UAV Mapping Method and System Based on 3D Point Cloud Data Analysis
By performing grid division and adaptive grid subdivision on 3D point cloud data in UAV mapping, the problem of unstable surface information extraction under non-uniform point cloud conditions was solved, and efficient 3D mapping model generation was achieved.
Patent Information
- Application Number
- CN202611146839.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-30
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies in UAV mapping face challenges such as unstable surface information extraction under non-uniform and noisy point cloud conditions, and the inability of uniform grid subdivision to balance model accuracy and data volume.
By dividing the 3D point cloud data into a spatial grid, selecting surface seed points and tracing the point cloud along the normal vector direction, an initial 3D surface model is constructed, and adaptive mesh subdivision is performed to generate a 3D mapping model.
It improves the reliability and model accuracy of surface feature extraction, reduces grid size, and lowers storage and processing pressure, making it suitable for UAV mapping of complex terrain and features.
Smart Images

Figure CN123118961A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of surveying and mapping technology, specifically to a UAV surveying and mapping method and system based on three-dimensional point cloud data analysis. Background Technology
[0002] When drones equipped with laser scanners conduct surface mapping, they collect massive amounts of 3D point cloud data. How to reconstruct accurate 3D surface models from these discrete, unstructured point clouds is a core problem in the surveying field. Existing 3D reconstruction methods mostly employ a global fitting strategy, uniformly constructing surfaces or triangulating the entire point cloud data. This approach struggles to effectively distinguish between real surface points and interference points when dealing with drone point cloud data that has uneven density distribution, noise, and outliers. It easily produces errors such as distortion and holes in sparse or missing point cloud areas, making it difficult for the reconstructed model to accurately represent the true shape of terrain features. Simultaneously, the drastic changes in point cloud density mean that globally uniform processing parameters cannot accommodate the detailed needs of different regions. Features in high-density areas are smoothly lost, while low-density areas produce false surfaces due to insufficient information. After the mesh model is constructed, conventional methods generally use a uniform subdivision strategy to improve model accuracy, that is, inserting new vertices with the same granularity on all mesh patches. This processing method leads to over-refinement of the mesh in flat areas, generating a large number of redundant patches, increasing storage burden and post-processing computational overhead. Meanwhile, in feature regions with drastic curvature changes, such as concave and convex areas and edges, uniform subdivision fails to provide sufficient vertex density to accurately depict local details, resulting in insufficient model accuracy. How to stably extract effective surface information from raw data under conditions of uneven point cloud distribution and noise, and how to ensure sufficient resolution in feature-rich regions while controlling the overall mesh size, are pressing problems that current technologies need to address. Summary of the Invention
[0003] This invention provides a UAV mapping method and system based on 3D point cloud data analysis to solve the problems of unstable surface information extraction under non-uniform and noisy point cloud conditions, and the inability of uniform grid subdivision to balance model accuracy and data volume in existing technologies.
[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a UAV mapping method based on three-dimensional point cloud data analysis, comprising: acquiring raw three-dimensional point cloud data of a target area collected by a laser scanner mounted on a UAV; dividing the raw three-dimensional point cloud data into spatial grids to obtain several voxel grids; selecting surface seed points within each voxel grid according to the spatial distribution density of each point within the voxel grid; performing point cloud tracing along the normal vector direction of the voxel grid starting from the surface seed points to obtain surface point cloud bands of the voxel grids; constructing an initial three-dimensional surface model of the target area based on the spatial overlap relationship between the surface point cloud bands of adjacent voxel grids; and performing adaptive mesh subdivision of the initial three-dimensional surface model based on the curvature changes of each grid vertex in the initial three-dimensional surface model to generate a three-dimensional mapping model of the target area.
[0005] As a preferred embodiment of the present invention, when dividing the original 3D point cloud data into spatial grids, the extreme values of the coordinates of the original 3D point cloud data in the three coordinate axes are first obtained. The spans in the three coordinate axes are then calculated based on these extreme values. Each span is divided into several equally spaced intervals according to a preset spatial resolution. The intervals in the three coordinate axes are then combined using Cartesian products to obtain several voxel grids covering the spatial range of the original 3D point cloud data. Each voxel grid is a cubic spatial unit. This method can organize discrete point cloud data into ordered spatial units, facilitating subsequent parallel processing and improving computational efficiency.
[0006] As a preferred embodiment of the present invention, the process of selecting surface seed points is as follows: The number of three-dimensional points contained within each voxel grid is counted, and this number is used as the initial point density of the voxel grid. Voxel grids with an initial point density greater than a preset density threshold are marked as dense grids. Within the dense grid, the spatial distance from each three-dimensional point to the center point of the dense grid is calculated, and the three-dimensional point with the smallest spatial distance is selected as the surface seed point of the dense grid. This allows for accurate capture of representative seed points near the object surface in dense point cloud regions, providing a reliable starting point for subsequent point cloud tracking and avoiding tracking deviations caused by seed points deviating from the surface.
[0007] As a preferred embodiment of the present invention, the specific steps of the point cloud tracking include: calculating the covariance matrix of all three-dimensional points within the voxel grid where the surface seed point is located; using the eigenvector direction corresponding to the smallest eigenvalue of the covariance matrix as the initial normal vector direction of the voxel grid; starting from the surface seed point, searching for adjacent three-dimensional points along the positive and negative directions of the initial normal vector direction; including the searched adjacent three-dimensional points whose spatial distance from the surface seed point is less than a tracking distance threshold into a tracking point set; fitting the spatial positions of all three-dimensional points in the tracking point set into a spatial curve; and aggregating the three-dimensional points within a predetermined width range on both sides of the spatial curve into a surface point cloud band of the voxel grid. By tracking along the local normal vector direction and fitting a spatial curve, a coherent surface point cloud band can be robustly extracted under noise and occlusion environments, ensuring the continuity of the surface structure.
[0008] As a preferred embodiment of the present invention, the process of constructing the initial three-dimensional surface model includes: calculating the spatial overlap volume between the surface point cloud bands of two adjacent voxel grids, and determining whether the spatial overlap volume is greater than an overlap volume threshold; if so, fusing the surface point cloud bands of the two adjacent voxel grids to generate a fused point cloud band; if not, keeping the surface point cloud bands of the two adjacent voxel grids independent; and triangulating the fused point cloud band or the independent surface point cloud bands to generate a mesh patch of the initial three-dimensional surface model of the target region. Through fusion decisions based on spatial overlap relationships, the splicing gaps between adjacent grids can be effectively eliminated, making the generated mesh patch more watertight and smooth.
[0009] Preferably, during triangulation, all 3D points in the fused point cloud band or individual surface point cloud bands are projected onto a local 2D plane to obtain a projected point set. Two-dimensional Delaunay triangulation is then performed on this projected point set to obtain a projected triangular mesh. The vertex coordinates of each triangle in the projected triangular mesh are then back-projected into 3D space to obtain 3D triangular facets. All these 3D triangular facets are then stitched together according to their spatial adjacency relationships to form the mesh of the initial 3D surface model. This triangulation method can generate a topologically consistent triangular mesh while preserving the geometric features of the point cloud, thus improving the construction quality of the initial model.
[0010] As a preferred embodiment of the present invention, the adaptive mesh subdivision step includes: calculating the average curvature of each mesh vertex in the initial 3D surface model, and marking mesh vertices with an average curvature greater than a curvature threshold as high-curvature vertices; selecting neighborhood mesh patches centered on each high-curvature vertex; inserting new mesh vertices into the neighborhood mesh patches according to the subdivision ratio to generate refined local mesh patches; and merging the refined local mesh patches with the unrefined mesh patches to obtain a 3D mapping model of the target area. Performing local mesh refinement in areas with drastic curvature changes can significantly improve the model's ability to express terrain details with a relatively small data increment.
[0011] Preferably, when calculating the average curvature of the mesh vertex, the normal vectors of all mesh faces adjacent to the mesh vertex are obtained, the vector angle between the normal vector of each mesh face and the normal vector of the mesh vertex is calculated, and the cumulative value of all the vector angles is divided by the number of mesh faces adjacent to the mesh vertex to obtain the average curvature of the mesh vertex. This curvature calculation method is simple and efficient, and can accurately reflect the degree of curvature of the surface around the vertex.
[0012] Preferably, when inserting a new mesh vertex, the midpoints of each edge in the neighboring mesh patch are determined as candidate insertion vertex positions. The projection points of the candidate insertion vertex positions onto the surface of the neighboring mesh patch are calculated, and these projection points are used as the actual insertion vertex. The actual insertion vertex is then reconnected to the original mesh vertex to generate several sub-mesh patches. All of these sub-mesh patches are combined into a refined local mesh patch. Projecting the insertion point onto the original surface ensures that the subdivided mesh closely fits the real surface, avoiding model distortion.
[0013] This invention also provides an unmanned aerial vehicle (UAV) mapping system based on 3D point cloud data analysis, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the aforementioned UAV mapping method based on 3D point cloud data analysis. This system can automatically complete the entire process from raw point clouds to a high-precision 3D mapping model, balancing processing efficiency and model detail representation, and is suitable for UAV mapping operations on complex terrains and features.
[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By dividing the original 3D point cloud data into a spatial grid to obtain a voxel grid, and selecting surface seed points within each voxel grid based on the spatial distribution density of points, point cloud tracing is performed along the normal vector direction of the voxel grid starting from these seed points to form surface point cloud bands. This method transforms discrete point clouds into ordered point bands distributed along the surface direction. Utilizing local point density to select seed points effectively avoids low-density noise areas and prevents outliers from being misidentified as surface feature origins. During tracing along the normal vector direction, the spatial distance constraint between adjacent 3D points ensures that the growth of the point cloud band strictly conforms to the real surface, exhibiting adaptive capability to changes in local point density. Even in transitional regions where point cloud density changes drastically, point bands reflecting the surface contour can be extracted coherently without interruption or deviation due to density reduction. Compared to the conventional approach of directly fitting the entire point cloud to a surface, this strategy of first extracting locally and then constructing globally significantly reduces the interference of noise points and density abrupt changes on reconstruction quality, ensuring the reliability of surface feature extraction from the source. An initial 3D surface model is constructed by fusing the spatial overlap between surface point cloud bands of adjacent voxel grids, and then adaptively subdividing the model based on changes in grid vertex curvature. During subdivision, new vertices are inserted only into the neighborhood patches of grid vertices with an average curvature exceeding a threshold, while flat areas with lower curvature retain their original grid density. The position of the inserted vertex is determined by projecting it onto the surface from its midpoint, ensuring that the refined sub-patterns closely conform to the original surface morphology. This curvature-driven subdivision mechanism allows for sufficient refinement of the grid in areas rich in features such as undulations, edges, and corners, significantly improving the model's ability to reproduce feature edges and terrain details. In flat and open areas, redundant vertices are avoided, effectively controlling the grid size. Compared to uniform subdivision, the overall number of grid patches is significantly reduced while achieving the same level of detail, lowering the processing pressure for model storage, transmission, and subsequent visualization analysis. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0016] Figure 1 This is a flowchart of a UAV mapping method based on 3D point cloud data analysis; Figure 2 This is a flowchart of the surface point cloud band extraction method; Figure 3 This is a schematic diagram of the coordinate distribution and extreme values of each axis of the original 3D point cloud data; Figure 4 This is a diagram showing the initial point density distribution and dense grid markings of a voxel grid; Figure 5 This is a schematic diagram of point cloud tracing of surface seed points along the direction of the initial normal vector; Figure 6 This is a schematic diagram of the spatial overlap of point cloud bands on the surface of adjacent voxel grids; Figure 7 It is a grid vertex average curvature and high curvature vertex identification map. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] See Figure 1 This invention provides a UAV mapping method based on 3D point cloud data analysis, comprising: acquiring raw 3D point cloud data of a target area collected by a laser scanner mounted on a UAV; dividing the raw 3D point cloud data into spatial grids to obtain several voxel grids; selecting surface seed points within each voxel grid based on the spatial distribution density of points within each voxel grid; performing point cloud tracing along the normal vector direction of the voxel grid starting from the surface seed points to obtain surface point cloud bands of the voxel grid; constructing an initial 3D surface model of the target area based on the spatial overlap relationship between surface point cloud bands of adjacent voxel grids; and performing adaptive mesh subdivision of the initial 3D surface model based on the curvature changes of each grid vertex in the initial 3D surface model to generate a 3D mapping model of the target area.
[0019] Example 1:
[0020] In specific implementation, obtaining the extreme values of the original 3D point cloud data along the three coordinate axes involves traversing each 3D point in the original 3D point cloud data and recording the coordinate values of each 3D point along the X-axis, Y-axis, and Z-axis. The maximum and minimum values along the X-axis are then identified, along with the maximum and minimum values along the Y-axis and Z-axis. These maximum and minimum values are the extreme values along the X-axis, Y-axis, and Z-axis, respectively.
[0021] The span of the original 3D point cloud data in each of the three coordinate axes is calculated based on the extreme values of the coordinates. In the X-axis direction, the difference between the maximum and minimum X-axis values is used as the X-axis span; in the Y-axis direction, the difference between the maximum and minimum Y-axis values is used as the Y-axis span; and in the Z-axis direction, the difference between the maximum and minimum Z-axis values is used as the Z-axis span. The X-axis span, Y-axis span, and Z-axis span respectively characterize the spatial extension range of the original 3D point cloud data in the three coordinate axes.
[0022] Each span is divided into several equally spaced intervals according to a preset spatial resolution. The preset spatial resolution is denoted as... , This is a positive real number, and its value is determined based on the point cloud acquisition density of the laser scanner mounted on the UAV. In one feasible method, the average nearest neighbor distance of all 3D points in the original 3D point cloud data is calculated, and an integer multiple of this average nearest neighbor distance is set as the preset spatial resolution. This ensures that each voxel grid can accommodate a reasonable number of 3D points. For the X-axis span, the number of X-axis intervals is calculated. , The value is the X-axis span divided by Round up to the nearest integer; for the Y-axis span, calculate the number of Y-axis intervals. , The value is the Y-axis span divided by Round up to the nearest integer; for the Z-axis span, calculate the number of Z-axis intervals. , The value is the Z-axis span divided by The integer obtained by rounding up. Then, in the X-axis direction, starting from the minimum value of the X-axis, with... Divide into steps of 1 / 2 and 2 / 3. The X-axis is divided into several equally spaced intervals; in the Y-axis direction, starting from the minimum value of the Y-axis, ... Divide into steps of 1 / 2 and 2 / 3. A number of equally spaced Y-axis intervals; in the Z-axis direction, starting from the minimum value of the Z-axis, ... Divide into steps of 1 / 2 and 2 / 3. The Z-axis is divided into several equally spaced intervals. The left boundary of each interval is a closed interval, and the right boundary is an open interval, to avoid points on the interval boundary belonging to multiple intervals repeatedly.
[0023] The intervals along the three coordinate axes are combined using Cartesian products to obtain several voxel grids covering the spatial range of the original 3D point cloud data. Each voxel grid is a cubic spatial unit. Specifically, for the first voxel grid along the X-axis... The interval, the first in the Y-axis direction The interval and the first interval in the Z-axis direction The intervals are combined to obtain a voxel grid, where The value range is 1 to integers, The value range is 1 to integers, The value range is 1 to An integer. The coverage area of each voxel grid in the X-axis direction is the sum of the minimum value of the X-axis and... Add to the minimum value of the X-axis The coverage area in the Y-axis direction is the sum of the minimum value of the Y-axis and... Add to the minimum value of the Y-axis The coverage area in the Z-axis direction is the sum of the minimum value of the Z-axis and... Add to the minimum value of the Z-axis After the Cartesian product combination, a total of Individual voxel grids, when stitched together, completely cover the spatial region containing the original 3D point cloud data, with no overlap or gaps between the individual voxel grids. This is because the same spatial resolution is used in all three coordinate axes. The grid is divided into sections, so each voxel grid has a side length of... A cubic spatial unit.
[0024] See Figure 3The figure shows the probability density curves and corresponding extreme value positions along the X, Y, and Z coordinate axes obtained from the analysis of the original 3D point cloud data in Example 1. The horizontal axis represents the coordinate value in meters (m), and the vertical axis represents the probability density of the corresponding coordinate value. The solid blue line represents the probability density curve along the X coordinate axis, the dashed red line represents the probability density curve along the Y coordinate axis, and the dotted green line represents the probability density curve along the Z coordinate axis. The maximum and minimum values of each coordinate axis are indicated by dashed lines of corresponding colors, where the blue dashed lines represent the minimum and maximum values of the X-axis, the red dashed lines represent the minimum and maximum values of the Y-axis, and the green dashed lines represent the minimum and maximum values of the Z-axis.
[0025] As shown in the figure, the extreme values of the X-axis are located at approximately 3 meters (minimum) and 90 meters (maximum), the extreme values of the Y-axis are located at approximately 7 meters (minimum) and 80 meters (maximum), and the extreme values of the Z-axis are located at approximately -2 meters (minimum) and 62 meters (maximum). These extreme values clearly define the spatial range of the original 3D point cloud data along the three coordinate axes.
[0026] The probability density curves along each coordinate axis exhibit different distribution characteristics. The probability density curve along the X-axis is relatively broad, with a peak at approximately 57 meters, indicating that the point cloud data is spatially dispersed along the X-axis. The peak of the probability density curve along the Y-axis is concentrated at approximately 42 meters, indicating the main aggregation region of the point cloud in this direction. The probability density curve along the Z-axis is relatively sharp, with a peak at approximately 30 meters, showing a high concentration of the point cloud in the vertical direction.
[0027] Example 2:
[0028] In practice, for each voxel grid obtained after spatial grid division, the total number of 3D points falling within the spatial range of each voxel grid is counted. For a single voxel grid, all 3D points in the original 3D point cloud data are traversed, and it is determined whether the X, Y, and Z coordinate values of each 3D point are simultaneously within the X-axis, Y-axis, and Z-axis intervals of the voxel grid; if so, the 3D point is included in the voxel grid. After the traversal is completed, the total number of included 3D points is used as the initial point density of the voxel grid. The above process is repeated until the initial point density of all voxel grids is obtained.
[0029] Optionally, voxel grids with initial point densities greater than a preset density threshold are marked as dense grids. The preset density threshold is determined by calculating the arithmetic mean of the initial point densities of all voxel grids, and multiplying this arithmetic mean by a scaling factor. The scaling factor ranges from 0.5 to 1.0, and is specifically set to 0.8. The scaling factor is set to 0.8 because it requires that the point density of the voxel grids marked as dense grids be at least 80% of the average point density of all voxel grids. This ensures that the selected dense grids correspond to surface locations with sufficient point cloud coverage in the target area, avoiding seed point selection in areas with sparse point clouds and unreliable surface information. The initial point density of each voxel grid is compared with the preset density threshold. When the initial point density of a voxel grid is greater than the preset density threshold, the voxel grid is marked as a dense grid. When the initial point density of a voxel grid is less than or equal to the preset density threshold, the voxel grid is not marked as a dense grid and will not participate in the selection of surface seed points.
[0030] For each voxel grid labeled as a dense grid, the center point of the dense grid is determined. The coordinates of the center point of the dense grid are calculated based on the interval range of the dense grid along the three coordinate axes. Let the left boundary of the X-axis interval corresponding to the dense grid be denoted as... The right boundary is The X-coordinate value of the center point of the dense grid is... Let the left boundary of the Y-axis interval corresponding to the dense grid be . The right boundary is The Y-coordinate of the center point of the dense grid is then... Let the left boundary of the Z-axis interval corresponding to the dense grid be . The right boundary is The Z-coordinate value of the center point of the dense grid is... .
[0031] Within the dense grid, all 3D points falling within the grid are traversed, and the spatial distance from each 3D point to the center point of the grid is calculated. For a given 3D point, its spatial coordinates are denoted as... The center point coordinates of the dense grid are: The spatial distance from the three-dimensional point to the center point of the dense grid is calculated by the following formula:
[0032] in, This represents the spatial distance from the three-dimensional point to the center point of the dense grid. The coordinates of the three-dimensional point on the X, Y, and Z axes are given. The coordinates of the center point of the dense grid on the X, Y, and Z axes are given.
[0033] The spatial distances corresponding to all 3D points within the dense grid are obtained sequentially. All spatial distances are numerically compared, and the minimum spatial distance value is determined. The 3D point corresponding to the minimum spatial distance value is selected as the surface seed point of the dense grid. When multiple 3D points within the dense grid have the same minimum spatial distance, one of these 3D points is arbitrarily selected as the surface seed point. The above operation is performed on each voxel grid marked as a dense grid, thereby obtaining the surface seed point corresponding to each dense grid.
[0034] See Figure 4 In the figure, the horizontal axis represents the initial point density of the voxel grid, in points per grid cell, and the vertical axis represents the corresponding frequency. The bar chart shows the statistical frequency distribution of the initial point density of each voxel grid cell. The light blue bars correspond to the frequency of voxel grid cells with an initial point density less than a preset density threshold, while the light orange background area represents the dense grid cell area with an initial point density greater than the preset density threshold. The red dashed line in the figure indicates the specific value of the preset density threshold, which is approximately 12 points per grid cell.
[0035] As described in Example 2, the preset density threshold is obtained by multiplying the arithmetic mean of the initial point densities of all voxel grids by a scaling factor of 0.8. The figure shows that the initial point densities of most voxel grids are concentrated in the range of 0 to approximately 15 points / grid, with a clearly right-skewed frequency distribution. The highest frequency point density range is approximately 3 to 6 points / grid. Approximately 70% of the voxel grids in the area to the left of the red dashed line have initial point densities below the threshold, indicating that these voxel grids have sparse point clouds and do not meet the definition of a dense grid, thus not participating in the selection of surface seed points. The area to the right of the red dashed line, i.e., the orange highlighted area, covers a smaller number of voxel grids with point densities greater than the threshold. However, these grids have sufficient point cloud coverage, meeting the dense grid standard, thereby ensuring that the selected seed points come from a surface region with a dense and reliable point cloud.
[0036] Example 3:
[0037] In specific implementation, please refer to Figure 2 For each surface seed point located in a voxel grid, the spatial coordinates of all 3D points falling within the voxel grid are extracted to form a point set. Let the point set contain... A three-dimensional point, For integers greater than or equal to 1, the first... The spatial coordinates of a three-dimensional point are: ,in, The index of the three-dimensional point. The value range is from 1 to The integer. Calculate the stated The centroid coordinates of the voxel grid are obtained by averaging the coordinates of three-dimensional points along the three coordinate axes. ,in , , The covariance matrix of the three-dimensional point coordinates within the voxel grid is constructed using the centroid coordinates. The formula for calculating the covariance matrix is as follows:
[0038] in, The covariance matrix corresponding to the voxel grid is a A real symmetric matrix; This indicates the number of three-dimensional points within the voxel grid; , , They represent the first The coordinate values of a 3D point on the X, Y, and Z axes; , , These represent the coordinates of the centroid of the voxel grid on the X, Y, and Z axes, respectively.
[0039] Optionally, for the covariance matrix Eigenvalue decomposition is performed to obtain three eigenvalues and their corresponding eigenvectors. The three eigenvalues are arranged in ascending order of numerical value, and the eigenvector corresponding to the smallest eigenvalue is used as the initial normal vector direction of the voxel grid. The initial normal vector direction represents the orientation of the local surface represented by the point cloud within the voxel grid in space, and the initial normal vector direction is perpendicular to the local surface.
[0040] Starting from the surface seed point, a search for adjacent 3D points is performed along both the positive and negative directions of the initial normal vector. The initial normal vector direction is defined as a unit vector, denoted as […]. The positive direction is The opposite direction is During the forward search process, the spatial position of the surface seed point is used as a reference, along... The directional search progressively probes the neighborhood of 3D points. The probe method is as follows: At the current probe point location, a spherical neighborhood is constructed with the current probe point as the center and a tracking distance threshold as the radius. 3D points belonging to the voxel grid that fall within this spherical neighborhood are counted. If a 3D point exists within the spherical neighborhood, the 3D point with the closest spatial distance to the current probe point is selected as the forward tracking point and added to the tracking point set. Simultaneously, the current probe point is updated to the forward tracking point, and the above spherical neighborhood search process is repeated. If no 3D point exists within the spherical neighborhood, the forward search terminates. In the reverse search process, using the surface seed point as a reference, along... The directional search employs the same spherical neighborhood detection mechanism as the forward search, progressively searching for reverse tracking points and adding them to the tracking point set until no 3D points remain within the spherical neighborhood, at which point the reverse search terminates. The tracking distance threshold is set based on the point spacing of the laser scanner; in this embodiment, the tracking distance threshold is set to a preset spatial resolution. The tracking speed is twice that of the voxel grid, which is based on ensuring that the tracking process can cross the voxel grid boundary and connect surface points in adjacent areas, while avoiding jumping to non-associated surfaces due to excessive tracking step size.
[0041] Through the aforementioned forward and reverse searches, the tracking point set is obtained, which includes the surface seed points, all forward tracking points, and all reverse tracking points. Curve fitting is then performed on the spatial positions of all three-dimensional points in the tracking point set to generate a spatial curve. The specific process of curve fitting is as follows: using principal component analysis, the covariance matrix of the coordinates of all three-dimensional points in the tracking point set is calculated, and the eigenvector corresponding to the largest eigenvalue is extracted as the principal direction; using the centroid of the tracking point set as the base point, a straight line is drawn along the principal direction, and this straight line is the fitted spatial curve. It can be understood that when the surface curvature covered by the tracking point set is small, straight line fitting can accurately reflect the surface orientation; when the surface curvature is large, methods such as cubic spline interpolation can be used to achieve curve fitting, with the principal direction still used as the principal axis direction of the curve during fitting.
[0042] The predetermined width range on both sides of the spatial curve is determined. The predetermined width is set according to the lateral density of the point cloud of the laser scanner and is set as the preset spatial resolution. 1.5 times. Using the aforementioned space curve as the central axis, construct a region with a radius of... A cylindrical pipe region, 1.5 times the size of the original, is used. From all the three-dimensional points of the voxel grid, the three-dimensional points whose three-dimensional coordinates fall within the cylindrical pipe region are extracted and aggregated into a surface point cloud band of the voxel grid. The three-dimensional points in the surface point cloud band are distributed on both sides of the spatial curve, forming strip-shaped point cloud sets extending along the direction of the initial normal vector.
[0043] See Figure 5 The figure shows the results of point cloud tracing along the initial normal vector direction for the voxel grid containing the surface seed point in Example 3, both forward and backward. The horizontal axis represents the distance along the spatial curve direction in meters, ranging from approximately -5 meters to 5 meters; the vertical axis represents the vertical offset distance in meters, ranging from approximately -0.8 meters to 1.3 meters. In the legend, the "surface seed point" is marked with a yellow pentagram, located at the horizontal coordinate 0 with a vertical offset of 0, indicating the tracing starting point; the "forward tracing point" is represented by an orange dot, distributed in the horizontal coordinate range of 0 to 5 meters; and the "reverse tracing point" is represented by a green triangle, distributed in the horizontal coordinate range of -5 meters to 0 meters.
[0044] The spatial curve is indicated by a solid black line, which passes through the surface seed points along the horizontal axis, representing the principal direction of the spatial curve obtained by fitting the tracking point set through principal component analysis. The two gray dashed lines at the top and bottom of the figure represent the boundaries of the pipe region. The vertical distance between the boundary and the center line of the spatial curve is approximately ±0.75 meters, indicating that the width of the cylindrical pipe region constructed based on the spatial curve within the voxel grid is approximately 1.5 times the preset spatial resolution. This conforms to the setting of the predetermined width in Embodiment 3.
[0045] As shown in the figure, both forward and reverse tracking points extend approximately along the spatial curve, and most points are distributed within the boundary of the pipe region. This indicates that the extracted surface point cloud bands form a continuous and banded set of point clouds within the voxel grid, demonstrating an effective point cloud tracking process along the initial normal vector direction. The forward tracking point terminates near 5 meters on the horizontal axis, while the reverse tracking point terminates near -5 meters on the horizontal axis. This reflects the limitation of the tracking distance threshold on the tracking range, ensuring that the tracking process crosses the voxel grid boundary and avoids jumping to non-associated surfaces.
[0046] Example 4:
[0047] In practice, for any two adjacent voxel grids, the surface point cloud bands obtained by point cloud tracing of each voxel grid in the two adjacent voxel grids are acquired respectively. The first surface point cloud band corresponds to the first voxel grid, and the second surface point cloud band corresponds to the second voxel grid. The three-dimensional points contained in the first surface point cloud band constitute the first point set, and the three-dimensional points contained in the second surface point cloud band constitute the second point set.
[0048] Calculate the spatial overlap volume between surface point clouds of two adjacent voxel grids. The calculation method involves constructing the 3D convex hulls of the first and second point sets respectively. The construction of the 3D convex hulls is implemented using an incremental algorithm or a fast convex hull algorithm, resulting in a minimal convex polyhedron enclosed by triangular facets that contains all 3D points in the corresponding point set. The 3D convex hull corresponding to the first point set is denoted as the first convex hull, and the 3D convex hull corresponding to the second point set is denoted as the second convex hull. The intersection of the first and second convex hulls is calculated, resulting in an intersection polyhedron, the volume of which is the spatial overlap volume. The intersection polyhedron is calculated using 3D Boolean operations. Both the first and second convex hulls are represented as closed triangular meshes. Through facet intersection detection, facet clipping, and voxel classification, the spatial regions simultaneously located within the first and second convex hulls are extracted. The volume of the intersection polyhedron is obtained by integrating and summing the triangular meshes of the intersection polyhedron. Specifically, for each triangular facet of the intersection polyhedron, the directed volume of the tetrahedron with that facet as the top facet and the origin as the bottom facet is calculated. The algebraic sum of all directed volumes is the volume of the intersection polyhedron.
[0049] After obtaining the spatial overlap volume, it is determined whether the spatial overlap volume is greater than the overlap volume threshold. The overlap volume threshold is set according to the spatial scale of the surface point cloud bands. The setting method is as follows: calculate the first convex hull volume of the first convex hull corresponding to the first surface point cloud band, and the second convex hull volume of the second convex hull corresponding to the second surface point cloud band; take the smaller of the first and second convex hull volumes, and multiply the smaller convex hull volume by a volume ratio coefficient as the overlap volume threshold. The volume ratio coefficient is set to 0.1, that is, the overlap volume threshold is set to one-tenth of the smaller convex hull volume. The basis for setting the volume ratio coefficient to 0.1 is that when the spatial intersection area of two surface point cloud bands reaches at least 10% of the smaller convex hull volume, the two surface point cloud bands are considered to have a substantial spatial overlap relationship, indicating that the local surfaces covered by the two voxel grids are interconnected in space and should be fused.
[0050] The spatial overlap volume is numerically compared with an overlap volume threshold. When the spatial overlap volume is greater than the overlap volume threshold, the result is considered "yes," and the surface point cloud bands of two adjacent voxel grids, namely the first surface point cloud band and the second surface point cloud band, are fused to generate a fused point cloud band. The point cloud fusion process is as follows: all 3D points in the first surface point cloud band and all 3D points in the second surface point cloud band are merged into a union point set, and the union point set is downsampled using voxelization. The side length of the downsampled voxel unit is the preset spatial resolution. The coordinates of all 3D points within each voxel unit are averaged to 0.5 times the original value. The average coordinates of all 3D points within each voxel unit are used as the coordinates of the new 3D points, and all the new 3D points are merged to form a fused point cloud band. When the spatial overlap volume is less than or equal to the overlap volume threshold, the result is negative, and the surface point cloud bands of two adjacent voxel grids remain independent. That is, the first surface point cloud band and the second surface point cloud band are not merged, and they are treated as independent point cloud units for subsequent processing.
[0051] Triangulation is performed on the merged point cloud bands or individual surface point cloud bands to generate mesh patches of the initial 3D surface model of the target region. The specific implementation process of triangulation is as follows:
[0052] Projecting all 3D points in the merged point cloud or individual surface point cloud onto a local 2D plane yields the projected point set. The local 2D plane is determined as follows: Calculate the average coordinates of all 3D points in the point cloud to be projected (either the merged or individual surface point cloud) to obtain the point cloud centroid; calculate the covariance matrix of all 3D point coordinates in the point cloud to be projected, perform eigenvalue decomposition on the covariance matrix, and use the eigenvector corresponding to the smallest eigenvalue as the normal vector of the projection plane; construct a spatial plane with the point cloud centroid as a point on the plane and the normal vector of the projection plane as the normal line; this spatial plane is the local 2D plane. Establish a 2D Cartesian coordinate system on the local 2D plane, with the projection of the point cloud centroid onto the local 2D plane as the origin, the projection of the eigenvector corresponding to the largest eigenvalue of the covariance matrix onto the local 2D plane as the horizontal axis, and the projection of the eigenvector corresponding to the second largest eigenvalue of the covariance matrix onto the local 2D plane as the vertical axis. Each 3D point in the point cloud to be projected is projected onto a local 2D plane along the normal vector direction of the projection plane, thus obtaining the 2D coordinates of each 3D point on the local 2D plane. The 2D coordinates corresponding to all 3D points constitute the projection point set.
[0053] A two-dimensional Delaunay triangulation is performed on the projected point set to obtain a projected triangulation network. The two-dimensional Delaunay triangulation is implemented using a point-by-point insertion algorithm or a divide-and-conquer algorithm. The process is as follows: a super triangle enclosing all points in the projected point set is constructed. Each point in the projected point set is successively inserted into the current triangulation network. After each insertion, an edge flipping operation is performed to ensure that the triangulation network satisfies the Delaunay criterion, i.e., the circumcircle of any triangle does not contain any other points in the projected point set. After all points have been inserted, a projected triangulation network composed of several two-dimensional triangles is obtained. Triangles associated with vertices of the super triangle are removed, and the remaining triangles are the final result of the projected triangulation network.
[0054] The coordinates of each vertex of the projected triangulation are back-projected into 3D space to obtain a 3D triangular patch. The back-projection process is the inverse operation of the projection process: for each vertex of each triangle in the projected triangulation, the coordinates of that vertex on the local 2D plane are... ,in, The vertex index; The corresponding three-dimensional point lies on a local two-dimensional plane, and its three-dimensional coordinates in the spatial coordinate system are calculated using the coordinates of the origin of the two-dimensional coordinate system, the unit vector of the horizontal axis, and the unit vector of the vertical axis. Let the coordinates of the origin of the local two-dimensional coordinate system in the spatial coordinate system be... The horizontal axis unit vector is The vertical axis unit vector is Then the vertex The coordinates in the spatial coordinate system are ,in, , , Unit vector on the horizontal axis Components on the X, Y, and Z coordinate axes , , Unit vector along the vertical axis The components on the X, Y, and Z coordinate axes. Each triangle in the projected triangulation, after back projection, yields a three-dimensional triangular patch in space.
[0055] All 3D triangular facets are stitched together according to their spatial adjacency relationships to form the mesh of the initial 3D surface model. The spatial adjacency relationship is determined as follows: for any two 3D triangular facets, it is checked whether they share the same edge; if they share the same edge, the two 3D triangular facets are considered spatially adjacent. Based on the spatial adjacency relationship, all 3D triangular facets are organized into a connected triangular mesh, which is the mesh of the initial 3D surface model of the target region. During the stitching process, for 3D triangular facets that overlap or intersect, a half-edge data structure is used for mesh repair to eliminate non-manifold edges and self-intersecting facets, ensuring the topological correctness of the initial 3D surface model.
[0056] See Figure 6 The figure illustrates the distribution of surface point cloud bands and their 3D convex hulls corresponding to two adjacent voxel grids in Example 4 after projection onto a two-dimensional plane. The x and y coordinates are the X and Y coordinates after projection, respectively, in meters. In the figure, blue dots represent the surface point cloud bands of the first voxel grid, red squares represent the surface point cloud bands of the second voxel grid, and purple dots mark the spatial overlap area of the two surface point cloud bands. The solid blue lines and dashed red lines represent the projection contours of the 3D convex hulls corresponding to the point cloud bands of the first and second voxel grids on the two-dimensional plane, respectively.
[0057] As can be seen from the outlines of the two convex hulls in the figure, there is a large overlap between the first and second convex hulls. This overlap area is represented by the purple point cloud distribution area, covering the overlapping spatial regions of the two surface point cloud bands. This spatial overlap volume is used to determine whether there is a substantial spatial overlap relationship between the surface point cloud bands of adjacent voxel grids. Specifically, if the spatial overlap volume is greater than a preset overlap volume threshold with one-tenth of the smaller convex hull volume as the threshold, then point cloud fusion processing is performed on the two surface point cloud bands in Example 4; otherwise, the two surface point cloud bands remain independent.
[0058] Example 5:
[0059] In specific implementation, the average curvature of each mesh vertex in the initial 3D surface model is calculated. For a mesh vertex in the initial 3D surface model, it is denoted as the currently processed vertex, and all mesh patches adjacent to the currently processed vertex are obtained. The method for determining adjacent mesh patches is as follows: traverse all mesh patches of the initial 3D surface model. For each mesh patch, check whether the currently processed vertex is a vertex of that mesh patch. If so, the mesh patch is recorded as an adjacent mesh patch of the currently processed vertex. Assume that the currently processed vertex has multiple... Adjacent mesh patches, It is an integer greater than or equal to 1.
[0060] Obtain the normal vector of each adjacent mesh facet. For an adjacent mesh facet, the three vertices form a triangle in space. Let the spatial coordinates of the three vertices be... , , Then the normal vector of the adjacent mesh facet From vector with vector The cross product is normalized to obtain, i.e. ,in, The sequence number of the adjacent mesh facet. The value range is from 1 to An integer. Calculate the normal vector of the currently processed vertex. The normal vector of the currently processed vertex is obtained by the area-weighted average of the normal vectors of all mesh faces adjacent to the currently processed vertex, specifically as follows: ,in, For the first The area of each adjacent grid patch , For the first The normal vector of each adjacent mesh patch.
[0061] For each adjacent mesh patch, calculate the normal vector of the adjacent mesh patch. With the normal vector of the currently processed vertex The angle between the vectors. It is obtained by using the inverse cosine function, that is ,in, This represents the dot product of two unit normal vectors. The unit is radians. When the dot product result exceeds the numerical error... When dealing with an interval, truncate it to that interval to ensure that the arccosine can be calculated.
[0062] All Angle between vectors The vectors are summed to obtain the sum of their angles. Divide the accumulated value of the vector angle by the number of adjacent mesh patches. The average curvature of the currently processed vertex is obtained. The calculation formula is:
[0063] in, This represents the average curvature of the currently processed vertex, in radians; This represents the total number of mesh faces adjacent to the currently processed vertex; Indicates the first The angle between the normal vectors of adjacent mesh faces and the normal vector of the currently processed vertex, in radians. This calculation is performed on each mesh vertex in the initial 3D surface model to obtain the average curvature of each mesh vertex.
[0064] Optionally, mesh vertices with an average curvature greater than a curvature threshold are marked as high-curvature vertices. The curvature threshold is determined by: statistically analyzing the average curvature of all mesh vertices in the initial 3D surface model, and calculating the median and absolute deviation of the average curvature of all mesh vertices. The process of obtaining the absolute deviation of the median is as follows: calculating the absolute difference between the average curvature and the median of each mesh vertex, taking the median of all absolute differences, and then multiplying it by a constant factor of 1.4826 to convert it into an estimate consistent with the standard deviation. The curvature threshold is set to the median plus twice the absolute deviation of the median. The basis for setting the curvature threshold to this value is that, in statistics, the median plus twice the absolute deviation of the median can define the upper bound outlier values in the dataset, corresponding to high-curvature regions with curvature significantly higher than that of general surfaces, and can stably identify the locations of abrupt changes in surface curvature, avoiding the influence of a very small number of extreme values. For each grid vertex, the average curvature of the grid vertex is compared with a curvature threshold. When the average curvature of the grid vertex is greater than the curvature threshold, the grid vertex is marked as a high curvature vertex; when the average curvature of the grid vertex is less than or equal to the curvature threshold, the grid vertex is not marked as a high curvature vertex.
[0065] In practice, for each mesh vertex marked as a high-curvature vertex, neighboring mesh patches are selected centered on the high-curvature vertex. The selection method for these neighboring mesh patches is as follows: from all mesh patches of the initial 3D surface model, mesh patches containing at least one high-curvature vertex are extracted, and the set of these mesh patches is taken as the neighboring mesh patches of the high-curvature vertex. When multiple neighboring mesh patches of high-curvature vertices overlap, the overlapping mesh patches are merged into the same neighboring patch group, and subsequent mesh subdivision operations are performed uniformly on this neighboring patch group.
[0066] New mesh vertices are inserted into the neighboring mesh patches according to the subdivision ratio, generating refined local mesh patches. Specifically, the process of inserting new mesh vertices is as follows: the midpoints of each edge in the neighboring mesh patch are determined as candidate insertion vertex positions. For each edge in the neighboring mesh patch, the edge connects two original mesh vertices, which are denoted as the original mesh vertices. and the original mesh vertices If the spatial coordinates of the two original mesh vertices are known, then the candidate insertion vertex position is... for That is, the arithmetic mean of the coordinates of the two original grid vertices.
[0067] Calculate the projection point of the candidate insertion vertex position onto the surface of the neighboring mesh patch, and use the projection point as the actual insertion vertex. The projection process is as follows: at the candidate insertion vertex position... Within a local neighborhood, fit a local surface, and Project onto the local surface along the normal direction. The fitting method for the local surface is: using the candidate insertion vertex positions... Using the average side length of all edges in the neighboring mesh patches as the radius, construct a spherical region centered at the nearest sphere. Extract all mesh vertices from the initial 3D surface model that fall into this spherical region to form a neighboring point set. Perform moving least squares surface fitting on the neighboring point set, selecting a quadratic polynomial as the basis function and employing a Gaussian weighting function, with the weights varying with the distance from the neighboring points. The decay occurs as the distance increases. After obtaining the locally fitted surface, the calculation is performed. The point with the shortest distance to the locally fitted surface is the projection point of the candidate insertion vertex onto the surface of the neighboring mesh patch. This projection point is taken as the actual insertion vertex, and denoted as _____. Perform the above operation on each edge in the neighboring mesh patch to obtain the actual inserted vertex corresponding to each edge.
[0068] Each actual inserted vertex is reconnected to the original mesh vertex, generating several sub-mesh patches. The connection rule is: for a triangular patch in the original mesh patch, its three vertices are... , , The actual insertion vertex is inserted on each of the three sides. , , Connect the three actual inserted vertices inside the original triangle to form four sub-triangles. The vertex combinations of the four sub-triangles are as follows: ( , , ), ( , , ), ( , , )and( , , These four sub-triangles are the sub-mesh patches. All the sub-mesh patches obtained after the above subdivision of all original mesh patches within the neighboring mesh patch are combined into a refined local mesh patch.
[0069] The refined local mesh patches are merged with the unrefined mesh patches to obtain a 3D mapping model of the target area. Specifically, the merging method is as follows: from the initial 3D surface model's mesh patch set, the mesh patches originally included in the neighboring mesh patches are removed; all sub-mesh patches from the refined local mesh patches are added to the mesh patch set, maintaining the mesh topology, forming a merged triangular mesh. This triangular mesh is the 3D mapping model of the target area.
[0070] See Figure 7 The horizontal axis in the figure represents the average curvature of the mesh vertices in the initial 3D surface model, in radians, ranging from 0 to approximately 1.6; the vertical axis represents the frequency of mesh vertices at the corresponding curvature value. The bar chart shows the distribution of the number of mesh vertices within each curvature interval. The frequency generally decreases as curvature increases, with higher vertex frequencies in the lower curvature region on the left and a significant decrease in vertex frequencies in the higher curvature region on the right. The curvature threshold, indicated by the dashed line, is approximately 0.25 radians, dividing the curvature range into low-curvature and high-curvature vertex intervals. The area to the right of the dashed line is marked with a light red shading, indicating the region containing high-curvature vertices.
[0071] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.
Claims
1. A UAV mapping method based on three-dimensional point cloud data analysis, characterized in that, include: Acquire raw 3D point cloud data of the target area collected by a laser scanner mounted on a drone; The original 3D point cloud data is divided into spatial grids to obtain several individual pixel grids; Based on the spatial distribution density of each point within each voxel grid, surface seed points within the voxel grid are selected. Starting from the surface seed point, point cloud tracing is performed along the normal vector direction of the voxel grid to obtain the surface point cloud band of the voxel grid. Based on the spatial overlap relationship between the surface point cloud bands of adjacent voxel grids, an initial three-dimensional surface model of the target region is constructed. Based on the curvature changes of each grid vertex in the initial three-dimensional surface model, the initial three-dimensional surface model is adaptively subdivided into grids to generate a three-dimensional mapping model of the target region. 2.The UAV mapping method based on three-dimensional point cloud data analysis of claim 1, wherein, The step of dividing the original 3D point cloud data into a spatial grid to obtain several individual pixel grids includes: Obtain the extreme values of the original 3D point cloud data in the three coordinate axes; Calculate the span of the original 3D point cloud data in the three coordinate axes based on the coordinate extreme values; Each span is divided into several equally spaced intervals according to a preset spatial resolution; The intervals along the three coordinate axes are combined by Cartesian product to obtain several voxel grids covering the spatial range of the original three-dimensional point cloud data, and each voxel grid is a cubic spatial unit. 3.The UAV mapping method based on three-dimensional point cloud data analysis of claim 1, wherein, The step of selecting surface seed points within the voxel grid based on the spatial distribution density of each point within the voxel grid includes: The number of three-dimensional points contained in each voxel grid is counted, and the number is used as the initial point density of the voxel grid. The voxel grid with an initial point density greater than a preset density threshold is marked as a dense grid. Within the dense grid, calculate the spatial distance from each three-dimensional point within the dense grid to the center point of the dense grid; The three-dimensional point with the smallest spatial distance is selected as the surface seed point of the dense grid. 4.The UAV mapping method based on three-dimensional point cloud data analysis of claim 1, wherein, The step of performing point cloud tracing along the normal vector direction of the voxel grid, starting from the surface seed point, to obtain the surface point cloud band of the voxel grid includes: Calculate the covariance matrix of all three-dimensional points within the voxel grid where the surface seed point is located, and take the eigenvector direction corresponding to the smallest eigenvalue of the covariance matrix as the initial normal vector direction of the voxel grid. Starting from the surface seed point, search for adjacent 3D points in both the forward and reverse directions of the initial normal vector; The points among the searched adjacent 3D points whose spatial distance from the surface seed point is less than the tracking distance threshold are included in the tracking point set; The spatial positions of all three-dimensional points in the tracking point set are fitted into a spatial curve, and the three-dimensional points within a predetermined width range on both sides of the spatial curve are aggregated into a surface point cloud of the voxel grid.
5. The method of claim 1, wherein, The step of constructing an initial three-dimensional surface model of the target region based on the spatial overlap relationship between surface point cloud bands of adjacent voxel grids includes: Calculate the spatial overlap volume between the surface point cloud bands of two adjacent voxel grids; Determine whether the spatial overlap volume is greater than the overlap volume threshold; If so, the surface point cloud bands of two adjacent voxel grids are fused to generate a fused point cloud band; If not, then the surface point cloud bands of two adjacent voxel grids remain independent; The fused point cloud band or the individual surface point cloud bands are triangulated to generate a mesh patch of the initial three-dimensional surface model of the target region. 6.The UAV mapping method based on three-dimensional point cloud data analysis of claim 5, wherein, The step of triangulating the fused point cloud band or the individual surface point cloud bands to generate a mesh patch of the initial three-dimensional surface model of the target region includes: Project all three-dimensional points in the fused point cloud or the individual surface point cloud onto a local two-dimensional plane to obtain a projection point set; Perform two-dimensional Delaunay triangulation on the projection point set to obtain a projection triangulation network; The vertex coordinates of each triangle in the projected triangulation are back-projected into three-dimensional space to obtain a three-dimensional triangular patch; All the three-dimensional triangular facets are spliced together according to their spatial adjacency to form the mesh facets of the initial three-dimensional surface model.
7. The method of claim 1, wherein, The step of adaptively subdividing the initial three-dimensional surface model based on the curvature changes of each grid vertex in the initial three-dimensional surface model to generate a three-dimensional mapping model of the target region includes: Calculate the average curvature of each mesh vertex in the initial three-dimensional surface model; Mesh vertices with an average curvature greater than a curvature threshold are marked as high curvature vertices; Centered on each of the high curvature vertices, select the neighborhood mesh facets of the high curvature vertices; New mesh vertices are inserted into the neighboring mesh patches according to the subdivision ratio to generate refined local mesh patches; The refined local mesh patches are merged with the unrefined mesh patches to obtain a three-dimensional mapping model of the target area. 8.The UAV mapping method based on three-dimensional point cloud data analysis of claim 7, wherein, The step of calculating the average curvature of each mesh vertex in the initial three-dimensional surface model includes: Obtain the normal vectors of all mesh faces adjacent to the stated mesh vertex; Calculate the angle between the normal vector of each mesh patch and the normal vector of each mesh vertex; The average curvature of the grid vertex is obtained by dividing the sum of the angles of all the vectors by the number of grid patches adjacent to the grid vertex. 9.The UAV mapping method based on three-dimensional point cloud data analysis of claim 7, wherein, The step of inserting new mesh vertices into the neighboring mesh patches according to the subdivision ratio to generate refined local mesh patches includes: The midpoint positions of each edge in the neighborhood mesh are determined as candidate insertion vertex positions; Calculate the projection point of the candidate insertion vertex position onto the surface of the neighboring mesh patch, and use the projection point as the actual insertion vertex; The actual inserted vertex is reconnected to the original mesh vertex to generate several sub-mesh patches; All the sub-mesh patches are combined into a refined local mesh patch.
10. An unmanned aerial vehicle surveying system based on three-dimensional point cloud data analysis, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the UAV mapping method based on three-dimensional point cloud data analysis as described in any one of claims 1 to 9.
