Airborne array interference SAR point cloud building three-dimensional reconstruction method and device

Abstract

The invention discloses an airborne array interference SAR point cloud building three-dimensional reconstruction method and device, and belongs to the technical field of three-dimensional point cloud data processing, and the method comprises the steps: obtaining an original airborne array interference SAR point cloud, and carrying out the preprocessing, including point cloud denoising, point cloud registration, and point cloud filtering; based on the preprocessed point cloud data, building point cloud is extracted through a double-threshold cooperation mechanism of point cloud grid density and point cloud height features; based on the extracted building point cloud, building a building contour through an improved Alpha-Shape algorithm; and constructing a three-dimensional building model through an RANSAC (Random Sample Consensus) algorithm based on the building point cloud and the constructed building contour. The invention further relates to a three-dimensional building model and application thereof, the technical problem of airborne array interference SAR point cloud in building three-dimensional reconstruction is effectively solved, the reconstruction precision and efficiency are improved, and the method has important application value in the fields of digital city construction, building post-disaster reconstruction, city planning management and the like.

Description

Technical Field

[0001] This invention relates to a method and apparatus for three-dimensional reconstruction of buildings from airborne array interferometric SAR point clouds, belonging to the field of three-dimensional point cloud data processing technology. Background Technology

[0002] Interferometric Synthetic Aperture Radar (InSAR) technology possesses the core advantages of all-weather, all-time imaging and high resolution, playing an irreplaceable role in fields such as topographic mapping and digital city modeling. As an important extension of InSAR technology, Array InSAR 3D imaging technology uses multiple input / output (MIMO) technology to virtualize multiple equivalent antenna centers, receive ground echo signals, and generate multi-channel coherent SAR images. This ultimately achieves a breakthrough by enabling 3D imaging in a single flight, significantly improving the efficiency of 3D data acquisition.

[0003] However, due to objective factors such as differences in the characteristics of transmitting and receiving antennas and baseline errors, the three-dimensional imaging quality of array interferometric SAR faces significant challenges. The generated point clouds contain a large amount of noise, which directly restricts the application value of array interferometric SAR point clouds in practical scenarios and has become a key problem that urgently needs to be solved.

[0004] Interferometric SAR point clouds differ fundamentally from traditional LiDAR point clouds. Due to the unique imaging method and technical principles of interferometric SAR, its point clouds exhibit significant characteristics such as sparse point clouds perpendicular to the aircraft's flight direction and sparse point clouds on the rooftops of tall buildings, lower overall density, and inferior quality compared to LiDAR point clouds. These differences make it difficult for traditional 3D reconstruction methods to be directly adapted to interferometric SAR point clouds, greatly limiting their application scenarios in the field of 3D reconstruction.

[0005] While some point cloud-based 3D reconstruction methods exist in the current technology, these methods are mainly designed for LiDAR or TomoSAR point clouds and cannot effectively handle the noise problems and distribution characteristics unique to array interferometric SAR point clouds. In particular, existing technologies have significant shortcomings in dealing with multipath interference noise and triple scattering noise in array interferometric SAR point clouds. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a method and apparatus for three-dimensional reconstruction of buildings from airborne array interferometric SAR point clouds, which can solve the technical challenges faced by array interferometric SAR point clouds in three-dimensional reconstruction of buildings.

[0007] The technical solution adopted by this invention to solve its technical problem is as follows: In a first aspect, an embodiment of the present invention provides a method for three-dimensional reconstruction of buildings from airborne array interferometric SAR point clouds, comprising the following steps: Step S1: Obtain the original airborne array interferometric SAR point cloud and perform preprocessing, including point cloud denoising, point cloud registration, and point cloud filtering. Step S2: Based on the preprocessed point cloud data, extract the building point cloud through a dual threshold collaborative mechanism of point cloud grid density and point cloud height features; Step S3: Based on the extracted building point cloud, construct the building outline using an improved Alpha-Shape algorithm; Step S4: Based on the building point cloud and the constructed building outline, construct a 3D building model using the RANSAC algorithm.

[0008] As one possible implementation of this embodiment, the point cloud denoising includes: An adaptive threshold-based hybrid filtering algorithm (ATHFA) is used to remove outlier noise. The adaptive threshold-based hybrid filtering algorithm dynamically adjusts the filtering parameters according to the spatial distribution characteristics of the array interferometric SAR point cloud. A surface fitting based multipath noise filtering algorithm (SF-MNFA) is used to remove tri-scattering noise. The algorithm identifies and filters out noise points generated by multipath interference by fitting the geometric features of the building surface.

[0009] As one possible implementation of this embodiment, the point cloud registration includes: The Iterative Closest Point (ICP) algorithm is used to achieve point cloud registration in the same view direction, and accurate registration is achieved through feature point matching and optimization of the transformation matrix. Human-computer interaction is used to achieve point cloud registration on the opposite side, and spatial alignment of point clouds in different flight directions is achieved by selecting control points and feature regions.

[0010] As one possible implementation of this embodiment, the point cloud filtering includes: The Cloth Simulation Filter (CSF) algorithm is used to classify ground points and non-ground points. The filtering parameters are adaptively adjusted based on the terrain undulation characteristics to ensure the filtering effect under different terrain conditions; The classification results are optimized through post-processing steps to correct misclassified point clouds.

[0011] As one possible implementation of this embodiment, step S2 includes: The point cloud of building facades is extracted by using two parameters: the height percentage threshold of points within the grid and the density of points within the grid. Cluster analysis was performed on the extracted building facade point cloud using the 2D DBSCAN clustering algorithm; Optimize the clustering results using 2D morphological operations; Extract building point clouds using a mask to form point cloud data for a single building.

[0012] As one possible implementation of this embodiment, step S3 includes: Calculate the radius of the empty circumcircle of the Delaunay triangle mesh and determine the critical value of simple connectivity, then extract the initial edge points of the building; Contour key points were determined using the Douglas-Peucker Algorithm (DP) and regularization processing; Conduct inspection and optimization treatment of building grooves; Construct the final set of building outline points.

[0013] As one possible implementation of this embodiment, step S4 includes: Determine the top and bottom heights of a building based on its point cloud data; Fit the building roof plan using the RANSAC algorithm; A three-dimensional mesh model of the building was constructed using constrained Delaunay triangulation. Integrate the grids of the building's base, top, and side walls to form a complete closed 3D model.

[0014] Secondly, an embodiment of the present invention provides a device for three-dimensional reconstruction of buildings from airborne array interferometric SAR point clouds, comprising: The data preprocessing module is used to acquire the original airborne array interferometric SAR point cloud and perform preprocessing, including point cloud denoising, point cloud registration, and point cloud filtering. The building point cloud extraction module is used to extract building point clouds based on preprocessed point cloud data and through a dual-threshold collaborative mechanism of point cloud grid density and point cloud height features. The building outline construction module is used to construct building outlines based on the extracted building point cloud using an improved Alpha-Shape algorithm. The building model building module is used to construct 3D building models based on building point clouds and the constructed building outlines, using the RANSAC algorithm.

[0015] Thirdly, the three-dimensional building model provided in this embodiment of the invention is constructed using the method described above, and the three-dimensional building model includes: The base triangular mesh is formed by constrained Delaunay triangulation of the building outline key points; The top triangular mesh is copied from the bottom vertex and its z-coordinate is replaced with the building's top height. The sidewall triangular grid is formed by splitting the quadrilaterals composed of the bottom boundary edge and the corresponding top edge; The bottom triangular mesh, top triangular mesh, and sidewall triangular mesh are integrated through vertex data and triangular facet data to form a complete closed triangular mesh model.

[0016] Fourthly, the present invention provides an application of a three-dimensional building model as described above, for urban planning and management and / or digital city construction and / or post-disaster reconstruction of buildings.

[0017] One of the above technical solutions has the following advantages or beneficial effects: 1. A specialized point cloud preprocessing method is proposed to address the unique noise types and distribution characteristics of array interferometric SAR point clouds, effectively solving the problems of outlier noise and multipath interference noise in array interferometric SAR point clouds.

[0018] 2. A dual-threshold collaborative mechanism is adopted to extract building point clouds. Through the synergistic effect of point cloud grid density and point cloud height features, the accuracy and completeness of building point cloud extraction are effectively improved.

[0019] 3. An improved Alpha-Shape algorithm is proposed, which can accurately detect and represent the groove region in the building outline, thereby improving the accuracy of building outline reconstruction.

[0020] 4. The structured modeling method based on the RANSAC algorithm can automatically determine the top and bottom heights of a building and construct a complete 3D model of the building through constrained Delaunay triangulation.

[0021] 5. Through a systematic technical solution, an automated reconstruction process from raw point cloud to complete 3D model has been achieved, which greatly improves reconstruction efficiency and accuracy.

[0022] 6. The reconstructed 3D building models have wide application value in fields such as digital city construction, post-disaster reconstruction of buildings, and urban planning and management. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a method for three-dimensional reconstruction of buildings from airborne array interferometric SAR point clouds, according to an exemplary embodiment. Figure 2This is a schematic diagram of a device for three-dimensional reconstruction of building points from airborne array interferometric SAR, according to an exemplary embodiment. Figure 3 This is a single-view raw airborne array interferometric SAR point cloud map illustrated according to an exemplary embodiment; Figure 4 This is a preprocessed image of an airborne array interferometric SAR point cloud, illustrated according to an exemplary embodiment. Figure 5 This is a point cloud extraction map of a building according to an exemplary embodiment; Figure 6 This is an example of a three-dimensional reconstruction model and a preprocessed point cloud overlay. Detailed Implementation

[0024] To more clearly illustrate the technical features of the present invention, the present invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings.

[0025] Example 1 like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for three-dimensional reconstruction of buildings from airborne array interferometric SAR point clouds, comprising the following steps: Step S1: Obtain the original airborne array interferometric SAR point cloud and perform preprocessing, including point cloud denoising, point cloud registration, and point cloud filtering. Step S2: Based on the preprocessed point cloud data, extract the building point cloud through a dual threshold collaborative mechanism of point cloud grid density and point cloud height features; Step S3: Based on the extracted building point cloud, construct the building outline using an improved Alpha-Shape algorithm; Step S4: Based on the building point cloud and the constructed building outline, construct a 3D building model using the RANSAC algorithm.

[0026] As one possible implementation of this embodiment, the point cloud denoising includes: An adaptive threshold hybrid filtering algorithm is used to remove outlier noise. The adaptive threshold hybrid filtering algorithm dynamically adjusts the filtering parameters according to the spatial distribution characteristics of the array interferometric SAR point cloud. A multipath interference noise filtering algorithm based on surface fitting is used to remove triple scattering noise. The algorithm identifies and filters out noise points generated by multipath interference by fitting the geometric features of the building surface.

[0027] The adaptive threshold hybrid filtering algorithm calculates the local density features of the point cloud and determines the neighborhood density of each point based on the k-nearest neighbor statistical method. According to the neighborhood density distribution characteristics, a Gaussian mixture model is used to adaptively determine the filtering threshold. Combining the statistical outlier removal algorithm and the radius outlier removal algorithm, outlier noise points are effectively removed while maintaining the structural features of the building.

[0028] The multipath interference noise filtering algorithm based on planar fitting extracts the main planar features of buildings using the RANSAC algorithm; constructs geometric constraints based on the extracted planar features to identify noise points that do not conform to the planar geometric constraints; and distinguishes between real building points and multipath interference noise points by combining point cloud reflection intensity information and spatial distribution features.

[0029] As one possible implementation of this embodiment, the point cloud registration includes: An iterative nearest-point algorithm is used to achieve point cloud registration in the same-side view direction, and accurate registration is achieved through feature point matching and optimization of the transformation matrix. Human-computer interaction is used to achieve point cloud registration on the opposite side, and spatial alignment of point clouds in different flight directions is achieved by selecting control points and feature regions.

[0030] The iterative nearest point algorithm registration process is as follows: extract key feature points from the point cloud, including corner points, edge points, and planar points; perform initial matching based on feature descriptors to establish the correspondence between point clouds; optimize the rotation matrix and translation vector using the least squares method, and iteratively calculate until the registration error converges.

[0031] The human-computer interaction registration process is as follows: select no less than 3 evenly distributed control points in the point cloud; establish a spatial transformation model based on the control points and calculate the registration parameters; verify the registration accuracy through feature region matching, and make manual adjustments when necessary.

[0032] As one possible implementation of this embodiment, the point cloud filtering includes: A cloth-simulation filtering algorithm is used to classify ground points and non-ground points; The filtering parameters are adaptively adjusted based on the terrain undulation characteristics to ensure the filtering effect under different terrain conditions; The classification results are optimized through post-processing steps to correct misclassified point clouds.

[0033] The fabric simulation filtering algorithm projects the point cloud onto a two-dimensional plane to construct a virtual fabric model; determines the contact points between the fabric and the terrain surface through physical simulation; and distinguishes between ground points and non-ground points based on the height difference of the contact points to form an accurate ground point cloud.

[0034] In step S1, the execution order of point cloud denoising, point cloud registration, and point cloud filtering is as follows: first, point cloud denoising is performed to remove obvious outlier noise and multipath interference noise; then, point cloud registration is performed to achieve spatial alignment of point clouds from multiple perspectives; finally, point cloud filtering is performed to separate ground points and non-ground points to obtain a point cloud containing only ground feature information.

[0035] As one possible implementation of this embodiment, step S2 includes: The point cloud of building facades is extracted by using two parameters: the height percentage threshold of points within the grid and the density of points within the grid. Cluster analysis was performed on the extracted building facade point cloud using the 2D DBSCAN clustering algorithm; Optimize the clustering results using 2D morphological operations; Extract building point clouds using a mask to form point cloud data for a single building.

[0036] The method of extracting building facade point clouds using two parameters—the height proportion threshold and the density of points within the grid—in collaboration (dual threshold collaboration mechanism) includes: calculating the average height of the point cloud as a reference for the height proportion threshold; projecting the point cloud onto the XY plane and dividing it into regular grids; counting the number of points and the height distribution characteristics of the points in each grid cell; and determining the building facade point cloud based on the collaborative judgment of the height proportion threshold and the point density threshold.

[0037] The calculation of the height percentage threshold includes: calculating the average height of the entire point cloud dataset; for each grid cell, counting the proportion of points with heights higher than the average height; setting the height percentage threshold to filter out grid regions with obvious vertical structures.

[0038] The determination of the point density threshold includes: analyzing the overall distribution density characteristics of the point cloud; considering the low density of the point cloud in the array interferometric SAR, setting an appropriate density threshold; and adaptively adjusting the density threshold parameter according to the building size and the point cloud resolution.

[0039] The 2D DBSCAN clustering algorithm includes the following steps: performing cluster analysis on the set of building facade points in the projection plane; forming a point cloud of a single building facade based on the principles of spatial proximity and density reachability; and assigning a unique label to each building for easy subsequent individual processing.

[0040] The 2D morphological operation includes the following steps: performing binary rasterization on the clustered building point cloud; using a closing operation to connect adjacent building regions and fill small gaps; and using a filling operation to eliminate holes inside the building outline.

[0041] The binary rasterization process includes: setting the raster size to 1 to 2 times the average point spacing; setting the value of the raster to 1 if there is a point in the raster, otherwise setting it to 0; and forming a binary raster representation of the building.

[0042] The closing operation includes: using square structuring elements to perform an expansion and then erosion operation on the grid; connecting adjacent building areas to form a continuous building outline; and smoothing the outline boundary while maintaining the main shape features of the building.

[0043] The filling operation includes: calculating the connected components of the raster graph using connectivity criteria; identifying the largest connected component as the main building area; and filling the holes in the largest connected component to form a complete building mask.

[0044] The mask extraction includes: using the morphologically processed raster image as a mask; selectively extracting the preprocessed original point cloud on the projection plane; obtaining complete point cloud data of a single building, providing a foundation for subsequent contour construction.

[0045] As one possible implementation of this embodiment, step S3 includes: Calculate the radius of the empty circumcircle of the Delaunay triangle mesh and determine the critical value of simple connectivity, then extract the initial edge points of the building; Key points of the contour were determined using the Douglas-Puk algorithm and regularization processing. Conduct inspection and optimization treatment of building grooves; Construct the final set of building outline points.

[0046] The calculation of the radius of the empty circumcircle of the Delaunay triangle mesh and the determination of the critical value of single connectivity include the following steps: constructing a Delaunay triangle mesh of the point cloud; calculating the radius of the empty circumcircle of each triangle to form a radius spectrum; determining the critical value of single connectivity and extracting the outline shape that is closest to the main body of the building.

[0047] The Delaunay triangle mesh construction is based on the following criteria: the empty circle property, ensuring that no other point in the mesh is inside the circumcircle of any triangle; maximizing the minimum angle property, optimizing the quality of the triangle mesh; and providing a geometric basis for the Alpha-Shape algorithm.

[0048] The determination of the single connectivity threshold includes: sorting all empty circumcircle radii after deduplication to form a radius spectrum; finding the minimum value in the spectrum that first connects the Alpha shapes; and using this threshold value to extract the building edge points, balancing the integrity of the outline and the preservation of details.

[0049] The Douglas-Puk algorithm processing includes: simplifying the extracted edge points to find the key inflection points of the building; grouping the edge points by the inflection points; and performing least-squares line fitting on the edge points in different groups.

[0050] The least squares line fitting includes: performing line fitting on each group of edge points, with each endpoint not participating in the fitting; finding the intersection of two adjacent fitted lines as the initial key corner points of the building; judging the rationality of the intersection points through a distance threshold, and correcting the corner points if necessary.

[0051] The regularization process includes: determining the main direction vector of the building as the core reference for contour regularization; calculating the angle between each side vector and the main direction to classify the direction; and adjusting the building boundary according to the direction judgment threshold.

[0052] The determination of the main direction is based on the following criteria: calculate all vectors formed by adjacent initial key corner points and their magnitudes; select the vector with the largest magnitude as the main direction vector of the building; and establish the core reference direction for the regular outline.

[0053] The building recess detection includes: connecting the key points of the outline sequentially to form the building boundary; calculating the deviation of the interior angle of adjacent sides from 90 degrees to identify abnormal points; and detecting the recess area through a continuous abnormal point pattern.

[0054] The groove optimization process includes: projecting four points in the detected groove area; projecting the groove boundary points onto the corresponding straight lines to obtain a regular groove outline; and updating the set of key points of the building outline to form the final building outline.

[0055] As one possible implementation of this embodiment, step S4 includes: Determine the top and bottom heights of a building based on its point cloud data; Fit the building roof plan using the RANSAC algorithm; A three-dimensional mesh model of the building was constructed using constrained Delaunay triangulation. Integrate the grids of the building's base, top, and side walls to form a complete closed 3D model.

[0056] The building height determination (determining the top and bottom height of the building based on the building point cloud) includes: calculating the 90th quantile of the z-coordinate of the building point cloud as a height threshold; extracting points higher than the height threshold as roof candidate points; fitting the roof plane using the RANSAC algorithm and taking the z-value of the center point of the plane as the building top height; and taking the height of the lowest point in the building point cloud as the bottom height.

[0057] The RANSAC roof plane fitting (fitting the building roof plane using the RANSAC algorithm) includes: randomly selecting the minimum set of points from the roof candidate points to fit the plane model; calculating the distance from other points to the plane and counting the number of interior points; and finding the plane with the most interior points as the optimal roof plane through iterative optimization.

[0058] The constrained Delaunay triangulation (using constrained Delaunay triangulation to construct a 3D mesh model of a building) includes: constructing a bottom triangular mesh based on the key points of the building outline; copying the bottom vertices and replacing the z-coordinate with the top height to generate a top triangular mesh; and constructing a side wall triangular mesh along each boundary of the final outline.

[0059] The process of integrating the building's bottom, top, and side wall meshes to form a complete closed 3D model includes: constructing the building's bottom triangular mesh; constructing the building's top triangular mesh; constructing the building's side wall mesh; stitching together the vertex data of the building's bottom, top, and side walls; integrating all triangular facet data to form a complete triangular mesh model; ensuring the model's closure and geometric consistency.

[0060] The construction of the base triangular mesh includes: combining the x and y coordinates and the base height z value of the key points of the building outline; performing constrained Delaunay triangulation to ensure the integrity of the outline boundary; and forming the base triangular mesh representation of the building.

[0061] The construction of the top triangular mesh includes: copying the vertex coordinates of the bottom triangular mesh; replacing the z-coordinate of the top vertex with the building's top height; and maintaining the same triangular mesh topology as the bottom mesh.

[0062] The sidewall grid construction includes: each boundary along the final outline of the building; quadrilaterals formed by the bottom boundary edge and the corresponding top edge; and splitting each quadrilateral into two triangular faces to form the sidewall grid.

[0063] The method for constructing 3D building models also includes a quality assessment step: calculating point cloud coverage to assess the degree to which the reconstructed model covers the original point cloud; analyzing shape similarity factors, including area difference, perimeter difference, and height difference; and providing quantitative evaluation indicators for reconstruction accuracy.

[0064] Example 2 like Figure 2 As shown in the figure, an embodiment of the present invention provides a device for three-dimensional reconstruction of buildings from airborne array interferometric SAR point clouds, comprising: The data preprocessing module is used to acquire the original airborne array interferometric SAR point cloud and perform preprocessing, including point cloud denoising, point cloud registration, and point cloud filtering. The building point cloud extraction module is used to extract building point clouds based on preprocessed point cloud data and through a dual-threshold collaborative mechanism of point cloud grid density and point cloud height features. The building outline construction module is used to construct building outlines based on the extracted building point cloud using an improved Alpha-Shape algorithm. The building model building module is used to construct 3D building models based on building point clouds and the constructed building outlines, using the RANSAC algorithm.

[0065] The specific process of using the device described in this embodiment to perform three-dimensional reconstruction of buildings from airborne array interferometric SAR point clouds is as follows.

[0066] 1. Acquire the original airborne array interferometric SAR point cloud and perform preprocessing.

[0067] Preprocessing of the raw airborne array interferometric SAR point cloud includes point cloud denoising, registration, and filtering. An adaptive threshold-based hybrid filtering algorithm (ATHFA) is used to remove outlier noise, and a surface fitting-based multipath noise filtering algorithm (SF-MNFA) is used to remove tri-scattering noise. In Cloud Compare software, the Iterative Closest Point (ICP) algorithm is used for same-side view registration, and a human-computer interaction method is used for opposite-side point cloud registration. The Cloth Simulation filter (CSF) algorithm in Cloud Compare software is used to classify ground points and non-ground points, thus obtaining a point cloud containing only ground feature information. .

[0068] 2. Based on the preprocessed point cloud data, the building point cloud is extracted through a dual-threshold collaborative mechanism of point cloud grid density and point cloud height features.

[0069] Use the height percentage threshold of points within the grid and grid point density Two parameters are used to extract the point cloud of the building facade. First, the point cloud is calculated according to formula (1). average height : (1), in, Point cloud average height, express Total number of point clouds Indicates the first The height of each point This represents the summation operator; Then Projected to In the plane, with side length The projection plane is divided into regular grids, and the position of each cell grid is determined by... This indicates the number of points within each grid cell. The definition is shown in formula (2): (2), in, Represents the set of points within a grid cell. Operators that count the number of elements in a set; Use formula (3) to calculate the point in the cell that is higher than percentage : (3), Calculate each cell sequentially and Points within the grid that satisfy the condition of formula (4) are added to the building facade point set. middle: (4); On the projection plane Internal DBSCAN algorithm for Clustering is performed to form point clouds of individual building facades. 'k' is the label for each building. (On the projection plane) Internal Perform binary rasterization, with a raster size of [value missing]. Generally, the value is set to 1 to 2 times the average pixel spacing. If a pixel exists in the raster, the value of that raster is set to 1; if no pixel exists, the value of that raster is set to 0, resulting in a binary raster image. For binary raster images Perform morphological operations, including closing and filling operations. Utilize square structuring elements. The raster is first dilated and then eroded to obtain the raster image. Then, the 8-connectivity criterion is used to calculate its connected components, and the holes in the largest connected component are filled to obtain the final raster image. raster image As a mask, on the projection plane The above is a preprocessed point cloud. Point cloud extraction is performed, and the extracted point cloud It refers to point clouds of a single building.

[0070] 3. Based on the extracted building point cloud, the building outline is constructed by improving the Alpha-Shape algorithm.

[0071] The projected point cloud is constructed into a Delaunay triangle mesh according to the two criteria of empty circle property and maximizing minimum angle property. The radius of the empty circumcircle of each triangle in the mesh is calculated according to formula (5). : (5), in, Let E represent the lengths of the three sides of a triangle, i.e., the Euclidean distance between two points. The formula for calculating this distance is shown in formula (6): (6), in , , These are the coordinates of the three vertices of the triangle. The formula for calculating the area of ​​a triangle is shown in formula (7): (7); All After deduplication and sorting, we get Spectrum, when set Parameter greater than The shape of the alpha pattern formed during the spectrum changes. The smallest element in the spectrum that first connects the alpha shapes into a single block. Defined as a simply connected critical value Therefore The resulting Alpha shape is the outline that best fits the main body of the building, based on... Extract building edge points .

[0072] The Douglas-Peucker Algorithm (DP) is used to... Process the data, identify the inflection points of the buildings, and use these inflection points as boundaries for... The data is grouped, and least-squares line fitting is performed on the edge points within each group, excluding the endpoints of each segment from the fitting process. The intersection of two adjacent fitted lines is taken as the initial key corner point of the building. To avoid the intersection point deviating from the original contour, the minimum distance from the intersection point to the two groups of edge points is calculated. If the distance between two points is greater than three times the average distance between edge points, the intersection point is considered to be significantly off-center. In this case, the common endpoint of the two sets is used to replace the intersection point as the initial critical corner point.

[0073] To obtain a regularized building outline, a systematic processing of the initial key corner points is required. The specific process is as follows: First, calculate the vector formed by adjacent initial key corner points and its magnitude, and select the vector with the largest magnitude as the main direction vector of the building, thereby establishing the core reference direction for the regularized outline; then, simultaneously calculate the midpoint coordinates of each vector and the unit direction vector of each side. unit vector in the principal direction The absolute value of the cosine of the included angle , Represents the dot product of two vectors. This represents the absolute value operator. And sets... , Thresholds are determined in two directions. If... Determine if the line containing the vector is parallel to the principal direction, and unify its slope to the slope of the principal direction; if If the line is determined to be perpendicular to the principal direction, its slope is replaced with a slope perpendicular to the principal direction; if If the slope of the original line remains unchanged, then the regularized line for each side is determined using the point-slope form of the line equation, based on the midpoint of the vector and the regularized slope. (8), in Let these be the coordinates of the midpoint of the vector. The slope of the line.

[0074] Finally, the intersection points of adjacent regularized lines are calculated as the final key points of the building outline. The final set of key points of the building outline is denoted as... .

[0075] Some buildings may also have noticeable recessed areas, which can be identified solely by the aforementioned key points of the building's outline. It is difficult to form a regular groove profile. Taking advantage of the characteristic that "adjacent edges of buildings are usually perpendicular," this algorithm performs the following processing. Connect the points in the diagram sequentially to form the building boundary, and calculate the interior angles of adjacent sides. If the deviation is greater than the threshold The common points on both sides are then considered outliers; all key points are checked sequentially, and if four consecutive outliers are found, then a decision is made. There is a groove region in memory. Let the four points of the groove region be denoted as follows: , , , ; put point and points Projected onto points respectively and points Two new projection points are obtained on the formed straight line. and ;Will Points in and points Replace with dots and Update the final building outline key points If no groove area is detected, then Unchanged. Connect sequentially. The outline of a building can be obtained by using the points in the diagram.

[0076] The specific improvements to the Alpha-Shape algorithm include: 1) Automatic determination mechanism for critical values ​​of simply connected systems: The traditional Alpha-Shape algorithm requires manual setting of the α parameter, which introduces subjectivity. The improved algorithm constructs a Delaunay triangle mesh and calculates the radius of the empty circumcircle of each triangle; All radii are deduplicated and sorted to form a radius spectrum, and the minimum critical value that makes the Alpha shape connect into one piece is automatically determined; This threshold balances the need for both outline integrity and detail preservation.

[0077] 2) Contour optimization combined with the Douglas-Puk algorithm: After extracting the initial edge points, the Douglas-Puk algorithm is used to identify key inflection points; The edge points are grouped by the inflection point, and the least squares line fitting is performed on the edge points in different groups respectively. The intersection of two adjacent fitted lines is taken as the initial key corner point of the building; The rationality of the intersection point is determined by the distance threshold, and corner point correction is performed if necessary.

[0078] 3) Regularization based on the principal direction vector: Calculate all vectors formed by adjacent initial key corner points and their magnitudes; Select the vector with the largest magnitude as the main direction vector of the building; Calculate the cosine of the angle between each side vector and the main direction, and set the direction judgment threshold; The building boundaries are adjusted according to thresholds to ensure the geometric regularity of the outline.

[0079] 4) Building Groove Detection and Optimization Mechanism: Connect the key points of the outline in sequence and calculate the deviation of the interior angle of the adjacent side from 90 degrees; The groove region is detected by continuous anomaly point pattern, and four key points in the groove region are identified. Projecting the groove boundary points onto the corresponding straight line yields a regular groove profile. Update the set of key points for the building outline to form the final regularized building outline; A technical comparison between the improved Alpha-Shape algorithm and traditional algorithms reveals the following advantages: Compared to the traditional Alpha-Shape algorithm, the improved algorithm can automatically determine the optimal parameters, reducing manual intervention; By combining the advantages of multiple algorithms, complex structures such as grooves can be effectively handled while maintaining the main shape features of the building. Through regularization processing, standard outlines that conform to the geometric characteristics of actual buildings are generated; It can still maintain high contour reconstruction accuracy under sparse point cloud conditions.

[0080] IV. Based on the building point cloud and the constructed building outline, a 3D building model is constructed using the RANSAC algorithm.

[0081] Initially screen candidate rooftop points using the property that "rooftop points are usually the highest points in the building's point cloud": Calculation The 90th quantile of the z-coordinate is used as the height threshold. ,extract Medium to large Points as candidate points The RANSAC algorithm is used to fit the plane that best represents the roof, and the z-value of the center point of this plane is taken as the building's roof height. , The height z of the lowest point in the middle is taken as the base height. ,combination Midpoint coordinates as well as The bottom triangular mesh is constructed by constraining Delaunay triangulation, and then the bottom vertices are copied and their z-coordinates are replaced. Generate a top triangular mesh. Then, along each boundary of the final contour, split the quadrilateral formed by the bottom boundary edge and the corresponding top edge into two triangular faces to form a side wall mesh. Finally, stitch together the vertices and triangular faces of the bottom, top, and side walls to integrate them into a complete closed triangular mesh, thus completing the modeling of a single building. According to the DBSCAN clustering situation, process each building in sequence until all buildings are modeled.

[0082] The point cloud data used in this embodiment is airborne array interferometric SAR point cloud acquired using the array interferometric SAR system developed by the Aerospace Information Research Institute of the Chinese Academy of Sciences. Figure 3 It can be seen that the red box represents outlier noise, the green box represents triple scattering noise, and the blue box represents point clouds of buildings in the single-view direction, showing the characteristics of sparse point clouds on the side and top surfaces and uneven distribution of the overall point cloud.

[0083] pass Figure 3 and Figure 4 contrast, Figure 4 The results significantly eliminated outlier noise and third-scatter noise, and achieved registration of point clouds on the same side and opposite side, while also separating ground feature points from ground points.

[0084] Figure 5 The point cloud images of buildings were extracted. It can be seen that there are no other ground features around each building's point cloud, and the point cloud of each building is completely extracted. In order to quantitatively evaluate the building extraction results, the original point cloud was processed by human-computer interaction to obtain reference data for the building point cloud. The filtering error evaluation criteria proposed by ISPRS in 2003 are adopted, including Type I error, Type II error and total error. Type I error refers to the error of misclassifying building points as non-building points, Type II error refers to the error of misclassifying non-building points as building points, and total error refers to the proportion of Type I and Type II error points to the total number of points, as shown in Table 1.

[0085] Table 1 Error Definition Table

[0086] Where e and f represent the number of building points and non-building points in the reference data, g and h are the number of building points and non-building points in the extracted data, respectively, and n is the total number of scanned data points. The three error calculation formulas are shown in formula (9): (9); The results show that the Type I error is 2.99%, the Type II error is 0.86%, and the total error is 2.22%. Looking at the error breakdown, the Type II error is less than 1%, which is at a low level. This indicates that during the point cloud data extraction process, the probability of misclassifying non-building points such as background, vegetation, and ground as building points is low, effectively avoiding the introduction of false building features, ensuring the "purity" of the building point cloud data, and reducing the interference of redundant information on the true form of buildings in subsequent modeling. Although the Type I error is slightly higher than the Type II error, it is still at a low level, indicating that the proportion of points in the edge, detail, or local area of ​​the building itself that are misclassified as non-building points and thus missed is limited. There was no large-scale loss of building features, and the point cloud data's ability to capture the main structure and key features of buildings is relatively reliable, with a controllable impact on data integrity. The total error of 2.22% indicates that the extracted building point cloud data has a high overall consistency with the reference data, preserving the true features of buildings to the greatest extent while effectively filtering out interference from non-building points.

[0087] Figure 6This is an overlay image of the 3D reconstructed model and the extracted point cloud of the buildings. It can be seen that each building has been modeled using a streamlined process, the recessed areas of the buildings are represented, the model is regular and consistent with the actual building shapes. To enable quantitative evaluation of the modeling results, point cloud coverage was defined. The shape similarity factor is calculated using formula (10) to determine the percentage of points in the original point cloud that are within the reconstructed model. (Area difference) (perimeter difference) and (Height difference). Planar reference data is obtained by averaging multiple measurements on optical imagery, while height reference data is obtained by acquiring building floor data from publicly available map services and calculating it as 3m per floor. (10), in, The area obtained by the method of this invention; Area obtained for reference only; The perimeter obtained by the method of this invention; Perimeter obtained for reference data; The height obtained by the method of this invention; The height was obtained for reference data; the final accuracy results are shown in Table 2.

[0088] Table 2. Building Model Accuracy Evaluation Table

[0089] Table 2 shows the building point cloud coverage. Between 84.58% and 92.74%, the overall accuracy for this region is 89.53%. The reason why the success rate is less than 90% is because the model is a LOD1 model that only considers the main building, and the auxiliary point clouds are not included in the modeling. Figure 6 As can be seen, this model can wrap the point cloud of buildings. Area difference Perimeter difference Height difference The overall accuracies were 0.94%, 1.16%, and 1.10%, respectively, all of which were within the low deviation range. This indicates that the model has high accuracy in the restoration of shape and size, which fully demonstrates the reliability of the airborne array interferometric SAR point cloud building 3D reconstruction method.

[0090] The results above show that the airborne array interferometric SAR point cloud building 3D reconstruction method can achieve high-quality array interferometric SAR building point cloud systematic and procedural modeling, enrich the technical system of array SAR 3D reconstruction, and has good technical effect and practical value.

[0091] Example 3 This invention provides a three-dimensional building model constructed using the method described in this invention. The three-dimensional building model includes: The base triangular mesh is formed by constrained Delaunay triangulation of the building outline key points; The top triangular mesh is copied from the bottom vertex and its z-coordinate is replaced with the building's top height. The sidewall triangular grid is formed by splitting the quadrilaterals composed of the bottom boundary edge and the corresponding top edge; The bottom triangular mesh, top triangular mesh, and sidewall triangular mesh are integrated through vertex data and triangular facet data to form a complete closed triangular mesh model.

[0092] The airborne array interferometric SAR point cloud 3D building reconstruction method provided by this invention can accurately adapt to the typical data characteristics of array interferometric SAR point clouds, such as high noise ratio, uneven distribution, and low density. Through targeted point cloud processing, building point cloud extraction, and contour construction, this method ultimately achieves structured building modeling and can efficiently and rapidly generate 3D building models in urban scenes.

[0093] Example 4 The present invention provides an application of a three-dimensional building model in urban planning and management, which includes at least: The three-dimensional visualization of the regulatory detailed plan intuitively displays the planning intent and spatial layout through the model; Three-dimensional compliance review of architectural design schemes, and precise verification of building height, area and perimeter based on models; The city's colors and building facade styles are managed in a unified manner, and visual harmony is analyzed through models.

[0094] Example 5 The present invention provides an application of a three-dimensional building model in digital city construction, which includes at least: As the core data foundation of urban three-dimensional geographic information systems, it enables the construction of digital twins of real urban environments; It supports the integrated management of building attribute information and spatial location, enabling multi-dimensional data fusion analysis; Used for dynamic updating and version management of urban spatial data to ensure the timeliness of urban information; Spatial relationship analysis between municipal pipelines and buildings supports pipeline planning and maintenance decisions.

[0095] Example 6 The present invention provides an application of a three-dimensional building model in post-disaster reconstruction of buildings, which includes at least: Based on rapidly acquired array interferometric SAR point cloud data after the disaster, a three-dimensional model of the post-disaster buildings was reconstructed. The damage level is analyzed by comparing it with the pre-disaster model, and the buildings that have collapsed, are severely damaged, or are slightly damaged are automatically identified. Risk analysis of collapsed buildings, and structural safety assessment based on models; Emergency evacuation simulation and rescue route planning are conducted, and scientific emergency plans are developed using models.

[0096] The three-dimensional building models reconstructed by this invention, with their high precision, high efficiency, and wide coverage, demonstrate significant technical advantages in the aforementioned application areas, and can provide strong technical support for urban digital transformation and refined management.

[0097] 1. A specialized point cloud preprocessing method is proposed to address the unique noise types and distribution characteristics of array interferometric SAR point clouds, effectively solving the problems of outlier noise and multipath interference noise in array interferometric SAR point clouds.

[0098] 2. A dual-threshold collaborative mechanism is adopted to extract building point clouds. Through the synergistic effect of point cloud grid density and point cloud height features, the accuracy and completeness of building point cloud extraction are effectively improved.

[0099] 3. An improved Alpha-Shape algorithm is proposed, which can accurately detect and represent the groove region in the building outline, thereby improving the accuracy of building outline reconstruction.

[0100] 4. The structured modeling method based on the RANSAC algorithm can automatically determine the top and bottom heights of a building and construct a complete 3D model of the building through constrained Delaunay triangulation.

[0101] 5. Through a systematic technical solution, an automated reconstruction process from raw point cloud to complete 3D model has been achieved, which greatly improves reconstruction efficiency and accuracy.

[0102] 6. The reconstructed 3D building models have wide application value in fields such as digital city construction, post-disaster reconstruction of buildings, and urban planning and management.

[0103] 7. Through experimental verification, the present invention has achieved significant technical results in the three-dimensional reconstruction of buildings from array interferometric SAR point clouds, with a point cloud coverage of 89.53% and area difference, perimeter difference and height difference of only 0.94%, 1.16% and 1.10%, respectively.

[0104] 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 claims of the present invention.

Claims

1. A method for three-dimensional reconstruction of buildings from airborne array interferometric SAR point clouds, characterized in that, Includes the following steps: Step S1: Obtain the original airborne array interferometric SAR point cloud and perform preprocessing, including point cloud denoising, point cloud registration, and point cloud filtering. Step S2: Based on the preprocessed point cloud data, extract the building point cloud through a dual threshold collaborative mechanism of point cloud grid density and point cloud height features; Step S3: Based on the extracted building point cloud, construct the building outline using an improved Alpha-Shape algorithm; Step S4: Based on the building point cloud and the constructed building outline, construct a 3D building model using the RANSAC algorithm.

2. The method according to claim 1, characterized in that, The point cloud denoising includes: An adaptive threshold hybrid filtering algorithm is used to remove outlier noise. The adaptive threshold hybrid filtering algorithm dynamically adjusts the filtering parameters according to the spatial distribution characteristics of the array interferometric SAR point cloud. A multipath interference noise filtering algorithm based on surface fitting is used to remove triple scattering noise. The algorithm identifies and filters out noise points generated by multipath interference by fitting the geometric features of the building surface.

3. The method according to claim 1, characterized in that, The point cloud registration includes: An iterative nearest-point algorithm is used to achieve point cloud registration in the same-side view direction, and accurate registration is achieved through feature point matching and optimization of the transformation matrix. Human-computer interaction is used to achieve point cloud registration on the opposite side, and spatial alignment of point clouds in different flight directions is achieved by selecting control points and feature regions.

4. The method according to claim 1, characterized in that, The point cloud filtering includes: A cloth-simulation filtering algorithm is used to classify ground points and non-ground points; The filtering parameters are adaptively adjusted based on the terrain undulation characteristics to ensure the filtering effect under different terrain conditions; The classification results are optimized through post-processing steps to correct misclassified point clouds.

5. The method according to claim 1, characterized in that, Step S2 includes: The point cloud of building facades is extracted by using two parameters: the height percentage threshold of points within the grid and the density of points within the grid. Cluster analysis was performed on the extracted building facade point cloud using the 2D DBSCAN clustering algorithm; Optimize the clustering results using 2D morphological operations; Extract building point clouds using a mask to form point cloud data for a single building.

6. The method according to claim 1, characterized in that, Step S3 includes: Calculate the radius of the empty circumcircle of the Delaunay triangle mesh and determine the critical value of simple connectivity, then extract the initial edge points of the building; Key points of the contour were determined using the Douglas-Puk algorithm and regularization processing. Conduct inspection and optimization treatment of building grooves; Construct the final set of building outline points.

7. The method according to any one of claims 1-6, characterized in that, Step S4 includes: Determine the top and bottom heights of a building based on its point cloud data; Fit the building roof plan using the RANSAC algorithm; A three-dimensional mesh model of the building was constructed using constrained Delaunay triangulation. Integrate the grids of the building's base, top, and side walls to form a complete closed 3D model.

8. A device for three-dimensional reconstruction of buildings from airborne array interferometric SAR point clouds, characterized in that, include: The data preprocessing module is used to acquire the original airborne array interferometric SAR point cloud and perform preprocessing, including point cloud denoising, point cloud registration, and point cloud filtering. The building point cloud extraction module is used to extract building point clouds based on preprocessed point cloud data and through a dual-threshold collaborative mechanism of point cloud grid density and point cloud height features. The building outline construction module is used to construct building outlines based on the extracted building point cloud using an improved Alpha-Shape algorithm. The building model building module is used to construct 3D building models based on building point clouds and the constructed building outlines, using the RANSAC algorithm.

9. A three-dimensional building model, characterized in that, The three-dimensional building model is constructed using the method described in any one of claims 1-7, and includes: The base triangular mesh is formed by constrained Delaunay triangulation of the building outline key points; The top triangular mesh is copied from the bottom vertex and its z-coordinate is replaced with the building's top height. The sidewall triangular grid is formed by splitting the quadrilaterals composed of the bottom boundary edge and the corresponding top edge; The bottom triangular mesh, top triangular mesh, and sidewall triangular mesh are integrated through vertex data and triangular facet data to form a complete closed triangular mesh model.

10. An application of the three-dimensional building model as described in claim 9, characterized in that, Used for urban planning and management and / or digital city construction and / or post-disaster reconstruction of buildings.

Images