Unmanned aerial vehicle city lod1 model reconstruction method based on virtual control constraint of satellite image

Abstract

The application discloses a UAV city LOD1 model reconstruction method based on satellite image virtual control constraint, belongs to the technical field of remote sensing surveying and mapping and three-dimensional city modeling, and acquires satellite optical remote sensing images and UAV oblique photography images with absolute geographic codes; building outline vectors with absolute geographic coordinates are extracted from the satellite images, uncontrolled relative orientation and sparse three-dimensional reconstruction are performed on the UAV images, and an initial three-dimensional model of a local coordinate system is constructed; cross-view coarse registration is completed by taking the outline vectors as matching primitives, an initial rigid body transformation relationship is solved, and satellite virtual control points are generated; a joint bundle adjustment model containing the control points is constructed, soft constraint is fused into adjustment calculation, a textureless LOD1 city three-dimensional model is generated based on the optimized model, dependence on field image control points is completely eliminated, and cross-view matching difficulties of satellite and UAV images are effectively overcome, and finally, a full-process automatic solution of a simple city model suitable for lightweight application is output.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing mapping and 3D city modeling technology, specifically involving a method for reconstructing a UAV city LOD1 model based on virtual control constraints of satellite imagery. Background Technology

[0002] In recent years, with the booming development of the low-altitude economy, the demand for "large-scale, rapid, and low-cost" urban 3D models has become increasingly urgent. Unmanned aerial vehicle (UAV) oblique photogrammetry technology, with its advantages of mobility, high data acquisition efficiency, and high detail reproduction, has become the mainstream method for urban 3D modeling. This technology uses UAVs equipped with multi-lens cameras to acquire multi-view images and utilizes structure-of-motion (sfM) and multi-view stereo matching (MVS) techniques to automatically reconstruct textured real-world 3D models, realistically restoring the geometric structure and visual appearance of urban scenes.

[0003] However, traditional oblique photogrammetry 3D reconstruction methods have fundamental flaws in obtaining absolute geographic coordinates. Because the reconstruction process is based on a relative coordinate system, it is necessary to establish and measure ground control points (GCPs) within the survey area to incorporate the model into the geodetic coordinate system. Control point measurement requires professional surveyors to enter the survey area with RTK or total stations, which is not only time-consuming, labor-intensive, and costly in complex environments such as urban canyons, but also poses safety risks. Currently, existing technologies have not effectively solved the problem of oblique photogrammetry's high dependence on ground control points.

[0004] Furthermore, traditional methods output realistic 3D models containing high-resolution textures and dense triangular meshes, resulting in data volumes of several gigabytes or even tens of gigabytes per square kilometer. For applications such as drone navigation, base station occlusion analysis, and urban wind corridor simulation, which only require building geometry and location information, these models suffer from severe data redundancy and cannot meet the needs of lightweight applications.

[0005] To address the aforementioned issues, some attempts have emerged in the industry to implement image control-free or lightweight modeling. For example, using UAVs equipped with RTK / PPK modules to acquire high-precision POS data can reduce reliance on control points to some extent. However, this method only constrains camera position, has limited ability to suppress attitude angle errors, and incurs high hardware costs. Other research has attempted to use satellite orthophotos as a reference benchmark, establishing correspondences between images through feature matching algorithms to replace ground control points. However, due to the orthophoto perspective of satellite images and the oblique perspective of UAV images, the texture differences between the two are significant, resulting in extremely low matching accuracy of traditional feature matching algorithms and difficulty in establishing stable and reliable correspondence mappings. Furthermore, satellite images themselves have a 3-5 meter planar positioning error. If this is directly substituted as a "rigid constraint" into aerial triangulation, the error will be forcibly transmitted to the UAV model, causing the local relative geometric relationships to be disrupted and resulting in distortion and deformation within the model.

[0006] In summary, existing technologies have not yet provided a fully automated solution that can completely eliminate dependence on field control points, effectively overcome the difficulty of cross-view matching between satellite and UAV imagery, and ultimately output a simple city model suitable for lightweight applications. Summary of the Invention

[0007] The purpose of this invention is to overcome the problems in the existing technology of UAV oblique photogrammetry 3D reconstruction, such as high dependence on ground control points, difficulty in cross-view matching of satellite and UAV images, and serious redundancy of output model data. It proposes a UAV city LOD1 model reconstruction method based on virtual control constraints of satellite images.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for reconstructing a UAV city LOD1 model based on virtual control constraints of satellite imagery, comprising the following steps: Acquire satellite optical remote sensing images of the target area with absolute geocoding, as well as UAV oblique photogrammetry images; extract building roof outlines from the satellite optical remote sensing images to generate building outline vector data with absolute geographic coordinates; perform relative orientation and sparse 3D reconstruction without ground control on the UAV oblique photogrammetry images, solve for the initial relative orientation elements and sparse 3D point cloud of the UAV images, and construct the initial 3D model of the UAV in the local coordinate system. Using building outline vector data as matching primitives, cross-view coarse registration is performed between satellite optical remote sensing images and UAV initial 3D models to calculate the initial rigid body transformation relationship between satellite optical remote sensing images and initial 3D models. Based on the initial rigid body transformation relationship, contour feature points are selected in the building contour vector data, and the absolute geographic coordinates of the contour feature points are transformed into the local coordinate system of the initial three-dimensional model to generate satellite virtual control points. A joint bundle adjustment model including satellite virtual control points is constructed. Adjustment calculation is performed based on the joint bundle adjustment model. During the calculation process, satellite virtual control points are incorporated as soft-constraint observations to obtain an adjusted and optimized 3D model. Based on the adjusted and optimized 3D model, a textureless LOD1 level city 3D model is generated.

[0009] Furthermore, the satellite optical remote sensing imagery is geocoded imagery with a resolution within a preset range, and the UAV oblique photography imagery is a multi-view image sequence with a forward overlap rate and a lateral overlap rate both within a preset range.

[0010] Furthermore, the roof outlines of buildings are extracted from satellite optical remote sensing images using a cascaded outline extraction method, specifically: Multi-scale segmentation of satellite optical remote sensing images yields initial candidate building regions; A deep learning-based semantic segmentation network is used to initially identify candidate building regions and output a building probability map. A method combining morphological processing and edge detection is used to extract the initial building edge contours from the building probability map; A geometrically constrained contour regularization algorithm is used to optimize the initial building edge contour, including corner fitting, line segment approximation and contour closure, to generate building contour vector data.

[0011] Furthermore, cross-view coarse registration employs a point-line iterative nearest-point method; The absolute geographic coordinates of the contour feature points are transformed into the local coordinate system of the initial 3D model to generate satellite virtual control points, specifically: The initial 3D model is downsampled to extract the surface model and the matching 2D projection contour; The satellite imagery building outline vector data and the initial 3D model's 2D projection outline are unified to the same horizontal projection coordinate system, and a point-line correspondence is established: for each point on the satellite outline, the closest point on the UAV model's projection outline is found, and point-line pairing is established; Construct an objective function based on minimizing the distance between points and lines, and use the iterative nearest point method to minimize the objective function as a constraint to solve the rigid body transformation matrix and obtain the initial rigid body transformation relationship; Corner points or feature points uniformly sampled along the contour in the satellite image building outline vector data are selected, and the three-dimensional coordinates are transformed into the initial three-dimensional model coordinate system of the UAV image using the initial rigid body transformation relationship, which serve as virtual control points.

[0012] Furthermore, the specific method for incorporating satellite virtual control points as soft-constraint observations during the solution process is as follows: In the bundle adjustment error equation for UAV imagery regional network adjustment, a virtual observation equation for virtual control points is added, as shown in the following equation:

[0013] in, This is the correction vector for the virtual control points. Let X be the coefficient matrix corresponding to the virtual control points, and let X be the vector of unknowns containing corrections for the exterior orientation elements and corrections for the coordinates of the densified points. This represents the observation vector of the virtual control point; The initial weights of the virtual control points are set according to the nominal geolocation accuracy of the satellite imagery. The nominal geolocation accuracy is the planar error value recorded in the satellite imagery metadata. The initial weights are lower than the weights of the UAV image point observations.

[0014] Furthermore, during the adjustment process, the exterior orientation elements and 3D coordinates of the densified points of the UAV imagery are calculated simultaneously, and the adjusted and optimized 3D model is generated based on the exterior orientation elements and 3D coordinates of the densified points. An adaptive robust weight function is introduced during the adjustment process. The adaptive robust weight function adopts the IGG III scheme and the equivalent weight function. satisfy: when hour, ; when hour, ; when hour,

[0015] Where v is the standardized residual, The value range is 1.0 to 1.5. The value ranges from 2.5 to 3.0; after each adjustment iteration, the weights of each virtual control point are recalculated based on the calculated residuals.

[0016] Furthermore, a textureless LOD1-level city 3D model is generated using an elevation-robust road skeleton registration method, including: Extract the 3D skeleton lines of the road network from the adjusted and optimized 3D model; Acquire publicly available vector road network data for the same area as satellite optical remote sensing images, and convert the publicly available vector road network data into three-dimensional road skeleton lines; Calculate the three-dimensional translation and rotation deviations between the three-dimensional skeleton lines of the road network and the three-dimensional road skeleton lines of the road network; Using 3D translation and rotation deviations as system correction parameters, the adjusted 3D model is globally corrected to generate a textureless LOD1 level city 3D model.

[0017] Secondly, this invention provides a UAV city LOD1 model reconstruction system based on satellite imagery virtual control constraints, comprising: The initial model building module is used to acquire satellite optical remote sensing images with absolute geocoding of the target area, as well as UAV oblique photogrammetry images; extract building roof outlines from the satellite optical remote sensing images to generate building outline vector data with absolute geographic coordinates; perform relative orientation and sparse 3D reconstruction without ground control on the UAV oblique photogrammetry images, solve for the initial relative orientation elements and sparse 3D point cloud of the UAV images, and construct the initial 3D model of the UAV in the local coordinate system. The transformation relationship calculation module is used to use building outline vector data as matching primitives to perform cross-view coarse registration between satellite optical remote sensing images and UAV initial 3D models, and calculate the initial rigid body transformation relationship between satellite optical remote sensing images and initial 3D models. The virtual control point conversion module is used to select contour feature points in the building contour vector data based on the initial rigid body transformation relationship, and convert the absolute geographic coordinates of the contour feature points to the local coordinate system of the initial three-dimensional model to generate satellite virtual control points. The LOD1 model generation module is used to construct a joint bundle adjustment model including satellite virtual control points. The adjustment is performed based on the joint bundle adjustment model, and the satellite virtual control points are incorporated as soft-constraint observations during the solution process to obtain an adjusted and optimized 3D model. Based on the adjusted and optimized 3D model, a textureless LOD1 level city 3D model is generated.

[0018] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a method for reconstructing a UAV city LOD1 model based on virtual control constraints of satellite imagery.

[0019] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a method for reconstructing a UAV city LOD1 model based on virtual control constraints of satellite imagery.

[0020] Compared with the prior art, the present invention has the following beneficial technical effects: This invention proposes a UAV city LOD1 model reconstruction method based on satellite imagery virtual control constraints. It completely eliminates reliance on field control points, using building roof corners extracted from satellite imagery as virtual control points. A soft constraint mechanism is integrated into bundle adjustment, eliminating the need for setting up and measuring ground control points throughout the process. Robust registration of cross-view heterogeneous data is achieved using a point-line iterative nearest-point algorithm that leverages the cross-view invariant feature of building outlines, successfully overcoming the low matching rate problem of traditional texture feature matching algorithms under significant viewpoint differences. A high-precision soft constraint adjustment mechanism introduces an adaptive robust weight function, ensuring that satellite points only exert a traction effect on the overall model, preventing satellite imagery positioning errors from disrupting the rigorous relative geometric relationships between UAV images. Lightweight model output produces a textureless, simplified LOD1 city model, compressing the data volume to less than one percent of the original realistic 3D model. The entire process is automated with system correction, requiring no manual intervention from data input to model output. This invention provides a fully automated solution that completely eliminates the reliance on field control points, effectively overcomes the difficulty of cross-view matching between satellite and UAV imagery, and ultimately outputs a simple city model suitable for lightweight applications. Attached Figure Description

[0021] The accompanying drawings described herein are for illustrative purposes only and are not intended to limit the scope of the invention in any way. Furthermore, the shapes and proportions of the components in the drawings are merely schematic to aid in understanding the invention and do not specifically limit the shapes and proportions of the components. In the drawings: Figure 1 This is a flowchart of the UAV city LOD1 model reconstruction method based on satellite imagery virtual control constraints according to the present invention.

[0022] Figure 2 This is a simplified structural diagram of the UAV city LOD1 model reconstruction system based on satellite imagery virtual control constraints according to the present invention.

[0023] Figure 3 This is an electronic device diagram of the UAV city LOD1 model reconstruction method based on satellite imagery virtual control constraints according to the present invention.

[0024] Figure 4 This is an overall flowchart of the UAV city LOD1 model reconstruction method based on satellite imagery virtual control constraints in this embodiment of the invention.

[0025] Figure 5 This diagram illustrates the comparison between the principles of traditional hard-constraint adjustment and the soft-constraint adjustment proposed in this scheme.

[0026] Figure 6 This is a schematic diagram illustrating the principle of point-line iterative nearest point cross-view coarse registration.

[0027] Figure 7This is a schematic diagram illustrating the process and effect of the cascaded contour extraction algorithm.

[0028] Figure 8 This is a schematic diagram of the rigid body correction principle based on a public vector road network.

[0029] Figure 9(a) is a schematic diagram of the overall structure of the simplified 3D city model of the present invention, LOD1.

[0030] Figure 9(b) is a schematic diagram of the spatial relationship between the red no-fly zone and urban buildings in this invention.

[0031] Figure 9(c) is a top-view projection analysis of the no-fly zone coverage area of ​​the present invention.

[0032] Figure 9(d) is a schematic diagram comparing the size of multiple no-fly zones in the scenario of the present invention.

[0033] Figure 10 A schematic diagram of an electronic device structure for implementing the method of the embodiments of the present invention. Detailed Implementation

[0034] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of 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 should fall within the scope of protection of the present invention.

[0035] Example 1 See Figure 1 This paper provides a method for reconstructing LOD1 models of cities using unmanned aerial vehicles (UAVs) based on virtual control constraints using satellite imagery, including the following steps: Acquire satellite optical remote sensing images with absolute geocoding and UAV oblique photogrammetry images of the target area. The satellite optical remote sensing images are geocoded images with a resolution within a preset range, and the UAV oblique photogrammetry images are multi-view image sequences with preset forward and lateral overlap rates. Extract building roof outlines from the satellite optical remote sensing images to generate building outline vector data with absolute geographic coordinates. The extraction of building roof outlines from the satellite optical remote sensing images uses a cascaded outline extraction method, specifically: multi-scale segmentation of the satellite optical remote sensing images to obtain initial candidate building regions; a deep learning-based semantic segmentation network is used to perform preliminary discrimination of the initial candidate building regions, outputting a building probability map; a combination of morphological processing and edge detection is used to extract the initial building edge outlines from the building probability map; a geometrically constrained outline regularization algorithm is used to optimize the initial building edge outlines, including corner fitting, line segment approximation, and outline closure, generating building outline vector data. A ground-free relative orientation and sparse 3D reconstruction is performed on UAV oblique photogrammetric images to obtain the initial relative orientation elements and sparse 3D point cloud of the UAV images, and to construct the initial 3D model of the UAV in the local coordinate system.

[0036] Using building outline vector data as matching primitives, cross-view coarse registration is performed between satellite optical remote sensing imagery and UAV initial 3D model. The cross-view coarse registration adopts the point-line iterative nearest point method; the initial rigid body transformation relationship between satellite optical remote sensing imagery and initial 3D model is calculated. Based on the initial rigid body transformation relationship, contour feature points are selected from the building contour vector data. The absolute geographic coordinates of the contour feature points are transformed to the local coordinate system of the initial 3D model to generate satellite virtual control points. Specifically, the initial 3D model is downsampled to extract the surface model and the matching 2D projection contour. The satellite image building contour vector data and the 2D projection contour of the initial 3D model are unified to the same horizontal projection coordinate system, and a point-line correspondence relationship is established: for each point on the satellite contour, the closest point on the UAV model projection contour line is found to establish point-line pairing. An objective function based on minimizing the point-line distance is constructed, and the iterative nearest point method is used to minimize the objective function as a constraint to solve the rigid body transformation matrix, obtaining the initial rigid body transformation relationship. Corner points or feature points uniformly sampled along the contour in the satellite image building contour vector data are selected, and the 3D coordinates are transformed to the initial 3D model coordinate system of the UAV image using the initial rigid body transformation relationship, serving as virtual control points.

[0037] A joint bundle adjustment model incorporating satellite virtual control points is constructed. Adjustment calculations are then performed based on this model, incorporating satellite virtual control points as soft-constraint observations during the calculation process. Specifically, virtual observation equations for virtual control points are added to the bundle adjustment error equations of the UAV imagery regional network adjustment, as shown in the following equation:

[0038] in, This is the correction vector for the virtual control points. Let X be the coefficient matrix corresponding to the virtual control points, and let X be the vector of unknowns containing corrections for the exterior orientation elements and corrections for the coordinates of the densified points. This represents the observation vector of virtual control points. The initial weights of the virtual control points are set based on the nominal geolocation accuracy of the satellite imagery. The nominal geolocation accuracy is the planar error value recorded in the satellite imagery metadata, and the initial weights are lower than the weights of the UAV image point observations. During the adjustment calculation, the exterior orientation elements and the 3D coordinates of the densified points in the UAV imagery are calculated simultaneously. Based on the exterior orientation elements and the 3D coordinates of the densified points, an adjusted and optimized 3D model is generated. An adaptive robust weight function is introduced during the adjustment process. The adaptive robust weight function adopts the IGG III scheme and the equivalent weight function. satisfy: when hour, ; when hour, ; when hour,

[0039] Where v is the standardized residual, The value range is 1.0 to 1.5. The value ranges from 2.5 to 3.0; after each adjustment iteration, the weights of each virtual control point are recalculated based on the calculated residuals.

[0040] A textureless LOD1-level urban 3D model is generated based on the adjusted and optimized 3D model. A road skeleton registration method robust to elevation is employed, including: extracting the 3D skeleton lines of the road network from the adjusted and optimized 3D model; acquiring public vector road network data for the same area as the satellite optical remote sensing imagery and converting the public vector road network data into 3D road skeleton lines; calculating the 3D translation and rotation deviations between the 3D road skeleton lines and the 3D road skeleton lines; and using the 3D translation and rotation deviations as system correction parameters to globally correct the adjusted and optimized 3D model, thereby generating a textureless LOD1-level urban 3D model.

[0041] Example 2 See Figure 2 A UAV city LOD1 model reconstruction system based on satellite imagery virtual control constraints includes: The initial model building module is used to acquire satellite optical remote sensing images with absolute geocoding of the target area, as well as UAV oblique photogrammetry images; extract building roof outlines from the satellite optical remote sensing images to generate building outline vector data with absolute geographic coordinates; perform relative orientation and sparse 3D reconstruction without ground control on the UAV oblique photogrammetry images, solve for the initial relative orientation elements and sparse 3D point cloud of the UAV images, and construct the initial 3D model of the UAV in the local coordinate system. The transformation relationship calculation module is used to use building outline vector data as matching primitives to perform cross-view coarse registration between satellite optical remote sensing images and UAV initial 3D models, and calculate the initial rigid body transformation relationship between satellite optical remote sensing images and initial 3D models. The virtual control point conversion module is used to select contour feature points in the building contour vector data based on the initial rigid body transformation relationship, and convert the absolute geographic coordinates of the contour feature points to the local coordinate system of the initial three-dimensional model to generate satellite virtual control points. The LOD1 model generation module is used to construct a joint bundle adjustment model including satellite virtual control points. The adjustment is performed based on the joint bundle adjustment model, and the satellite virtual control points are incorporated as soft-constraint observations during the solution process to obtain an adjusted and optimized 3D model. Based on the adjusted and optimized 3D model, a textureless LOD1 level city 3D model is generated.

[0042] Example 3 See Figure 3 An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a method for reconstructing a UAV city LOD1 model based on virtual control constraints of satellite imagery.

[0043] Example 4 A computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for reconstructing a UAV city LOD1 model based on virtual control constraints of satellite imagery.

[0044] Example 5 See Figure 4A method for reconstructing a UAV city LOD1 model based on satellite imagery virtual control constraints is disclosed, comprising: Step 1, acquiring satellite optical remote sensing images and UAV oblique photogrammetry images of the target area; Step 2, performing free-network aerial triangulation on the UAV oblique photogrammetry images to generate an initial 3D model and extract 3D line features; Step 3, extracting building roof contours from the satellite optical remote sensing images using a cascaded contour extraction algorithm; Step 4, performing cross-view coarse registration using a point-line iterative nearest point algorithm to obtain initial rigid body transformation relationships; Step 5, generating satellite virtual control points based on the initial rigid body transformation relationships, constructing a soft-constraint joint adjustment model, and solving for high-precision exterior orientation elements; Step 6, generating a simplified LOD1 model and eliminating system offset based on a public vector road network. This invention solves the problems of high dependence on ground control points in UAV oblique photogrammetry, difficulty in cross-view matching between satellite and UAV images, and severe data redundancy in the output model by using a soft-constraint adjustment mechanism of satellite virtual control points and cross-view geometric registration. It achieves lightweight city LOD1 model reconstruction that completely eliminates the need for field control points and is fully automated. Specifically, it includes the following steps: Step 1: Acquire satellite optical remote sensing images and UAV oblique photography images of the target area; Satellite optical remote sensing images are high-resolution optical images that have undergone orthorectification and are accompanied by rational polynomial coefficient models. Their spatial resolution is better than 1 meter, and they can map the image points on the image to the object space coordinates in the geodetic coordinate system. The UAV oblique photography images are acquired through an oblique photography system, with a forward overlap of no less than 70% and a lateral overlap of no less than 50%. Each image carries the initial exterior orientation element observations recorded by the airborne positioning and attitude determination system. Extract geocoding model parameters from satellite imagery to establish a mapping relationship between image point coordinates and geodetic coordinates; extract positioning and attitude data and camera calibration parameters from UAV imagery for initial value setting in subsequent free-network aerial triangulation.

[0045] Step 1 also includes a coordinate system sub-step: Extract coordinate system information from the geocoding model of satellite imagery, and determine whether the data uses a projected coordinate system or a geographic coordinate system by analyzing its coordinate reference string and geographic range; Extract the coordinate system information recorded by the UAV image positioning and attitude determination system, and determine whether it is consistent with the coordinate system of the satellite image. If the two coordinate systems are inconsistent, the initial positioning data of the UAV imagery will be converted to the coordinate system of the satellite imagery to establish a unified coordinate reference. If a geographic coordinate system is determined to exist, the metric conversion factor between the longitude and latitude directions is calculated based on the latitude of the center of the survey area. All subsequent distance-related calculations are then unified to the metric space to ensure that cross-source data are geometrically registered and adjusted under a unified metric framework.

[0046] Step 2: Perform free-network aerial triangulation on the UAV oblique photogrammetric images to generate an initial 3D model and extract 3D line features; For the UAV oblique photography sequence images obtained in step 1, motion restoration structure calculation is performed under the condition of no ground control points; feature points are extracted and cross-image matching is performed on the UAV sequence images, and a bundle adjustment model is constructed based on the collinearity condition equation. The relative exterior orientation elements and sparse 3D point cloud of each image are iteratively solved to form the initial 3D model. A rough depth map is generated based on multi-view stereo matching, and three-dimensional line segments such as the intersection of building facade and ground and vertical edge of wall are fitted from it to form a three-dimensional line feature set of UAV; The initial model is located in a local coordinate system and has the correct internal relative geometric relationships, but lacks absolute scale and absolute geographic coordinates; Step 2 also includes a free network adjustment initialization sub-step: Extract the initial exterior orientation element observations recorded by the positioning and attitude determination system carried by the UAV imagery in step 1, including the spatial coordinates of the photography center and attitude angle parameters; The above observations are introduced as prior constraints into the bundle adjustment model and participate in the adjustment solution in a weighted manner, which accelerates iterative convergence and improves solution stability. If the positioning and attitude determination system data is unavailable or insufficient in accuracy, a purely visual motion reconstruction structure solution is adopted, with the minimization of image point reprojection error as the sole objective function for free network adjustment.

[0047] Step 3: Extract the roof outline of the building from the satellite optical remote sensing image using a cascaded contour extraction algorithm; See the flowchart of the cascaded contour extraction algorithm. Figure 7 Based on the satellite optical remote sensing imagery obtained in step 1, the satellite imagery is segmented at multiple scales and semantically discriminated. Combined with edge detection and geometric regularization, regular building outline vector data is generated. Through the cascaded processing of multi-scale segmentation, semantic discrimination, edge detection and geometric regularization, the building roof outline vector data is automatically extracted. The cascaded outline extraction algorithm includes the following sub-steps. S3.1 Multi-scale segmentation and candidate region generation: The satellite image is segmented at multiple scales to divide the image into multiple homogeneous regions and obtain initial candidate regions for buildings. The segmentation scale is adaptively adjusted according to the image resolution so that the segmented regions match the roof size of the buildings. Each segmented region is used as a candidate unit for subsequent processing.

[0048] S3.2 Semantic discrimination and probability map generation: Based on the candidate regions generated in 3.1, a deep learning-based semantic segmentation network is used to discriminate the initial building candidate regions and output a building probability map. The value of each pixel in the probability map represents the probability that the pixel belongs to the roof of the building, and the value range is from 0 to 1. The semantic segmentation network optimizes its parameters using a pre-trained architectural sample dataset to adapt to the roof morphology features in different urban scenarios. For shadowed or low-contrast areas, it uses neighborhood context information for inference and completion to improve the completeness of the probability map.

[0049] S3.3 Edge Detection and Initial Contour Extraction: Based on the probability map generated in 3.2, the building probability map is used as a mask constraint to limit the search space of the edge detection operator, performing edge detection only in high-probability regions to extract the initial building edge contours. Edge detection employs a combination of gradient magnitude calculation and non-maximum suppression. The gradient magnitude calculation formula is as follows:

[0050] in and These are the gradient components of the image in the x and y directions, respectively.

[0051] Using the probability map as a mask for the region of interest, the edge detection operator is applied only to candidate building regions, effectively eliminating interference from non-building regions such as shadows, roads, and squares. If the confidence level of a region in the probability map is lower than a preset threshold, a dual-threshold edge detection strategy is adopted for that region to retain weak edges and maintain the continuity of the contour.

[0052] S3.4, Geometric Constraints and Contour Regularization: Based on the initial edge contour extracted in 3.3, the initial edge contour is optimized using a contour regularization algorithm based on geometric constraints to generate regular building contour vector data and record the absolute geodetic coordinates of each corner point; The initial edge contour may have problems such as jagged boundaries, broken line segments, or unclear corner points. The contour regularization process in step 3.4 includes edge point simplification, merging of collinear edge segments, corner point orthogonalization constraints, and contour closure. The regularization process includes the following: Edge point simplification involves eliminating collinear redundant points through curvature analysis while retaining key inflection points of the contour. Curvature calculation is performed on the edge point sequence, and when the rate of change of curvature of multiple consecutive points is lower than a threshold, only the beginning and end endpoints are retained. The threshold for the rate of change of curvature is selected based on the geometric complexity of the building contour and is usually set to 0.1 to 0.3.

[0053] Collinear edge segments are merged by least-squares fitting of adjacent approximately collinear edge segments into a single straight line segment; the fitted line satisfies the condition that the sum of the squares of the perpendicular distances from each point to the line is minimized, and the line is considered collinear when the average distance from each point to the fitted line is less than a preset threshold.

[0054] Corner orthogonalization constraint utilizes the geometric prior that adjacent boundaries of a building roof tend to be orthogonal. When the deviation of the included angle between adjacent line segments from 90° is less than a preset threshold, it is forcibly adjusted to 90° so that the extracted contour conforms to the actual geometric shape of the building. The deviation threshold is selected according to the regularity of the building type, and is usually set to 5° to 10°.

[0055] Contour closure processing extends the contour gaps caused by occlusion or extraction failure along the direction of adjacent line segments until they intersect to form a closed polygon.

[0056] The regularized building outline vector data contains the absolute geodetic coordinates of each corner point. For each building outline, the absolute geodetic coordinates of its polygon corner points are recorded and a unique building identifier is assigned. The extracted building outlines satisfy the characteristics of closure, orthogonality, and straight line segments, and conform to the actual geometric shape of the building.

[0057] Step 4: Use the point-line iterative nearest point algorithm to perform cross-view coarse registration and obtain the initial rigid body transformation relationship; The principle of the point-line iterative nearest point algorithm is as follows: Figure 6 Based on the initial 3D model of the UAV generated in step 2 and the building roof contour vector data extracted from satellite imagery in step 3, the 3D line features of the UAV are projected onto the horizontal plane to establish a point-line correspondence with the building contours extracted from satellite imagery. Iterative optimization is then performed using the bidirectional chamfer distance as the objective function to solve for the optimal rigid body transformation matrix. This process includes the following sub-steps: S4.1 Initial Projection and 2D Contour Generation: Using the positioning and attitude data recorded by the UAV imagery, the initial 3D model generated in step 2 undergoes an initial rigid transformation, including rotation, translation, and scaling, to ensure it roughly falls within the coverage area of ​​the satellite imagery. During the initial transformation, the centroids of the UAV's 3D line feature point set and the satellite contour corner point set are calculated separately. The UAV model's centroid is then translated to the satellite contour's centroid position, achieving initial alignment. The purpose of this initial alignment step is to initially eliminate the scale and orientation differences between the two sets of data, thereby improving the convergence speed and algorithm robustness of subsequent iterations. The transformed UAV 3D line features are then vertically projected onto the horizontal plane to generate a 2D projected contour. The projected contour retains the geometric shape information of the building on the horizontal plane. S4.2 Establishment of point-line correspondence: Unify the satellite image building outline vector data and the UAV two-dimensional projection outline to the same horizontal projection coordinate system. For each corner point or sampling point on the satellite outline, search for the point with the closest Euclidean distance on the UAV projection outline and establish a point-line pairing relationship; the point on the satellite outline is denoted as p, and its corresponding point on the UAV projection outline is denoted as q.

[0058] S4.3 Solving for the Rigid Body Transformation Matrix: Using the bidirectional chamfer distance as the objective function, iterative optimization is employed to find the rigid body transformation matrix that minimizes the objective function. The bidirectional chamfer distance is defined as the sum of the squares of the closest bidirectional distances between two point sets, and the objective function expression is: ; Where S represents the set of building roof outline corner points extracted from satellite imagery, and P represents the set of UAV model projection outline points. Let R represent the Euclidean distance, where R is the rotation matrix and T is the translation vector; or R is a 3×3 rotation matrix and T is a 3×1 translation vector. The first term of the objective function represents the sum of squares of the shortest distances from satellite contour points to UAV projected contours, and the second term represents the sum of squares of the shortest distances from UAV projected contour points to satellite contours. Bidirectional symmetric calculation ensures global consistency of registration. In each iteration, the optimal transformation parameters are solved by fixing the point-line correspondence. Then, the correspondence is updated based on the new transformation. The iteration is repeated until the contour overlap error is less than the preset threshold, and the initial rigid body transformation relationship is output. Convergence conditions include: the contour overlap error is less than a preset threshold, which is determined empirically based on satellite image resolution and building size characteristics, and is usually set to 0.5 meters to 1.0 meters; or when the error change in multiple consecutive iterations is less than the set value, it is considered convergence and the iteration is terminated; the error change threshold can be adjusted according to the registration accuracy requirements, and can be set to the order of several millimeters for general application scenarios.

[0059] After iterative convergence, the output satisfies The initial rigid body transformation relationship; in, This represents the coordinates of the satellite image in the geodetic coordinate system. This represents the coordinates in the local coordinate system of the UAV model. This transformation relationship provides reliable initial values ​​for subsequent virtual control point generation and joint adjustment.

[0060] Step 5: Generate satellite virtual control points based on the initial rigid body transformation relationships; Based on the initial rigid body transformation relationship obtained in step 4, corner points or feature points uniformly sampled at fixed intervals along the contour from the satellite image building outline vector data extracted in step 3 are selected. Using the obtained initial rigid body transformation relationship, their absolute geodetic coordinates are transformed to the coordinate system where the UAV initial 3D model is located, and used as virtual control point coordinates. Feature points can be selected by selecting the corner points of the building outline polygon, or by sampling uniformly at fixed intervals along the building outline. The sampling interval is determined according to the building size and outline complexity, and is usually set to 2 to 5 meters. For each selected feature point, coordinate transformation is performed using the initial rigid body transformation relationship, which satisfies the following conditions: ,in These are the coordinates of the feature points in the geodetic coordinate system of the satellite image. The coordinates are the transformed coordinates in the local coordinate system of the UAV model. R and T are the rotation matrix and translation vector obtained in step 4. Each virtual control point records its three-dimensional coordinates in the local coordinate system and its corresponding absolute geodetic coordinates, which are used as the input of the virtual observation equation in the subsequent joint adjustment. Step 5 also includes a virtual control point quality check sub-step: Calculate the Euclidean distance between the transformed local coordinates of each virtual control point and the nearest point of the corresponding 3D line feature in step 2, and use it as the transformation residual; If the transformation residual of a virtual control point exceeds the preset threshold, it is determined that the point may have contour extraction error or registration deviation, and it is removed or downweighted. After quality inspection, the retained virtual control points should cover multiple buildings in different locations within the survey area to ensure the uniformity of spatial distribution. The number of virtual control points is determined according to the building density in the survey area, and is usually no less than 3.

[0061] Step 6: Construct a soft-constraint joint adjustment model and solve for high-precision exterior orientation elements and densification point coordinates; Based on the initial 3D UAV model generated in step 2, the satellite virtual control points generated in step 5, and the UAV image point observations, a joint adjustment model incorporating virtual control point constraints is constructed. The satellite virtual control points are integrated into the bundle adjustment using a soft constraint mechanism to solve for the high-precision exterior orientation elements and densification point coordinates of the UAV imagery. A comparison is made between soft-constraint joint adjustment and traditional hard-constraint adjustment. Figure 5 As shown; by adding virtual observation equations for satellite virtual control points to the traditional bundle adjustment method for regional networks, and assigning small weights to the virtual control points based on nominal positioning accuracy, the satellite reference only generates an overall traction effect without disrupting local relative geometric relationships. This includes the following sub-steps: S6.1 Construction of Joint Error Equations: Based on the traditional bundle adjustment method for regional networks, virtual observation equations for satellite virtual control points are added to form a set of joint error equations.

[0062] S6.2. Set the weight strategy for image point observation and virtual control points, and introduce an adaptive robustness mechanism; S6.3 Perform global optimization calculation, iteratively update all unknowns until convergence, and output the refined UAV absolute coordinate model; The joint error equation set in step 6.1 consists of two parts: the image point reprojection error equation and the virtual control point observation equation. The image point reprojection error equation is:

[0063] in Represents the external orientation of the camera. Represents the three-dimensional coordinates of the object. These are the coordinates of the image points actually observed. The virtual control point observation equation is:

[0064] in Let be the coordinates of the object-side corner point to be solved. The coordinates of the virtual control points generated in step 5; In step 6.2, the weighting strategy is as follows: the mean square error of the image point observation is approximately 0.5 pixels, corresponding to a weight. Initial weights of satellite virtual control points The nominal geolocation accuracy is set according to the satellite imagery, and this accuracy is read from the satellite imagery metadata as a planar error value. Corresponding to the initial weight

[0065] The value ranges from 1.0 to 5.0 meters; in this embodiment, 3.0 meters is used. much smaller Virtual control points only exert a traction effect on the overall translation and rotation of the adjustment results without changing the local relative geometric relationships between UAV images; After each bundle adjustment iteration, the adaptive robust weight function recalculates the weights of each virtual control point based on the calculated residuals, and the weights of residual abnormal points are automatically reduced or set to zero. The equivalent weight function Satisfy: When hour, ;when hour, ;when hour, Where v is the standardized residual, The value range is 1.0 to 1.5. The value range is 2.5 to 3.0; Step 7: Generate a simplified LOD1 model and eliminate system offsets; Based on the refined 3D model after adjustment in step 6 and the building roof outline vector data extracted in step 3, a textureless LOD1 level simplified city model is generated. By reusing the adjusted and optimized 2D polygon of the roof as the base, combining digital surface model sampling to calculate the building height, and vertically extruding to generate a textureless LOD1 white model, and using publicly available vector road network data for rigid body correction to eliminate system offset, a lightweight simplified urban model is finally output, including the following sub-steps: S7.1 The two-dimensional polygon of the roof extracted in step 3 and optimized by adjustment is directly reused on the bottom surface of the building; S7.2. Sample and calculate the building height in the digital surface model corresponding to the refined model; S7.3. Extrude the two-dimensional polygon vertically along the height H to generate a LOD1 white model. The output is in a general three-dimensional model data format and does not contain any texture mapping coordinates. S7.4. Use a rigid body correction method based on a public vector road network to eliminate system offset; The formula for calculating height in step 7.2 is as follows: ,in This represents the maximum elevation value within the polygonal region. This represents the average ground elevation of the surrounding area. The rigid body correction principle in step 7.4 is described in [reference needed]. Figure 8 As a rigid object, the road's geometry should remain stable under different times and different sensors. The road skeleton line, after elevation smoothing filtering, is extracted from the digital surface model and rigidly registered with the public vector road network data. The transformation only allows horizontal translation and rotation around the vertical axis. The geometric rigidity constraint of the road network is used to correct the overall model to the accurate geographic coordinate system. After solving the transformation parameters, they are applied to the vertex coordinates of all building LOD1 models to complete the global system offset correction. The final output of the simplified LOD1 city model is shown in Figures 9(a), 9(b), 9(c), and 9(d), with the data volume compressed to less than one percent of the original real-world 3D data. As shown in Figure 9(a), the simplified 3D city model of LOD1 provided by this invention is displayed from an isometric perspective. In the model, city buildings are simplified into cuboid prism structures with different heights, and each prism is arranged on a horizontal reference plane according to a preset grid spacing rule. Different shades of blue are used to distinguish the prisms to represent the differences in building volume. This model expresses the spatial distribution characteristics and density relationships of city buildings with minimal geometric detail.

[0066] As shown in Figure 9(b), this invention superimposes a spatial representation of a no-fly zone onto the LOD1 city model. The red no-fly zone is presented in the form of a semi-transparent cylinder. This cylinder has its radial and vertical boundaries determined by a preset control radius and flight restriction height, with the center point of the no-fly zone as its axis. The architectural prisms inside the cylinder are still visible through the light, representing the spatial inclusion relationship between the no-fly zone and urban buildings.

[0067] As shown in Figure 9(c), this invention uses an orthographic projection top-down view to display the coverage of no-fly zones on a horizontal plane. The yellow and red areas in the image represent the horizontal projections of two different control levels of no-fly zones onto a reference plane, while the blue square blocks represent the top projections of building prisms. This perspective is used to analyze the coverage of each no-fly zone on the geographical plane, their spacing, and their spatial overlap with building clusters.

[0068] As shown in Figure 9(d), this invention demonstrates the three-dimensional spatial relationship of multiple no-fly zones coexisting from a low-angle perspective. The left side of the image shows a yellow no-fly zone cylinder, and the right side shows a red no-fly zone cylinder; the difference in their vertical height and their horizontal spacing are clearly presented. The contrast in height between the architectural prisms in the foreground and background and the no-fly zone cylinders characterizes the vertical spatial scale of the no-fly zone relative to urban buildings.

[0069] Step 7 also includes a height adaptive correction sub-step: The system verifies the reasonableness of the generated building heights. If a building height H exceeds the building height limit of the area or is significantly abnormal compared to similar buildings in the surrounding area, the system automatically triggers a correction mechanism. First, it adjusts the ground elevation sampling range and recalculates the height. If it is still abnormal, it uses the average height of neighboring similar buildings for correction to ensure that the height information of the output model conforms to the actual situation.

[0070] This invention offers superior image control-free performance. By using building roof corners extracted from satellite imagery as virtual control points and incorporating them into bundle adjustment using a soft constraint method, it achieves effective transfer of absolute geodetic reference without compromising the local geometric accuracy of UAV imagery. The entire process eliminates the need to set up and measure ground control points, significantly reducing operating costs and time, and fundamentally freeing traditional oblique photogrammetry from its high dependence on field control points.

[0071] This invention exhibits strong robustness in cross-view registration. By integrating multi-scale segmentation, semantic discrimination, and geometric regularization processing through a cascaded contour extraction algorithm, it significantly improves the integrity and regularity of roof contour extraction in complex urban environments, effectively overcoming the contour breakage problem caused by shadow occlusion and temporal differences. The point-line iterative nearest point algorithm utilizes the cross-view invariant feature of building contours to achieve accurate coarse registration of heterogeneous images, successfully solving the problem of extremely low matching rate of traditional texture feature matching algorithms under conditions of orthophoto and oblique viewpoint differences.

[0072] The present invention provides an adaptive robust adjustment mechanism. It introduces an adaptive robust weight function based on residual statistics in the joint adjustment, which can automatically identify and suppress gross errors in satellite virtual control points. This ensures that satellite points only exert translational and rotational traction on the model as a whole, avoiding the destruction of the strict relative geometric relationship between UAV images by the positioning error of the satellite images themselves, and realizing unbiased transfer from low-precision benchmarks to high-precision models.

[0073] The invention achieves significant lightweighting of the model, ultimately outputting a textureless LOD1 level simplified city model. The data volume is compressed to less than one percent of the original real-world 3D data. While ensuring that the planar and elevation accuracy meets the application requirements of low-altitude navigation, spatial analysis, and other applications, it effectively solves the problems of serious data redundancy and difficulty in meeting the requirements of real-time performance and lightweighting in traditional real-world models.

[0074] This invention provides accurate and reliable global correction. By using a road skeleton registration method with robust elevation, it utilizes publicly available vector road network data as rigid landmarks to eliminate system offsets, thereby further improving the absolute positioning accuracy of the model.

[0075] This invention features a high degree of automation throughout the entire process, requiring no manual intervention from satellite and UAV data input to LOD1 model output. It supports exporting multiple standard 3D model formats, seamlessly integrates with mainstream geographic information systems and 3D visualization platforms, and is suitable for the rapid reconstruction and updating of simple urban models over large areas.

[0076] In step 3, the semantic segmentation network can be replaced with a network architecture that adopts a hollow spatial pyramid pooling structure, which captures multi-scale contextual information through parallel convolutional kernels with different expansion rates, in order to better adapt to urban areas with large differences in roof shapes. If the satellite image quality is high and there is little shadow interference, the deep learning discrimination step in step 3.2 can be skipped. The candidate region obtained by multi-scale segmentation in step 3.1 can be used as the basis to extract the initial contour through edge detection in step 3.3, and then the final contour can be generated through geometric regularization in step 3.4, so as to improve processing efficiency. In step 4, the objective function can be replaced by the bidirectional Hausdorff distance instead of the bidirectional chamfer distance; the bidirectional Hausdorff distance is defined as the maximum of two unidirectional Hausdorff distances, and its expression is: ; Where S represents the set of building roof outline corner points extracted from satellite imagery, and P represents the set of UAV model projection outline points. This represents the maximum value among the shortest distances from each point on contour S to contour P; The bidirectional Hausdorff distance is highly sensitive to extreme outliers and can effectively suppress contour shifts caused by building occlusion or local shadows. It is suitable for scenarios where there are large differences in building contour shapes or local outliers. In step 6, the adaptive robust weighting function can be replaced by the Hubel kernel function, whose expression is: when hour, ; when hour, ; in, To adjust the parameters, this embodiment takes... The Hubel kernel function maintains high weights for small residuals and appropriately reduces the weights for large residuals, making it simple to implement, computationally efficient, and suitable for scenarios with limited computing resources. In step 7, if there is no publicly available vector road network data coverage in the survey area or the road network data is incomplete, the rigid body correction step based on the road network in step 7.4 can be skipped, and the adjusted model can be used directly as the final output. At this time, the absolute positioning accuracy of the model mainly depends on the nominal geographic positioning accuracy of the satellite image itself, which is usually in the range of 2 to 5 meters. This degradation strategy is suitable for application scenarios with relatively relaxed requirements for absolute positioning accuracy, or as an emergency handling solution when road network data is missing.

[0077] Example 6 See Figure 10 This embodiment illustrates an electronic device that includes a memory, a processor, and a set of computer instructions stored in the memory and accessible to the processor. When the processor executes the set of instructions, it can run the UAV city LOD1 model reconstruction process based on satellite imagery virtual control constraints as described in Embodiment 5.

[0078] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, read-only optical discs, optical storage, etc.) containing computer-usable program code.

[0079] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0080] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for reconstructing a UAV city LOD1 model based on satellite imagery virtual control constraints, characterized in that, Includes the following steps: Acquire satellite optical remote sensing images of the target area with absolute geocoding, as well as UAV oblique photography images; extract building roof outlines from the satellite optical remote sensing images to generate building outline vector data with absolute geographic coordinates; perform relative orientation and sparse 3D reconstruction without ground control on the UAV oblique photography images to calculate the initial relative orientation elements and sparse 3D point cloud of the UAV images, and construct the initial 3D model of the UAV in the local coordinate system. Using the building outline vector data as matching primitives, cross-view coarse registration is performed between the satellite optical remote sensing image and the initial 3D model of the UAV, and the initial rigid body transformation relationship between the satellite optical remote sensing image and the initial 3D model is calculated. Based on the initial rigid body transformation relationship, contour feature points are selected from the building contour vector data, and the absolute geographic coordinates of the contour feature points are transformed into the local coordinate system of the initial three-dimensional model to generate satellite virtual control points. A joint bundle adjustment model including the satellite virtual control points is constructed. Adjustment calculation is performed based on the joint bundle adjustment model. During the calculation, the satellite virtual control points are incorporated as soft-constraint observations to obtain an adjusted and optimized 3D model. Based on the adjusted and optimized 3D model, a textureless LOD1 level city 3D model is generated.

2. The method for reconstructing a UAV city LOD1 model based on satellite imagery virtual control constraints according to claim 1, characterized in that, The satellite optical remote sensing image is a geocoded image with a resolution within a preset range, and the UAV oblique photography image is a multi-view image sequence with a forward overlap rate and a lateral overlap rate both reaching preset values.

3. The method for reconstructing a UAV city LOD1 model based on satellite imagery virtual control constraints according to claim 1, characterized in that, The roof outlines of buildings are extracted from the satellite optical remote sensing images using a cascaded outline extraction method, specifically: The satellite optical remote sensing image is segmented at multiple scales to obtain initial candidate building regions; A deep learning-based semantic segmentation network is used to initially identify candidate building regions and output a building probability map. A method combining morphological processing and edge detection is used to extract the initial building edge contours from the building probability map; A geometrically constrained contour regularization algorithm is used to optimize the initial building edge contour, including corner fitting, line segment approximation and contour closure, to generate building contour vector data.

4. The method for reconstructing a UAV city LOD1 model based on satellite imagery virtual control constraints according to claim 1, characterized in that, The cross-view coarse registration adopts the point-line iterative nearest point method; The step of transforming the absolute geographic coordinates of the contour feature points to the local coordinate system of the initial 3D model to generate satellite virtual control points specifically involves: The initial 3D model is downsampled to extract the surface model and the matching 2D projection contour; The satellite image building outline vector data and the two-dimensional projection outline of the initial three-dimensional model are unified to the same horizontal projection coordinate system to establish a point-line correspondence: for each point on the satellite outline, find the closest point on the UAV model projection outline and establish point-line pairing; Construct an objective function based on minimizing the distance between points and lines, and use the iterative nearest point method to minimize the objective function as a constraint to solve the rigid body transformation matrix, thereby obtaining the initial rigid body transformation relationship; Corner points or feature points uniformly sampled along the contour in the satellite image building outline vector data are selected, and the three-dimensional coordinates are transformed into the initial three-dimensional model coordinate system of the UAV image using the initial rigid body transformation relationship, serving as virtual control points.

5. The method for reconstructing a UAV city LOD1 model based on satellite imagery virtual control constraints according to claim 1, characterized in that, The specific method for incorporating the satellite virtual control points as soft-constraint observations during the solution process is as follows: In the bundle adjustment error equation for UAV imagery regional network adjustment, a virtual observation equation for virtual control points is added, as shown in the following equation: in, This is the correction vector for the virtual control points. Let X be the coefficient matrix corresponding to the virtual control points, and let X be the vector of unknowns containing corrections for the exterior orientation elements and corrections for the coordinates of the densified points. This represents the observation vector of the virtual control point; The initial weights of the virtual control points are set according to the nominal geolocation accuracy of the satellite imagery, which is the planar error value recorded in the satellite imagery metadata. The initial weights are lower than the weights of the UAV image point observations.

6. The method for reconstructing a UAV city LOD1 model based on satellite imagery virtual control constraints according to claim 1, characterized in that, During the adjustment calculation process, the exterior orientation elements and three-dimensional coordinates of the densified points of the UAV image are calculated simultaneously, and the adjusted and optimized three-dimensional model is generated based on the exterior orientation elements and three-dimensional coordinates of the densified points. An adaptive robust weight function is introduced during the adjustment process. This adaptive robust weight function adopts the IGG III scheme and has an equivalent weight function. satisfy: when hour, ; when hour, ; when hour, Where v is the standardized residual, The value range is 1.0 to 1.

5. The value ranges from 2.5 to 3.0; after each adjustment iteration, the weights of each virtual control point are recalculated based on the calculated residuals.

7. The method for reconstructing a UAV city LOD1 model based on satellite imagery virtual control constraints according to claim 1, characterized in that, The generation of a textureless LOD1 level city 3D model employs an elevation-robust road skeleton registration method, including: Extract the three-dimensional skeleton lines of the road network from the adjusted and optimized three-dimensional model; Obtain publicly available vector road network data for the same area as the satellite optical remote sensing image, and convert the publicly available vector road network data into a three-dimensional road skeleton line; Calculate the three-dimensional translation and rotation deviations between the three-dimensional skeleton lines of the road network and the three-dimensional road skeleton lines of the road network; Using the three-dimensional translation and rotation deviations as system correction parameters, the adjusted and optimized three-dimensional model is globally corrected to generate a textureless LOD1 level city three-dimensional model.

8. A UAV city LOD1 model reconstruction system based on satellite imagery virtual control constraints, characterized in that, include: The initial model building module is used to acquire satellite optical remote sensing images with absolute geocoding of the target area and UAV oblique photography images; extract building roof outlines from the satellite optical remote sensing images to generate building outline vector data with absolute geographic coordinates; perform relative orientation and sparse 3D reconstruction without ground control on the UAV oblique photography images to solve the initial relative orientation elements and sparse 3D point cloud of the UAV images, and construct the initial 3D model of the UAV in the local coordinate system. The transformation relationship calculation module is used to use the building outline vector data as matching primitives to perform cross-view coarse registration between the satellite optical remote sensing image and the UAV initial three-dimensional model, and calculate the initial rigid body transformation relationship between the satellite optical remote sensing image and the initial three-dimensional model. The virtual control point conversion module is used to select contour feature points in the building contour vector data based on the initial rigid body transformation relationship, convert the absolute geographic coordinates of the contour feature points to the local coordinate system of the initial three-dimensional model, and generate satellite virtual control points. The LOD1 model generation module is used to construct a joint bundle adjustment model including the satellite virtual control points, perform adjustment calculations based on the joint bundle adjustment model, incorporate the satellite virtual control points as soft-constraint observations during the calculation process, obtain an adjusted and optimized 3D model, and generate a textureless LOD1 level city 3D model based on the adjusted and optimized 3D model.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the UAV city LOD1 model reconstruction method based on satellite imagery virtual control constraints as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the UAV city LOD1 model reconstruction method based on satellite imagery virtual control constraints as described in any one of claims 1-7.

Images