparse point clouds obtained in Step 1, the entire scene of the concrete dam is coarsely trained using the planar Gaussian splashing algorithm. A morphology detection mechanism is added to the original training process to determine the morphology of Gaussian elements and restrict abnormally stretched Gaussian elements from further densifying, thereby generating a coarsely trained 3D Gaussian model covering the entire scene of the concrete dam. Step 3: Based on the 3D Gaussian model generated in Step 2, project all Gaussian primitive center points onto a horizontal plane, count their number distribution, divide the concrete dam scene into multiple blocks along the direction perpendicular to the river flow, and determine the boundary position of each block based on the relatively balanced number of Gaussian primitives contained in each block; combine the camera center position and the visibility of image rendering perturbation to determine the visibility of each aerial image to each block, assign images with high visibility to the corresponding blocks, and generate the image set used for training each block; Step 4: Using the coarse-trained 3D Gaussian model obtained in Step 2 as the unified initial model for each block, and combining it with the block image set obtained in Step 3, perform fine training on the local scene of the concrete dam for each block, so that the visible Gaussian elements in the training image corresponding to the current block are optimized; during the training process, the Gaussian element morphology detection mechanism is also used to generate multiple refined 3D Gaussian models corresponding to different blocks. Step 5: Read the refined 3D Gaussian model output from each block in Step 4, perform boundary trimming based on the block boundary information in Step 3, retain the Gaussian elements and their complete parameters within the corresponding block range in each model, merge the retained Gaussian elements in each block under a unified global coordinate system to form a 3D Gaussian expression covering the entire scene of the concrete dam, and output a high-fidelity refined 3D Gaussian model of the entire scene of the concrete dam with the scale of the real scene.
2. The high-fidelity 3D Gaussian reconstruction method for UAV aerial images of concrete dam scenes according to claim 1, characterized in that, Step 1 specifically includes: Step 1-1: Collect multi-view image data of the concrete dam scene using a drone equipped with an RTK module. During shooting, fly at a preset height above the dam crest along a preset flight path and capture images of the overall dam scene vertically downwards. Collect images of the dam surface, water guide wall, and spillway at different heights along the elevation direction. Then adjust the flight height and shooting angle to capture close-up shots of the gate piers and gate structures. Record the RTK position information corresponding to each image during the shooting process, and keep the image shooting distance and overlap rate consistent. Step 1-2: Calculate the mean values of longitude, latitude, and geodetic height corresponding to the aerial images from Step 1-1, and use them as the reference origin of the local spatial rectangular coordinate system. Convert the location information of the aerial images and the reference origin from WGS84 geodetic coordinates to WGS84 geocentric-earth-fixed coordinates. The conversion formula is as follows: (1); (2); (3); (4); In the formula, N represents the radius of curvature of the ramusoidal circle, a represents the semi-major axis of the reference ellipsoid (for the WGS84 ellipsoid, a is 6378137.0m), e represents the first eccentricity, L represents the geodetic longitude, B represents the geodetic latitude, H represents the geodetic height, and X, Y, and Z represent the WGS84 geocentric-geofixed coordinates. Subtract the geocentric-geocentric coordinates (X0, Y0, Z0) of the reference origin from the geocentric-geocentric coordinates (X, Y, Z0) of each image. Then, construct a rotation matrix based on the longitude L0 and latitude B0 of the reference point to perform a rotation transformation, converting the geocentric-geocentric coordinate expression to the local spatial rectangular coordinate expression of the Northeast Sky ENU. The formula is as follows: (5); (6); Steps 1-3: Based on the aerial images acquired in Step 1-1, the SfM method is used to establish matching relationships between images. Based on these matching relationships, the camera parameters and sparse 3D points corresponding to each image are initially estimated. Simultaneously, the local spatial Cartesian coordinates obtained in Step 1-2 are used as a priori camera center position. The camera parameters and sparse 3D point positions are jointly optimized using bundle adjustment with the camera center position prior error, resulting in a sparse 3D point cloud and camera parameters with a real-world scene scale. The objective function for bundle adjustment optimization with camera center position prior constraints is as follows: (7); In the formula K i R is the camera intrinsic parameter corresponding to the i-th image. i Let t be the rotation matrix in the camera pose corresponding to the i-th image. i Let X be the translation vector. j For the j-th sparse 3D point, x ij Let j be the two-dimensional observation coordinates of the j-th 3D point in the i-th image. Let be the set of observable 3D points in the i-th image. C is the projection function of a 3D point onto the image plane. i The camera center is calculated from the camera pose. The prior position of the camera center after conversion from RTK latitude, longitude, and altitude to the ENU coordinate system. Let be the covariance matrix of the prior location of the i-th image. The prior error weights for the camera center position.
3. The high-fidelity 3D Gaussian reconstruction method for UAV aerial images of concrete dam scenes according to claim 1, characterized in that, Step 2 specifically includes: Step 2-1: Based on the sparse point cloud in Step 1, construct an initial three-dimensional Gaussian representation for each three-dimensional point, including the Gaussian center position, covariance matrix, color parameters, and opacity, to form an initial three-dimensional Gaussian field. Step 2-2: Select an image as the current training viewpoint, project all Gaussian primitives onto the image plane, determine the set of Gaussian primitives corresponding to each pixel, use a tile-based differentiable Gaussian splash renderer, perform block traversal and visibility sorting of the planar Gaussian primitives in the screen space, and accumulate the color contribution of each Gaussian primitive layer by layer through the α mixing method to generate the RGB image of the current viewpoint. Steps 2-3: Construct a scale compression loss constraint to minimize the scale along the principal axis of the Gaussian primitive, making the Gaussian primitive gradually planarize during the optimization process; construct a single-view local geometric consistency constraint based on the unbiased depth map, recover 3D points through backprojection of neighboring pixels and fit the local plane, and constrain consistency with the rendering normal; construct a multi-view geometric consistency constraint, realize cross-view pixel mapping through the homography relationship between cameras, and calculate the front and back projection errors to constrain geometric consistency; Backpropagation based on total loss updates the position, opacity, covariance matrix and color features of each Gaussian element, making the rendered image closer to the real aerial image and making the Gaussian elements closer to the real scene surface. Steps 2-4: Calculate the gradient and absolute gradient of each Gaussian primitive on the two-dimensional image plane, determine whether there is a rendering error in the region where the Gaussian primitive is located, and retain the magnitude information of the error change; when the gradient or absolute gradient of a Gaussian primitive is greater than a preset threshold, the Gaussian primitive is regarded as a candidate for densification. Step 2-5: Introduce a Gaussian cell morphology detection mechanism before performing Gaussian point cloud densification, read the scale size of the candidate Gaussian cell in the direction of the three principal axes to obtain the minimum scale s min , the intermediate scale s mid , and the maximum scale s max , where s min ≤ s mid ≤ s max ; use the ratio of the maximum scale to the intermediate scale to judge the degree of elongation in the plane: (8); In the formula, R represents the in-plane dimensional elongation of the candidate Gaussian element, and ε is a minimal constant to prevent division by zero; By setting a threshold, it is determined whether Gaussian units are overstretched. If the ratio of the maximum scale to the intermediate scale is greater than the preset scale threshold, it is judged as an extreme abnormal Gaussian and is no longer allowed to participate in cloning or splitting. Steps 2-6: For the densification candidate Gaussian units that pass the morphological detection, if the maximum scale value is not greater than the preset scale threshold, the Gaussian unit is cloned; if the maximum scale value is greater than the preset scale threshold, the Gaussian unit is split, and multiple new Gaussian units with smaller scales are generated in the neighborhood of the original Gaussian unit. Steps 2-7: Prune Gaussian primitives based on their opacity, 2D projection radius, 3D scale, and number of multi-view projection observations during training; delete Gaussian primitives with low opacity, large projection range, large 3D scale, and few multi-view observations based on thresholds. Steps 2-8: Repeat the image selection, differentiable rendering, loss calculation, backpropagation, densification, and pruning process until the preset number of iterations is reached, at which point the training ends and the coarse-trained 3D Gaussian model of the entire concrete dam scene is saved.
4. The high-fidelity 3D Gaussian reconstruction method for UAV aerial images of concrete dam scenes according to claim 3, characterized in that, In step 2-2, during the rendering process, the normal vectors of each Gaussian plane and their signed plane distances to the camera are fused with consistent weights to output normal maps and plane distance maps; pixel depth is calculated through the geometric relationship between rays and planes to achieve unbiased depth recovery, so that the depth is strictly constrained to the plane geometry and an unbiased depth map is output.
5. The high-fidelity 3D Gaussian reconstruction method for UAV aerial images of concrete dam scenes according to claim 1, characterized in that, In steps 2-3, multi-view photometric consistency is further calculated, and cross-view appearance consistency is ensured through patch alignment and normalized cross-correlation metrics; the total loss function is constructed as follows: (9); wherein L total is the total loss function, L rgb is the image reconstruction loss, L scale is the scale compression loss, L normal is the single-view normal consistency loss, L geo is the multi-view geometry consistency loss, L ncc is the multi-view NCC photometric consistency loss, and λs, λn, λg, λp are loss weights for each loss.
6. The high-fidelity 3D Gaussian reconstruction method for UAV aerial images of concrete dam scenes according to claim 2, characterized in that, Step 3 specifically includes: Step 3-1: Read the center point position of Gaussian units in the coarse-trained model of the whole scene, and calculate the camera center position by combining the camera parameters obtained in Step 1-3, and determine the coverage area of the target scene in the horizontal plane. Step 3-2: Project the center points of all Gaussian elements onto a horizontal plane, with the direction parallel to the river flow as the y-axis and the direction perpendicular to the river flow as the x-axis. According to the preset number of blocks, count the distribution of Gaussian elements, determine the boundary of each block according to the proportion of the number, and form a spatial block between adjacent boundaries to obtain a scene block with a relatively balanced number of Gaussian elements. Step 3-3: Determine the basic attribution relationship of the image based on the camera center position corresponding to each training image. When the camera center projection position of the training image is located within the two-dimensional plane boundary of a certain spatial block, the training image is assigned to that spatial block. Steps 3-4: Based on the rendering results of the full-scene coarse-trained model under each training viewpoint, calculate the visible content score of each spatial block in the training image; the formula is as follows: (10); In the formula, Let I represent the visibility contribution of the c-th camera to the (j,k)-th block; c This represents the original rendered image; This indicates that block B will be... j,k The image obtained by re-rendering after setting the transparency of the Ingoski elements to zero; When the visible content score is greater than a preset threshold, it is determined that the training image has a valid visible relationship with the spatial block, and the image is assigned to the training image set of the spatial block; Step 3-5: Merge the set of location-attributed images obtained in Step 3-3 based on the camera center position with the set of content-visible images obtained in Step 3-4 based on the visibility of image rendering perturbation, and remove duplicate images to obtain the final training image set corresponding to each spatial block.
7. A high-fidelity 3D Gaussian reconstruction method for UAV aerial images of concrete dam scenes according to claim 3, characterized in that, Step 4 specifically includes: Step 4-1: Load the full-scene coarse-trained Gaussian model obtained in Step 2 as the unified initial model for each scene block; Step 4-2: During the local scene fine training process, the training total loss function in Step 2-3 is continued. The image reconstruction loss is included in the calculation of each iteration. The single-view normal consistency loss is calculated once every 2 iterations, and the multi-view geometric consistency loss and multi-view NCC photometric consistency loss are calculated once every 4 iterations. Step 4-3: Perform morphological detection on Gaussian elements before densification; Step 4-4: When pruning Gaussian elements, a rendering contribution-observability joint pruning mechanism is adopted; by calculating the corresponding retention score for each Gaussian element and sorting them in descending order, a set proportion of low-scoring Gaussian elements are deleted. Steps 4-5: Update Gaussian point cloud parameters using a learning rate lower than that used in the coarse training phase to reduce parameter fluctuations in Gaussian point cloud position, scale, and color representations during iterative optimization; increase the number of iterations while slowing down the frequency of Gaussian point cloud densification pruning. Steps 4-6: Perform fine-tuning on each block separately to obtain multiple refined Gaussian models corresponding to each block.
8. A high-fidelity 3D Gaussian reconstruction method for UAV aerial images of concrete dam scenes according to claim 7, characterized in that, In step 4-4, the rendering contribution-observability joint pruning mechanism calculates the retention score based on the rendering contribution and observability of Gaussian elements in the current block of the training image. The calculation formula is as follows: (11); (12); In the formula, S i Let C be the overall score of the i-th Gaussian element. i The score contributes to the rendering of the i-th Gaussian pixel; the larger the value, the more important the Gaussian pixel is to the rendering result. i The observability score represents how many views in the training view set actually observe a Gaussian element, w. c and w o α represents the weight; K represents the number of perspectives participating in the final contribution statistics; α represents the weight. i,k T represents the effective opacity of the i-th Gaussian element at the k-th viewpoint. i,k It represents the cumulative transmission weight of the i-th Gaussian element at the k-th viewpoint, which is used to characterize the effective visibility of the Gaussian element after it is blocked by the preceding Gaussian element along the current line of sight.
9. A high-fidelity 3D Gaussian reconstruction method for UAV aerial images of concrete dam scenes according to claim 1, characterized in that, Step 5 specifically includes: Step 5-1: Read the spatial boundary information of each scene block generated in Step 3 and the information of each local scene fine-trained 3D Gaussian model obtained in Step 4, and extract the center position coordinates of each Gaussian element. Step 5-2: For any scene block, only retain the Gaussian elements whose center point falls within the horizontal boundary of the block; the elevation direction is not included in the clipping. Step 5-3: Process the common boundary of adjacent scene blocks using a unified boundary attribution rule; use a closed judgment for the lower boundary of the block and an open judgment for the upper boundary of the block, so that Gaussian elements located near the common boundary belong to only one scene block; Step 5-4: Perform a field consistency check on the 3D Gaussian models of each local scene after cropping to confirm that they all contain the necessary fields such as position parameters, color spherical harmonic feature parameters, opacity parameters, scale parameters, and rotation parameters; for blocks with complete fields and consistent format, merge their Gaussian metadata according to the spatial block order; for blocks with missing fields or inconsistent format, record the exception information and prompt for reprocessing. Step 5-5: Output the merged Gaussian metaset as a ply format file to obtain a refined 3D Gaussian model of the entire concrete dam scene.
