Unmanned aerial vehicle point cloud fusion refined modeling method for slope structural surface

Abstract

The invention relates to the technical field of slope engineering modeling, and discloses a slope structural surface unmanned aerial vehicle point cloud fusion refined modeling method, which comprises the following steps: acquiring slope global point cloud data and optical images through an unmanned aerial vehicle carried high-precision positioning system, combining local point cloud data acquired by a handheld laser scanner, calibrating and pre-processing, and obtaining a slope global point cloud model; a point cloud normal vector is determined based on a local plane fitting technology, structural boundary information is extracted through a normal vector difference boundary detection algorithm, structural plane clustering segmentation is completed by adopting a region growing algorithm, finally, a triangular mesh model is generated, texture mapping and hierarchical simplification are performed, and refined modeling of a slope structural plane is realized. According to the method, the modeling precision and efficiency are improved, and reliable model support is provided for stability analysis and risk assessment of slope engineering.

Description

Technical Field

[0001] This invention relates to the field of slope engineering modeling technology, and in particular to a method for refined modeling of slope structural surfaces using UAV point cloud fusion. Background Technology

[0002] The geometric morphology of slope structural surfaces is a key factor affecting slope stability. Accurately obtaining a three-dimensional model of the slope structural surfaces is of great significance for the design, construction, and risk assessment of slope engineering. Traditional slope modeling methods often use contact measuring equipment such as total stations and levels, which are not only inefficient and labor-intensive, but also difficult to guarantee measurement accuracy in slope areas with complex terrain or difficult access.

[0003] With the development of drone technology and laser scanning technology, drone point cloud measurement technology has been widely used in slope modeling due to its advantages such as high efficiency, non-contact operation, and wide coverage. However, single drone point cloud data suffers from insufficient accuracy in local details, making it difficult to accurately reflect the subtle features of slope structures. While handheld laser scanners can acquire high-precision local point cloud data, their measurement range is limited, preventing rapid scanning of the entire slope area. Furthermore, in the point cloud data processing, effectively removing noise, achieving accurate point cloud matching and simplification, and accurately segmenting structural surfaces based on point cloud data to generate high-quality 3D models remain the main challenges facing current slope modeling technology.

[0004] Therefore, there is an urgent need for a method that can integrate the advantages of UAV global point cloud data and handheld laser scanning local point cloud data to achieve refined modeling of slope structure surfaces, in order to solve the problems of difficulty in balancing measurement accuracy and efficiency and insufficient model detail expression in existing technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a method for refined modeling of slope structure surfaces using UAV point cloud fusion, which aims to solve one or more of the problems mentioned above.

[0006] This invention provides a method for refined modeling of slope structure surfaces using UAV point cloud fusion, including: By using a drone equipped with a high-precision positioning system, the slope area is scanned according to a preset route to obtain full-area point cloud data and optical images of the slope area. A handheld laser scanner is then used to perform a detailed scan of the slope area to obtain local point cloud data of the slope area. The global point cloud data and local point cloud data are calibrated to obtain point cloud data of the slope area, and the point cloud data is preprocessed, including noise reduction, point cloud matching and point cloud simplification. Based on local plane fitting technology, the normal vector of each point in the point cloud data is determined, and based on the normal vector difference boundary detection algorithm, the structural boundary information of the slope area is determined. Based on the structural boundary information, the structural surface of the slope area is clustered and segmented using a region growing algorithm. Based on the point cloud data of the segmented structural surface, a triangular mesh model is generated. The optical image is then texture-mapped to the triangular mesh model to establish a pixel-level correspondence between the optical image and the triangular mesh model. The mapped triangular mesh model is then hierarchically simplified to achieve the modeling of the slope structural surface.

[0007] Preferably, the global point cloud data and the local point cloud data are calibrated to obtain the point cloud data of the slope area, specifically as follows: The high-precision positioning system includes an optical camera and a lidar; A standard target calibration range is used, and a reflective target array is employed to calculate the installation attitude and ranging error of the lidar and handheld laser scanner based on the known target position, thereby achieving calibration.

[0008] Preferably, when performing noise reduction processing, a neighborhood radius is set, the point density of each point in the neighborhood is calculated, and if the number of points of a certain point is lower than the point density threshold, the corresponding point is determined to be a noise point and deleted; wherein, the neighborhood radius is 0.5 meters.

[0009] Preferably, when performing point cloud matching processing, the iterative nearest point algorithm is adopted, an initial corresponding point set is selected, no less than 50 pairs of point cloud feature points are randomly sampled, the maximum step size of each iteration is limited to 0.1 meters, the convergence error tolerance is set to 1 millimeter, and a global optimization is performed every 5 iterations.

[0010] Preferably, when performing point cloud simplification, the voxel mesh method is used to simplify the point cloud. The size of the voxel mesh is set to 0.2m × 0.2m × 0.2m, and the point cloud space is divided into uniform cubic units. Only one representative point is retained in each voxel. Based on the curvature criterion, the curvature is calculated using a quadratic surface fitting method. The curvature threshold is divided into three levels: high, medium, and low, each corresponding to a different simplification ratio, in order to achieve point cloud simplification.

[0011] Preferably, based on local plane fitting technology, the normal vector of each point in the point cloud data is determined as follows: taking the target point as the center, a set of neighboring points within a certain radius is selected. The initial value of the neighborhood radius of the local plane fitting is 1 meter, which decreases exponentially as the curvature increases, and is reduced to a minimum of 0.3 meters; a required plane is fitted by the least squares method, and the normal vector of the plane is set as the normal vector of the target point.

[0012] Preferably, based on the structural boundary information, a region growing algorithm is used to cluster and segment the structural surfaces of the slope area. Specifically, based on the structural boundary information and combined with the spatial connectivity and similarity of the point cloud, a region growing algorithm is used to cluster and segment the structural surfaces. Starting from the seed point, according to the preset growth criteria, neighboring points that meet the conditions are gradually included in the same structural surface area. During the growth process, the statistical characteristics of the area are recalculated every 20 newly added points, and the statistical characteristics of the area are updated in real time to correct the growth direction. The distance threshold of the growth criteria is 0.3 meters, and the maximum angular deviation of the normal vector is 15 degrees.

[0013] Preferably, a triangular mesh model is generated based on the point cloud data of the segmented structural surfaces. Specifically, a triangular mesh model is generated using the Delaunay triangulation algorithm based on the point cloud data of the segmented structural surfaces. The constraints of the Delaunay triangulation algorithm include the empty circle property and the side length restriction, wherein the length of the longest side shall not exceed three times that of the shortest side.

[0014] Preferably, the optical image is texture-mapped to the triangular mesh model to establish a pixel-level correspondence between the optical image and the triangular mesh model. Specifically, the texture coordinates of the optical image are back-projected onto the surface of the triangular mesh model, and Poisson editing or Laplacian weighted image processing technology is used to seamlessly fuse and correct the texture to eliminate seams and blurring. The accuracy of solving the Laplacian equation in Poisson editing is controlled within 0.001, and the Laplacian weight coefficient is set to 1.

[0015] Preferably, the mapped triangular mesh model is simplified hierarchically, specifically by using an edge folding algorithm to determine the folding priority based on the edge length, the difference in normal vectors between the two endpoints, and the area change of adjacent triangles, and setting the weight allocation to 40%, 30%, and 30% respectively, and merging redundant triangles. Vertex caching technology is used to organize the vertex storage order. The cache block size is set to 16 bytes. The LRU replacement strategy is used to manage the cache content and invalid data is cleaned up periodically to improve the cache hit rate during graphics rendering and accelerate model loading and display speed.

[0016] Compared with existing technologies, the advantages of this invention lie in its integration of the wide-area coverage advantage of UAV global point cloud data and the high-precision advantage of handheld laser scanning local point cloud data. Through precise calibration, the two types of data are organically combined, ensuring both modeling efficiency and improving the accuracy of local details. In the data preprocessing process, targeted denoising, matching, and simplification algorithms are employed to effectively remove point cloud noise, achieving accurate matching and reasonable simplification of point cloud data, laying a solid foundation for subsequent structural surface segmentation and model construction. Based on local plane fitting technology and normal vector difference boundary detection algorithm, slope structure boundary information can be accurately extracted. Combined with an improved region growing algorithm, accurate clustering and segmentation of structural surfaces are achieved, accurately reflecting the geometric morphology of the slope structural surfaces. A uniform and reasonable triangular mesh model is generated using the Delaunay triangulation algorithm, and texture mapping technology achieves a realistic representation of the model. Simultaneously, a hierarchical simplification algorithm optimizes the model structure while ensuring model accuracy, improving loading and display speed and meeting the needs of slope engineering analysis and application. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the method for refining the modeling of slope structure surfaces using UAV point cloud fusion according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0020] like Figure 1 As shown, this invention provides a method for refined modeling of slope structure surfaces using UAV point cloud fusion, including: By using a drone equipped with a high-precision positioning system, the slope area is scanned according to a preset route to obtain full-area point cloud data and optical images of the slope area. A handheld laser scanner is then used to perform a detailed scan of the slope area to obtain local point cloud data of the slope area. The global point cloud data and local point cloud data are calibrated to obtain point cloud data of the slope area, and the point cloud data is preprocessed, including noise reduction, point cloud matching and point cloud simplification. Based on local plane fitting technology, the normal vector of each point in the point cloud data is determined, and based on the normal vector difference boundary detection algorithm, the structural boundary information of the slope area is determined. Based on the structural boundary information, the structural surface of the slope area is clustered and segmented using a region growing algorithm. Based on the point cloud data of the segmented structural surface, a triangular mesh model is generated. The optical image is then texture-mapped to the triangular mesh model to establish a pixel-level correspondence between the optical image and the triangular mesh model. The mapped triangular mesh model is then hierarchically simplified to achieve the modeling of the slope structural surface.

[0021] This invention achieves refined modeling of slope structural surfaces. On one hand, by acquiring global point cloud data and optical images using a high-precision positioning system mounted on a UAV, combined with local point cloud data acquired by a handheld laser scanner, it can simultaneously capture both the overall information and local details of the slope area, making the acquired data more comprehensive and accurate, providing a high-quality data foundation for subsequent modeling. On the other hand, the global and local point cloud data are calibrated and preprocessed to remove noise, improve the matching degree and simplification of the point cloud data, and effectively enhance the efficiency and quality of data processing. When determining structural boundary information and performing structural surface clustering and segmentation, local plane fitting technology and normal vector difference boundary detection algorithms are used to accurately identify the structural boundaries of the slope area. Then, a region growing algorithm is used to perform reasonable clustering and segmentation of the structural surfaces, enabling the modeling results to more realistically reflect the actual structural condition of the slope. After generating a triangular mesh model and performing texture mapping, the mapped triangular mesh model is hierarchically simplified, which ensures the visualization effect of the model while reducing the data volume of the model, facilitating storage and processing. This modeling method has significant application value in engineering practice, providing reliable model basis for slope stability analysis and geological hazard assessment, which helps improve the scientificity and accuracy of engineering decisions and reduce safety hazards and economic losses caused by slope problems.

[0022] In this embodiment, Unmanned aerial vehicle (UAV) system: Select a multi-rotor UAV (such as DJI M300RTK) equipped with a high-precision positioning system (including a 1-inch optical camera and a 16-line LiDAR), with a LiDAR ranging accuracy of ±2mm@100m and an optical camera with more than 20 million pixels. Handheld laser scanner: A phase-detection scanner (such as the Faro Focus S70) is used, with a scanning accuracy of ±0.1mm@1m and adjustable point cloud density (up to 1000 points / cm²). 2 ).

[0023] Regarding drone scanning: Flight route planning: A zigzag flight path will be adopted, with a forward overlap of 80% and a lateral overlap of 60%. The flight altitude will be determined based on the slope height (e.g., 50-150m) to ensure that the lidar point cloud density is ≥50 points / m². 2 ; Data acquisition: Simultaneously record the full-domain point cloud of the lidar (format .las) and optical images (format .jpg, resolution 3840×2160), and calibrate the IMU (inertial measurement unit) every 10 minutes of flight to reduce drift error.

[0024] Regarding handheld laser scanning: Region selection: Mark the scanning range for key local areas such as slope joint development zones, slope toe fracture zones, and areas with dense anchor bolts and cables; Scanning operation: Arrange 3-5 reflective targets (accuracy ±0.05mm) around the scanning area as coordinate transformation references. Scan each local area 2-3 times (at different angles) to avoid occlusion.

[0025] In some embodiments of this application, the global point cloud data and local point cloud data are calibrated to obtain point cloud data for the slope area. Specifically, the high-precision positioning system includes an optical camera and a lidar; a standard target calibration field is adopted, and a reflective target array is used to calculate the installation attitude and ranging error of the lidar and handheld laser scanner based on the known target position to achieve calibration.

[0026] Specifically, before data acquisition, the optical camera and lidar sensors mounted on the UAV need to undergo rigorous intrinsic and extrinsic parameter calibration to ensure precise spatial matching of the data collected by the two sensors. During flight, the optical image acquisition and lidar point cloud acquisition are synchronized through hardware triggering or software timestamp marking to ensure the consistency of their time series during subsequent fusion processing, so that each frame of image can accurately correspond to the corresponding point cloud slice.

[0027] Specifically, a standard target calibration field is set up: 10 reflective targets are arranged in a gentle slope area, and the three-dimensional coordinates of the targets (geodetic coordinate system, such as CGCS2000) are measured using a total station (accuracy ±0.5mm). Error back calculation: Compare the target coordinates in the UAV lidar point cloud and handheld lidar point cloud with the total station measured coordinates, establish error equations, and back calculate the installation attitude (roll / pitch / yaw, accuracy ±0.1°) and ranging error (±1-2mm) of the two types of equipment. Coordinate unification: Based on the inverse calculation error, the local point cloud is rotated and translated to unify the global and local point clouds to the geodetic coordinate system. After calibration, the overlap between the two types of point clouds is ≥98%.

[0028] Understandably, this calibration method can significantly improve the accuracy and consistency of point cloud data. Utilizing a standard target calibration field and reflective target array, with known target positions as a benchmark, the installation posture and ranging errors of the lidar and handheld laser scanner are calculated, effectively eliminating deviations caused by equipment installation and measurement processes. This allows for precise matching of global and local point cloud data, laying a solid foundation for subsequent point cloud data preprocessing and modeling. Accurately calibrated point cloud data achieves better results in preprocessing stages such as noise reduction, point cloud matching, and point cloud simplification. Because the inherent errors in the data are effectively controlled, outliers can be more accurately identified and removed during noise removal, corresponding points can be found more accurately during point cloud matching, and key information can be better preserved during point cloud simplification. Clearly defining the high-precision positioning system as including an optical camera and lidar during calibration helps to comprehensively utilize the advantages of both devices. Optical cameras can acquire rich optical image information, while lidar provides high-precision 3D point cloud data; their combination can provide more comprehensive and accurate data support for slope structure modeling.

[0029] In some embodiments of this application, when performing noise reduction processing, a neighborhood radius is set, the point density of each point in the neighborhood is calculated, and if the number of points of a certain point is lower than the point density threshold, the corresponding point is determined to be a noise point and deleted; wherein, the neighborhood radius is 0.5 meters.

[0030] Specifically, the algorithm principle of noise reduction is: statistical filtering based on neighborhood point density to remove isolated noise points (such as birds, leaves, and measurement error points). Parameter settings: Neighborhood radius: fixed at 0.5m (verified through multiple experiments, this radius can effectively distinguish between noise and valid points); Density threshold: count the number of points in the 0.5m neighborhood of each point, and set the threshold to 20 (i.e., points with less than 20 points in the neighborhood are judged as noise).

[0031] Understandably, this denoising method effectively removes noisy points from point cloud data, significantly improving data quality and usability. By setting specific neighborhood radii and point density thresholds, noisy points can be identified scientifically and objectively. A neighborhood radius of 0.5 meters is a suitable value verified in practice, balancing recognition accuracy with processing efficiency. When the number of points in a neighborhood is below the point density threshold, it is identified as a noise point and deleted, thus preventing noise points from interfering with subsequent modeling. After removing noise points, the point cloud data is cleaner and more accurately reflects the true morphology of the slope structure. In subsequent point cloud matching, the cleaner data makes the matching of corresponding points more accurate, reducing matching errors and increasing the success rate. During point cloud simplification, it also better preserves key feature information of the slope structure, avoiding the accidental deletion of crucial information due to the presence of noise points.

[0032] In some embodiments of this application, when performing point cloud matching processing, an iterative nearest point algorithm is used, an initial corresponding point set is selected, no less than 50 pairs of point cloud feature points are randomly sampled, the maximum step size of each iteration is limited to 0.1 meters, the convergence error tolerance is set to 1 millimeter, and a global optimization is performed every 5 iterations.

[0033] Specifically, the algorithm selection for point cloud matching processing is as follows: the improved Iterative Closest Point (ICP) algorithm is adopted to solve the problems of slow convergence and easy getting trapped in local optima in the traditional ICP algorithm; Key steps: Initial registration: The SIFT (Scale Invariant Feature Transform) algorithm is used to extract feature points (such as corner points and edge points) from the global and local point clouds, and 50-80 pairs of initial feature points are matched to reduce initial bias; Iterative optimization: Step size control: The maximum step size for each iteration is 0.1m to avoid over-adjustment; Error control: Convergence error tolerance is 1mm (i.e., stop when the average distance between the point clouds in two iterations is <1mm); Global optimization: A global optimal solution search (based on particle swarm optimization) is performed every 5 iterations to prevent local optima; Matching accuracy: The average matching error of the final point cloud is <0.8mm, which meets the requirements of fine modeling.

[0034] Understandably, employing the iterative nearest-point algorithm for point cloud matching effectively achieves precise registration of point cloud data. Randomly sampling at least 50 pairs of point cloud feature points provides the algorithm with a rich and representative data foundation, enabling the matching process to more comprehensively capture the correspondences between point clouds and improve matching accuracy. The maximum step size for each iteration is limited to 0.1 meters; this setting prevents excessive jumps during iteration, ensuring the stability and convergence of the iteration. A convergence error tolerance of 1 millimeter clarifies the accuracy requirements of the matching, ensuring that the final matching result meets the high-precision needs of actual modeling. A global optimization is performed every 5 iterations, helping to promptly correct potential deviations in local matching and optimize the matching result globally. This approach effectively reduces accumulated errors during the matching process, improving the overall quality of point cloud matching. In subsequent refined modeling, accurate point cloud matching results provide accurate basic data for model construction, allowing the model to more realistically reflect the actual condition of the slope structure.

[0035] In some embodiments of this application, when performing point cloud simplification, a voxel mesh method is used to simplify the point cloud. The size of the voxel mesh is set to 0.2 m × 0.2 m × 0.2 m, and the point cloud space is divided into uniform cubic units. Only one representative point is retained in each voxel. Combined with the curvature criterion, the curvature is calculated using a quadratic surface fitting method. The curvature threshold is divided into three levels: high, medium, and low, which correspond to different simplification ratios to achieve point cloud simplification.

[0036] Specifically, point cloud simplification includes: a two-step simplification method: Initial simplification of the voxel mesh: Voxel size: 0.2m × 0.2m × 0.2m (Testing has shown that this size can reduce the amount of data by 60% while preserving structural features). Operation: Divide the point cloud space into uniform cubes, calculate the centroid of all points within each voxel, and retain the centroid as a representative point; Secondary simplification of curvature criterion: Using quadratic surface fitting (such as a parabola), the principal curvatures (κ1, κ2) are calculated for each point, with the curvature value κ = √(κ1 + κ2). 2 +κ2 2 ); and threshold grading is performed based on curvature.

[0037] Understandably, using the voxel mesh method for point cloud simplification, dividing the point cloud space into uniform cubic units and retaining only one representative point within each voxel, can effectively reduce the amount of point cloud data, lower the complexity of data processing, and reduce computational costs. Setting the voxel mesh size to 0.2m × 0.2m × 0.2m ensures simplification while preserving the overall characteristics and spatial distribution information of the point cloud data, avoiding the loss of important structural details due to excessive simplification. Combining curvature criteria and using a quadratic surface fitting method to calculate curvature allows for targeted simplification based on the local geometric features of the point cloud. Dividing the curvature threshold into high, medium, and low levels, each corresponding to a different simplification ratio, allows for the retention of more points in areas with large curvature changes (such as the edges of slopes, protrusions, or depressions) to accurately capture complex geometries; while appropriately reducing the number of points in flat areas with smaller curvature changes, achieving maximum simplification of point cloud data while maintaining model accuracy.

[0038] In some embodiments of this application, the normal vector of each point in the point cloud data is determined based on the local plane fitting technique. Specifically, a set of neighboring points within a certain radius is selected with the target point as the center. The initial value of the neighborhood radius of the local plane fitting is 1 meter, which decreases exponentially as the curvature increases, and is reduced to a minimum of 0.3 meters. A required plane is fitted by the least squares method, and the normal vector of the plane is set as the normal vector of the target point.

[0039] Understandably, determining the normal vector of each point in point cloud data based on local plane fitting is an effective method. Selecting a neighborhood point set centered on the target point, and dynamically adjusting the neighborhood radius according to curvature, offers several advantages. Setting the initial value to 1 meter ensures sufficient neighborhood information within a certain range, guaranteeing the accuracy of normal vector calculation. The strategy of decreasing the radius exponentially to a minimum of 0.3 meters as curvature increases is quite ingenious. In areas with high curvature, i.e., where slope surface changes drastically, a smaller neighborhood radius allows the normal vector to more accurately reflect local geometric features, as the local changes are more complex and require more refined calculations. In flat areas with low curvature, a larger neighborhood radius can integrate more point information, making the normal vector more representative of the overall direction of the area. Fitting the required plane using the least squares method and using its normal vector as the normal vector of the target point ensures the stability and reliability of the normal vector calculation. The least squares method is a mature and effective mathematical method that can find the plane that best fits the data distribution in point cloud data with certain errors and noise, thereby obtaining accurate normal vectors.

[0040] In some embodiments of this application, a region growing algorithm is used to cluster and segment the structural surfaces of the slope area based on structural boundary information. Specifically, based on structural boundary information and combined with the spatial connectivity and similarity of the point cloud, a region growing algorithm is used to cluster and segment the structural surfaces. Starting from the seed point, neighboring points that meet the conditions are gradually included in the same structural surface area according to a preset growth criterion. During the growth process, the statistical characteristics of the area are recalculated every 20 newly added points, and the statistical characteristics of the area are updated in real time to correct the growth direction. The distance threshold of the growth criterion is 0.3 meters, and the maximum angular deviation of the normal vector is 15 degrees.

[0041] It is understandable that using a region growing algorithm combined with structural boundary information to cluster and segment the structural surfaces of slope areas is a scientific and efficient method. Based on structural boundary information, and considering the spatial connectivity and similarity of the point cloud, the clustering and segmentation can better reflect the actual structure of the slope. Starting from the seed point and gradually expanding the region according to preset growth criteria, this approach is highly logical and feasible. The distance threshold of 0.3 meters and the maximum deviation of the normal vector angle of 15 degrees in the growth criteria are carefully considered. The distance threshold of 0.3 meters reasonably defines which points should be included in the same structural surface region, avoiding incorrectly classifying points that are too far away, thus ensuring the accuracy of structural surface region segmentation. The maximum deviation of the normal vector angle of 15 degrees further constrains the point assignment from a geometric perspective, making the points within the same structural surface region have high consistency in the normal vector direction, better reflecting the true characteristics of the structural surface. During the growth process, the statistical characteristics of the region are recalculated and the growth direction is updated and corrected in real time every 20 newly added points. This strategy greatly improves the accuracy of clustering and segmentation. As the region expands, the distribution and characteristics of points will change. Timely updates to statistical properties allow the algorithm to better adapt to these changes, avoid deviations in the growth direction, and make the final segmented structural surface regions more accurate and reasonable.

[0042] Specifically, the process of calculating the normal vector (reflecting the orientation of the point cloud surface) is as follows: Adaptive adjustment of neighborhood radius: Initial radius: 1m (suitable for flat areas); Dynamic adjustment: If the curvature κ of a certain point is greater than 0.31 / m, the neighborhood radius decreases according to the formula r=1×e^(-2κ), and is reduced to a minimum of 0.3m (to ensure that the neighborhood point set of the high curvature region fits the local surface). Plane fitting: For the neighborhood point set of each point, the least squares method is used to fit the plane equation ax+by+cz+d=0, and the plane normal vector is (a,b,c). After normalization (magnitude=1), the normal vector calculation accuracy is ±0.5°.

[0043] The specific steps for clustering and segmenting the structural surfaces of the slope area include: Structural boundary detection: Algorithm principle: The normal vector of the point cloud at the structural boundary will change abruptly. By calculating the angle between the normal vectors of adjacent points (distance < 0.2m), points with an angle > 15° are determined to be boundary points; Boundary extraction: Connect the boundary points to form a closed structural boundary polygon, which is marked as the "constraint line" for structural surface segmentation.

[0044] Region Growing Algorithm Clustering Segmentation: Seed Point Selection: Randomly select seed points (every 100m) in non-boundary regions. 2 Choose one), requiring the variance of the seed point normal vector to be <5° (ensure that the region where the seed point is located is the same structural plane); Growth guidelines: Distance constraint: The Euclidean distance between the point to be grown and the current region point is < 0.3m; Normal vector constraint: The angle between the growth point and the average normal vector of the current region is <15°; Dynamic optimization: For every 20 new points, the average normal vector and centroid coordinates of the current region are recalculated, the growth direction is corrected, and growth across structural surfaces is avoided; Segmentation results: Output a subset of point clouds for each structural plane (format .ply), labeled with the structural plane number, orientation, and dip angle (error < 0.5°).

[0045] In some embodiments of this application, a triangular mesh model is generated based on the point cloud data of the segmented structural surfaces. Specifically, the triangular mesh model is generated using the Delaunay triangulation algorithm based on the point cloud data of the segmented structural surfaces. The constraints of the Delaunay triangulation algorithm include the empty circle property and the side length restriction, wherein the length of the longest side shall not exceed three times that of the shortest side.

[0046] Specifically, the algorithm chosen for generating the triangular mesh model is the Delaunay triangulation algorithm, whose "empty circle characteristic" can avoid generating elongated triangles and ensure mesh uniformity. Constraints: Property of an empty circle: The circumcircle of any triangle does not contain any other points; Side length restriction: The length of the longest side ≤ 3 × the length of the shortest side (to avoid triangle distortion); Mesh quality: The average side length of the generated triangular mesh is 0.1-0.3m, and the proportion of triangles with interior angles between 30° and 120° is ≥90%.

[0047] Understandably, using the Delaunay triangulation algorithm combined with specific constraints to generate triangular mesh models can effectively improve the quality and practicality of the model. The empty circle property is one of the core advantages of the Delaunay triangulation algorithm, ensuring that the circumcircle of any triangle in the generated mesh does not contain any other points. This results in a well-defined geometric shape and topological structure, avoiding overly long or distorted triangles, thus improving the model's stability and reliability. The side length constraint further optimizes the quality of the triangular mesh. Specifying that the longest side cannot exceed three times the shortest side ensures a more uniform distribution of triangle side lengths, guaranteeing that each triangle has a relatively regular shape. Such a mesh structure exhibits better convergence and accuracy in numerical calculations and analysis, contributing to improved accuracy in subsequent processing. Generating high-quality triangular mesh models is crucial for the refined modeling of slope structure surfaces using UAV point cloud fusion. Regular and stable triangular meshes can more accurately represent the geometry of the structural surface, providing a more reliable foundation for subsequent model analysis and processing. When performing stress analysis, deformation simulation, and other tasks, high-quality triangular meshes can more accurately simulate the mechanical behavior of structural surfaces, making the analysis results closer to reality.

[0048] In some embodiments of this application, the optical image is texture-mapped to the triangular mesh model to establish a pixel-level correspondence between the optical image and the triangular mesh model. Specifically, the texture coordinates of the optical image are back-projected onto the surface of the triangular mesh model, and Poisson editing or Laplacian weighted image processing technology is used to seamlessly fuse and correct the texture to eliminate seams and blurring. The accuracy of solving the Laplacian equation in Poisson editing is controlled within 0.001, and the Laplacian weight coefficient is set to 1.

[0049] Specifically, texture mapping (enhancing the realism of the model). Texture coordinate projection: Camera parameter calibration: Using Zhang Zhengyou's calibration method, obtain the intrinsic parameters (focal length, principal point coordinates) and extrinsic parameters (attitude, position) of the optical camera; Back projection: The pixel coordinates of the optical image are back projected onto the surface of the triangular mesh using camera parameters to establish the correspondence between pixels and mesh vertices; Texture blending correction: Poisson editing: For overlapping texture regions, solve the Laplace equation (accuracy 0.001) to smoothly transition texture gray values ​​and eliminate seams; Laplacian weights: When using the Laplacian weights technique, set the weight coefficient to 1 to ensure that texture details are not lost; Mapping effect: Texture resolution ≥ 10 pixels / cm, no obvious seams or blurring, and clear identification of structural surface texture and crack width.

[0050] Understandably, mapping optical images to triangular mesh models and establishing pixel-level correspondences significantly enhances the realism and visualization of the model. By projecting the texture coordinates of the optical images back onto the surface of the triangular mesh model, real-world texture information can be accurately assigned to the model, making the actual appearance of the slope structure more intuitive. Using Poisson editing or Laplacian weighted image processing techniques for seamless texture fusion and correction eliminates seams and blurring at texture joints, resulting in natural and clear texture transitions on the model surface, as if the real slope were right before your eyes. The accuracy of solving the Laplacian equation in Poisson editing is controlled within 0.001, ensuring high precision in texture fusion and accurately presenting texture details, avoiding texture distortion caused by calculation errors. Setting the Laplacian weight coefficient to 1 reasonably balances the influence of each pixel during texture fusion, ensuring the fusion effect is neither too harsh nor too blurry, achieving the best visual effect.

[0051] In some embodiments of this application, the mapped triangular mesh model is simplified hierarchically, specifically by: employing an edge folding algorithm to determine the folding priority based on the edge length, the difference in normal vectors between the two endpoints, and the area change of adjacent triangles, and setting the weight allocation to 40%, 30%, and 30% respectively, and merging redundant triangles; employing vertex caching technology to organize the vertex storage order, setting the cache block size to 16 bytes, using an LRU replacement strategy to manage cache content, and periodically cleaning up invalid data to improve the cache hit rate during graphics rendering and accelerate model loading and display speed.

[0052] Specifically, model hierarchy simplification is used to improve loading speed.

[0053] Edge folding algorithm: Folding priority calculation: For each edge, calculate the priority P = 0.4 × L + 0.3 × N + 0.3 × A, where: L: Length of the edge (normalized to 0-1, the longer the edge, the higher the priority). N: Angle between the normal vectors of the two endpoints of the edge (normalized to 0-1, the larger the angle, the higher the priority). A: The area change of two adjacent triangles (normalized to 0-1, the smaller the change, the higher the priority). Redundancy merging: Fold edges from high to low priority, merge adjacent triangles until the number of model faces is reduced by 50%-70% (adjust according to the application scenario, such as when the number of faces is ≤1 million for real-time monitoring). Vertex caching optimization: Storage organization: Vertex coordinates, normal vectors, and texture coordinates are stored in 16-byte blocks (e.g., coordinates 3×4 bytes, normal vectors 3×4 bytes, texture coordinates 2×4 bytes). Cache management: The LRU (Least Recently Used) replacement strategy is adopted. When the cache is full, the least recently used block is deleted. Invalid vertices (not referenced by any triangles) are cleaned up periodically (every 10 seconds), and the cache hit rate is improved to over 85%. Simplified effect: Model loading time is reduced from 30 seconds to less than 5 seconds, and it can be rendered in real time on a regular PC (i7 processor, 16G memory) (frame rate ≥30fps).

[0054] Understandably, simplifying the mapped triangular mesh model hierarchically can effectively reduce model complexity and improve processing efficiency. Employing an edge-folding algorithm determines folding priority based on edge length, the difference in normal vectors between endpoints, and the area changes of adjacent triangles, and allocates weights appropriately, allowing for the targeted merging of redundant triangles. This reduces unnecessary details in the model while significantly reducing the model's data volume while preserving its basic features. For example, small triangles with minimal impact on the overall structure can be merged using this algorithm, resulting in a simpler model. Vertex caching further optimizes the model processing. Organizing vertex storage order and setting appropriate cache block sizes allows for more efficient access to vertex data during graphics rendering. Using an LRU replacement strategy to manage cache content and periodically clean up invalid data ensures that the most frequently used data is always stored in the cache, greatly improving cache hit rate. This means that when loading and displaying the model, the system can retrieve the required data from the cache more quickly, without frequent reads from disk or other storage devices, significantly accelerating model loading and display speed.

[0055] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0056] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0057] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0058] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for refined modeling of slope structure surfaces using UAV point cloud fusion, characterized in that, include: The drone is equipped with a high-precision positioning system to scan the slope area according to a preset route, and obtain the full-area point cloud data and optical images of the slope area. A handheld laser scanner is used to perform a fine scan of the slope area to obtain the local point cloud data of the slope area. The global point cloud data and local point cloud data are calibrated to obtain point cloud data of the slope area, and the point cloud data is preprocessed, including noise reduction, point cloud matching and point cloud simplification. Based on local plane fitting technology, the normal vector of each point in the point cloud data is determined, and based on the normal vector difference boundary detection algorithm, the structural boundary information of the slope area is determined. Based on the structural boundary information, the structural surface of the slope area is clustered and segmented using a region growing algorithm. Based on the point cloud data of the segmented structural surface, a triangular mesh model is generated. The optical image is then texture-mapped to the triangular mesh model to establish a pixel-level correspondence between the optical image and the triangular mesh model. The mapped triangular mesh model is then hierarchically simplified to achieve the modeling of the slope structural surface.

2. The method for refined modeling of slope structure surfaces using UAV point cloud fusion according to claim 1, characterized in that, The global point cloud data and local point cloud data are calibrated to obtain the point cloud data of the slope area, specifically: The high-precision positioning system includes an optical camera and a lidar; A standard target calibration range is used, and a reflective target array is employed to calculate the installation attitude and ranging error of the lidar and handheld laser scanner based on the known target position, thereby achieving calibration.

3. The method for refined modeling of slope structure surfaces using UAV point cloud fusion according to claim 1, characterized in that, During the denoising process, a neighborhood radius is set, and the point density of each point in the neighborhood is calculated. If the number of points of a certain point is lower than the point density threshold, the corresponding point is determined to be a noise point and deleted. The neighborhood radius is 0.5 meters.

4. The method for refined modeling of slope structure surfaces using UAV point cloud fusion according to claim 1, characterized in that, When performing point cloud matching, the iterative nearest point algorithm is adopted. An initial corresponding point set is selected, and no less than 50 pairs of point cloud feature points are randomly sampled. The maximum step size of each iteration is limited to 0.1 meters, the convergence error tolerance is set to 1 millimeter, and a global optimization is performed every 5 iterations.

5. The method for refined modeling of slope structure surfaces using UAV point cloud fusion according to claim 1, characterized in that, When performing point cloud simplification, the voxel mesh method is used to simplify the point cloud. The size of the voxel mesh is set to 0.2m × 0.2m × 0.2m, and the point cloud space is divided into uniform cubic units. Only one representative point is retained in each voxel. Based on the curvature criterion, the curvature is calculated using a quadratic surface fitting method. The curvature threshold is divided into three levels: high, medium, and low, each corresponding to a different simplification ratio, in order to achieve point cloud simplification.

6. The method for refined modeling of slope structure surfaces using UAV point cloud fusion according to claim 1, characterized in that, Based on the local plane fitting technique, the normal vector of each point in the point cloud data is determined. Specifically, with the target point as the center, a set of neighboring points within a certain radius is selected. The initial value of the neighborhood radius of the local plane fitting is 1 meter, which decreases exponentially as the curvature increases, and is reduced to a minimum of 0.3 meters. A required plane is fitted by the least squares method, and the normal vector of the plane is set as the normal vector of the target point.

7. The method for refined modeling of slope structure surfaces using UAV point cloud fusion according to claim 1, characterized in that, Based on the structural boundary information, a region growing algorithm is used to cluster and segment the structural surfaces of the slope area. Specifically, based on the structural boundary information and combined with the spatial connectivity and similarity of the point cloud, a region growing algorithm is used to cluster and segment the structural surfaces. Starting from the seed point, neighboring points that meet the conditions are gradually included into the same structural surface area according to the preset growth criteria. During the growth process, the statistical characteristics of the region are recalculated every 20 new points, and the statistical characteristics of the region are updated in real time to correct the growth direction. The distance threshold of the growth criterion is 0.3 meters, and the maximum angular deviation of the normal vector is 15 degrees.

8. The method for refined modeling of slope structure surfaces using UAV point cloud fusion according to claim 1, characterized in that, Based on the point cloud data of the segmented structural surfaces, a triangular mesh model is generated. Specifically, based on the point cloud data of the segmented structural surfaces, a triangular mesh model is generated using the Delaunay triangulation algorithm. The constraints of the Delaunay triangulation algorithm include the empty circle property and the side length restriction, wherein the length of the longest side shall not exceed three times that of the shortest side.

9. The method for refined modeling of slope structure surfaces using UAV point cloud fusion according to claim 1, characterized in that, The optical image is texture-mapped to the triangular mesh model to establish a pixel-level correspondence between the optical image and the triangular mesh model. Specifically, the texture coordinates of the optical image are back-projected onto the surface of the triangular mesh model, and Poisson editing or Laplacian weighted image processing technology is used to seamlessly fuse and correct the texture to eliminate seams and blurring. The accuracy of the Laplace equation solution edited by Poisson is controlled within 0.001, and the Laplace weight coefficient is set to 1.

10. The method for refined modeling of slope structure surfaces using UAV point cloud fusion according to claim 1, characterized in that, The mapped triangular mesh model is simplified hierarchically. Specifically, an edge folding algorithm is used to determine the folding priority based on the length of the edge, the difference in the normal vectors of the two endpoints, and the area change of adjacent triangles. The weights are set to 40%, 30%, and 30% respectively, and redundant triangles are merged. Vertex caching technology is used to organize the vertex storage order. The cache block size is set to 16 bytes. The LRU replacement strategy is used to manage the cache content and invalid data is cleaned up periodically to improve the cache hit rate during graphics rendering and accelerate model loading and display speed.

Images