Terrain and live-action three-dimensional modeling method for new energy unmanned aerial vehicle

Abstract

The invention discloses a terrain live-action three-dimensional modeling method for a new energy unmanned aerial vehicle, and belongs to the field of unmanned aerial vehicle modeling, the new energy unmanned aerial vehicle which is high in cruising ability, high in flight stability and flexible in load adaptation is adopted in the terrain live-action three-dimensional modeling method for the new energy unmanned aerial vehicle, and the new energy unmanned aerial vehicle carries a five-eye oblique camera system. The five-eye oblique camera system comprises an orthographic lens and four oblique lenses, and the inclination angle of each oblique lens is 30-45 degrees. Through feature point weighted selection and dynamic sparse point cloud adjustment, the point cloud distribution and geometric quality are optimized, the terrain precision and detail reduction capability of the three-dimensional model are improved, and the modeling requirements of complex terrains and key areas are met.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle modeling, and more specifically, to a new energy unmanned aerial vehicle terrain real scene three-dimensional modeling method. Background Art

[0002] With the development of drones and 3D modeling technology, traditional terrain mapping methods can no longer meet the needs of fast and accurate modeling of complex terrain due to low efficiency, high cost and limited coverage. The rise of new energy drones, combined with multi-view oblique photography technology, provides an efficient and flexible solution for real-scene 3D modeling of terrain. However, the generation of high-precision models faces many challenges, including insufficient density of sparse point clouds, uneven distribution of feature points, missing texture information, and high complexity of multi-view image data processing. Summary of the invention

[0003] In order to overcome these bottlenecks, it is necessary to explore an innovative modeling method based on new energy drones. By equipping with a high-precision oblique camera, combined with intelligent image processing software and high-performance computing servers, the geometric quality and point distribution uniformity of sparse point clouds can be improved, and the dense point cloud generation process can be optimized, ultimately achieving high-precision and realistic three-dimensional terrain modeling. This method is not only suitable for a wide range of terrain surveying and mapping, but also can support urban planning, disaster monitoring, ecological protection and other fields, laying the foundation for the development of intelligent geographic information systems (GIS). The object of the present invention is to provide a new energy drone terrain real-scene three-dimensional modeling method, the new energy drone terrain real-scene three-dimensional modeling method adopts a new energy drone with strong endurance, high flight stability, and flexible load adaptation, the new energy drone is equipped with a five-eye oblique camera system, the five-eye oblique camera system includes 1 orthophoto lens and 4 oblique lenses, the inclination angle of the oblique lens is 30° to 45°;

[0004] The new energy UAV terrain real scene 3D modeling method comprises the following steps:

[0005] S1. Obtain topographic maps, satellite images and DEM data of the target area online, analyze terrain complexity and coverage, set the flight path of the drone using aerial survey planning software, the aerial survey planning software includes DJIPilot and Pix4DCapture, determine the flight altitude, overlap rate and tilt camera angle, generate a route file and upload it to the drone, the overlap rate is 70%-85% in the forward direction and 60%-80% in the lateral direction;

[0006] S2, the drone automatically flies according to the set route and collects images. The collection process is monitored in real time. After the drone completes the route, it returns to the base station;

[0007] S3. Check the quality of the collected image data, including whether there is blur, underexposure or insufficient overlap, and then back up the qualified image data to the cloud to prevent data loss;

[0008] S4, number and classify the collected images to ensure the integrity of the image data, associate the images with the drone GPS data, and obtain the spatial location information of each image;

[0009] S5. Use 3D modeling software to automatically detect conjugate points between images, complete image registration and alignment, generate sparse point clouds, and determine the basic geometric structure of the model. The 3D modeling software includes Zhitu and Agisoft Metashape.

[0010] S6. Generate high-density point cloud data based on sparse point cloud to construct detailed terrain structure, convert dense point cloud into 3D mesh model to form 3D surface, perform texture mapping on mesh surface, project collected images onto the model, and generate realistic 3D real scene model;

[0011] S7. Use ground control points and high-precision coordinate data to adjust the model position and accuracy, remove noise in the point cloud, optimize the model boundary and texture details, and then convert the model to the local projection coordinate system;

[0012] S8. Use the 3D display software Cesium to display the results and upload the model data to the cloud platform to facilitate team collaboration or customer access.

[0013] Preferably, the S2 drone uses ground control points to improve the geographic accuracy of the model during data collection, as follows:

[0014] Step 1: Select the location of the ground control points. The distribution principle is as follows: the ground control points are evenly distributed in the modeling area, with key points at the edge and center. The number of points is determined according to the size of the modeling area and the accuracy requirements, with 510 points per square kilometer. In areas with complex terrain changes and high-precision models, the number of control points is increased by 12. Avoid placing points in vegetation-covered areas and on the surface of water bodies, and choose flat, open and clearly marked locations.

[0015] Step 2: Lay out and mark ground control points. The control points are marked as follows: Use white cross targets to ensure easy identification when drones are shooting. The mark size is 0.51 meters in diameter. In areas that require long-term mapping, use metal nails and concrete markers;

[0016] Step 3: Use real-time dynamic differential technology to receive differential signals through GNSS base stations to measure the coordinates of each GCP with high precision. The coordinates must include three-dimensional positions (X, Y, Z) and be converted to the target coordinate system. Record the number, coordinates, photos and description of each control point to ensure correct post-processing.

[0017] Step 4: Import the GCP coordinate data into the 3D modeling software, match the control points with the corresponding positions on the image, and after calibration, check the error report to ensure that the geographic accuracy of the model meets the requirements.

[0018] Preferably, during the process of data collection by the S2 drone, RTK and PPK positioning technologies are used to collect high-precision geographic coordinates, the RTK base station is placed at a known coordinate point, signal transmission is turned on, and the drone equipped with the RTK module receives the base station signal in real time, recording the high-precision flight path and image shooting location.

[0019] Preferably, in the process of generating high-density point cloud data from sparse point cloud in S6, a weighted selection method of feature points is adopted.

[0020] Step 1: Use the SuperPoint algorithm to extract key points from the image, perform preliminary screening on the extracted points, remove duplicate and redundant points, and retain evenly distributed key points in each image;

[0021] Step 2: Weight the feature points according to the following four dimensions: Geometric Importance (W g ): Calculate the edge density of feature points in the image, and the feature points near the boundary have higher weights; distribution uniformity (W d ): Use grid partitioning to ensure uniform distribution of feature points in space; View coverage (W v ): Analyze the visibility of feature points in multiple perspectives. Feature points covering multiple perspectives have higher weights. Texture clarity (W t ): Use the local texture information of the shadow to evaluate the clarity of the feature points;

[0022] Total weight calculation formula:

[0023] W total =αW g +βW d +γW v +δW t

[0024] Among them, α, β, γ, and δ are adjustable parameters, which are set according to application requirements;

[0025] Step 3: According to the total weight, high-weighted points are retained from all feature points first, and these weighted feature points are used for matching in the image alignment stage to ensure the geometric structure integrity when generating sparse point clouds and the basic reliability when generating dense point clouds;

[0026] Step 4: Calculate the distribution density of the sparse point cloud in three-dimensional space. Use the voxel grid segmentation method to divide the space into a uniform voxel grid and count the number of points in each voxel. If the point cloud density of some areas is lower than the set threshold, mark it as a sparse area.

[0027] Step 5: Use the interpolation algorithm to generate supplementary points in the low-density area, combine the image information, re-extract the feature points of the low-density area from the original multi-view image, and fill them into the sparse point cloud;

[0028] Step 6: Use a distance-based filtering algorithm for high-density areas to remove redundant points and prioritize points with prominent local features;

[0029] Step 7: In steep terrain areas, analyze the normal vector distribution of the point cloud, identify areas with large slopes and curvatures, and increase the point cloud density in these areas; use the local geometric information of the sparse point cloud in the boundary area to identify the edge of the building and the boundary of the road, and strengthen the point cloud density in these areas; in areas with high texture changes, use image texture change detection to increase the number of sparse point cloud points in areas with high texture changes;

[0030] Step 8: Adjust the local uniform distribution of the sparse point cloud to avoid point cloud aggregation and excessive sparseness. Use the Voronoi diagram partitioning algorithm to optimize the spatial distribution of points so that the sparse point cloud presents reasonable uniformity in key areas.

[0031] Step 9: Correct the boundary area of ​​the sparse point cloud, supplement the missing points of the boundary through interpolation and bundle adjustment, and enhance the subsequent point cloud densification effect.

[0032] Preferably, the geometric importance calculation steps are as follows:

[0033] Step 1: Edge detection: Use Canny to detect the edges in the image and calculate the edge strength value E(p) for each feature point. The higher the edge strength, the greater the geometric importance.

[0034] Step 2: Local curvature: Calculate the local curvature around the feature point in the point cloud: select the neighborhood point set N(p) of the feature point p, fit the local plane of the neighborhood point, and calculate the plane normal vector;

[0035] The curvature is defined as:

[0036]

[0037] Among them, λ1, λ2, λ3 are the eigenvalues ​​of the neighborhood covariance matrix of the feature point, and λ1>λ2>λ3;

[0038] Step 3: Geometric importance weight: Combine edge strength and curvature to calculate geometric importance:

[0039] W g (p) = a·E(p) + b·Curvature(p)

[0040] Among them, a and b are adjustment parameters, which are adjusted according to the complexity of the specific terrain.

[0041] Preferably, the steps of calculating the distribution uniformity are as follows:

[0042] Step 1: Grid segmentation method: Divide the image into grids of fixed size and count the number of feature points N in each grid. g , define the distribution uniformity weight of the grid where each feature point is located as:

[0043]

[0044] That is, the fewer points in the grid, the higher the weight of distribution uniformity;

[0045] Step 2: Global uniformity evaluation: Calculate the point cloud distribution variance σ of all grids in the image 2 :

[0046]

[0047] in, is the average number of points in each grid;

[0048] Using σ 2 Dynamically adjust the impact weight of distribution uniformity;

[0049] Preferably, the viewing angle coverage calculation steps are as follows:

[0050] Step 1: Number of visible viewpoints: Determine how many images the feature point p is visible in based on the camera position information. Let the image set be {I1 , I2 , … , I k}, define the number of visible viewing angles V(p) as:

[0051] V(p)=|{I i :p∈View(I i )}|

[0052] That is, the number of images to which the feature points are projected;

[0053] Step 2: View overlap analysis: Calculate the overlap area of ​​the visible images of the feature points:

[0054] Extract the overlapping area of ​​the perspective between the images and calculate the area A of the overlapping area overlap ;

[0055] The overlap weight of point p is defined as:

[0056]

[0057] Among them, A total is the total projected area of ​​all images;

[0058] Step 3: View coverage weight: Based on the number of visible viewpoints and overlap analysis, define the view coverage weight:

[0059] W v (p) = c·V(p) + d·W overlap (p)

[0060] c and d are adjustment parameters.

[0061] Preferably, the viewing angle coverage calculation steps are as follows:

[0062] Step 1: Local grayscale variance: Calculate the grayscale variance of the area where the feature points are located:

[0063]

[0064] Among them, I i (p) is the gray value of the i-th pixel in the window, is the average gray value of the window. The larger the variance, the richer the texture.

[0065] Step 2: Calculate the local entropy value: Convert the grayscale value of the window where the feature point is located into a histogram and calculate the entropy value:

[0066]

[0067] Among them, P i is the probability of gray value i in the gray histogram. The larger the entropy value, the more texture details;

[0068] Step 3: Texture clarity weight: Combine variance and entropy to define texture clarity:

[0069] W t (p) = η·σ 2 (p)+κ·H(p)

[0070] η and κ are adjustment parameters.

[0071] Preferably, the total weight of feature points is:

[0072] W total (p) = α·W g (p)+β·W d (p)+γ·W v (p)+δ·W t (p)

[0073] By dynamically adjusting the weight parameters α, β, γ, and δ, the feature point selection strategy can be adjusted according to the actual needs of terrain modeling.

[0074] Compared with the prior art, the advantages of the present invention are:

[0075] (1) The present invention optimizes point cloud distribution and geometric quality through weighted selection of feature points and dynamic sparse point cloud adjustment, thereby improving the terrain accuracy and detail restoration capability of the three-dimensional model and meeting the modeling requirements of complex terrain and key areas.

[0076] (2) The present invention utilizes a high-performance computing server cluster and intelligent image processing software, combined with the flexibility of new energy drones, to achieve rapid data collection and efficient modeling, significantly shortening processing time and being applicable to a variety of terrain and environmental scenarios.

[0077] (3) The high-precision three-dimensional real-scene model generated by the present invention can be used in many fields such as urban planning, environmental protection, disaster assessment, etc., providing reliable data support and decision-making basis for related industries, and promoting the application and development of intelligent geographic information systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 It is a schematic diagram of the overall method of a method for three-dimensional modeling of terrain real scene of a new energy UAV of the present invention; DETAILED DESCRIPTION

[0079] For examples, see Figure 1 A method for real-scene 3D modeling of terrain for a new energy UAV, the method adopts a new energy UAV with strong endurance, high flight stability and flexible load adaptation, the new energy UAV is equipped with a five-eye oblique camera system, the five-eye oblique camera system includes one orthographic lens and four oblique lenses, and the oblique lens has an inclination angle of 30° to 45°;

[0080] The new energy UAV terrain real scene 3D modeling method comprises the following steps:

[0081] S1. Obtain topographic maps, satellite images and DEM data of the target area online, analyze terrain complexity and coverage, set the flight path of the drone using aerial survey planning software, the aerial survey planning software includes DJIPilot and Pix4DCapture, determine the flight altitude, overlap rate and tilt camera angle, generate a route file and upload it to the drone, the overlap rate is 70%-85% in the forward direction and 60%-80% in the lateral direction;

[0082] S2, the drone automatically flies according to the set route and collects images. The collection process is monitored in real time. After the drone completes the route, it returns to the base station;

[0083] S3. Check the quality of the collected image data, including whether there is blur, underexposure or insufficient overlap, and then back up the qualified image data to the cloud to prevent data loss;

[0084] S4, number and classify the collected images to ensure the integrity of the image data, associate the images with the drone GPS data, and obtain the spatial location information of each image;

[0085] S5. Use 3D modeling software to automatically detect conjugate points between images, complete image registration and alignment, generate sparse point clouds, and determine the basic geometric structure of the model. The 3D modeling software includes Zhitu and Agisoft Metashape.

[0086] S6. Generate high-density point cloud data based on sparse point cloud to construct detailed terrain structure, convert dense point cloud into 3D mesh model to form 3D surface, perform texture mapping on mesh surface, project collected images onto the model, and generate realistic 3D real scene model;

[0087] S7. Use ground control points and high-precision coordinate data to adjust the model position and accuracy, remove noise in the point cloud, optimize the model boundary and texture details, and then convert the model to the local projection coordinate system;

[0088] S8. Use the 3D display software Cesium to display the results and upload the model data to the cloud platform to facilitate team collaboration or customer access.

[0089] Using new energy drones and high-performance computing systems, we can quickly generate a 3D real-life model of the target terrain. Through high-resolution images and precise point cloud data, we can ensure that the model accuracy meets the geographic information requirements of 1:500 or even 1:1000 scale. We summarize the adaptation solutions for drones, oblique cameras, software, and computing systems, and provide a standardized process for large-scale 3D modeling projects.

[0090] The S2 drone uses ground control points to improve the geographic accuracy of the model during data collection. The method is as follows:

[0091] Step 1: Select the location of the ground control points. The distribution principle is as follows: the ground control points are evenly distributed in the modeling area, with key points at the edge and center. The number of points is determined according to the size of the modeling area and the accuracy requirements, with 510 points per square kilometer. In areas with complex terrain changes and high-precision models, the number of control points is increased by 12. Avoid placing points in vegetation-covered areas and on the surface of water bodies, and choose flat, open and clearly marked locations.

[0092] Step 2: Lay out and mark ground control points. The control points are marked as follows: Use white cross targets to ensure easy identification when drones are shooting. The mark size is 0.51 meters in diameter. In areas that require long-term mapping, use metal nails and concrete markers;

[0093] Step 3: Use real-time dynamic differential technology to receive differential signals through GNSS base stations to measure the coordinates of each GCP with high precision. The coordinates must include three-dimensional positions (X, Y, Z) and be converted to the target coordinate system. Record the number, coordinates, photos and description of each control point to ensure correct post-processing.

[0094] Step 4: Import the GCP coordinate data into the 3D modeling software, match the control points with the corresponding positions on the image, and after calibration, check the error report to ensure that the geographic accuracy of the model meets the requirements.

[0095] During the data collection process of the S2 drone, RTK and PPK positioning technologies are used to collect high-precision geographic coordinates. The RTK base station is placed at a known coordinate point, and signal transmission is turned on. The drone equipped with the RTK module receives the base station signal in real time and records the high-precision flight path and image shooting location.

[0096] In the process of generating high-density point cloud data from the S6 sparse point cloud, a weighted selection method of feature points is adopted.

[0097] Step 1: Use the SuperPoint algorithm to extract key points from the image, perform preliminary screening on the extracted points, remove duplicate and redundant points, and retain evenly distributed key points in each image;

[0098] Step 2: Weight the feature points according to the following four dimensions:

[0099] Geometric importance (W g ): Calculates the edge density of feature points in the image. Feature points near the boundary have higher weights because these points are more helpful in defining the shape of the object.

[0100] Distribution uniformity (W d): Analyze the spatial distribution of feature points to avoid the point cloud being concentrated in a certain area and ignoring other areas. Use the grid partitioning method to ensure the uniform distribution of feature points in space.

[0101] View coverage (W v ): Analyze the visibility of feature points in multiple perspectives, and feature points covering multiple perspectives have higher weights.

[0102] Texture clarity (W t ): Use the local texture information of the shadow to evaluate the clarity of feature points.

[0103] Total weight calculation formula:

[0104] W total =αW g +βW d +γW v +δW t

[0105] Among them, α, β, γ, and δ are adjustable parameters, which are set according to application requirements;

[0106] Step 3: According to the total weight, high-weighted points are retained from all feature points first, and these weighted feature points are used for matching in the image alignment stage to ensure the geometric structure integrity when generating sparse point clouds and the basic reliability when generating dense point clouds;

[0107] Step 4: Calculate the distribution density of the sparse point cloud in three-dimensional space. Use the voxel grid segmentation method to divide the space into a uniform voxel grid and count the number of points in each voxel. If the point cloud density of some areas is lower than the set threshold, mark it as a sparse area.

[0108] Step 5: Use the interpolation algorithm to generate supplementary points in the low-density area, combine the image information, re-extract the feature points of the low-density area from the original multi-view image, and fill them into the sparse point cloud;

[0109] Step 6: Use a distance-based filtering algorithm for high-density areas to remove redundant points and prioritize points with prominent local features;

[0110] Step 7: In steep terrain areas, analyze the normal vector distribution of the point cloud, identify areas with large slopes and curvatures, and increase the point cloud density in these areas; use the local geometric information of the sparse point cloud in the boundary area to identify the edge of the building and the boundary of the road, and strengthen the point cloud density in these areas; in areas with high texture changes, use image texture change detection to increase the number of sparse point cloud points in areas with high texture changes;

[0111] Step 8: Adjust the local uniform distribution of the sparse point cloud to avoid point cloud aggregation and excessive sparseness. Use the Voronoi diagram partitioning algorithm to optimize the spatial distribution of points so that the sparse point cloud presents reasonable uniformity in key areas.

[0112] Step 9: Correct the boundary area of ​​the sparse point cloud, supplement the missing points of the boundary through interpolation and bundle adjustment, and enhance the subsequent point cloud densification effect.

[0113] The geometric importance calculation steps are as follows:

[0114] Step 1: Edge detection: Use Canny to detect the edges in the image and calculate the edge strength value E(p) for each feature point. The higher the edge strength, the greater the geometric importance.

[0115] Step 2: Local curvature: Calculate the local curvature around the feature point in the point cloud: select the neighborhood point set N(p) of the feature point p, fit the local plane of the neighborhood point, and calculate the plane normal vector;

[0116] The curvature is defined as:

[0117]

[0118] Among them, λ1, λ2, λ3 are the eigenvalues ​​of the neighborhood covariance matrix of the feature point, and λ1>λ2>λ3;

[0119] Step 3: Geometric importance weight: Combine edge strength and curvature to calculate geometric importance:

[0120] W g (p) = a·E(p) + b·Curvature(p)

[0121] Among them, a and b are adjustment parameters, which are adjusted according to the complexity of the specific terrain.

[0122] The steps for calculating the distribution uniformity are as follows:

[0123] Step 1: Grid segmentation method: Divide the image into grids of fixed size and count the number of feature points N in each grid. g , define the distribution uniformity weight of the grid where each feature point is located as:

[0124]

[0125] That is, the fewer points in the grid, the higher the weight of distribution uniformity;

[0126] Step 2: Global uniformity evaluation: Calculate the point cloud distribution variance σ of all grids in the image 2 :

[0127]

[0128] in, is the average number of points in each grid;

[0129] Using σ 2 Dynamically adjust the impact weight of distribution uniformity;

[0130] The viewing angle coverage calculation steps are as follows:

[0131] Step 1: Number of visible viewpoints: Determine how many images the feature point p is visible in based on the camera position information. Let the image set be {I1 , I2 , … , I k}, define the number of visible viewing angles V(p) as:

[0132] V(p)=|{I i :p∈View(I i )}|

[0133] That is, the number of images to which the feature points are projected;

[0134] Step 2: View overlap analysis: Calculate the overlap area of ​​the visible images of the feature points:

[0135] Extract the overlapping area of ​​the perspective between the images and calculate the area A of the overlapping area overlap ;

[0136] The overlap weight of point p is defined as:

[0137]

[0138] Among them, A total is the total projected area of ​​all images;

[0139] Step 3: View coverage weight: Based on the number of visible viewpoints and overlap analysis, define the view coverage weight:

[0140] W v (p) = c·V(p) + d·W overlap (p)

[0141] c and d are adjustment parameters.

[0142] The viewing angle coverage calculation steps are as follows:

[0143] Step 1: Local grayscale variance: Calculate the grayscale variance of the area where the feature points are located:

[0144]

[0145] Among them, I i(p) is the gray value of the i-th pixel in the window, is the average gray value of the window. The larger the variance, the richer the texture.

[0146] Step 2: Calculate the local entropy value: Convert the grayscale value of the window where the feature point is located into a histogram and calculate the entropy value:

[0147]

[0148] Among them, P i is the probability of gray value i in the gray histogram. The larger the entropy value, the more texture details;

[0149] Step 3: Texture clarity weight: Combine variance and entropy to define texture clarity:

[0150] W t (p) = η·σ 2 (p)+κ·H(p)

[0151] η and κ are adjustment parameters.

[0152] Total weight of feature points:

[0153] W total (p) = α·W g (p)+β·W d (p)+γ·W v (p)+δ·W t (p)

[0154] By dynamically adjusting the weight parameters α, β, γ, and δ, the feature point selection strategy can be adjusted according to the actual needs of terrain modeling.

[0155] The above shows and describes the basic principles, main features and advantages of the present invention; those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only preferred examples of the present invention and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected; the scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for three-dimensional modeling of real-scene terrain of new energy UAV, characterized in that: The method for real-scene 3D modeling of terrain of new energy UAV adopts a new energy UAV with strong endurance, high flight stability and flexible load adaptation. The new energy UAV is equipped with a five-eye oblique camera system, which includes one orthographic lens and four oblique lenses, and the inclination angle of the oblique lens is 30° to 45°. The new energy UAV terrain real scene 3D modeling method comprises the following steps: S1. Obtain topographic maps, satellite images and DEM data of the target area online, analyze terrain complexity and coverage, set the flight path of the drone using aerial survey planning software, the aerial survey planning software includes DJIPilot and Pix4DCapture, determine the flight altitude, overlap rate and tilt camera angle, generate a route file and upload it to the drone, the overlap rate is 70%-85% in the forward direction and 60%-80% in the lateral direction; S2, the drone automatically flies according to the set route and collects images. The collection process is monitored in real time. After the drone completes the route, it returns to the base station; S3. Check the quality of the collected image data, including whether there is blur, underexposure or insufficient overlap, and then back up the qualified image data to the cloud to prevent data loss; S4, number and classify the collected images to ensure the integrity of the image data, associate the images with the drone GPS data, and obtain the spatial location information of each image; S5. Use 3D modeling software to automatically detect conjugate points between images, complete image registration and alignment, generate sparse point clouds, and determine the basic geometric structure of the model. The 3D modeling software includes Zhitu and Agisoft Metashape. S6. Generate high-density point cloud data based on sparse point cloud to construct detailed terrain structure, convert dense point cloud into 3D mesh model to form 3D surface, perform texture mapping on mesh surface, project collected images onto the model, and generate realistic 3D real scene model; S7. Use ground control points and high-precision coordinate data to adjust the model position and accuracy, remove noise in the point cloud, optimize the model boundary and texture details, and then convert the model to the local projection coordinate system; S8. Use the 3D display software Cesium to display the results and upload the model data to the cloud platform to facilitate team collaboration or customer access.

2. According to the method of claim 1, the method is characterized in that: The S2 drone uses ground control points to improve the geographic accuracy of the model during data collection. The method is as follows: Step 1: Select the location of the ground control points. The distribution principle is as follows: the ground control points are evenly distributed in the modeling area, with key points at the edge and center. The number of points is determined according to the size of the modeling area and the accuracy requirements, with 510 points per square kilometer. In areas with complex terrain changes and high-precision models, the number of control points is increased by 12. Avoid placing points in vegetation-covered areas and on the surface of water bodies, and choose flat, open and clearly marked locations. Step 2: Lay out and mark ground control points. The control points are marked as follows: Use white cross targets to ensure easy identification when drones are shooting. The mark size is 0.51 meters in diameter. In areas that require long-term mapping, use metal nails and concrete markers; Step 3: Use real-time dynamic differential technology to receive differential signals through GNSS base stations to measure the coordinates of each GCP with high precision. The coordinates must include three-dimensional positions (X, Y, Z) and be converted to the target coordinate system. Record the number, coordinates, photos and description of each control point to ensure correct post-processing. Step 4: Import the GCP coordinate data into the 3D modeling software, match the control points with the corresponding positions on the image, and after calibration, check the error report to ensure that the geographic accuracy of the model meets the requirements.

3. According to the method for three-dimensional modeling of real-scene terrain of new energy UAV according to claim 2, it is characterized in that: During the data collection process of the S2 drone, RTK and PPK positioning technologies are used to collect high-precision geographic coordinates. The RTK base station is placed at a known coordinate point, and signal transmission is turned on. The drone equipped with the RTK module receives the base station signal in real time and records the high-precision flight path and image shooting location.

4. According to the method for three-dimensional modeling of terrain real scene of new energy UAV according to claim 1, it is characterized in that: In the process of generating high-density point cloud data from the S6 sparse point cloud, a weighted selection method of feature points is adopted. Step 1: Use the SuperPoint algorithm to extract key points from the image, perform preliminary screening on the extracted points, remove duplicate and redundant points, and retain evenly distributed key points in each image; Step 2: Weight the feature points according to the following four dimensions: Geometric Importance (W g ): Calculate the edge density of feature points in the image, and the feature points near the boundary have higher weights; distribution uniformity (W d ): Use grid partitioning to ensure uniform distribution of feature points in space; View coverage (W v ): Analyze the visibility of feature points in multiple perspectives. Feature points covering multiple perspectives have higher weights. Texture clarity (W t ): Use the local texture information of the shadow to evaluate the clarity of the feature points; Total weight calculation formula: W total =αW g +βW d +γW v +δW t Among them, α, β, γ, and δ are adjustable parameters, which are set according to application requirements; Step 3: According to the total weight, high-weighted points are retained from all feature points first, and these weighted feature points are used for matching in the image alignment stage to ensure the geometric structure integrity when generating sparse point clouds and the basic reliability when generating dense point clouds; Step 4: Calculate the distribution density of the sparse point cloud in three-dimensional space. Use the voxel grid segmentation method to divide the space into a uniform voxel grid and count the number of points in each voxel. If the point cloud density of some areas is lower than the set threshold, mark it as a sparse area. Step 5: Use the interpolation algorithm to generate supplementary points in the low-density area, combine the image information, re-extract the feature points of the low-density area from the original multi-view image, and fill them into the sparse point cloud; Step 6: Use a distance-based filtering algorithm for high-density areas to remove redundant points and prioritize points with prominent local features; Step 7: In steep terrain areas, analyze the normal vector distribution of the point cloud, identify areas with large slopes and curvatures, and increase the point cloud density in these areas; use the local geometric information of the sparse point cloud in the boundary area to identify the edge of the building and the boundary of the road, and strengthen the point cloud density in these areas; in areas with high texture changes, use image texture change detection to increase the number of sparse point cloud points in areas with high texture changes; Step 8: Adjust the local uniform distribution of the sparse point cloud to avoid point cloud aggregation and excessive sparseness. Use the Voronoi diagram partitioning algorithm to optimize the spatial distribution of points so that the sparse point cloud presents reasonable uniformity in key areas. Step 9: Correct the boundary area of ​​the sparse point cloud, supplement the missing points of the boundary through interpolation and bundle adjustment, and enhance the subsequent point cloud densification effect.

5. According to the method of claim 4, a new energy UAV terrain real scene 3D modeling method is characterized in that: The geometric importance calculation steps are as follows: Step 1: Edge detection: Use Canny to detect the edges in the image and calculate the edge strength value E(p) for each feature point. The higher the edge strength, the greater the geometric importance. Step 2: Local curvature: Calculate the local curvature around the feature point in the point cloud: select the neighborhood point set N(p) of the feature point p, fit the local plane of the neighborhood point, and calculate the plane normal vector; The curvature is defined as: Among them, λ1, λ2, λ3 are the eigenvalues ​​of the neighborhood covariance matrix of the feature point, and λ1>λ2>λ3; Step 3: Geometric importance weight: Combine edge strength and curvature to calculate geometric importance: W g (p)=a·E(p)+b·Curvature(p) Among them, a and b are adjustment parameters, which are adjusted according to the complexity of the specific terrain.

6. According to the method of claim 5, the method is characterized in that: The steps for calculating the distribution uniformity are as follows: Step 1: Grid segmentation method: Divide the image into grids of fixed size and count the number of feature points N in each grid. g , define the distribution uniformity weight of the grid where each feature point is located as: That is, the fewer points in the grid, the higher the weight of distribution uniformity; Step 2: Global uniformity evaluation: Calculate the point cloud distribution variance σ of all grids in the image 2 : in, is the average number of points in each grid; Using σ 2 Dynamically adjust the impact weight of distribution uniformity.

7. According to claim 6, a new energy UAV terrain real scene 3D modeling method is characterized in that: The viewing angle coverage calculation steps are as follows: Step 1: Number of visible viewpoints: Determine how many images the feature point p is visible in based on the camera position information. Let the image set be {I1 , I2 , … , I k }, define the number of visible viewing angles V(p) as: V(p)=|{I i :p∈View(I i )}| That is, the number of images to which the feature points are projected; Step 2: View overlap analysis: Calculate the overlap area of ​​the visible images of the feature points: Extract the overlapping area of ​​the perspective between the images and calculate the area A of the overlapping area overlap ; The overlap weight of point p is defined as: Among them, A total is the total projected area of ​​all images; Step 3: View coverage weight: Based on the number of visible viewpoints and overlap analysis, define the view coverage weight: W v (p)=c·V(p)+d·W overlap (p) c and d are adjustment parameters.

8. According to claim 7, a new energy UAV terrain real scene 3D modeling method is characterized in that: The viewing angle coverage calculation steps are as follows: Step 1: Local grayscale variance: Calculate the grayscale variance of the area where the feature points are located: Among them, I i (p) is the gray value of the i-th pixel in the window, is the average gray value of the window. The larger the variance, the richer the texture. Step 2: Calculate the local entropy value: Convert the grayscale value of the window where the feature point is located into a histogram and calculate the entropy value: Among them, P i is the probability of gray value i in the gray histogram. The larger the entropy value, the more texture details; Step 3: Texture clarity weight: Combine variance and entropy to define texture clarity: W t (p)=η·σ 2 (p)+κ·H(p) η and κ are adjustment parameters.

9. According to claim 8, a new energy UAV terrain real scene 3D modeling method is characterized in that: Total weight of feature points: W total (p)=α·W g (p)+β·W d (p)+γ·W v (p)+δ·W t (p) By dynamically adjusting the weight parameters α, β, γ, and δ, the feature point selection strategy can be adjusted according to the actual needs of terrain modeling.

Images