Pine forest three-dimensional model rapid reconstruction method based on unmanned aerial vehicle photography

Abstract

The invention relates to the technical field of computer graphics and three-dimensional modeling, and discloses a pine forest three-dimensional model rapid reconstruction method based on unmanned aerial vehicle photography, which comprises the following steps: acquiring a multi-view image sequence, extracting a single pine topology skeleton line, and calling a parameterized fractal generation program to construct a virtual three-dimensional geometry; performing Euclidean distance transformation on the contour of the image entity to generate a distance field graph, and calculating the projection consistency deviation of the virtual geometry in the distance field graph; and according to the deviation gradient, iteratively correcting the geometric growth control parameters until convergence, and finally instantiating and loading a preset leaf cluster model at the tail end of the skeleton to generate a three-dimensional grid. The problems of reconstruction damage and non-convergence caused by rigid matching failure of a discrete discontinuous object in a dynamic environment are solved, reverse recovery of a high-fidelity topological structure under low data density is realized, and the topological structure closure and geometric attitude convergence stability of a reconstruction model are ensured.

Description

Technical Field

[0001] This invention relates to a method for rapid reconstruction of a 3D model of a pine forest based on drone photography, belonging to the field of computer graphics and 3D modeling technology. Background Technology

[0002] In current large-scale outdoor scene 3D digitization operations, drones equipped with visible light cameras are used to acquire multi-view image sequences. Based on multi-view stereo geometry algorithms, the 3D structure of the scene is restored. This has become a general paradigm for obtaining geospatial information. Such technologies usually follow a processing chain from motion-based structure restoration, dense point cloud matching to surface mesh reconstruction. The convergence of the mathematical model depends on the prior assumptions that the manifold continuity and photometric consistency are satisfied in the local neighborhood of the surface of the photographed object.

[0003] Simply relying on improved hardware accuracy or optimization of general algorithms faces technical bottlenecks when dealing with complex forestry scenarios. For example, Chinese invention patent CN112560623B discloses a method for rapid identification of mangrove plant species based on drones. This solution introduces a combination of Digital Surface Model (DSM) and deep learning, using height information to assist species identification. However, its core technology is limited to two-dimensional or 2.5-dimensional analysis of the visual features of the canopy surface. Such methods focus on species identification rather than structural reconstruction and cannot delve into the internal topology of individual trees. When the reconstruction object is transformed into a non-rigid object like a pine forest, which has both high-frequency discrete coniferous features and high canopy closure, the modeling path based on DSM or pixel-level dense matching faces the risk of fundamental failure. When the reconstruction object is a building or land... When the shape transforms into a discrete voxel set with high-frequency spatial frequency characteristics, such as a pine forest, the traditional pixel-level dense matching modeling path faces the risk of fundamental failure. The pine canopy is composed of a massive number of discontinuous needle clusters, and sub-pixel-level geometric features are easily lost in the conventional downsampling or feature matching process, resulting in sparse point clouds or failure in depth map calculation. There is a time difference in the acquisition of UAV sequence images, and pine branches and leaves undergo non-rigid transient displacement in uncontrolled outdoor wind fields. There is an inherent contradiction between physical deformation and rigid geometric matching constraints, which makes the photometric consistency optimization function unable to converge or trapped in local extrema. Simply increasing the image resolution or increasing the shooting overlap increases the computational load exponentially, introduces more high-frequency noise and matching ambiguity, and cannot solve the fundamental problems of field of view occlusion and dynamic deformation.

[0004] Therefore, how to overcome the bottleneck of discrete non-rigid object reconstruction by relying solely on conventional two-dimensional image sequences, and how to achieve high-fidelity three-dimensional model construction that conforms to the topological laws of plants under low data density and dynamic interference conditions, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A method for rapid reconstruction of a 3D model of a pine forest based on UAV photography, comprising the following steps: Step 101: Obtain a multi-view two-dimensional image sequence for the target pine forest area, calculate a sparse spatial point cloud based on the multi-view two-dimensional image sequence, and cluster the sparse spatial point cloud for a single pine tree target to extract the topological skeleton line representing the spatial orientation of its trunk and main branches. Step 102: Call the preset parametric fractal generation program, use the topological skeleton line as the growth reference, and construct a virtual three-dimensional geometry based on a set of geometric growth control parameters. The geometric growth control parameters at least limit the branch iteration depth and branch deflection angle. Step 103: Perform Euclidean distance transformation on the pine tree entity outline in the multi-view 2D image sequence to generate a distance field map representing the distance gradient from the pixel to the outline boundary. The pixel value in the distance field map defines the Euclidean distance from the pixel to the nearest outline boundary. Step 104: Project the virtual 3D geometry onto the distance field map, determine the distance sampling value of each projected pixel in the distance field map, and calculate the projection consistency deviation based on the weighted sum of the distance sampling values; Step 105: Determine whether the projection consistency deviation meets the preset convergence condition. If the projection consistency deviation does not meet the preset convergence condition, iteratively correct the geometric growth control parameters according to the gradient direction of the projection consistency deviation, and update the virtual three-dimensional geometry based on the corrected geometric growth control parameters until the preset convergence condition is met. Step 106: Lock the geometric growth control parameters and topology skeleton line that meet the preset convergence conditions, and load the preset leaf cluster geometric model by instantiating the reference at the end node position of the topology skeleton line to generate the target three-dimensional mesh model.

[0006] Preferably, a virtual three-dimensional geometry is constructed based on a set of geometric growth control parameters, including: calculating the normalized height value of each growth node on the topological skeleton line relative to the starting root node; constructing a parameter mapping function, which defines a monotonic functional relationship between the branch deflection angle and the normalized height value; and when constructing the virtual three-dimensional geometry, calculating and assigning the branch deflection angle of the current node based on the normalized height value of the current growth node through the parameter mapping function to form a non-uniform geometric topology structure that varies with height.

[0007] Preferably, the projection consistency deviation is calculated based on the weighted sum of distance sample values, including: determining the topology growth level corresponding to each projected pixel in the virtual 3D geometry, where the topology growth level represents the iteration depth of the geometric unit to which the projected pixel belongs in the fractal generation process; constructing a hierarchy weight function, which limits the negative correlation between the weighting coefficients and the topology growth level; assigning weighting coefficients according to the topology growth level corresponding to each projected pixel, wherein regions with lower topology growth levels are assigned larger weighting coefficients; and calculating the projection consistency deviation by weighted summation of the distance sample values ​​based on the weighting coefficients.

[0008] Preferably, the method for clustering sparse spatial point clouds of a single pine tree to extract topological skeleton lines representing the spatial orientation of its trunk and main branches includes: detecting whether there are discrete canopy regions in the sparse spatial point cloud that lack trunk point cloud data; if discrete canopy regions exist, constructing a surface mesh model for the top point cloud of the discrete canopy regions; calculating the Gaussian curvature distribution of the surface mesh model and identifying curvature extrema and curvature ridge features; using the curvature extrema as the starting endpoint, drawing virtual trunk paths along the direction of gravity, and inferring virtual branch paths based on the spatial mapping relationship of curvature ridge features; and combining the virtual trunk paths and virtual branch paths as topological skeleton lines to complete the topological structure of visual blind spots.

[0009] Preferably, the process of iteratively correcting the geometric growth control parameters based on the gradient direction of the projection consistency deviation further includes a non-rigid correction step based on the residual space distribution: generating a binary difference map representing the difference region between the projection area of ​​the virtual 3D geometry and the outline of the pine tree entity; calculating the geometric centroid of the non-zero pixel region in the binary difference map and determining the offset vector of the geometric centroid relative to the projection center of the virtual 3D geometry; using the offset vector as a deformation control parameter to apply non-rigid deformation based on Bézier curves to the topological skeleton line; and re-executing the construction and projection calculation steps of the virtual 3D geometry based on the deformed topological skeleton line to eliminate the systematic deviation of the trunk attitude.

[0010] Preferably, before constructing a virtual three-dimensional geometry based on a set of geometric growth control parameters, the method further includes: constructing a spatial index structure based on all extracted topological skeleton lines, calculating the neighborhood density value of the topological skeleton line of each individual pine tree within a preset spatial radius; constructing an iterative depth control function, which limits the negative correlation between the branch iteration depth and the neighborhood density value; using the branch iteration depth calculated based on the neighborhood density value as the fractal generation termination threshold for constructing the corresponding virtual three-dimensional geometry, and performing geometric simplification processing on the invisible internal structure located in the high-density region.

[0011] Preferably, the distance sample values ​​are weighted and summed based on weighting coefficients, following the weighted bias calculation formula: ,in, For projection consistency deviation, This represents the total number of projected pixels. For the first The distance sample value of each projected pixel in the distance field map. For the first The topology growth level corresponding to each projected pixel. The corresponding level is determined based on the hierarchical weight function. The weighting coefficients.

[0012] Preferably, acquiring a multi-viewpoint two-dimensional image sequence for the target pine forest area includes: performing semantic segmentation on the multi-viewpoint two-dimensional image sequence and extracting a binarized semantic mask containing the pine tree target; removing discrete connected components with an area smaller than a preset noise threshold from the binarized semantic mask; and using the noise-removed binarized semantic mask as the pine tree entity outline for subsequent Euclidean distance transformation.

[0013] Preferably, loading a pre-defined leaf cluster geometric model by instantiation at the end node position of the topology skeleton line includes: establishing a prefab library containing leaf cluster geometric models of various morphological specifications; determining the model index value of each end node according to the end distribution density parameter in the geometric growth control parameters; calling the corresponding leaf cluster geometric model from the prefab library according to the model index value, and binding its coordinate transformation matrix to the coordinate system of the end node of the topology skeleton line.

[0014] Preferably, determining whether the projection consistency deviation meets the preset convergence condition includes: calculating the rate of change of the projection consistency deviation of the current iteration round and the projection consistency deviation of the previous iteration round; when the projection consistency deviation is less than the first absolute threshold or the rate of change is less than the second relative threshold and is maintained for a preset number of consecutive iterations, it is determined that the preset convergence condition is met; if the above condition is not met even after reaching the preset maximum number of iterations, the iteration is terminated and the current optimal parameter set is output.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In a pine forest captured by UAV photography, a two-dimensional projection contour binary mask is transformed into a continuous distance transformation field, and a rigid weight function negatively correlated with the topological growth level of the geometric unit is constructed. When calculating projection consistency, the gradient potential energy of the distance transformation field is used to make the contour edge of the optimization algorithm differentiable. The rigid weight function is used to apply differentiated constraints to geometric units with different topological heights, so that the optimizer iteration process prioritizes locking the low-frequency steady-state trunk structure parameters and automatically reduces the sensitivity to high-frequency transient end branch displacement. The spatial anisotropy energy function is constructed to eliminate the parameter oscillation phenomenon caused by the small deformation of the target object when processing asynchronously acquired image sequences, ensuring the topological closure and geometric attitude convergence stability of the reconstruction model.

[0016] 2. In the skeleton extraction stage, a spatial index is constructed and the neighborhood density field of a single tree skeleton is calculated. A dynamic adjustment mechanism for generation parameters constrained by environmental space is established. The spatial distribution information of forest stand density is used to control the fractal iteration depth of the L-System. In areas where the neighborhood density is higher than the preset threshold, the geometric subdivision of secondary branches and mesh generation are automatically terminated, and the pre-set leaf cluster proxy is called for filling. Based on the visibility prior computing power allocation strategy, the geometric artifact accumulation problem caused by the lack of internal texture and insufficient lighting in high canopy closure areas is avoided, the number of invalid patches generated is reduced, and the visual consistency and rendering efficiency of the model's visible surface are ensured.

[0017] 3. Calculate the geometric centroid of the difference map between the virtual projection and the real contour and extract the offset vector of the centroid relative to the projection center. Introduce a negative feedback adjustment loop based on the spatial distribution characteristics of the error. Transform the spatial distribution pattern of the projection residual into displacement commands of control points with clear directionality, driving the topological skeleton to undergo Bézier bending deformation in accordance with physical laws. This enables the parametric model to break through the rigidity limitation of the preset growth rules and automatically adapt to the irregular natural posture of the target object caused by gravity or environmental stress through geometric deduction, eliminating the systematic positional deviation between the reconstructed model and the real image. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the pine forest reconstruction process that integrates distance field optimization and parameterized growth according to the present invention. Figure 2 This is a graph showing the projection bias weighting mechanism of the topological hierarchy in this invention. Figure 3 This is a diagram illustrating the key technical elements of the pine forest 3D reconstruction system of this invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0020] This embodiment provides a method for rapid reconstruction of a 3D model of a pine forest based on UAV photography, including four core processing stages: multi-viewpoint data acquisition, sparse skeleton extraction and completion, parametric growth based on environmental perception, distance field-based projection optimization, and instantiated model loading. It acquires a multi-viewpoint 2D image sequence for the target pine forest area, which contains the camera intrinsic and extrinsic parameter position information of each frame. The multi-viewpoint 2D image sequence is processed using the Structure for Motion Reconstruction (SMR) algorithm to calculate a sparse spatial point cloud. For a single pine tree target, the sparse spatial point cloud is clustered and segmented according to its spatial location, dividing the clustered point cloud space into... For a voxel mesh, calculate the geometric centroid within each valid voxel; construct a weighted undirected graph with edge weights equal to Euclidean distance using the centroids as nodes, and execute Kruskal's algorithm to generate a minimum spanning tree (MST); set a pruning threshold of [value missing]. Remove dangling paths with lengths below the threshold from the MST; use a third-order B-spline function to fit and interpolate the remaining main path node sequence, according to... Step-size resampling outputs a continuous curvature skeleton centerline. For each cluster, topological skeleton lines representing the spatial orientation of its trunk and branches are extracted. For discrete canopy regions in high-canopy closure areas, i.e., regions lacking trunk point cloud data, a topological completion operation is performed to construct a local Delaunay triangular mesh model of the top point cloud of the discrete canopy region. The Gaussian curvature distribution of this surface mesh model is calculated, and curvature extrema points with Gaussian curvature greater than a preset positive threshold and curvature ridge features with the largest rate of curvature change are identified. The curvature extrema points are determined as treetop feature points. Starting from this point, a virtual trunk path is derived along the direction of gravity, i.e., the negative Z-axis direction of the world coordinate system. Based on the projection direction of the curvature ridge features on the horizontal plane, virtual branch paths are deduced. The virtual trunk path and virtual branch path are combined to form a complete topological skeleton line.

[0021] After obtaining the topological skeleton line, a preset parametric fractal generation program is invoked. Using the topological skeleton line as the growth reference, a virtual 3D geometry is constructed based on a set of geometric growth control parameters. These geometric growth control parameters at least limit the branch iteration depth and branch deflection angle. When constructing the virtual 3D geometry, the normalized height value of each growth node on the topological skeleton line relative to the starting root node is calculated. Construct a parameter mapping function to limit the branch deflection angle. with normalized height value The monotonic functional relationship between them is defined as a linear function in this embodiment. ,in, From a basic perspective, The droop coefficient is calculated and assigned to the branch deflection angle of the current node based on the normalized height value of the current growth node through a parameter mapping function, forming a non-uniform geometric topology structure that varies with height. To optimize the allocation of computing power in large-scale scenarios, this embodiment adopts a dynamic adjustment mechanism for generation parameters based on environmental spatial constraints. Based on the geometric centroids of all extracted topological skeleton lines, a Kd-Tree spatial index structure is constructed. For the currently processed single pine tree, the geometric centroid is queried within a preset spatial radius. The number of neighboring skeletons within a given area, i.e., the neighborhood density value. In this embodiment, a preset spatial radius is used. Set to 5.0 meters, construct an iteration depth control function, and limit the branch iteration depth to have a negative correlation with the neighborhood density value. When the value is less than 3, the branch iteration depth is set to 5 levels to preserve the edge details of isolated trees; when the neighborhood density value is less than 3, the branch iteration depth is set to 5 levels to preserve the edge details of isolated trees. When the value is greater than 10, the branch iteration depth is set to level 2. The invisible internal structure of the high-density region is geometrically simplified. The calculated branch iteration depth is used as the termination threshold for fractal generation to construct a virtual three-dimensional geometry.

[0022] Semantic segmentation and denoising are performed on the pine tree entity contours in a multi-view 2D image sequence. A binarized semantic mask is extracted as the pine tree entity contour. Euclidean distance transform is performed on the pine tree entity contour to generate a distance field map, in which the pixel values ​​are... Defined as the Euclidean distance from the pixel to the nearest contour boundary, the virtual 3D geometry is projected onto the distance field map, and the distance sampling value of each projected pixel in the distance field map is determined. A topology-level weighted mechanism is introduced to calculate projection consistency deviation and determine the topology growth level of each projected pixel in the virtual 3D geometry. The main trunk corresponds to level 0, and the terminal branches and leaves correspond to level 0. Construct hierarchical weight functions This function defines a negative correlation between the weighting coefficients and the topology growth level; the lower the topology growth level, the larger the assigned weighting coefficients. In this embodiment, the formula is used. Calculate the weighting factors, where As a preset constant, weighting coefficients are assigned according to the topology growth level corresponding to each projected pixel. The sum of the products of the distance sample value and the weighting coefficients is calculated to obtain the projection consistency deviation. That is, following the formula .

[0023] Determine whether the projection consistency deviation meets the preset convergence condition. The convergence condition is set to the rate of change of the projection consistency deviation being less than a certain value for three consecutive iterations. When the convergence condition is not met, a reference angle including branch deflection is constructed. With droop coefficient control parameter vector The numerical differential center difference method is used to set the perturbation step size. rad, in turn Positive and negative perturbations are applied to each component to trigger geometry reconstruction and range field sampling, and the projection deviation is calculated. Construct the gradient vector relative to the approximate values ​​of the partial derivatives of each parameter component. Introducing the learning rate With momentum decay factor According to the formula Calculate the update amount; iteratively correct the geometric growth control parameters based on the gradient direction of the projection consistency deviation, and simultaneously perform non-rigid correction based on the residual space distribution to generate a binary difference map representing the difference between the virtual 3D geometry projection area and the pine tree entity outline. Calculate the geometric centroid of the non-zero pixel region in the binary difference map and determine the offset vector of this geometric centroid relative to the virtual 3D geometry projection center. Using the offset vector as the deformation control parameter, a non-rigid deformation based on Bézier curves is applied to the topology skeleton line. Based on the deformed topology skeleton line and the corrected geometric growth control parameters, the virtual 3D geometry is updated until the preset convergence condition is met. Finally, the geometric growth control parameters and the topology skeleton line when the preset convergence condition is met are locked. At the end node position of the topology skeleton line, a preset leaf cluster geometry model is loaded by instantiation reference to generate the target 3D mesh model. A prefab library containing leaf cluster geometry models of various shapes and specifications is established. According to the end distribution density parameter in the geometric growth control parameters, the model index value of each end node is determined, and the corresponding leaf cluster geometry model is called from the prefab library, and its coordinate transformation matrix is ​​bound to the coordinate system of the end node of the topology skeleton line.

[0024] Example 1: This example selects a pine forest area with a canopy closure greater than 0.8 and continuous natural wind interference as the reconstruction target to verify the reconstruction capability of this method under complex occlusion and dynamic environments. A multi-viewpoint 2D image sequence for the target area is obtained. The image sequence is processed using the Structure for Motion Recovery algorithm to generate a sparse spatial point cloud, and single-tree clustering is performed. For the high canopy closure central area, due to mutual occlusion, a large amount of trunk point cloud data is missing; only the canopy top point cloud is retained. A local Delaunay triangular mesh is constructed for this discrete canopy area, and the Gaussian curvature of the mesh surface vertices is calculated. Values ​​with a Gaussian curvature greater than 0.05 are identified. The extreme point of curvature is identified as the treetop feature point. Starting from this treetop feature point, a virtual trunk path is drawn along the negative direction of the Z-axis of the world coordinate system. Based on the projection direction of the ridge feature with the largest rate of curvature change on the horizontal plane, the virtual first-level branch path is deduced, thereby completing the initial topological skeleton line without visual data.

[0025] Based on all extracted topological skeleton lines, a Kd-Tree spatial index was constructed. For a single pine tree located at the center of the stand, the neighborhood density value within a 5.0-meter radius of its geometric centroid was calculated. The neighborhood density value of this region was measured. The value is 12, which is greater than the preset high-density threshold of 10. Based on the iteration depth control function, the branching iteration depth of this plant is determined. Set to level 2, only the main trunk and first-level branches are generated, while invisible internal secondary branches are ignored. Neighborhood density values ​​are measured for a single pine tree located at the forest edge. Set the branch iteration depth to 2. Setting the level to 5 to fully preserve edge details, during the parametric growth and projection optimization stages, an Euclidean distance transformation is performed on the pine tree outline in the multi-view images to generate a distance field map. The constructed virtual 3D geometry is then projected onto the distance field map. For the terminal branches and leaves affected by the wind field, a topology-based weighted mechanism is used to calculate the projection consistency deviation. Set the hierarchical weight function as follows: For the backbone region, its topological growth level The value is 0, and a weighting coefficient of 1.0 is assigned to apply a rigid constraint on the positional deviation; for the terminal branch and leaf region, its topological growth level is... The value is 4, and the weighting coefficient is 0.2.

[0026] During the iteration process, it was found that the projection consistency deviation did not meet the convergence condition, and a systematic lateral deviation was shown in the binary difference map. The geometric centroid of the non-zero region in the difference map was calculated, and the offset vector of the centroid relative to the projection center of the virtual geometry was determined. The offset vector is used as the deformation control parameter to drive the topological skeleton line to undergo non-rigid bending based on the third-order Bézier curve until the projected profile coincides with the trunk region of the real profile. Finally, the converged geometric growth control parameters are locked. At the end node position of the topological skeleton line, based on the spatial attitude matrix of the current node, the pre-set leaf cluster geometric model is loaded through the instantiation reference technique to generate the final three-dimensional mesh model. This model contains only the trunk and outer leaf clusters in the forest stand center area, and contains a complete branching structure in the forest edge area.

[0027] Example 2: This example constructs a systematic multi-dimensional control experiment. The experimental scenario is set as a pine plantation in a mountainous terrain. The average slope of the area is about 20 degrees and the average canopy closure is 0.85. During the data acquisition period, the canopy layer is continuously disturbed by gusts of wind with a speed of 3.0 m / s to 5.5 m / s due to the valley wind effect. This condition of high-density shading and non-rigid displacement is a typical failure scenario faced by traditional multi-view geometric reconstruction algorithms. Data acquisition is carried out using a quadcopter industrial-grade drone. Its full-frame CMOS camera is positioned at a relative height of about 80 meters above the top of the canopy and acquires 512 multi-view two-dimensional image sequences with centimeter-level positioning accuracy with an 80% overlap rate in the heading and lateral directions.

[0028] To ensure the scientific validity of the core parameter settings, the reference stiffness coefficient in the geometric growth control parameters is addressed. The parameter directly determines the hierarchical weight function when the standardized calibration procedure is executed. The constraint strength for geometric units at different topological levels involves a technical trade-off between the topological rigidity of the model and the goodness of fit of the projected profile: if If the value is too low, the optimization process will excessively cater to wind noise in the image, resulting in non-physical distortions in the generated topological skeleton lines; if... If the value is too high, the model will lose its ability to fit the actual tilt posture of the tree trunk, resulting in the projection consistency deviation failing to converge. Therefore, while keeping other parameters constant, a gradient test group from 0.1 to 5.0 was set up, using the projection residual convergence value and the skeleton curvature anomaly (a second derivative statistic characterizing whether the skeleton is excessively bent) as evaluation indicators. See Table 1. The data shows that when... When the value is between 0.8 and 1.2, the system achieves the minimum value of the projection residual while maintaining the skeleton's conformity to the plant's growth pattern. This range is the optimal working window. Based on this result, this experiment will... Locked to 1.0.

[0029] Table 1: Reference Stiffness Coefficients Calibration test data table Based on this parameter benchmark, a control experiment system with three groups was constructed. Control group A adopted the traditional multi-view stereo geometry (MVS) dense reconstruction algorithm (based on the Colmap open-source framework); control group B (partially missing control group) adopted the scheme of this invention, but removed the projection optimization step based on the distance field and only used binary IoU as the loss function; the experimental group (sample group of this invention) adopted the complete technical scheme including complete distance field optimization and instantiation loading. For intermediate features in the data processing process, the distance field map generated by the experimental group clearly shows the gradient potential field that linearly increases from the contour boundary outward. This gradient field provides continuous differentiability for parameter correction of the virtual 3D geometry during the optimization process. The descent direction effectively avoids the gradient vanishing problem of binary masks when processing small displacements. Experimental results show that when processing forest edge data containing wind noise, the model generated by control group A has a large number of non-closed fragmented surfaces in the treetop area, and geometric holes caused by occlusion exist in the trunk area, making it impossible to form a complete solid model. Although control group B generates a continuous skeleton, at high wind speeds, due to the lack of flexible buffering of the distance field and hierarchical weighting constraints, its optimization algorithm oscillates when trying to forcibly match the swaying treetop contours, resulting in a systematic deviation of an average of 0.35 meters between the final converged skeleton position and the actual tree trunk. In contrast, the experimental group introduced a hierarchical weighting function to improve the performance of the model. In calculating projection consistency deviation Automatically reduce high-frequency shaking of terminal branches and leaves (topology level) The weight contribution of ) thus locks the optimizer to the low-frequency steady-state backbone structure (topology level). ).

[0030] Table 2: Comparison of Performance Indicators of Each Experimental Group Referring to Table 2, under the premise of ensuring the geometric integrity rate of the trunk reaches 100.0%, the intersection-over-union ratio (IoU) of the projected outline of the experimental group reaches 0.89, which is better than 0.65 of the control group A and 0.71 of the control group B. In particular, in terms of the average time for single tree reconstruction, the experimental group only needs 19.1 seconds, which is about 15% of the control group A (125.6 seconds).

[0031] Example 3: This example combines Figures 1 to 3 This paper describes a method for rapid reconstruction of a 3D model of a pine forest based on UAV photography. Figure 1As shown, the overall process begins with the input of a multi-viewpoint 2D image sequence, including semantic mask and pose information. The process is divided into two parallel processing paths. One path performs key topological structure completion and curvature extremum identification based on sparse point cloud clustering and skeleton extraction. The other path performs Euclidean distance transformation on the pine tree entity outline to output the distance gradient potential energy field. Based on this, the system calls the parameterized fractal generation program to construct a virtual 3D geometry containing non-uniform geometric topology according to the geometric growth control parameters. The core convergence point of the process is the calculation of projection consistency deviation. This step uses a weight function based on the topological growth level, combined with the distance field map and the virtual 3D geometry to calculate the deviation value. Entering the judgment stage, if the preset convergence condition is not met, the parameter iteration correction and calibration stage is entered. The growth parameters are corrected according to the deviation gradient and non-rigid correction based on Bessel is performed. The updated parameters are fed back to the fractal generation program for iterative iteration. If the convergence condition is met, the leaf cluster instantiation and loading stage is entered. By calling the prefab library and binding the coordinate system, the target 3D mesh model that achieves high-fidelity topological inverse reconstruction is finally output.

[0032] like Figure 2 As shown, the horizontal axis represents the topology growth level. The vertical axis represents the normalized measure of weights or distance values; the solid line represents the hierarchical weight function W( Its value varies with the level. The increase in [value] shows a monotonically decreasing trend, indicating that lower-level regions are assigned higher weights; the dashed line represents the distance sampling value. The example distribution, whose values ​​vary with hierarchy. The increase shows an upward trend; the dotted line represents the weighted contribution. ×W( );like Figure 3 As shown, a fishbone diagram illustrates the system's technical architecture and key factor attribution for reconstructing a high-fidelity 3D model of a pine forest. The main axis points to the final reconstruction goal. The data acquisition and preprocessing branch covers multi-viewpoint image sequences, binarized semantic masks, and sparse spatial point cloud elements. The skeleton extraction and completion branch details curvature feature completion of blind spots, topological skeleton line extraction, and single pine tree clustering operations. The parametric fractal growth branch integrates the L-System generation program, normalized height mapping, and neighborhood density control functions. The projection consistency optimization branch includes projection consistency deviation calculation and hierarchical weight functions. The Euclidean distance transformation DistanceField, in the closed-loop feedback correction branch, involves non-rigid Bezier deformation, calculation of the centroid of the binary difference map, and parameter iterative optimization. In the model instantiation branch, it includes coordinate transformation matrix binding, end node indexing, and calling the leaf cluster prefab library.

[0033] Example 4: This example provides targeted supplementation and transparent description of the core mechanism, key parameter calibration procedure, and numerical calculation boundary of the distance field weighted projection optimization algorithm. It addresses the potential algorithmic path black boxes and parameter setting black boxes regarding the details of distance field generation, the specific calculation logic of weighting coefficients, and the quantization criteria for convergence judgment in the previous examples. During the construction of the distance field map, the system receives a binarized pine tree outline mask image as the original input. For each background pixel in the mask image... The system needs to calculate the pixel points of the nearest foreground element, i.e., the outline of the pine tree. To ensure real-time performance in large-scale image processing, a chamfered distance transformation algorithm is employed, defining the distance metric between pixels as Euclidean distance. The distance value of each pixel is propagated and updated through two raster scans: a forward scan and a backward scan. The forward scan mask uses four neighboring points: top left, top, top right, and left; the backward scan mask uses four neighboring points: bottom left, bottom, and bottom right. The resulting distance field map... In the middle, pixel value Store the normalized distance values, whose range is: , where 1.0 corresponds to the preset ratio of the image diagonal length.

[0034] Regarding the calculation logic of the weighting coefficients, this embodiment defines explicit topological hierarchy mapping rules and a weight decay model. In the fractal generation process, each geometric unit, such as a branch segment or leaf cluster, is based on a unique topological hierarchy index. The main stem is level 0, the first-level branches are level 1, and so on down to the terminal leaf clusters. Level, weight function The design follows the physical constraint principle of rigid core-flexible end, specifically a variant of the inverse hyperbolic tangent function, i.e. ,in, The overall constraint strength factor. As the hierarchical decay rate factor, in this embodiment, it is determined through offline calibration experiments when... and At that time, the model can most effectively tolerate the random wind displacement of the terminal branches while maintaining the rigidity of the main trunk; regarding the quantification criteria for convergence judgment, the system calculates the projection consistency deviation in each iteration. relative rate of change Set the judgment window size to The next iteration, when continuous The next iteration All are less than the convergence threshold (This embodiment is set as) Only when the condition is met will the system determine that the optimization process has converged and terminate the iteration.

[0035] Example 5: This example addresses the issue of missing key parameter settings during system initialization by constructing an offline calibration and data filling procedure. It ensures that the core parameters in the parameter mapping function and weight decay model have reproducible engineering sources. This procedure establishes a standardized parameter library construction process for different pine species, such as Masson pine, Chinese pine, and Korean pine, based on their morphological differences. The procedure sets a standardized data acquisition environment. In a controlled experimental forest, representative mature pine trees are selected as calibration samples. Point cloud data of the samples is acquired using a terrestrial 3D lidar (TLS), and their true topological skeleton structure is manually extracted, including the length, angle, and hierarchical relationship of the trunk and branches at all levels. For each pine species, data from no fewer than 50 samples are collected. The extracted true topological skeleton is normalized, and the statistical characteristics of the deflection angle distribution of branches at all levels relative to their parent branches are calculated. The parameter mapping function is determined by least squares fitting. The basic perspective With droop coefficient The optimal value, for example, for Masson pine, the calibration result is , .

[0036] And the hierarchical weight function parameters in distance field optimization This specification is designed based on dynamic response testing using wind tunnel simulation. A constructed virtual three-dimensional geometry is placed in a virtual wind tunnel environment, and different levels of wind loads, such as winds from level 3 to level 8, are applied. The displacement deviation of the virtual geometry's terminal branches and leaves relative to actual observation data is monitored, and the layer attenuation rate factor is adjusted accordingly. To find the weighted projection consistency deviation The parameter configuration, which maintains the minimum value under all wind levels, is the final calibration parameter for each pine species. The parameters are stored in the system's preset parameter library. In actual applications, the system directly calls the corresponding parameter group from the library based on the tree species information entered by the user or the tree species category automatically determined by image recognition.

[0037] Example 6: This example constructs a standardized pre-deployment calibration and model building procedure to address the problem of adaptive adjustment of system initialization parameters and determination of the optimal operating point under different geographical regions, seasons, and sensor configurations. This procedure is designed as the initialization process executed after the system's initial deployment or major hardware changes. The procedure performs a dynamic calibration step for sensor intrinsic parameters. An open area with rich geometric features, such as a calibration board or a man-made structure of known size, is selected at the edge of the target forest area as the calibration field. The UAV is controlled to collect at least 50 calibration image sequences at different altitudes (30 meters to 100 meters) and different pitch angles (-90 degrees to -45 degrees). The camera's focal length is calculated using the Zhang Zhengyou calibration method based on checkerboard detection. Principal point coordinates and radial distortion coefficient The system automatically evaluates the calibration reprojection error. If the average error exceeds 0.5 pixels, the operator is prompted to re-execute the acquisition task until the accuracy requirements are met. This step ensures the accuracy of the geometric reference for subsequent sparse point cloud generation and distance field projection calculation.

[0038] The system also performs an environmental light adaptability threshold calibration step. Samples of the tree canopy are collected under different lighting conditions, such as early morning, noon, and dusk. For each lighting condition, color histogram features of the vegetation area in the image are extracted, and the dynamic binarization threshold for semantic segmentation is automatically calculated using Otsu's method. The system constructs a lookup table mapping between light intensity and segmentation threshold. In actual operation, the system uses the ambient light intensity value read in real time by the airborne light sensor to interpolate from the lookup table to obtain the current optimal segmentation threshold. To ensure the integrity and edge stability of the extracted pine tree outlines in the field under drastically changing lighting conditions, a regional adaptive fine-tuning step for fractal growth parameters is performed. Three to five representative individual pine trees are selected as reference samples within the area to be reconstructed. High-precision point cloud ground truth data is acquired using an airborne LiDAR. The system automatically runs a parameterized fractal generation program, using a Bayesian optimization algorithm as its core and the chamfer distance between the reference sample point cloud and the virtual geometry as the objective function. Within a preset parameter space, the optimal set of geometric growth control parameters (including branching iteration depth, branching angle, length attenuation rate, etc.) for the pine tree population in the area is searched. When the objective function value converges to a preset tolerance range, this set of parameters is locked as the global default configuration for the current survey area.

[0039] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for rapid reconstruction of a 3D model of a pine forest based on UAV photography, characterized in that, Includes the following steps: Step 101: Obtain a multi-view two-dimensional image sequence for the target pine forest area, calculate a sparse spatial point cloud based on the multi-view two-dimensional image sequence, and cluster the sparse spatial point cloud for a single pine tree target to extract the topological skeleton line representing the spatial orientation of its trunk and main branches. Step 102: Call the preset parametric fractal generation program, use the topological skeleton line as the growth reference, and construct a virtual three-dimensional geometry based on a set of geometric growth control parameters. The geometric growth control parameters at least limit the branch iteration depth and branch deflection angle. Step 103: Perform Euclidean distance transformation on the pine tree entity outline in the multi-view 2D image sequence to generate a distance field map representing the distance gradient from the pixel to the outline boundary. The pixel value in the distance field map defines the Euclidean distance from the pixel to the nearest outline boundary. Step 104: Project the virtual 3D geometry onto the distance field map, determine the distance sampling value of each projected pixel in the distance field map, and calculate the projection consistency deviation based on the weighted sum of the distance sampling values; Step 105: Determine whether the projection consistency deviation meets the preset convergence condition. If the projection consistency deviation does not meet the preset convergence condition, iteratively correct the geometric growth control parameters according to the gradient direction of the projection consistency deviation, and update the virtual three-dimensional geometry based on the corrected geometric growth control parameters until the preset convergence condition is met. Step 106: Lock the geometric growth control parameters and topology skeleton line that meet the preset convergence conditions, and load the preset leaf cluster geometric model by instantiating the reference at the end node position of the topology skeleton line to generate the target three-dimensional mesh model.

2. The method for rapid reconstruction of a 3D model of a pine forest based on UAV photography according to claim 1, characterized in that, A virtual 3D geometry is constructed based on a set of geometric growth control parameters, including: calculating the normalized height value of each growth node on the topological skeleton line relative to the starting root node; constructing a parameter mapping function, which defines the monotonic functional relationship between the branch deflection angle and the normalized height value; when constructing the virtual 3D geometry, the branch deflection angle of the current node is calculated and assigned by the parameter mapping function based on the normalized height value of the current growth node, so as to form a non-uniform geometric topology structure that varies with height.

3. The method for rapid reconstruction of a 3D model of a pine forest based on UAV photography according to claim 1, characterized in that, The projection consistency deviation is calculated based on the weighted sum of distance sample values, including: determining the topology growth level of each projected pixel in the virtual 3D geometry, where the topology growth level represents the iteration depth of the geometric unit to which the projected pixel belongs in the fractal generation process; constructing a hierarchy weight function, which limits the negative correlation between the weighting coefficients and the topology growth level; assigning weighting coefficients according to the topology growth level corresponding to each projected pixel, where regions with lower topology growth levels are assigned larger weighting coefficients; and calculating the projection consistency deviation by weighted summing of the distance sample values ​​based on the weighting coefficients.

4. The method for rapid reconstruction of a 3D model of a pine forest based on UAV photography according to claim 1, characterized in that, To extract topological skeleton lines representing the spatial orientation of the trunk and main branches of a single pine tree by clustering sparse spatial point clouds, the method includes: detecting whether there are discrete canopy regions with missing trunk point cloud data in the sparse spatial point cloud; if discrete canopy regions exist, constructing a surface mesh model for the top point cloud of the discrete canopy regions; calculating the Gaussian curvature distribution of the surface mesh model and identifying curvature extrema and curvature ridge features; using the curvature extrema as the starting endpoint, drawing virtual trunk paths along the direction of gravity, and inferring virtual branch paths based on the spatial mapping relationship of curvature ridge features; and combining the virtual trunk paths and virtual branch paths as topological skeleton lines to complete the topological structure of visual blind spots.

5. The method for rapid reconstruction of a 3D model of a pine forest based on UAV photography according to claim 1, characterized in that, The process of iteratively correcting the geometric growth control parameters based on the gradient direction of the projection consistency deviation also includes a non-rigid correction step based on the residual space distribution: generating a binary difference map representing the difference between the projection area of ​​the virtual 3D geometry and the outline of the pine tree entity; calculating the geometric centroid of the non-zero pixel area in the binary difference map and determining the offset vector of the geometric centroid relative to the projection center of the virtual 3D geometry; using the offset vector as a deformation control parameter to apply non-rigid deformation based on Bézier curves to the topological skeleton line; and re-executing the construction and projection calculation steps of the virtual 3D geometry based on the deformed topological skeleton line.

6. The method for rapid reconstruction of a 3D model of a pine forest based on UAV photography according to claim 1, characterized in that, Before constructing a virtual three-dimensional geometry based on a set of geometric growth control parameters, the process includes: constructing a spatial index structure based on all extracted topological skeleton lines; calculating the neighborhood density value of the topological skeleton line of each individual pine tree within a preset spatial radius; constructing an iterative depth control function, which limits the negative correlation between the branch iteration depth and the neighborhood density value; using the branch iteration depth calculated based on the neighborhood density value as the fractal generation termination threshold for constructing the corresponding virtual three-dimensional geometry; and performing geometric simplification on the invisible internal structure located in the high-density region.

7. The method for rapid reconstruction of a 3D model of a pine forest based on UAV photography according to claim 3, characterized in that, The distance sample values ​​are weighted and summed based on weighting coefficients, following the weighted bias calculation formula: ,in, For projection consistency deviation, This represents the total number of projected pixels. For the first The distance sample value of each projected pixel in the distance field map. For the first The topology growth level corresponding to each projected pixel. The corresponding level is determined based on the hierarchical weight function. The weighting coefficients.

8. The method for rapid reconstruction of a 3D model of a pine forest based on UAV photography according to claim 1, characterized in that, Acquiring a multi-view 2D image sequence for the target pine forest region includes: performing semantic segmentation on the multi-view 2D image sequence and extracting a binarized semantic mask containing pine tree targets; removing discrete connected components with an area smaller than a preset noise threshold from the binarized semantic mask; and using the noise-removed binarized semantic mask as the outline of the pine tree entity for subsequent Euclidean distance transformation.

9. The method for rapid reconstruction of a 3D model of a pine forest based on UAV photography according to claim 1, characterized in that, At the end node position of the topology skeleton line, a pre-defined leaf cluster geometric model is loaded by instantiation and reference, including: establishing a prefab library containing leaf cluster geometric models of various morphological specifications; determining the model index value of each end node according to the end distribution density parameter in the geometric growth control parameters; calling the corresponding leaf cluster geometric model from the prefab library according to the model index value, and binding its coordinate transformation matrix to the coordinate system of the end node of the topology skeleton line.

10. A method for rapid reconstruction of a 3D model of a pine forest based on UAV photography according to claim 1, characterized in that, Determining whether the projection consistency deviation meets the preset convergence condition includes: calculating the rate of change of the projection consistency deviation of the current iteration round compared with the projection consistency deviation of the previous iteration round; when the projection consistency deviation is less than the first absolute threshold or the rate of change is less than the second relative threshold and is maintained for a preset number of consecutive iterations, it is determined that the preset convergence condition is met; if the above condition is not met even after reaching the preset maximum number of iterations, the iteration is terminated and the current optimal parameter set is output.

Images