A method for generating a three-dimensional model based on unmanned aerial vehicle oblique photogrammetry

Abstract

The application relates to the technical field of unmanned aerial vehicle photogrammetry, and discloses a three-dimensional model generation method based on unmanned aerial vehicle oblique photogrammetry. The method acquires an original oblique photogrammetry image set of an unmanned aerial vehicle and a topographic parameter set of a region to be modeled; performs multi-scale preprocessing on the original image set to obtain a first and a second scale feature image set; extracts a texture feature evolution sequence from the second scale feature image set, combines the topographic parameter set to complete shooting constraint initialization, and outputs an initial control constraint group; performs spatial feature reconstruction on the first scale feature image set according to the constraint group to obtain a three-dimensional topological feature vector set; performs multi-dimensional registration constraint on the vector set based on a similarity threshold to locate an optimal registration path; performs topological correction in combination with the path and the topographic parameter set to output a corrected three-dimensional feature set; performs precision normalization processing on the three-dimensional feature set to generate a precision coefficient; if a preset condition is not met, the texture feature evolution sequence is incrementally updated; and if the condition is met, a target three-dimensional mesh model is output.

Description

A method for generating 3D models based on UAV oblique photogrammetry Technical Field

[0001] This invention relates to the field of UAV photogrammetry technology, specifically a method for generating three-dimensional models based on UAV oblique photogrammetry. Background Technology

[0002] In the field of 3D spatial modeling, UAV oblique photogrammetry technology has been widely applied in scenarios such as topographic mapping, architectural planning, and disaster assessment due to its flexible operation and efficient image acquisition capabilities. Currently, 3D model generation methods based on this technology typically involve first acquiring the original oblique images taken by the UAV, and then constructing the model through steps such as feature extraction and spatial registration. However, there are still many limitations in practical applications.

[0003] Existing methods for preprocessing raw oblique photogrammetric images often employ a single-scale approach, making it difficult to simultaneously capture both detailed features and overall structure. Focusing solely on detailed features can lead to insufficient understanding of the overall image contours, resulting in localized deviations during subsequent spatial reconstruction. Conversely, prioritizing overall structure can result in the loss of crucial details, rendering the generated 3D model inadequate in terms of precision. This issue is particularly pronounced in scenarios requiring high detail in architectural facade textures and micro-topography.

[0004] Traditional methods often rely solely on imagery information or simple terrain parameters for initial image constraint setting, neglecting to consider the evolution of texture features. This leads to insufficient accuracy in the initial control constraint set. During subsequent spatial feature reconstruction and multi-dimensional registration, the initial constraints' bias necessitate multiple adjustments to the registration path, increasing computational costs and potentially causing a decline in overall model consistency due to repeated corrections. Furthermore, existing methods often rely on a single metric for model accuracy verification, lacking a dynamic accuracy adjustment mechanism. When model accuracy fails to meet preset conditions, targeted incremental updates to key aspects are impossible, necessitating a complete re-execution of the modeling process, significantly reducing modeling efficiency.

[0005] In complex terrains or scenes, existing methods have weak multidimensional registration constraints. Due to large terrain undulations or numerous occlusions in the scene, the overlapping areas between images vary greatly. Traditional registration methods struggle to accurately find the optimal registration path, easily leading to registration misalignment. This results in topological defects in the generated 3D model, such as terrain breaks and misaligned building components, severely impacting the model's usability. Summary of the Invention

[0006] The purpose of this invention is to provide a method for generating three-dimensional models based on UAV oblique photogrammetry, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a method for generating a 3D model based on UAV oblique photogrammetry, the method comprising:

[0008] Acquire the raw oblique photographic image set from the UAV and the terrain parameter set of the area to be modeled;

[0009] Perform multi-scale preprocessing on the original oblique photogrammetric image set to obtain a first-scale feature image set and a second-scale feature image set;

[0010] Extract texture feature evolution sequences from the second-scale feature image set; perform shooting constraint initialization based on the texture feature evolution sequences and terrain parameter set, and output an initial control constraint set; perform spatial feature reconstruction on the first-scale feature image set according to the initial control constraint set to obtain a three-dimensional topological feature vector set; perform multi-dimensional registration constraints on the three-dimensional topological feature vector set based on a similarity threshold to locate the optimal registration path; perform topological correction according to the optimal registration path and terrain parameter set, and output a corrected three-dimensional feature set.

[0011] The accuracy normalization process is performed on the corrected 3D feature set to generate 3D model accuracy coefficients; when the 3D model accuracy coefficients do not meet the preset model update conditions, the incremental update of the texture feature evolution sequence is triggered; when the 3D model accuracy coefficients meet the preset model update conditions, the target 3D mesh model is output.

[0012] Preferably, the method for performing multi-scale preprocessing on the original oblique photogrammetric image set is as follows:

[0013] The original oblique photogrammetric image set is divided into multiple image blocks; a resolution dimensionality reduction operation is performed on each image block to generate a basic scale image unit set; an edge enhancement operation is performed on the basic scale image unit set to obtain the first scale feature image set; spatial pyramid pooling is performed on the first scale feature image set to output the second scale feature image set.

[0014] Preferably, the method for extracting the texture feature evolution sequence from the second-scale feature image set is as follows:

[0015] Calculate the texture gradient change of adjacent image blocks in the second-scale feature image set; construct a spatiotemporal evolution matrix based on the texture gradient change; perform feature trajectory tracking on the spatiotemporal evolution matrix to generate the texture feature evolution sequence.

[0016] Preferably, the method for performing shooting constraint initialization based on the texture feature evolution sequence and terrain parameter set is as follows:

[0017] Elevation constraints and illumination constraints are determined based on the terrain parameter set; the texture feature evolution sequence is mapped and matched with the elevation constraints to generate a first constraint parameter set; the texture feature evolution sequence is mapped and matched with the illumination constraints to generate a second constraint parameter set; the first constraint parameter set and the second constraint parameter set are fused to output the initial control constraint set.

[0018] Preferably, the method for performing spatial feature reconstruction on the first scale feature image set based on the initial control constraint group is as follows:

[0019] The initial control constraint set is decomposed into a geometric constraint subset and a texture constraint subset; the geometric constraint subset is used to perform vertex topology reconstruction on the first scale feature image set to obtain an initial three-dimensional vertex set; the texture constraint subset is used to perform feature fuzzy clustering on the initial three-dimensional vertex set to generate the three-dimensional topological feature vector set.

[0020] Preferably, the method for performing multidimensional registration constraints on the three-dimensional topological feature vector set based on a similarity threshold is as follows:

[0021] Calculate the Euclidean distance matrix of each vector in the three-dimensional topological feature vector set; filter the effective matching pairs of the Euclidean distance matrix according to the similarity threshold; perform affine transformation iteration on the effective matching pairs to generate a candidate registration path set; select the path with the minimum error from the candidate registration path set as the optimal registration path.

[0022] Preferably, the method for performing topology correction based on the optimal registration path and terrain parameter set is as follows:

[0023] Extract the coordinate transformation parameters of the optimal registration path; generate an elevation correction value based on the terrain parameter set; perform weighted fusion of the coordinate transformation parameters and the elevation correction value to obtain a spatial correction operator; use the spatial correction operator to perform remapping on the three-dimensional topological feature vector set to output the corrected three-dimensional feature set.

[0024] Preferably, the method for performing accuracy normalization processing on the corrected three-dimensional feature set is as follows:

[0025] Calculate the local overlap distribution of the corrected 3D feature set; determine the maximum and minimum overlap deviations based on the local overlap distribution; perform normalization mapping based on the maximum and minimum overlap deviations to generate the accuracy coefficients of the 3D model.

[0026] Preferably, the method for triggering incremental updates of the texture feature evolution sequence is as follows:

[0027] Locate the low-confidence feature segments in the texture feature evolution sequence; generate new sampling point coordinates based on the terrain parameter set; collect supplementary oblique image blocks according to the new sampling point coordinates; and fuse the supplementary oblique image blocks into the texture feature evolution sequence.

[0028] The preferred method for outputting the target 3D mesh model is as follows:

[0029] Perform Delaunay triangulation on the corrected 3D feature set; adjust the density of triangular facets according to the accuracy coefficient of the 3D model; perform feature boundary optimization on the triangulation results to generate the target 3D mesh model.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] By preprocessing the original oblique photogrammetric image set at multiple scales, a first-scale feature image set and a second-scale feature image set can be obtained separately, effectively balancing the extraction requirements of image detail features and overall structure. The feature image sets at different scales can serve different subsequent processing stages. The first-scale feature image set is more suitable for spatial feature reconstruction to ensure the accuracy of the overall model structure, while the second-scale feature image set helps extract more refined texture feature evolution sequences, providing reliable data support for subsequent shooting constraint initialization. This avoids the problem of single-scale preprocessing failing to balance detail and overall structure, allowing subsequent modeling stages to be carried out based on adapted image data, improving modeling quality from the data source.

[0032] In the initialization phase of the shooting constraints, this method combines the texture feature evolution sequence and the terrain parameter set to perform the initialization operation. Compared with the traditional method that relies on only a single information source, it can make full use of the evolution law of texture features with shooting angle and terrain changes, making the initial control constraint set more consistent with the actual shooting scene and terrain conditions. An accurate initial control constraint set can directly improve the accuracy of subsequent spatial feature reconstruction, reduce local deviations in the spatial reconstruction process, and make the constructed 3D topological feature vector set more realistically reflect the spatial relationship of the area to be modeled. This lays a good foundation for subsequent multi-dimensional registration constraints and avoids the problem of repeated adjustments in subsequent steps due to initial constraint deviations.

[0033] Based on the initial control constraint set, spatial feature reconstruction is performed on the first-scale feature image set. This allows for the construction of a 3D topological feature vector set using reliable constraints. Then, a similarity threshold is used to apply multi-dimensional registration constraints to this vector set to locate the optimal registration path, effectively improving the accuracy and efficiency of registration. In complex terrain or occluded scenes, similarity threshold filtering can accurately identify the registration relationship with the highest matching degree, avoiding registration misalignment and ensuring the integrity and consistency of the 3D model's topological structure. This solves the problem of traditional registration methods struggling to find the optimal registration path in complex scenes.

[0034] Topology correction is performed based on the optimal registration path and terrain parameter set, outputting a corrected 3D feature set. This further corrects subtle deviations that may exist during spatial reconstruction and registration, ensuring the rationality of the 3D feature set in terms of topological structure. Subsequent accuracy normalization processing generates intuitive 3D model accuracy coefficients. A dynamic adjustment mechanism based on these coefficients can trigger incremental updates of the texture feature evolution sequence when the accuracy does not meet preset conditions. This eliminates the need to re-execute the entire modeling process, optimizing only the key texture feature stages, significantly reducing computational costs and improving modeling efficiency. When the accuracy meets the conditions, the target 3D mesh model is output, ensuring rapid model generation while maintaining accuracy standards, meeting the dual requirements of modeling efficiency and quality in practical applications.

[0035] The entire methodology forms a complete closed-loop processing mechanism. From image preprocessing to model output, each step is interconnected and mutually supportive. The processing results of the previous step provide a reliable basis for the next step, and the feedback from the next step can be used for targeted optimization when the accuracy of the previous step is insufficient. This closed-loop mechanism makes the modeling process highly adaptable and flexible. Whether in simple geometric scenes or complex terrain scenes, whether in rough modeling with low accuracy requirements or professional scenes with extremely high precision requirements, it can stably generate 3D models that meet the requirements, thus broadening the application scope of UAV oblique photogrammetry technology in the field of 3D modeling. Attached Figure Description

[0036] Figure 1 is a schematic diagram illustrating the working principle of the three-dimensional model generation method based on UAV oblique photogrammetry described in this invention;

[0037] Figure 2 is a flowchart of the multi-scale preprocessing method;

[0038] Figure 3 is a flowchart of the spatial feature reconstruction method;

[0039] Figure 4 is a flowchart of the multidimensional registration constraint method;

[0040] Figure 5 is a flowchart of the incremental update method for texture feature evolution sequence. Detailed Implementation

[0041] 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, 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.

[0042] Please refer to Figure 1. This invention provides a method for generating a 3D model based on UAV oblique photogrammetry, the method comprising:

[0043] The system acquires the original oblique photogrammetric image set from the UAV and the terrain parameter set of the area to be modeled; performs multi-scale preprocessing on the original oblique photogrammetric image set to obtain a first-scale feature image set and a second-scale feature image set; extracts the texture feature evolution sequence from the second-scale feature image set; performs shooting constraint initialization based on the texture feature evolution sequence and the terrain parameter set, and outputs an initial control constraint set; performs spatial feature reconstruction on the first-scale feature image set according to the initial control constraint set to obtain a three-dimensional topological feature vector set; performs multi-dimensional registration constraints on the three-dimensional topological feature vector set based on a similarity threshold to locate the optimal registration path; performs topological correction based on the optimal registration path and the terrain parameter set, and outputs a corrected three-dimensional feature set; performs accuracy normalization processing on the corrected three-dimensional feature set to generate three-dimensional model accuracy coefficients; when the three-dimensional model accuracy coefficients do not meet the preset model update conditions, triggers incremental updates of the texture feature evolution sequence; when the three-dimensional model accuracy coefficients meet the preset model update conditions, outputs the target three-dimensional mesh model.

[0044] Example 1: Referring to Figure 2, a method for generating a 3D model based on UAV oblique photogrammetry is described. In practice, it requires acquiring the original oblique photogrammetry image set from the UAV and the terrain parameter set of the area to be modeled. The original oblique photogrammetry image set is collected by multi-angle sensors onboard the UAV during flight operations, containing highly overlapping digital images of the target area taken from different angles. The terrain parameter set typically originates from prior geographic information system data or auxiliary data collected during flight, and includes key geographic parameters such as elevation information, land cover type, and illumination angle.

[0045] Multi-scale preprocessing of the original oblique photogrammetric image set is the fundamental step in the entire workflow. After the original image set is imported into the processing system, it is automatically segmented into multiple image blocks with regular grid divisions based on the image's geographical location and capture timestamp information. This segmentation process considers image resolution, the continuity of ground features, and the efficiency of subsequent processing; the size of each image block is optimized to balance computational resources and accuracy. Subsequently, a resolution reduction operation is performed on each individual image block. This operation is achieved through image resampling technology, using a bilinear interpolation algorithm to convert high-resolution image pixels into lower-resolution pixel units, significantly reducing the data volume while preserving necessary texture details, thereby generating a basic scale image unit set.

[0046] Building upon this foundation, edge enhancement operations are performed on the base-scale image unit set. This operation employs the Laplacian edge detection operator, using convolution calculations to enhance high-frequency information such as feature boundaries and building outlines in the image. The edge enhancement process not only highlights structural features in the image but also effectively suppresses illumination noise in uniform regions, resulting in a first-scale feature image set. This first-scale feature image set combines low data complexity with enhanced feature representation.

[0047] Furthermore, spatial pyramid pooling is performed on the first-scale feature image set. Spatial pyramid pooling extracts scale-invariant feature representations by pooling the feature images across multiple windows of different scales. Max pooling is then used to extract the most salient feature responses from each local region, thus transforming the first-scale feature image set into a second-scale feature image set in the form of a multi-resolution feature pyramid. The second-scale feature image set not only preserves the spatial structure of the original images but also enhances the representation of features at different scales.

[0048] Extracting the texture feature evolution sequence from the second-scale feature image set is crucial for subsequent geometric modeling. This process first calculates the texture gradient changes between adjacent image blocks. The Sobel operator is used to calculate the gradient components of each pixel in the horizontal and vertical directions, thus obtaining the gradient magnitude and direction. The gradient changes between adjacent blocks reflect the spatial continuity and evolution trend of texture features. Based on the calculated texture gradient changes, a spatiotemporal evolution matrix is ​​constructed. The rows and columns of this matrix correspond to the spatial coordinates of the image blocks and the acquisition time series, respectively, while the matrix elements record the spatiotemporal intensity of texture feature changes.

[0049] Feature trajectory tracing is performed on the spatiotemporal evolution matrix to generate a texture feature evolution sequence. An eigenvalue decomposition-based tracing method is employed to extract key eigenvectors from the matrix. These eigenvectors describe the evolution path of texture features in the spatiotemporal dimension. The evolution sequence records how key texture features change with spatial location and time in the form of sequence data, providing a quantitative basis for describing the dynamic changes of textures for subsequent initialization of shooting constraints.

[0050] Example 2: Referring to Figure 3, the process of initializing shooting constraints based on the texture feature evolution sequence and terrain parameter set is a key step connecting image feature analysis and 3D geometric reconstruction. This process begins with the parsing of the terrain parameter set, which includes geographical and lighting parameters such as elevation data, surface material information, solar azimuth and elevation angles of the area to be modeled. Elevation constraints are extracted from the terrain parameter set, and an elevation distribution map is generated through a digital elevation model. This map records the elevation values ​​of each point on the surface in a grid format. Illumination constraints are calculated based on the solar azimuth and elevation angles, combined with the shooting timestamp and geographical location, to calculate the illumination intensity distribution model of each image block at the time of acquisition. This model considers the combined effects of direct and diffused light.

[0051] When mapping and matching the texture feature evolution sequence with elevation constraints, a least-squares fitting algorithm is employed. The texture feature evolution sequence provides the spatiotemporal variation patterns of image texture, while the elevation constraints describe the topographic relief characteristics. By finding the statistical correlation between texture changes and topographic relief, a first set of constraint parameters is generated. This set of parameters includes an elevation adjustment coefficient to correct texture distortion caused by topographic relief, and a position offset to compensate for spatial position deviations caused by changes in shooting angle. The mapping and matching process iteratively optimizes the process to achieve the best statistical fit between texture features and elevation data.

[0052] When mapping and matching texture feature evolution sequences with lighting constraints, a photometric stereo vision method is applied. The texture feature evolution sequence reflects the changes in texture appearance under different lighting conditions, while the lighting constraints provide the theoretical distribution of lighting intensity. By comparing the differences between actual texture changes and the theoretical lighting model, a second set of constraint parameters is generated. This set of parameters includes lighting compensation coefficients to adjust for differences in texture brightness caused by uneven lighting, and shadow correction parameters to eliminate shadow effects caused by terrain occlusion. The matching process normalizes the lighting conditions by establishing a correspondence between texture brightness and lighting intensity.

[0053] When fusing the first and second constraint parameter sets, a weighted average method is used. Weights are assigned based on the confidence index of the parameter sets, with the confidence level calculated using matching error and parameter stability. The elevation adjustment coefficient and illumination compensation coefficient are linearly weighted, and the position offset and shadow correction parameters are vector-synthesized to finally output the initial control constraint set. This constraint set comprehensively expresses the geometric and photometric constraints of terrain and illumination conditions on image acquisition, providing control parameters for subsequent spatial feature reconstruction.

[0054] When performing spatial feature reconstruction on the first-scale feature image set based on the initial control constraint set, the initial control constraint set is first decomposed into a geometric constraint subset and a texture constraint subset. The geometric constraint subset includes vertex position constraints, edge connectivity constraints, and patch topological constraints, which are derived from the geometric transformation of elevation adjustment coefficients and position offsets. The texture constraint subset includes texture mapping parameters, color correction coefficients, and feature weights, which are derived from illumination compensation coefficients and shadow correction parameters through coordinate transformation.

[0055] When performing vertex topology reconstruction on the first-scale feature image set using a geometrically constrained subset, the Delaunay triangulation algorithm is employed. The first-scale feature image set has already undergone edge enhancement processing and possesses clear feature boundary information. The geometrically constrained subset provides initial estimates of vertex positions and connectivity constraints, generating an initial 3D vertex set through constrained Delaunay triangulation. This vertex set preserves edge features in the image while satisfying terrain geometric constraints.

[0056] When performing feature fuzzy clustering on the initial 3D vertex set using a texture constraint subset, the K-means clustering algorithm is applied. The texture constraint subset provides the classification criteria and weight parameters for texture features, while the initial 3D vertex set carries positional and texture information. Through iterative clustering in the feature space, vertices are divided into categories with similar texture features, and each category generates a representative feature vector. These feature vectors constitute a 3D topological feature vector set, which expresses both the geometric distribution characteristics of the vertices and captures the statistical regularities of texture attributes.

[0057] Through rigorous parameter mapping and fusion, image texture features are organically combined with geographic environmental parameters, establishing a bridge for the transformation from two-dimensional image features to three-dimensional geometric features. The spatial feature reconstruction results are faithful to the texture information of the original image and conform to the physical constraints of terrain and lighting, laying a data foundation for subsequent registration and correction. All calculation processes are completed automatically by algorithms, ensuring the objectivity and consistency of the processing results.

[0058] Example 3: Referring to Figure 4, the process of performing multidimensional registration constraints on a 3D topological feature vector set based on a similarity threshold is a crucial stage in establishing accurate spatial correspondences. The 3D topological feature vector set originates from the output of spatial feature reconstruction, with each vector containing the geometric coordinates and texture feature information of its vertices. A standard distance metric is used to calculate the Euclidean distance matrix between vectors in this set. This matrix is ​​a symmetric square matrix, where the row and column indices correspond to the sequence numbers of the vectors, and the matrix element values ​​represent the Euclidean distance between the corresponding vectors in the feature space. The distance calculation encompasses a comprehensive measure of geometric coordinate differences and texture feature differences, reflecting the overall similarity between vectors.

[0059] When filtering valid matching pairs in the Euclidean distance matrix based on a preset similarity threshold, an upper limit value for distance is set as the filtering condition. This threshold is determined by analyzing the statistical distribution characteristics of the vector set, typically taking the median or a specific quantile value of the distance distribution. All vector pairs corresponding to elements in the matrix whose distances are less than the threshold are marked as candidate matching pairs, forming the initial matching set. This filtering process eliminates obviously irrelevant vector pairs, reducing subsequent computational complexity.

[0060] When performing affine transformation iterations on valid matching pairs, the transformation parameters are estimated using a random sampling consensus algorithm. In each iteration, a minimum sample set is randomly selected to calculate the affine transformation model, and the transformation error for all matching pairs is calculated using this model. The transformation error for each matching pair is calculated using the following formula:

[0061]

[0062] in: Indicates the error of the affine transformation. To match the number of pairs, Describes the affine transformation function. and They represent the first The source vector and target vector in each matching pair. This represents the Euclidean norm. Through multiple iterations, the transformation model with the most interior points is selected, generating a set of candidate registration paths. Each candidate path corresponds to a set of transformation parameters, describing the spatial mapping relationship between two sets of feature vectors.

[0063] When selecting the path with the minimum error from the candidate registration path set, the total reprojection error of each path is calculated. The error calculation comprehensively considers geometric position deviation and texture feature differences, and a comprehensive error index is obtained through weighted summation. The path with the minimum comprehensive error is selected as the optimal registration path, which represents the optimal spatial correspondence between feature vectors.

[0064] When performing topology correction based on the optimal registration path and terrain parameter set, the coordinate transformation parameters of the optimal registration path are first extracted. These parameters include the Euler angles of the rotation matrix, the three components of the translation vector, and the scale factor, which fully describe the transformation relationship of spatial coordinates. When generating elevation correction values ​​based on the terrain parameter set, the elevation correction value for each vertex is calculated through interpolation using the accuracy index of the digital elevation model and terrain complexity data. The elevation correction values ​​compensate for geometric distortions caused by terrain undulations and shooting angles.

[0065] When weighted fusion of coordinate transformation parameters and elevation corrections, weights are assigned based on parameter confidence levels. The confidence level of the coordinate transformation parameters is derived from the reciprocal of the registration error, while the confidence level of the elevation corrections is based on the accuracy rating of the terrain data. The fusion process generates a spatial correction operator that contains complete geometric transformation parameters and elevation correction terms. When remapping the 3D topological feature vector set using the spatial correction operator, coordinate transformation and elevation correction are applied simultaneously to each vector. The remapping process preserves the texture features of the vectors, adjusting only their geometric coordinate positions, ultimately outputting a corrected 3D feature set.

[0066] Through rigorous mathematical calculations and parameter optimization, precise registration and geometric correction between feature vectors were achieved. The registration constraints considered the comprehensive similarity of multi-dimensional features, while the topology correction integrated coordinate transformation and terrain constraints, resulting in a final output feature set with higher geometric consistency and topological accuracy. All computational steps were executed automatically by the algorithm, ensuring the objectivity of the processing and the repeatability of the results.

[0067] Example 4: Referring to Figure 5, when performing accuracy normalization processing on the corrected 3D feature set, it is first necessary to calculate the local overlap distribution of the feature set. Taking a 3D reconstruction project of a city building complex as an example, the corrected 3D feature set contains 125,643 feature points. These point cloud data have been processed through the aforementioned registration and correction. The local overlap distribution is calculated by dividing the entire feature set space into uniform grid cells, with each grid cell set to a cube volume of 0.5m × 0.5m × 0.5m. The system counts the number of feature points within each grid cell and calculates the spatial consistency of feature points between adjacent grid cells. Spatial consistency is obtained by calculating the point cloud density change rate and positional deviation value, reflecting the matching accuracy of the 3D features at different spatial locations.

[0068] The maximum and minimum overlap deviations are determined based on the local overlap distribution. The maximum overlap deviation occurs in areas with sparse feature points and poor spatial consistency, typically located at building edges, vegetation cover, or abrupt topographic changes. The minimum overlap deviation occurs in areas with dense feature points and high spatial consistency, such as building facades or flat roads. Table 1 shows the overlap analysis data for a specific area's grid cells.

[0069] Table 1: Local Mesh Element Coincidence Analysis Data Table.

[0070]

[0071] A normalization mapping is performed based on the maximum and minimum coincidence deviations to convert the actual deviation values ​​into standardized accuracy coefficients. This mapping process uses a linear scaling method to map the deviation value range to the interval 0-1, where 0 represents the lowest accuracy and 1 represents the highest accuracy. Through this mapping relationship, 3D model accuracy coefficients are generated, which comprehensively reflect the spatial consistency quality of the entire 3D feature set.

[0072] When the accuracy coefficient of the 3D model does not meet the preset model update conditions, an incremental update of the texture feature evolution sequence is triggered. The preset model update conditions are usually set as an accuracy coefficient threshold, such as 0.85. If the calculated accuracy coefficient is lower than this threshold, the system automatically starts the update process. When locating low-confidence feature segments in the texture feature evolution sequence, a confidence scoring mechanism is used. The score is calculated based on factors such as the spatial consistency of feature points, texture clarity, and geometric stability. Low-confidence feature segments usually correspond to regions with low accuracy coefficients, such as mesh cells G-0127 and G-0129 in the table.

[0073] When generating new sampling point coordinates based on the terrain parameter set, the system references the digital elevation model and surface classification data. In low-confidence areas, the system automatically generates the spatial coordinates of supplementary sampling points based on terrain complexity and feature point distribution patterns. The distribution density of these coordinate points is positively correlated with terrain complexity, with denser sampling points placed at building edges and in areas with dramatic terrain changes. When acquiring supplementary oblique image blocks based on the coordinates of the new sampling points, the system controls the UAV flight system to re-capture multi-angle images of the designated area. The capture parameters are optimized and adjusted according to the geometric and photometric characteristics of the original image set to ensure consistency between the newly acquired images and the original image set.

[0074] When fusing supplementary tilted image blocks into the texture feature evolution sequence, a time-series stitching method is employed. After feature extraction and texture analysis, the texture feature evolution data of newly acquired image blocks are integrated into the original sequence. The fusion process maintains spatiotemporal continuity, and interpolation and smoothing ensure seamless integration between the old and new data. The updated texture feature evolution sequence contains more complete feature information, providing an improved data foundation for subsequent iterative processing.

[0075] Example 5: The process of outputting the target 3D mesh model is the final stage of 3D reconstruction. This process transforms the calibrated 3D feature set into a 3D mesh model with complete geometric structure and texture information. The calibrated 3D feature set contains spatial coordinates and texture feature data, which have already achieved geometric consistency and topological accuracy through the aforementioned processing steps.

[0076] When performing Delaunay triangulation on the corrected 3D feature set, an incremental insertion algorithm is used to construct the triangular mesh. This algorithm selects initial points from the feature point set, progressively inserts remaining points, and maintains the Delaunay properties. During the triangulation process, each feature point is treated as a mesh vertex, and the algorithm automatically establishes the topological connections between vertices, forming a set of triangular facets. Delaunay triangulation ensures the empty circle property of the mesh, avoids the generation of elongated triangles, and makes the mesh structure more uniform and reasonable. The triangulation results initially establish the geometric framework of the 3D model, but the mesh density and boundary quality still need further optimization.

[0077] When adjusting the density of triangular faces based on the accuracy coefficient of the 3D model, the accuracy coefficient serves as the control parameter for density adjustment. The accuracy coefficient originates from the normalization result of the preprocessing and ranges from 0 to 1; a higher value indicates better feature quality in that region. In regions with high accuracy coefficients, the density of triangular faces is increased to retain more detailed features; in regions with low accuracy coefficients, the density of triangular faces is decreased to simplify the model structure. Density adjustment is achieved through vertex insertion and deletion operations: in high-density regions, new vertices are inserted using edge bisection; in low-density regions, adjacent vertices are merged and redundant triangular faces are removed. This adaptive density adjustment method optimizes the data efficiency of the mesh while ensuring the overall accuracy of the model.

[0078] When performing feature boundary optimization on the triangulation results, the focus is on handling sharp edges and discontinuous boundaries in the model. Feature edges such as eaves and window frames in the architectural model need to maintain clear geometric features. The Laplacian smoothing algorithm is used to optimize the vertex positions of non-feature regions, reducing irregular undulations on the mesh surface. For feature boundaries, constraint optimization methods are used to maintain edge sharpness. The boundary optimization process also includes mesh hole detection and repair, automatically identifying and filling mesh voids caused by missing data to ensure the integrity and watertightness of the model.

[0079] When generating the target 3D mesh model, the optimized triangular mesh is fused with texture information. The texture mapping process applies the image texture to the corresponding triangular facets based on the texture coordinates of the feature points. The final output target 3D mesh model contains complete geometric information and texture data, stored in a standard 3D file format such as OBJ or PLY. This model can be directly used for 3D visualization, spatial analysis, and engineering applications, exhibiting high geometric accuracy and visual realism.

[0080] Through systematic mesh generation and optimization methods, discrete feature point sets are transformed into continuous 3D mesh models. Delaunay triangulation establishes the basic mesh structure, adaptive density adjustment optimizes the mesh data distribution, and feature boundary optimization improves the model's visual quality. All processing steps are executed automatically based on the algorithm, ensuring the consistency and reliability of the output model and meeting the quality requirements of practical applications for 3D models.

[0081] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0082] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for generating a 3D model based on UAV oblique photogrammetry, characterized in that, The process includes the following steps: acquiring the original oblique photogrammetry image set from the UAV and the terrain parameter set of the area to be modeled; performing multi-scale preprocessing on the original oblique photogrammetry image set to obtain a first-scale feature image set and a second-scale feature image set; Extract texture feature evolution sequences from the second-scale feature image set; perform initialization of shooting constraints based on the texture feature evolution sequences and terrain parameter sets, and output an initial control constraint set; perform spatial feature reconstruction on the first-scale feature image set according to the initial control constraint set to obtain a three-dimensional topological feature vector set; perform multi-dimensional registration constraints on the three-dimensional topological feature vector set based on a similarity threshold to locate the optimal registration path; perform topological correction according to the optimal registration path and terrain parameter set, and output a corrected three-dimensional feature set; perform accuracy normalization processing on the corrected three-dimensional feature set to generate three-dimensional model accuracy coefficients; when the three-dimensional model accuracy coefficients do not meet the preset model update conditions, trigger incremental updates of the texture feature evolution sequences; when the three-dimensional model accuracy coefficients meet the preset model update conditions, output the target three-dimensional mesh model.

2. The method for generating a 3D model based on UAV oblique photogrammetry according to claim 1, characterized in that, The method for performing multi-scale preprocessing on the original oblique photogrammetric image set is as follows: the original oblique photogrammetric image set is divided into multiple image blocks; a resolution dimensionality reduction operation is performed on each image block to generate a basic scale image unit set; an edge enhancement operation is performed on the basic scale image unit set to obtain the first scale feature image set; spatial pyramid pooling is performed on the first scale feature image set to output the second scale feature image set.

3. The method for generating a 3D model based on UAV oblique photogrammetry according to claim 1, characterized in that, The method for extracting the texture feature evolution sequence from the second-scale feature image set is as follows: calculate the texture gradient change of adjacent image blocks in the second-scale feature image set; construct a spatiotemporal evolution matrix based on the texture gradient change; perform feature trajectory tracking on the spatiotemporal evolution matrix to generate the texture feature evolution sequence.

4. The method for generating a 3D model based on UAV oblique photogrammetry according to claim 1, characterized in that, The method for initializing shooting constraints based on the texture feature evolution sequence and terrain parameter set is as follows: determine elevation constraints and illumination constraints according to the terrain parameter set; map and match the texture feature evolution sequence with the elevation constraints to generate a first constraint parameter set; map and match the texture feature evolution sequence with the illumination constraints to generate a second constraint parameter set; fuse the first constraint parameter set and the second constraint parameter set to output the initial control constraint set.

5. The method for generating a 3D model based on UAV oblique photogrammetry according to claim 1, characterized in that, The method for performing spatial feature reconstruction on the first scale feature image set based on the initial control constraint group is as follows: decompose the initial control constraint group into a geometric constraint subset and a texture constraint subset; use the geometric constraint subset to perform vertex topology reconstruction on the first scale feature image set to obtain an initial three-dimensional vertex set; use the texture constraint subset to perform feature fuzzy clustering on the initial three-dimensional vertex set to generate the three-dimensional topological feature vector set.

6. The method for generating a 3D model based on UAV oblique photogrammetry according to claim 1, characterized in that, The method for performing multidimensional registration constraints on the three-dimensional topological feature vector set based on a similarity threshold is as follows: calculate the Euclidean distance matrix of each vector in the three-dimensional topological feature vector set; Valid matching pairs of the Euclidean distance matrix are selected based on the similarity threshold; affine transformation iteration is performed on the valid matching pairs to generate a candidate registration path set; the path with the smallest error is selected from the candidate registration path set as the optimal registration path.

7. The method for generating a 3D model based on UAV oblique photogrammetry according to claim 1, characterized in that, The method for performing topology correction based on the optimal registration path and terrain parameter set is as follows: extract the coordinate transformation parameters of the optimal registration path; generate an elevation correction amount based on the terrain parameter set; perform weighted fusion of the coordinate transformation parameters and the elevation correction amount to obtain a spatial correction operator; and use the spatial correction operator to perform remapping on the three-dimensional topological feature vector set to output the corrected three-dimensional feature set.

8. The method for generating a 3D model based on UAV oblique photogrammetry according to claim 1, characterized in that, The method for performing accuracy normalization processing on the corrected three-dimensional feature set is as follows: calculate the local overlap distribution of the corrected three-dimensional feature set; determine the maximum overlap deviation and the minimum overlap deviation based on the local overlap distribution; perform normalization mapping based on the maximum overlap deviation and the minimum overlap deviation to generate the accuracy coefficient of the three-dimensional model.

9. A method for generating a 3D model based on UAV oblique photogrammetry according to claim 1, characterized in that, The method for triggering incremental updates of the texture feature evolution sequence is as follows: locating low-confidence feature segments in the texture feature evolution sequence; generating new sampling point coordinates based on the terrain parameter set; acquiring supplementary tilted image blocks according to the new sampling point coordinates; and fusing the supplementary tilted image blocks into the texture feature evolution sequence.

10. A method for generating a 3D model based on UAV oblique photogrammetry according to claim 9, characterized in that, The method for outputting the target 3D mesh model is as follows: performing Delaunay triangulation on the corrected 3D feature set; adjusting the density of triangular patches according to the accuracy coefficient of the 3D model; and performing feature boundary optimization on the triangulation results to generate the target 3D mesh model.

Images