Method for rapid generation of three-dimensional temperature field model based on multi-source data of unmanned aerial vehicle

Abstract

The application belongs to the technical field of three-dimensional modeling, and provides a three-dimensional temperature field model fast generation method based on unmanned aerial vehicle multi-source data, which comprises the following steps: firstly, obtaining target area laser point cloud and thermal infrared image data, octree blocking of the laser point cloud and calculation of density comprehensive index, hierarchical filtering of the density balanced, voxel and proportion controlled point cloud to obtain simplified point cloud; secondly, extracting high-curvature feature point cloud, merging and deduplicating, and then constructing a three-dimensional grid model through Poisson reconstruction. The point cloud and the thermal infrared image are spatially aligned, the temperature information is mapped to the model surface through coordinate conversion and perspective projection, and a three-dimensional temperature field model is generated. The application considers both efficiency and accuracy through adaptive point cloud simplification and feature reservation, simplifies the data alignment process through sensor external parameter calibration, accurately fuses temperature and three-dimensional structure information, and efficiently solves the problems of high cost, low efficiency, insufficient resolution and fusion accuracy in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of 3D modeling technology and provides a method for rapidly generating 3D temperature field models based on multi-source data from UAVs. Background Technology

[0002] In fields such as environmental monitoring, industrial inspection, geological disaster early warning, and agricultural production regulation, the core need is to rapidly acquire fused information on the three-dimensional spatial morphology and surface temperature distribution of large-scale scenes to support key applications such as temperature anomaly identification and spatial-temperature correlation analysis. Therefore, the industry urgently needs a three-dimensional temperature field model generation technology that balances large-scale coverage, high-precision modeling, rapid processing, and low-cost application. However, traditional monitoring methods have many limitations and cannot meet the needs of precise, efficient, and three-dimensional monitoring.

[0003] Currently, aerial photogrammetry is the primary method for acquiring three-dimensional temperature field models. The core technical process involves: planning and setting flight paths based on thermal infrared camera parameters, acquiring area array thermal infrared images of the study area, and then performing automatic three-dimensional reconstruction using image matching technology to ultimately obtain the three-dimensional temperature field model data for the study area. However, traditional temperature field modeling methods have significant drawbacks and limited supporting capabilities. Specifically, these are reflected in the following two aspects:

[0004] First, the data processing is complex and the reliability is low. Conventional methods require matching the acquired two-dimensional thermal infrared images of the temperature field before automatically reconstructing the three-dimensional temperature field based on the matching results. If the study area is located near the snow line at high altitudes or other areas covered by snow and ice year-round, conventional matching methods are unable to extract stable, reliable, and uniformly distributed feature points, which can easily lead to image matching failure and prevent the generation of the three-dimensional temperature field.

[0005] Secondly, the data processing efficiency is low, making it difficult to meet timeliness requirements. Due to the limited effective image size of thermal infrared cameras, the amount of image data required per unit area of ​​the study area far exceeds that of conventional visible light data, reaching tens of times greater. Furthermore, geological disaster monitoring based on three-dimensional temperature fields (such as ice avalanches, landslides, debris flows, and unstable rock masses) is mostly an emergency application, demanding extremely high timeliness in data processing and model generation. The cumbersome processing procedures of conventional aerial photogrammetry methods, coupled with the massive amount of data, create a significant contradiction, resulting in excessively long model generation cycles and an inability to quickly provide effective support for disaster monitoring and early warning. Summary of the Invention

[0006] To address the aforementioned technical challenges, this invention provides a rapid method for generating 3D temperature field models based on multi-source data from unmanned aerial vehicles (UAVs). It employs octree-based block adaptive point cloud simplification and feature point preservation, balancing modeling efficiency and 3D accuracy. Through joint calibration of sensor extrinsic parameters, it achieves direct alignment of multi-source data, simplifying the process and improving accuracy and stability. Furthermore, it accurately maps thermal infrared temperature information onto a 3D mesh, rapidly generating a high-precision 3D temperature field model, effectively solving the problems of high cost, low efficiency, poor resolution, and insufficient fusion accuracy inherent in existing methods.

[0007] The technical solution of the present invention includes: By using a drone equipped with a lidar camera and a thermal infrared imager, laser point cloud datasets and thermal infrared image data are simultaneously collected in the target area.

[0008] The laser point cloud dataset is divided into several blocks using an octree partitioning method. The local density of each point is obtained, and the block density, block density standard deviation, and global density are obtained based on the local density of each point to obtain a comprehensive index of the block. Based on the comprehensive index, the blocks are divided into five levels: extremely sparse, sparse, medium, dense, and extremely dense. Density equalization filtering and voxel filtering are performed on the dense and extremely dense blocks to obtain the target number of point clouds. Then, proportional control filtering is performed on the blocks of the five levels to ensure that the number of point clouds in each block is within the target range, resulting in a simplified point cloud dataset.

[0009] Obtain the curvature of each point cloud in the laser point cloud dataset, remove point clouds with curvature less than a threshold, and obtain the feature point cloud dataset.

[0010] The simplified point cloud dataset and the feature point cloud dataset are merged, and a hash table is constructed to map voxels. Duplicate points are removed to obtain the target point cloud dataset. Based on the point cloud position and normal information in the target point cloud dataset, Poisson reconstruction is performed to obtain the three-dimensional mesh model of the target region.

[0011] The laser point cloud dataset and thermal infrared image data are spatially aligned. Then, the thermal infrared image data aligned with the laser point cloud dataset is mapped onto the surface of the 3D mesh model through coordinate transformation and perspective projection to generate a 3D temperature field model with temperature information.

[0012] Furthermore, before dividing the octree block method into several blocks, the laser point cloud dataset is preprocessed to remove outliers and noise points.

[0013] Furthermore, the method for obtaining comprehensive indicators is as follows: Obtain each point cloud in the laser point cloud dataset Local density The formula is: Get block density Standard deviation of block density and global density The formula is: Obtain comprehensive indicators The formula is: In the formula, K is a point The number of nearest neighbors, It is a constant, with a value of 1e. -6 , ||·|| is the Euclidean distance. For point To all neighboring points The average distance, n is the number of point clouds in the block, N is the total number of point clouds in the laser point cloud dataset, α is an adjustment coefficient and α>0, generally taken as the standard deviation of the density of the entire block. The reciprocal of the average.

[0014] Furthermore, based on the comprehensive index r, the levels are divided as follows: r < 0.1 is extremely sparse, 0.1 ≤ r < 0.5 is sparse, 0.5 ≤ r < 2.0 is medium, 2.0 ≤ r < 10.0 is dense, and r ≥ 10.0 is extremely dense.

[0015] Furthermore, the method for obtaining the simplified point cloud dataset is as follows: Obtain the number of point cloud targets after dense and extremely dense block equalization filtering. The formula is: Number of point cloud targets Voxel filtering is performed to obtain the number of target point clouds. .

[0016] The number of point cloud targets is obtained by performing proportional control filtering on all blocks. The formula is: In the formula, This represents the number of original point clouds within the block. The target density after density equalization filtering. The value is set to 0.3 when the comprehensive index r ≥ 20. When 10 ≤ comprehensive index r < 20, take When 2.0 ≤ comprehensive index r < 10, take 5. t is the preset dilution ratio.

[0017] Number of target point clouds and the number of point cloud targets In comparison, if the target point cloud quantity Number of point cloud targets Within the error range, the number of target point clouds is output. If the target point cloud number Not in point cloud target quantity Within the error range, then the number of target point clouds is further adjusted. Perform voxel filtering until the target point cloud number is reached. Number of point cloud targets Within the error range.

[0018] The total number of target point clouds in all blocks This constitutes a simplified point cloud dataset.

[0019] Furthermore, the method for obtaining the feature point cloud dataset is as follows: Select the k nearest neighbors of any point P in the laser point cloud dataset and construct the covariance matrix. : In the formula, Let the point be the centroid of the neighborhood of that point. It is the j-th neighbor point.

[0020] For covariance matrix Eigenvalue decomposition yields three eigenvalues. , , ,and ≤ ≤ , corresponding feature vector Let P be the direction of the normal to point P.

[0021] Based on three eigenvalues , , Obtaining curvature The formula is: curvature By comparing with a threshold, point clouds with curvature less than the threshold are removed, resulting in point clouds with curvature greater than or equal to the threshold, which is the feature point cloud dataset.

[0022] Furthermore, the method for constructing the three-dimensional mesh model is as follows: The set of all points in the target point cloud dataset S is {s1, s2, ..., s}.N}, where N is the number of laser points, and a closed geometric body M is constructed in the point set space, with an indicator function applied to all points in the point set space. definition: For any point s in the laser point cloud i Let i be a natural number less than N, and s i In geometry If the function is internal, the value of the indicator function is 1; if it is external, the value of the indicator function is 0, where the domain of x is the point set S.

[0023] The gradient field ∇χ of the indicator function is zero both inside and outside the conceived geometry M, but a huge pulse is formed on the surface of M, with the pulse direction perpendicular to the surface of geometry M and pointing outward.

[0024] All points s in the laser point cloud i and its normal vector s i . This can be viewed as a sparse sampling of the gradient field ∇χ of the indicator function on the surface of the target geometry M, thereby inferring the original indicator function, specifically: Take any point q in the point set space, and construct a smooth and continuous vector field by weighted sum of the normals of all points in the target point cloud dataset S. ,Right now: In the formula, For a smooth convolution function, express Displacement along the p direction.

[0025] Vector Field For the ideal gradient field ∇χ, the approximation is: ∇χ= Differentiating both sides of the above equation, we have: Δχ=∇· In the formula, Δ is the Laplace operator, Δχ=∇·(∇χ) is the divergence of the gradient, and ∇· For vector fields The divergence.

[0026] Therefore, the known vector field can be calculated. Then the indicator function is obtained. This allows us to obtain the surface of the target geometry M.

[0027] Furthermore, the method for spatial alignment is as follows: The external parameters of the lidar camera and the thermal infrared imager are jointly calibrated, and a calibration matrix is ​​constructed based on the rigid transformation relationship obtained from the calibration. .

[0028] In the formula It is a 3×3 rotation matrix. It is a 3×1 translation vector.

[0029] The calibration matrix is ​​used to perform spatial position transformation on the thermal infrared image data, achieving spatial alignment between the laser point cloud data and the thermal infrared image data. The formula is as follows: In the formula, This represents the pose of the lidar in the world coordinate system. This represents the pose of the calibrated infrared camera in the world coordinate system.

[0030] Furthermore, the method for constructing the three-dimensional temperature field model is as follows: For any vertex P of a triangular facet in a 3D mesh model, the world coordinates are: The calibration matrix is ​​used to convert it into the world coordinate system coordinates of the thermal infrared imager. The formula is: Then, using perspective projection, the pixel position (u, v) of point P in the two-dimensional image plane is obtained, using the formula: In the formula, The physical dimension of the camera's focal length in the horizontal direction. The physical dimension of the camera's focal length in the vertical direction. The horizontal coordinates of the optical axis and the image plane. represents the perpendicular coordinates of the optical axis to the image plane.

[0031] Using the above conversion process, the positions of each vertex of the triangular facet in the three-dimensional mesh model in the two-dimensional image plane are obtained. Then, the thermal infrared elements corresponding to each facet can be extracted, all thermal infrared elements can be stored in the texture atlas, and the texture atlas can be fused with the three-dimensional mesh model to obtain the three-dimensional temperature field model.

[0032] The technical solution provided by this invention has the following advantages compared with the prior art: This invention effectively solves the problem that traditional fixed thinning cannot balance modeling efficiency and 3D structural accuracy by using octree segmentation and density gradation to adaptively simplify point clouds and retain key terrain details by combining curvature features. It adopts joint calibration of extrinsic parameters of lidar and thermal infrared camera to achieve direct spatial alignment, replacing the time-consuming and error-prone feature matching process, which greatly improves the speed and alignment accuracy of multi-source data fusion. Relying on Poisson reconstruction and perspective projection texture mapping, it quickly generates a 3D temperature field model with temperature information, overcoming the shortcomings of high ground deployment cost, low satellite remote sensing resolution, and lack of spatial-temperature correlation of single data source, and achieves high-precision, high-efficiency, and low-cost rapid construction of 3D temperature fields in large-scale complex scenes.

[0033] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Detailed Implementation

[0034] The following detailed description is based on a specific embodiment, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiment.

[0035] In the description of the embodiments of the present invention, unless otherwise stated, "a plurality of" means two or more.

[0036] This invention provides a method for rapidly generating a three-dimensional temperature field model based on multi-source data from unmanned aerial vehicles (UAVs), including: Acquire laser point cloud datasets and thermal infrared imagery data for the target area. Specifically, use a UAV equipped with a standard aerial survey system, including a LiDAR camera, thermal infrared imager, and inertial navigation positioning system, to fly over the target area and collect large-scale laser point cloud data and thermal infrared imagery data. The laser point cloud data provides high-precision terrain information, while the thermal infrared imagery data provides surface temperature textures for a temperature field model.

[0037] The laser point cloud dataset is divided into several blocks using an octree partitioning method. The local density of each point is obtained, and the block density, block density standard deviation, and global density are obtained based on the local density of each point to obtain a comprehensive index of the block. Based on the comprehensive index, the blocks are divided into five levels: extremely sparse, sparse, medium, dense, and extremely dense. Density equalization filtering and voxel filtering are performed on the dense and extremely dense blocks to obtain the target number of point clouds. Then, proportional control filtering is performed on the blocks of the five levels to ensure that the number of point clouds in each block is within the target range, resulting in a simplified point cloud dataset.

[0038] Obtain the curvature of each point cloud in the laser point cloud dataset, remove point clouds with curvature less than a threshold, and obtain the feature point cloud dataset.

[0039] The simplified point cloud dataset and the feature point cloud dataset are merged, and a hash table is constructed to map voxels. Duplicate points are removed to obtain the target point cloud dataset. Based on the point cloud position and normal information in the target point cloud dataset, Poisson reconstruction is performed to obtain the three-dimensional mesh model of the target region.

[0040] The laser point cloud dataset and thermal infrared image data are spatially aligned. Then, the thermal infrared image data aligned with the laser point cloud dataset is mapped onto the surface of the 3D mesh model through coordinate transformation and perspective projection to generate a 3D temperature field model with temperature information.

[0041] Furthermore, the extrinsic parameters of the lidar camera and the thermal infrared imager are jointly calibrated, and the spatial position of the thermal infrared image data is transformed based on the obtained parameters to ensure that the spatial positions of the acquired multi-source data coincide, thus eliminating the need for complex and time-consuming feature matching of the multi-source data.

[0042] Furthermore, before dividing the octree block method into several blocks, the laser point cloud dataset is preprocessed to remove outliers and noise points.

[0043] Furthermore, the method for obtaining comprehensive indicators is as follows: Obtain each point cloud in the laser point cloud dataset Local density The formula is: Get block density Standard deviation of block density and global density The formula is: Obtain comprehensive indicators The formula is: In the formula, K is a point The number of nearest neighbors, It is a constant, with a value of 1e. -6 , ||·|| is the Euclidean distance. For point To all neighboring points The average distance, n is the number of point clouds in the block, N is the total number of point clouds in the laser point cloud dataset, α is an adjustment coefficient and α>0, generally taken as the standard deviation of the density of the entire block. The reciprocal of the average.

[0044] Furthermore, the greater the average density of the block nodes and the more uneven the point cloud distribution density, the greater the comprehensive index r value. According to the comprehensive index r, the levels are divided as follows: r < 0.1 is extremely sparse, 0.1 ≤ r < 0.5 is sparse, 0.5 ≤ r < 2.0 is medium, 2.0 ≤ r < 10.0 is dense, and r ≥ 10.0 is extremely dense.

[0045] Furthermore, the point cloud after block processing is equalized in two stages. The purpose is to significantly reduce the number of point clouds and make the number of point clouds in each block more similar, thereby accelerating the subsequent 3D model reconstruction. The method for obtaining the simplified point cloud dataset is as follows: Obtain the number of point cloud targets after dense and extremely dense block equalization filtering. The formula is: Number of point cloud targets Voxel filtering is performed to obtain the number of target point clouds. .

[0046] The number of point cloud targets is obtained by performing proportional control filtering on all blocks. The formula is: In the formula, This represents the number of original point clouds within the block. The target density after density equalization filtering. The value is set to 0.3 when the comprehensive index r ≥ 20. When 10 ≤ comprehensive index r < 20, take When 2.0 ≤ comprehensive index r < 10, take 5. t is the preset dilution ratio.

[0047] Number of target point clouds and the number of point cloud targets In comparison, if the target point cloud quantity Number of point cloud targets Within the error range, the number of target point clouds is output. If the target point cloud number Not in point cloud target quantity Within the error range, then the number of target point clouds is further adjusted. Perform voxel filtering until the target point cloud number is reached. Number of point cloud targets Within the error range.

[0048] The total number of target point clouds in all blocks This constitutes a simplified point cloud dataset.

[0049] Furthermore, for terrain point clouds, to avoid voxel filtering removing too many feature points, feature points are extracted from the classified blocks to ensure that feature details are not lost as much as possible. The method for obtaining the feature point cloud dataset is as follows: Select the k nearest neighbors of any point P in the laser point cloud dataset and construct the covariance matrix. : In the formula, Let the point be the centroid of the neighborhood of that point. It is the j-th neighbor point.

[0050] For covariance matrix Eigenvalue decomposition yields three eigenvalues. , , ,and ≤ ≤ , corresponding feature vector Let P be the direction of the normal to point P.

[0051] Based on three eigenvalues , , Obtaining curvature The formula is: curvature By comparing with a threshold, point clouds with curvature less than the threshold are removed, resulting in point clouds with curvature greater than or equal to the threshold, which is the feature point cloud dataset.

[0052] Curvature The larger the value compared to the threshold, the greater the change in the normal direction, and the greater the curvature of the ground at that point, generally indicating an edge or corner point. Therefore, different thresholds can be set based on the point cloud terrain overview and the density classification mentioned earlier. If the curvature exceeds this value, it is extracted as a feature point.

[0053] Furthermore, the method for constructing the three-dimensional mesh model is as follows: The set of all points in the target point cloud dataset S is {s1, s2, ..., s}. N}, where N is the number of laser points, and a closed geometric body M is constructed in the point set space, with an indicator function applied to all points in the point set space. definition: For any point s in the laser point cloud i Let i be a natural number less than N, and s iIn geometry If the function is internal, the value of the indicator function is 1; if it is external, the value of the indicator function is 0, where the domain of x is the point set S.

[0054] The gradient field ∇χ of the indicator function is zero both inside and outside the conceived geometry M, but a huge pulse is formed on the surface of M, with the pulse direction perpendicular to the surface of geometry M and pointing outward.

[0055] All points s in the laser point cloud i and its normal vector s i . This can be viewed as a sparse sampling of the gradient field ∇χ of the indicator function on the surface of the target geometry M, thereby inferring the original indicator function, specifically: Take any point q in the point set space, and construct a smooth and continuous vector field by weighted sum of the normals of all points in the target point cloud dataset S. ,Right now: In the formula, For a smooth convolution function, express Displacement along the p direction.

[0056] Vector Field For the ideal gradient field ∇χ, the approximation is: ∇χ= Differentiating both sides of the above equation, we have: Δχ=∇· In the formula, Δ is the Laplace operator, Δχ=∇·(∇χ) is the divergence of the gradient, and ∇· For vector fields The divergence.

[0057] Therefore, the known vector field can be calculated. Then the indicator function is obtained. This allows us to obtain the surface of the target geometry M.

[0058] Furthermore, the method for spatial alignment is as follows: A drone equipped with a lidar camera and a thermal infrared imager is used to simultaneously collect lidar point cloud data and thermal infrared image data in the target area.

[0059] The external parameters of the lidar camera and the thermal infrared imager are jointly calibrated, and a calibration matrix is ​​constructed based on the rigid transformation relationship obtained from the calibration. .

[0060] In the formula It is a 3×3 rotation matrix. It is a 3×1 translation vector.

[0061] The calibration matrix is ​​used to perform spatial position transformation on the thermal infrared image data, achieving spatial alignment between the laser point cloud data and the thermal infrared image data. The formula is as follows: In the formula, This represents the pose of the lidar in the world coordinate system. This represents the pose of the calibrated infrared camera in the world coordinate system.

[0062] Furthermore, the method for constructing the three-dimensional temperature field model is as follows: For any vertex P of a triangular facet in a 3D mesh model, the world coordinates are: The calibration matrix is ​​used to convert it into the world coordinate system coordinates of the thermal infrared imager. The formula is: Then, using perspective projection, the pixel position (u, v) of point P in the two-dimensional image plane is obtained, using the formula: In the formula, The physical dimension of the camera's focal length in the horizontal direction. The physical dimension of the camera's focal length in the vertical direction. The horizontal coordinates of the optical axis and the image plane. represents the perpendicular coordinates of the optical axis to the image plane.

[0063] Using the above conversion process, the positions of each vertex of the triangular facet in the three-dimensional mesh model in the two-dimensional image plane are obtained. Then, the thermal infrared elements corresponding to each facet can be extracted, all thermal infrared elements can be stored in the texture atlas, and the texture atlas can be fused with the three-dimensional mesh model to obtain the three-dimensional temperature field model.

[0064] It should be noted that any parts not disclosed or specifically described in this invention are existing technologies or conventional configurations, and their specific structures and working principles will not be elaborated further. In this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0065] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily implemented by those skilled in the art. Therefore, the present invention is not limited to the specific details without departing from the general concept defined by the claims and their equivalents.

Claims

1. A method for rapidly generating a three-dimensional temperature field model based on multi-source data from unmanned aerial vehicles (UAVs), characterized in that, include: By using a drone equipped with a lidar camera and a thermal infrared imager, laser point cloud datasets and thermal infrared image data are collected simultaneously in the target area. The laser point cloud dataset is divided into several blocks using the octree partitioning method. The local density of each point is obtained, and the block density, block density standard deviation, and global density are obtained based on the local density of each point to obtain the comprehensive index of the block. Based on the comprehensive index, the blocks are divided into five levels: extremely sparse, sparse, medium, dense, and extremely dense. Density equalization filtering and voxel filtering are performed on the dense and extremely dense blocks to obtain the target number of point clouds. Then, proportional control filtering is performed on the blocks of the five levels to ensure that the number of point clouds in each block is within the target range, resulting in a simplified point cloud dataset. Obtain the curvature of each point cloud data in the laser point cloud dataset, remove point clouds with curvature less than the threshold, and obtain the feature point cloud dataset; The simplified point cloud dataset and the feature point cloud dataset are merged, and a hash table is constructed to map voxels. Duplicate points are removed to obtain the target point cloud dataset. Based on the point cloud position and normal information in the target point cloud dataset, Poisson reconstruction is performed to obtain the three-dimensional mesh model of the target region. The laser point cloud dataset and thermal infrared image data are spatially aligned. Then, the thermal infrared image data aligned with the laser point cloud dataset is mapped onto the surface of the 3D mesh model through coordinate transformation and perspective projection to generate a 3D temperature field model with temperature information.

2. The method for rapid generation of a three-dimensional temperature field model based on multi-source UAV data according to claim 1, characterized in that, Before dividing the laser point cloud dataset into several blocks using the octree block method, the dataset is preprocessed to remove outliers and noise points.

3. The method for rapidly generating a three-dimensional temperature field model based on multi-source UAV data according to claim 1, characterized in that, The method for obtaining comprehensive indicators is as follows: Obtain each point cloud in the laser point cloud dataset Local density The formula is: Get block density Standard deviation of block density and global density The formula is: Obtain comprehensive indicators The formula is: In the formula, K is a point The number of nearest neighbors, It is a constant, with a value of 1e. -6 , ||·|| is the Euclidean distance. For point To all neighboring points The average distance, n is the number of point clouds in the block, N is the total number of point clouds in the laser point cloud dataset, α is an adjustment coefficient and α>0, generally taken as the standard deviation of the density of the entire block. The reciprocal of the average.

4. The method for rapidly generating a three-dimensional temperature field model based on multi-source UAV data according to claim 3, characterized in that, Based on the comprehensive index r, the levels are divided as follows: r < 0.1 is extremely sparse, 0.1 ≤ r < 0.5 is sparse, 0.5 ≤ r < 2.0 is medium, 2.0 ≤ r < 10.0 is dense, and r ≥ 10.0 is extremely dense.

5. The method for rapidly generating a three-dimensional temperature field model based on multi-source UAV data according to claim 1, characterized in that, The method for obtaining the simplified point cloud dataset is as follows: Obtain the number of point cloud targets after dense and extremely dense block equalization filtering. The formula is: Number of point cloud targets Voxel filtering is performed to obtain the number of target point clouds. ; The number of point cloud targets is obtained by performing proportional control filtering on all blocks. The formula is: In the formula, This represents the number of original point clouds within the block. The target density after density equalization filtering. The value is set to 0.3 when the comprehensive index r ≥ 20. When 10 ≤ comprehensive index r < 20, take When 2.0 ≤ comprehensive index r < 10, take 5. t is the preset thinning ratio; Number of target point clouds and the number of point cloud targets In comparison, if the target point cloud quantity Number of point cloud targets Within the error range, the number of target point clouds is output. If the target point cloud number Not in point cloud target quantity Within the error range, then the number of target point clouds is further adjusted. Perform voxel filtering until the target point cloud number is reached. Number of point cloud targets Within the error range; The total number of target point clouds in all blocks This constitutes a simplified point cloud dataset.

6. The method for rapidly generating a three-dimensional temperature field model based on multi-source UAV data according to claim 1, characterized in that, The method for obtaining the feature point cloud dataset is as follows: Select the k nearest neighbors of any point P in the laser point cloud dataset and construct the covariance matrix. : In the formula, Let the point be the centroid of the neighborhood of that point. It is the j-th neighbor point; For covariance matrix Eigenvalue decomposition yields three eigenvalues. , , ,and ≤ ≤ , corresponding feature vector Let P be the direction of the normal to point P; Based on three eigenvalues , , Obtaining curvature The formula is: curvature By comparing with a threshold, point clouds with curvature less than the threshold are removed, resulting in point clouds with curvature greater than or equal to the threshold, which is the feature point cloud dataset.

7. The method for rapidly generating a three-dimensional temperature field model based on multi-source UAV data according to claim 1, characterized in that, The method for constructing the three-dimensional mesh model is as follows: The set of all points in the target point cloud dataset S is {s1, s2, ..., s}. N }, where N is the number of laser points, and a closed geometric body M is constructed in the point set space, with an indicator function applied to all points in the point set space. definition: For any point s in the laser point cloud i Let i be a natural number less than N, and s i In geometry If the value of the indicator function is 1 inside the point set S, and 0 outside the point set S, then the domain of x is the point set S. The gradient field ∇χ of the indicator function is zero both inside and outside the conceived geometry M, but a huge pulse is formed on the surface of M, with the pulse direction perpendicular to the surface of geometry M and pointing outwards; All points s in the laser point cloud i and its normal vector s i . This can be viewed as a sparse sampling of the gradient field ∇χ of the indicator function on the surface of the target geometry M, thereby inferring the original indicator function, specifically: Take any point q in the point set space, and construct a smooth and continuous vector field by weighted sum of the normals of all points in the target point cloud dataset S. ,Right now: In the formula, For a smooth convolution function, express Displacement along the p direction; Vector Field For the ideal gradient field ∇χ, the approximation is: ∇χ= Differentiating both sides of the above equation, we have: Δχ=∇· In the formula, Δ is the Laplace operator, Δχ=∇·(∇χ) is the divergence of the gradient, and ∇· For vector fields The divergence; Therefore, the known vector field can be calculated. Then the indicator function is obtained. This allows us to obtain the surface of the target geometry M.

8. The method for rapidly generating a three-dimensional temperature field model based on multi-source UAV data according to claim 1, characterized in that, The method for spatial alignment is as follows: The external parameters of the lidar camera and the thermal infrared imager are jointly calibrated, and a calibration matrix is ​​constructed based on the rigid transformation relationship obtained from the calibration. , In the formula It is a 3×3 rotation matrix. It is a 3×1 translation vector; The calibration matrix is ​​used to perform spatial position transformation on the thermal infrared image data, achieving spatial alignment between the laser point cloud data and the thermal infrared image data. The formula is as follows: In the formula, This represents the pose of the lidar in the world coordinate system. This represents the pose of the calibrated infrared camera in the world coordinate system.

9. The method for rapidly generating a three-dimensional temperature field model based on multi-source UAV data according to claim 8, characterized in that, The method for constructing the three-dimensional temperature field model is as follows: For any vertex P of a triangular facet in a 3D mesh model, the world coordinates are: , Convert it to the world coordinate system coordinates of the thermal infrared imager using the calibration matrix. The formula is: Then, using perspective projection, the pixel position (u, v) of point P in the two-dimensional image plane is obtained, using the formula: In the formula, The physical dimension of the camera's focal length in the horizontal direction. The physical dimension of the camera's focal length in the vertical direction. The horizontal coordinates of the optical axis and the image plane. The perpendicular coordinates of the optical axis to the image plane; Using the above conversion process, the positions of each vertex of the triangular facet in the three-dimensional mesh model in the two-dimensional image plane are obtained. Then, the thermal infrared elements corresponding to each facet can be extracted, all thermal infrared elements can be stored in the texture atlas, and the texture atlas can be fused with the three-dimensional mesh model to obtain the three-dimensional temperature field model.

Images