High-precision three-dimensional terrain reconstruction method based on deep learning

Abstract

The invention relates to the technical field of three-dimensional terrain reconstruction, and discloses a high-precision three-dimensional terrain reconstruction method based on deep learning. The method comprises the following steps: firstly, constructing a final image preprocessing model based on a neural network and an attention mechanism, identifying and removing an interference region, and outputting an interference-removed three-dimensional topographic image set; secondly, performing point cloud matching on the interference-removed three-dimensional topographic image set to generate a three-dimensional topographic point cloud data set, identifying coincident point cloud data in the three-dimensional topographic point cloud data set, determining topographic features to construct a three-dimensional model, and generating an initial three-dimensional topographic model; detecting and deleting discrete points in the initial three-dimensional terrain model, and simplifying the point cloud data by using a three-dimensional grid method to obtain a simplified three-dimensional terrain model; and finally, solving the color of the point cloud data in the simplified three-dimensional terrain model, and coloring the three-dimensional terrain. According to the method, the point cloud data is processed and modeled, the purpose of high-precision three-dimensional terrain reconstruction is achieved, and the method is objective and accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional terrain reconstruction, and specifically to a high-precision three-dimensional terrain reconstruction method based on deep learning. Background Art

[0002] With the rapid development of remote sensing technology, computer vision, and surveying and mapping, three-dimensional terrain reconstruction has gained widespread application in fields such as smart cities, disaster monitoring, military mapping, and autonomous driving. Current mainstream technologies rely primarily on a combination of multi-source data collection methods, including satellite remote sensing imagery, lidar, and drone oblique photography. Among these, stereo matching methods based on optical images have attracted considerable attention due to their low cost and wide coverage. In recent years, deep learning technology, through its end-to-end feature learning mechanism, has demonstrated breakthrough potential in areas such as stereo vision and point cloud processing, offering a new technological path to improving the accuracy and efficiency of terrain reconstruction.

[0003] Traditional 3D terrain reconstruction is primarily based on multi-view geometry principles, using dense set matching algorithms from photogrammetry to generate digital surface models. Specifically, this involves obtaining a 3D point cloud through epipolar correction and local / global optimization, recovering a sparse 3D point cloud from multi-view image sequences, achieving dense reconstruction through multi-view stereo algorithms, and generating a high-precision point cloud using direct LiDAR scanning. This is then combined with ground control points for coordinate registration.

[0004] Traditional three-dimensional terrain reconstruction methods use the multi-view geometry principle, which is prone to discontinuities in the reconstructed surface in weak texture areas and under interference from shadows, clouds, etc., resulting in a significant decrease in reconstruction accuracy. It is difficult to achieve dynamic and high-precision reconstruction, which restricts the engineering application of three-dimensional terrain reconstruction technology in complex scenes. Summary of the Invention

[0005] In response to the problems in the related art, the present invention provides a high-precision three-dimensional terrain reconstruction method based on deep learning to overcome the technical problems existing in the existing related art.

[0006] To solve the above technical problems, the present invention is achieved through the following technical solutions: The present invention provides a high-precision three-dimensional terrain reconstruction method based on deep learning, comprising the following steps: Step S1: Preprocessing: Build an image preprocessing model based on a neural network and an attention mechanism, input an initial terrain image set, and output a de-interferenced three-dimensional terrain image set; Step S2: point cloud generation and matching: converting the interference-free 3D terrain image set into a 3D terrain point cloud data set, identifying overlapping point cloud data and constructing an initial 3D terrain model; Step S3: Model optimization: detecting and deleting discrete points and simplifying point cloud data to generate a simplified three-dimensional terrain model; Step S4: Coloring and reconstruction: solving the point cloud color of the simplified three-dimensional terrain model and performing coloring processing to output a high-precision three-dimensional terrain model.

[0007] Preferably, the image preprocessing model constructed based on a neural network and an attention mechanism includes: Obtain a set of terrain image samples, add random noise and generate a dynamic shadow mask to generate an enhanced terrain image sample set; The backbone network is set as U-Net network, ResNet-34 pre-trained weights are used, and a multi-scale spatial attention module is introduced in the skip connection of the U-Net network; the PatchGAN network structure is used as the discriminator of the image preprocessing model; a joint loss function is set to construct the image preprocessing model; Dividing the enhanced terrain image sample set into a sample training set and a sample test set, inputting the sample training set into a generator of an image preprocessing model, and training in combination with a joint loss function until a maximum number of training rounds is reached, stopping the training, and obtaining a pre-trained image preprocessing model; The sample test set is then input into the pre-trained image preprocessing model for adversarial training, the learning rate is linearly decayed to the maximum number of training rounds, and the training is stopped to obtain the final image preprocessing model.

[0008] Preferably, the inputting of the initial terrain image set and the outputting of the interference-free three-dimensional terrain image set comprises: The area to be reconstructed is selected, and a sequence of images of the area to be reconstructed is collected to obtain an initial terrain image set, which is input into the final image preprocessing model. After denoising and shadow correction, a de-interferenced three-dimensional terrain image set is output.

[0009] Preferably, converting the interference-removed three-dimensional terrain image set into a three-dimensional terrain point cloud data set comprises: A plane coordinate system is established based on the interference-free three-dimensional terrain image set, and the two-dimensional terrain coordinates and heights of the pixels are obtained based on the grayscale values ​​of the pixels, and mapped into a three-dimensional terrain point cloud data set.

[0010] Preferably, the identifying of overlapping point cloud data includes: Set the initial neighborhood radius and count the number of 3D terrain point cloud data within the initial neighborhood radius; Dynamically adjust the neighborhood radius until the number of 3D terrain point cloud data within the initial neighborhood radius converges, obtain the final neighborhood radius, and calculate the comprehensive similarity coefficient; An overlap threshold is set. When the comprehensive similarity coefficient within the final neighborhood radius is greater than the overlap threshold, the corresponding three-dimensional terrain point cloud data is regarded as overlapped point cloud data and deleted. The three-dimensional terrain point cloud data set is detected in sequence to obtain a deduplicated three-dimensional terrain point cloud data set.

[0011] Preferably, the constructing of the initial three-dimensional terrain model includes: Mapping the deduplicated three-dimensional terrain point cloud data set into a three-dimensional model space to form a two-dimensional terrain; creating a super triangle containing the two-dimensional terrain, segmenting the super triangle and performing quadratic surface fitting, and calculating the curvature coefficient; The terrain features are marked as flat, convex, concave and saddle areas according to the curvature coefficient; After determining the terrain features, they are marked in the processed three-dimensional terrain to generate an initial three-dimensional terrain model.

[0012] Preferably, forming a two-dimensional terrain includes: Split the super triangle and generate a circumscribed circle, traversing the 3D terrain point cloud data in the super triangle; Delete the triangles containing 3D terrain point cloud data inside the circumscribed circle to generate a void area and reconstruct the mesh; Delete the triangles containing the super triangle vertices to generate the processed two-dimensional terrain.

[0013] Preferably, the detecting and deleting discrete points includes: Select the point cloud data to be detected in the initial 3D terrain model, and calculate the discrete threshold of the k point cloud data closest to the point cloud data to be detected; When the point cloud data to be detected is greater than the discrete upper threshold or less than the discrete lower threshold, the point cloud data to be detected is regarded as a discrete point, and the discrete point is deleted to obtain a filtered three-dimensional terrain model.

[0014] Preferably, simplifying the point cloud data to generate a simplified three-dimensional terrain model includes: Set the cube side length and divide the filtered 3D terrain model into several cubes according to the cube side length; calculate the centroid of each cube and use the cube centroid to replace all point cloud data in the cube to generate a simplified 3D terrain model.

[0015] Preferably, the step of solving the point cloud color of the simplified three-dimensional terrain model and performing coloring processing includes: Get the color values ​​of the cube vertices and calculate the weighted average color value of the cube vertex set; The weighted average color value of the cube vertex set is used as the color value of the cube centroid, and the simplified three-dimensional terrain model is colored to obtain the final three-dimensional terrain model.

[0016] The present invention has the following beneficial effects: 1. This invention obtains an initial set of terrain images, integrates the U-Net network and the attention mechanism to construct an image preprocessing model, and identifies and removes interference areas in the initial set of terrain images. By rationally designing data, networks, and loss functions, this method can improve the robustness of the model in complex environments, restore surface details in the image, eliminate shadow interference, and overcome interference from weak texture areas, shadows, clouds, and fog.

[0017] 2. This invention generates a three-dimensional terrain point cloud data set through point cloud matching, identifies overlapping point cloud data, and maps brightness to height. It can build a conversion from terrain images to point cloud data without relying on complex technology, overcoming the complexity of traditional point cloud matching methods. At the same time, it uses a dynamic neighborhood detection algorithm to identify overlapping point cloud data, dynamically calculates the neighborhood radius and uses multi-feature fusion similarity judgment, effectively reducing the accidental deletion of terrain features and significantly improving the accuracy of point cloud data deduplication. It is suitable for complex terrain reconstruction.

[0018] 3. The invention constructs a three-dimensional model by determining terrain features, generates an initial three-dimensional terrain model, projects point cloud data onto a two-dimensional plane, constructs a triangular network and determines terrain features to establish a continuous surface, ensuring geometric consistency and feature integrity, and improving the accuracy of three-dimensional terrain reconstruction.

[0019] 4. This invention detects and deletes discrete points in the initial three-dimensional terrain model and uses a three-dimensional grid method to simplify the point cloud data. This method overcomes the problem that discrete points are prone to appear in low-density areas. While ensuring the accuracy of the model, it greatly reduces the amount of point cloud data, reduces the computational memory and time consumption of terrain reconstruction, and obtains a simplified three-dimensional terrain model.

[0020] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the invention. For ordinary technicians in this field, they can also obtain drawings based on these drawings without paying any creative work.

[0022] Figure 1 The present invention provides a flowchart of a high-precision three-dimensional terrain reconstruction method based on deep learning. DETAILED DESCRIPTION

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0024] Traditional three-dimensional terrain reconstruction methods use the multi-view geometry principle, which is prone to discontinuities in the reconstructed surface in weak texture areas and under interference from shadows, clouds, etc., resulting in a significant decrease in reconstruction accuracy. It is difficult to achieve dynamic and high-precision reconstruction, which restricts the engineering application of three-dimensional terrain reconstruction technology in complex scenes.

[0025] To solve the above technical problems, a specific embodiment is that, during the three-dimensional reconstruction process, a multi-rotor drone equipped with a high-resolution camera is used to capture a sequence of aerial images of the target mountain area along a planned route, and the heading overlap is specified to be ≥80%, the lateral overlap is ≥70%, and the terrain following mode is used in complex terrain to maintain a constant altitude. In periods of uniform lighting and low wind speed, good terrain images and point cloud data are provided for the present invention.

[0026] In the specific implementation process of the above embodiment, first, an initial terrain image set is obtained, and an image preprocessing model is constructed by integrating the U-Net network and the attention mechanism to identify and remove interference areas in the initial terrain image set; the method realizes data enhancement through synthetic interference and geometric transformation, and the network architecture design uses the U-Net network, the attention mechanism and the adversarial generative network. By reasonably designing the data, network and loss function, the robustness of the model in complex environments can be improved, and surface details (such as rivers, exposed rocks and other weak texture areas) in the image can be restored, and shadow interference (such as mountain shadows) can be eliminated, overcoming interference from weak texture areas and shadows, clouds and fog; secondly, point cloud matching is performed to generate a three-dimensional terrain point cloud data set, identify overlapping point cloud data, and then determine the terrain features to construct a three-dimensional model to generate an initial three-dimensional terrain model; the method can construct a map from terrain image to point cloud data by mapping brightness to height without relying on complex technology. This method overcomes the complexity of traditional point cloud matching methods and uses a dynamic neighborhood detection algorithm to identify overlapping point cloud data. The core concept is to dynamically calculate the neighborhood radius and use multi-feature fusion similarity determination, effectively reducing the accidental deletion of terrain features and significantly improving the accuracy of point cloud data deduplication. It is suitable for complex terrain reconstruction. The point cloud data is then projected onto a two-dimensional plane, a triangular network is constructed, and terrain features are identified to establish a continuous surface, improving the accuracy of three-dimensional terrain reconstruction. Discrete points in the initial three-dimensional terrain model are detected and deleted, and the point cloud data is simplified using a stereo mesh method to obtain a simplified three-dimensional terrain model. This method overcomes the problem of discrete points being easily present in low-density areas. While ensuring model accuracy, it significantly reduces the amount of point cloud data, reducing the computational memory and time required for terrain reconstruction. Finally, the three-dimensional terrain is shaded, and the color information of the two-dimensional image is accurately mapped to the model through linear interpolation, achieving high-precision three-dimensional terrain reconstruction. The entire process relies on rigorous mathematical modeling methods and is suitable for scenarios such as scientific research and geological hazard assessment.

[0027] Furthermore, in order to better introduce the technical solutions of the embodiments of the present invention, Figure 1 As shown, an embodiment of the present invention provides a high-precision three-dimensional terrain reconstruction method based on deep learning, which specifically includes the following contents: S1. Obtain an initial terrain image set, construct a final image preprocessing model based on a neural network and an attention mechanism, input the initial terrain image set into the final image preprocessing model, identify and remove interference areas, and output a de-interferenced three-dimensional terrain image set; Said S1 comprises the following steps: S11. Selecting a region to be reconstructed, collecting a sequence of images of the region to be reconstructed, obtaining an initial terrain image set, and forming an initial terrain image set; again obtaining terrain images of other reconstructed regions to obtain a terrain image sample set; adding random noise to the terrain image sample set, generating a dynamic shadow mask, and superimposing the mask on the terrain image sample set; then stretching and rotating the terrain images in the terrain image sample set to generate a final terrain image sample set; S12. According to the final terrain image sample set, a final image preprocessing model is established based on a neural network and an attention mechanism. The specific steps are as follows: The backbone network is set as the U-Net network, the ResNet-34 pre-trained weights are used, and a multi-scale spatial attention module is introduced in the skip connection of the U-Net network. The multi-scale spatial attention module includes global average pooling, global maximum pooling, a fully connected layer, a convolutional layer, a ReLU function, and a Sigmoid function to obtain the generator of the image preprocessing model; the PatchGAN network structure is used as the discriminator of the image preprocessing model; a joint loss function is set, which includes L1 loss, perceptual loss, adversarial loss, and edge loss, and weights are assigned to each of them to obtain the loss function and loss weight of the image preprocessing model, and thus the image preprocessing model is constructed; S122, setting a learning rate, batch size, and maximum number of training rounds, dividing the final terrain image sample set into a sample training set and a sample test set, inputting the sample training set into a generator of an image preprocessing model, and training with a combination of L1 loss and edge loss until the maximum number of training rounds is reached, stopping training, and obtaining a pre-trained image preprocessing model; S123, inputting the sample test set into the pre-trained image preprocessing model, performing adversarial training, reducing the L1 loss weight, increasing the adversarial loss weight, dynamically adjusting the loss weight, linearly decaying the learning rate to the maximum number of training rounds, stopping training, and obtaining the final image preprocessing model; S13, inputting the initial terrain image set into the final image preprocessing model, the final image preprocessing model identifying and removing interference areas, and outputting a de-interferenced three-dimensional terrain image set; In this embodiment, by obtaining an initial terrain image set, integrating the U-Net network and the attention mechanism to construct an image preprocessing model, identifying and removing interference areas in the initial terrain image set, and realizing data enhancement through synthetic interference and geometric transformation, the network architecture design uses the U-Net network, the attention mechanism and the adversarial generative network. By reasonably designing the data, network and loss function, the robustness of the model in complex environments can be improved, and surface details (such as rivers, exposed rocks and other weak texture areas) can be restored in the image, and shadow interference (such as mountain shadows) can be eliminated. Specifically, for example, a dynamic shadow mask (simulating cloud layers and vegetation projections) is generated by random cropping or rendering tools, superimposed on the image, and Gaussian noise and salt and pepper noise are added to simulate sensor interference. The backbone network uses pre-trained ResNet-34 to extract deep features, and a multi-scale spatial attention module is introduced in the jump connection to make the model focus on terrain contours (such as ridges and gullies) rather than vegetation / shadow areas. The discriminator is used for local receptive field design to judge the local authenticity of the image. Function joint loss weight setting method: randomly select 100 training images, calculate the initial value of each loss item respectively, set the weight according to the mean ratio, when the L1 loss mean: perceptual loss mean: adversarial loss mean ≈ 10:2:1, then set the initial weights to 1.0, 0.2 and 0.1, and the edge loss weight is fixed at 0.1, and get L1 loss (1.0) + perceptual loss (0.2) + adversarial loss (0.1) + edge loss (0.1), to ensure clear texture and accurate structure; in the pre-training stage , set the learning rate to 0.001, the batch size to 8, the maximum number of epochs to 100, and used the sample training set (8000 groups) to train the generator, prioritizing the optimization of L1 loss (global terrain reconstruction) and edge loss (contour sharpening); during the adversarial training phase, the L1 weight was reduced to 0.3 and the adversarial loss was increased to 0.4 to improve the authenticity of details. The learning rate was linearly decayed to 0.0001 to avoid overfitting. The model successfully distinguished and removed more than 90% of shadows and vegetation artifacts in the test images, overcoming interference from weak texture areas, shadows, clouds, etc.; S2. performing point cloud matching on the interference-free three-dimensional terrain image set to generate a three-dimensional terrain point cloud data set, identifying overlapping point cloud data in the three-dimensional terrain point cloud data set, and then determining terrain features to construct a three-dimensional model to generate an initial three-dimensional terrain model; The S2 comprises the following steps: S21. Select any terrain image from the set of de-interferenced three-dimensional terrain images, convert it to grayscale, and record it as a first terrain image. Establish a plane coordinate system with the upper left corner of the first terrain image as the coordinate origin. Determine the coordinates and grayscale values ​​of pixels in the first terrain image. Measure the horizontal and vertical dimensions of the pixels in the first terrain image. The two-dimensional terrain coordinates are equal to the product of the coordinates of the pixel in the first terrain image and the horizontal and vertical dimensions of the pixel in the first terrain image, thereby obtaining the two-dimensional terrain coordinates of the pixel in the first terrain image. Set the terrain height range and calculate the height of the pixel in the first terrain image according to the gray value of the pixel in the first terrain image ,in Indicates the position of the first terrain image The gray value at Indicates the lower limit of terrain height, Indicates the upper limit of terrain height; combining the two-dimensional terrain coordinates of the pixel point in the first terrain image and the height of the pixel point in the first terrain image to obtain the three-dimensional terrain coordinates of the pixel point in the first terrain image, and recording the pixel point in the first terrain image as the first terrain image point cloud data; S22: Select a terrain image adjacent to the first terrain image from the interference-removed three-dimensional terrain image set, record it as a second terrain image, calculate the three-dimensional terrain coordinates of the pixel points in the second terrain image to obtain point cloud data of the second terrain image, and sequentially calculate the point cloud data of the interference-removed three-dimensional terrain image set to form a three-dimensional terrain point cloud data set; S23, using a dynamic neighborhood detection algorithm to identify overlapping point cloud data in the three-dimensional terrain point cloud data set, and deleting the overlapping point cloud data to obtain a deduplicated three-dimensional terrain point cloud data set. The specific steps are as follows: S231, set the initial neighborhood radius, select any three-dimensional terrain point cloud data from the three-dimensional terrain point cloud data set, record it as core point cloud data, and count the number of three-dimensional terrain point cloud data within the initial neighborhood radius with the core point cloud data as the center; set a high density threshold and a low density threshold; when the number of three-dimensional terrain point cloud data within the initial neighborhood radius is greater than the high density threshold, reduce the initial neighborhood radius to obtain a reduced neighborhood radius ,in represents the initial neighborhood radius, Represents the high density threshold, m represents the number of three-dimensional terrain point cloud data within the initial neighborhood radius; when the number of three-dimensional terrain point cloud data within the initial neighborhood radius is less than the high density threshold, the initial neighborhood radius is expanded to obtain the expanded neighborhood radius. ,in represents the low density threshold; until the number of 3D terrain point cloud data within the initial neighborhood radius converges, the final neighborhood radius is obtained; S232, obtaining the grayscale value of the three-dimensional terrain point cloud data within the final neighborhood radius, calculating the difference in grayscale value between the core point cloud data and the other three-dimensional terrain point cloud data, recording it as a grayscale value similarity coefficient; calculating the Euclidean distance between the core point cloud data and the other three-dimensional terrain point cloud data, recording it as a distance similarity coefficient; then calculating the curvature of the three-dimensional terrain point cloud data within the final neighborhood radius, recording the difference in curvature between the core point cloud data and the other three-dimensional terrain point cloud data as a curvature similarity coefficient; According to the gray value similarity coefficient , distance similarity coefficient and curvature similarity coefficient , use multi-feature weighted fusion to calculate the comprehensive similarity coefficient within the final neighborhood radius ,in 、 and represents similarity weight; S233, setting an overlap threshold, and when the comprehensive similarity coefficient within the final neighborhood radius is greater than the overlap threshold, treating the corresponding 3D terrain point cloud data as overlapped point cloud data and deleting the overlapped point cloud data; sequentially detecting all 3D terrain point cloud data in the 3D terrain point cloud data set to obtain a deduplicated 3D terrain point cloud data set; S24, setting a three-dimensional model space, wherein the three-dimensional terrain point cloud data in the deduplicated three-dimensional terrain point cloud data set includes X values, Y values, and Z values, and mapping the X values ​​and Y values ​​to the three-dimensional model space to form a two-dimensional terrain; creating a super triangle containing the two-dimensional terrain, segmenting the super triangle to obtain a triangular mesh, making a circumscribed circle of all triangles in the triangular mesh, traversing the three-dimensional terrain point cloud data in the super triangle, and when three-dimensional terrain point cloud data exists within the circumscribed circle of the triangle, deleting the corresponding triangle to generate a hole domain, and then connecting the three-dimensional terrain point cloud data within the circumscribed circle of the triangle with the hole domain to obtain a new triangle; until no three-dimensional terrain point cloud data exists within the circumscribed circles of all triangles, then deleting the triangle containing the super triangle vertex to obtain an inner triangle, thereby generating a processed two-dimensional terrain; Then map the Z value of the three-dimensional terrain point cloud data corresponding to the processed two-dimensional terrain to the three-dimensional model space to obtain the processed three-dimensional terrain; perform quadratic surface fitting on any three-dimensional terrain point cloud data in the processed three-dimensional terrain, calculate the three-dimensional terrain curvature, and calculate the curvature coefficient ,in represents the three-dimensional terrain curvature, Indicates the Z value of 3D terrain point cloud data; when and When , the corresponding area in the processed three-dimensional terrain is a flat area; when and When , the corresponding area in the processed three-dimensional terrain is a convex area; when and When , the corresponding area in the processed three-dimensional terrain is a concave area; otherwise, the corresponding area in the processed three-dimensional terrain is a saddle area, and the terrain features are determined and marked in the processed three-dimensional terrain to generate an initial three-dimensional terrain model; In this embodiment, by performing point cloud matching, a three-dimensional terrain point cloud data set is generated, overlapping point cloud data is identified, and then the terrain features are determined to construct a three-dimensional model to generate an initial three-dimensional terrain model; this method maps brightness to height and can construct a conversion from terrain image to point cloud data without relying on complex technology, overcoming the complexity of traditional point cloud matching methods, and using a dynamic neighborhood detection algorithm to identify overlapping point cloud data. The core idea is to dynamically calculate the neighborhood radius and use multi-feature fusion similarity judgment, which effectively reduces the accidental deletion of terrain features and significantly improves the accuracy of point cloud data deduplication. It is suitable for complex terrain reconstruction and then the point cloud data is projected onto the ground. It projects the image onto a two-dimensional plane, constructs a triangular network, and determines the terrain features to establish a continuous surface, thereby improving the accuracy of three-dimensional terrain reconstruction. Specifically, for example, the image of the mountain top area is selected, a plane coordinate system is established, and the two-dimensional coordinates of each pixel are calculated through the horizontal / vertical dimensions and the pixel position, and the pixel grayscale value is converted into a height value. The neighborhood radius is adaptively adjusted: with a certain point cloud (such as a point on the ridge line) as the core, the neighborhood radius is initially set to 1 meter. The density threshold is determined, the image of the mountain top area is blocked, the density values ​​of all blocks are counted, and a histogram is drawn. The 90% quantile of the density distribution is taken as the high density threshold, and the 10% quantile is taken as the low density threshold. If the density of the point cloud in the neighborhood is too high (exceeding the high-density threshold of 50 points), reduce the radius to 0.8 meters; if the density is too low (less than the low-density threshold of 10 points), expand it to 1.5 meters until the density stabilizes; Multi-feature fusion deduplication: calculate the grayscale similarity (to avoid repeated textures), distance similarity (excluding overlapping projection points) and curvature similarity (distinguishing terrain mutation areas) of the point cloud in the neighborhood, and assign weights based on feature importance (e.g., the spatial distance weight in terrain data is 0.6, and the curvature is 0.3); after comprehensive weighting, delete the overlapping points whose similarity exceeds the threshold (0.9) and retain the unique terrain feature points; project the deduplicated point cloud In the 3D model space, a super triangle covering the entire terrain (covering the entire mountain range) is constructed, a triangular mesh is generated, triangles containing external vertices are deleted, and the internal mesh (mountain surface) is retained. A quadratic surface is fitted to each triangular mesh, and curvature is calculated. Flat areas have a curvature close to 0 (mountaintop platform), raised areas have a positive curvature coefficient (independent peaks), concave areas have a negative curvature coefficient (craters or water-filled valleys), and saddle areas have a mixed curvature sign (low-lying connections between ridges). Multi-feature fusion improves the ability to retain terrain details, making it suitable for feature extraction and analysis of complex landforms and facilitating the description of scenes in 3D terrain models. S3, detecting and deleting discrete points in the initial three-dimensional terrain model to obtain a filtered three-dimensional terrain model, and simplifying the point cloud data in the filtered three-dimensional terrain model based on a stereo grid method to obtain a simplified three-dimensional terrain model; The S3 includes the following steps: S31. The initial three-dimensional terrain model includes a number of point cloud data. Any point cloud data is selected from the initial three-dimensional terrain model and recorded as the point cloud data to be detected. The k point cloud data closest to the point cloud data to be detected are selected to form a set of adjacent three-dimensional terrain point cloud data. The Euclidean distance and the average distance between the point cloud data to be detected and the set of adjacent three-dimensional terrain point cloud data are calculated. The variance and standard deviation of the point cloud data to be detected and the set of adjacent three-dimensional terrain point cloud data are calculated based on the Euclidean distance and the average distance. The discrete lower bound threshold is obtained by calculation. , discrete upper threshold ,in Represents the average distance between the point cloud data to be detected and the adjacent 3D terrain point cloud data set, Represents the standard deviation of the point cloud data to be detected and the adjacent 3D terrain point cloud data set; When the point cloud data to be detected is greater than the discrete upper threshold or less than the discrete lower threshold, the point cloud data to be detected is regarded as a discrete point and deleted; otherwise, the point cloud data to be detected is retained; all discrete points in the initial three-dimensional terrain model are deleted in sequence to obtain a filtered three-dimensional terrain model; S32: Set a cube in the filtered three-dimensional terrain model, use a three-dimensional grid method to simplify the point cloud data, and generate a simplified three-dimensional terrain model. The specific steps are as follows: S321, set the cube side length l, find the minimum point cloud data in the filtered three-dimensional terrain model, establish a spatial coordinate system with the minimum point cloud data as the coordinate origin, set the coordinates of the maximum point cloud data in the filtered three-dimensional terrain model as , the minimum point cloud data coordinates in the filtered 3D terrain model are , divide the filtered 3D terrain model into several cubes according to the side length of the cube, then the number of cubes ; S322, counting the number of point cloud data in the cube, calculating the centroid of each cube, and using the cube centroid to replace all point cloud data in the cube until all point cloud data in the filtered three-dimensional terrain model is simplified to generate a simplified three-dimensional terrain model; In this embodiment, a simplified three-dimensional terrain model is obtained by detecting and deleting discrete points in the initial three-dimensional terrain model and simplifying the point cloud data using a stereo grid method. This method overcomes the problem that discrete points are prone to appear in low-density areas. While ensuring model accuracy, it greatly reduces the amount of point cloud data and reduces the computational memory and time consumption of terrain reconstruction. Specifically, for example, the number of neighboring points k=15 is selected, a point P to be detected is selected, its 15 nearest neighboring points are found, and the Euclidean distance between P and the neighboring points is calculated. The average distance d=0.85 meters and the standard deviation σ=0.2 meters are obtained. The discrete threshold is calculated with a lower bound of 0.45 meters and an upper bound of 1.25 meters. If the average distance from P to the neighboring points is 1.3 meters (exceeding the upper bound), Determine if P is a discrete point and delete it. Traverse all 1.2 million points and delete approximately 50,000 discrete points (isolated vegetation points and equipment noise points), resulting in a filtered model containing 1.15 million points. A cube with a side length of 1.0 meter is selected. At this point, each cube covers a 10×10 pixel area, and color fusion is stable. The model boundary is obtained: the minimum coordinates are (0, 0, 1020m) and the maximum coordinates are (500m, 300m, 1350m). Traverse each cube. If the cube contains a point cloud (one cube contains 20 points), calculate its centroid coordinates, delete the original 20 points, and retain only the centroid point. The number of point clouds is reduced from 1.15 million to approximately 180,000, an 84% reduction in data volume. This lightweight data lays the foundation for subsequent colorization processing. S4, solving the color of the point cloud data in the simplified three-dimensional terrain model, coloring the three-dimensional terrain, and realizing high-precision reconstruction of the three-dimensional terrain; The S4 comprises the following steps: S41, the simplified three-dimensional terrain model includes several cubes, and the cube vertices are selected and numbered to obtain a cube vertex set ,in Indicate the 8th cube vertex and record the coordinates of the cube vertex. Obtain the color value of the cube vertex according to the interference-free 3D terrain image set and calculate the weighted average color value of the 8 cube vertices. ,in Represents the color value of the i-th cube vertex, , get the point cloud data color; S42, taking the weighted average color value of the eight cube vertices as the color value of the cube centroid, and sequentially coloring the color values ​​of all cube centroids in the simplified three-dimensional terrain model to obtain a final three-dimensional terrain model; In this embodiment, the three-dimensional terrain is shaded, and the color information of the two-dimensional image is accurately mapped to the model through linear interpolation, thereby achieving high-precision reconstruction of the three-dimensional terrain. Specifically, for example, the eight vertices (B1-B8) of each cube correspond to sampling points in the actual terrain. Each vertex is assigned a color, and the arithmetic average of the eight vertex colors is taken. The average result is RGB (120, 160, 90). The center of mass color of the cube is a soft grass green, and the colors of the center of mass of adjacent cubes gradually change to form a smooth transition. The color fusion reflects the local landform characteristics and presents a realistic three-dimensional terrain. The entire process relies on rigorous mathematical modeling methods and is suitable for complex scenarios such as scientific research and geological disaster assessment.

[0028] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0029] The preferred embodiments of the invention disclosed above are intended only to help illustrate the invention. These preferred embodiments do not exhaust all details, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A high-precision three-dimensional terrain reconstruction method based on deep learning, characterized in that: include: Step S1: Preprocessing: Build an image preprocessing model based on a neural network and an attention mechanism, input an initial terrain image set, and output a de-interferenced three-dimensional terrain image set; Step S2: point cloud generation and matching: converting the interference-free 3D terrain image set into a 3D terrain point cloud data set, identifying overlapping point cloud data and constructing an initial 3D terrain model; Step S3: Model optimization: detecting and deleting discrete points and simplifying point cloud data to generate a simplified three-dimensional terrain model; Step S4: Coloring and reconstruction: solving the point cloud color of the simplified three-dimensional terrain model and performing coloring processing to output a high-precision three-dimensional terrain model.

2. The high-precision three-dimensional terrain reconstruction method based on deep learning according to claim 1, characterized in that: The image preprocessing model based on the neural network and attention mechanism includes: Obtain a set of terrain image samples, add random noise and generate a dynamic shadow mask to generate an enhanced terrain image sample set; The backbone network is set as U-Net network, ResNet-34 pre-trained weights are used, and a multi-scale spatial attention module is introduced in the skip connection of the U-Net network; the PatchGAN network structure is used as the discriminator of the image preprocessing model; a joint loss function is set to construct the image preprocessing model; Dividing the enhanced terrain image sample set into a sample training set and a sample test set, inputting the sample training set into a generator of an image preprocessing model, and training in combination with a joint loss function until a maximum number of training rounds is reached, stopping the training, and obtaining a pre-trained image preprocessing model; The sample test set is then input into the pre-trained image preprocessing model for adversarial training, the learning rate is linearly decayed to the maximum number of training rounds, and the training is stopped to obtain the final image preprocessing model.

3. The high-precision three-dimensional terrain reconstruction method based on deep learning according to claim 2, characterized in that: The inputting of the initial terrain image set and the outputting of the interference-free three-dimensional terrain image set comprises: The area to be reconstructed is selected, and a sequence of images of the area to be reconstructed is collected to obtain an initial terrain image set, which is input into the final image preprocessing model. After denoising and shadow correction, a de-interferenced three-dimensional terrain image set is output.

4. The operation and maintenance visualization platform system according to claim 3, characterized in that: The converting the interference-removed three-dimensional terrain image set into a three-dimensional terrain point cloud data set comprises: A plane coordinate system is established based on the interference-free three-dimensional terrain image set, and the two-dimensional terrain coordinates and heights of the pixels are obtained based on the grayscale values ​​of the pixels, and mapped into a three-dimensional terrain point cloud data set.

5. The high-precision three-dimensional terrain reconstruction method based on deep learning according to claim 4, characterized in that: The identifying of coincident point cloud data comprises: Set the initial neighborhood radius and count the number of 3D terrain point cloud data within the initial neighborhood radius; Dynamically adjust the neighborhood radius until the number of 3D terrain point cloud data within the initial neighborhood radius converges, obtain the final neighborhood radius, and calculate the comprehensive similarity coefficient; An overlap threshold is set. When the comprehensive similarity coefficient within the final neighborhood radius is greater than the overlap threshold, the corresponding three-dimensional terrain point cloud data is regarded as overlapped point cloud data and deleted. The three-dimensional terrain point cloud data set is detected in sequence to obtain a deduplicated three-dimensional terrain point cloud data set.

6. The high-precision three-dimensional terrain reconstruction method based on deep learning according to claim 5, characterized in that: The constructing of the initial three-dimensional terrain model comprises: Mapping the deduplicated three-dimensional terrain point cloud data set into a three-dimensional model space to form a two-dimensional terrain; creating a super triangle containing the two-dimensional terrain, segmenting the super triangle and performing quadratic surface fitting, and calculating the curvature coefficient; The terrain features are marked as flat, convex, concave and saddle areas according to the curvature coefficient; After determining the terrain features, they are marked in the processed three-dimensional terrain to generate an initial three-dimensional terrain model.

7. The high-precision three-dimensional terrain reconstruction method based on deep learning according to claim 6, characterized in that: The forming of the two-dimensional terrain comprises: Split the super triangle and generate a circumscribed circle, traversing the 3D terrain point cloud data in the super triangle; Delete the triangles containing 3D terrain point cloud data inside the circumscribed circle to generate a void area and reconstruct the mesh; Delete the triangles containing the super triangle vertices to generate the processed two-dimensional terrain.

8. The high-precision three-dimensional terrain reconstruction method based on deep learning according to claim 7, characterized in that: The detecting and deleting of discrete points includes: Select the point cloud data to be detected in the initial 3D terrain model, and calculate the discrete threshold of the k point cloud data closest to the point cloud data to be detected; When the point cloud data to be detected is greater than the discrete upper threshold or less than the discrete lower threshold, the point cloud data to be detected is regarded as a discrete point, and the discrete point is deleted to obtain a filtered three-dimensional terrain model.

9. The high-precision three-dimensional terrain reconstruction method based on deep learning according to claim 8, characterized in that: The simplified point cloud data to generate a simplified three-dimensional terrain model includes: Set the cube side length and divide the filtered 3D terrain model into several cubes according to the cube side length; calculate the centroid of each cube and use the cube centroid to replace all point cloud data in the cube to generate a simplified 3D terrain model.

10. The high-precision three-dimensional terrain reconstruction method based on deep learning according to claim 9, characterized in that: The method of solving the point cloud color of the simplified three-dimensional terrain model and performing coloring processing includes: Get the color values ​​of the cube vertices and calculate the weighted average color value of the cube vertex set; The weighted average color value of the cube vertex set is used as the color value of the cube centroid, and the simplified three-dimensional terrain model is colored to obtain the final three-dimensional terrain model.

Images