Ground-air point cloud registration method in urban scene

Abstract

The invention discloses a ground-air point cloud registration method in an urban scene, and relates to a point cloud registration technology, and the method comprises the steps: obtaining ground point cloud two-dimensional building angular points of a target region based on point density projection and Hough transform, and constructing a ground angular point topological relation atlas; obtaining airborne point cloud two-dimensional building angular points of the target area based on morphological operation and Hough transform, and constructing an airborne point cloud angular point topological relation atlas; calculating the similarity between each ground angular point topological relation graph in the ground angular point topological relation graph set and each airborne point cloud angular point topological relation graph in the airborne point cloud angular point topological relation graph set, and determining an initial corresponding point; screening and optimizing the initial corresponding points by using a random sample consistency algorithm; and performing horizontal point cloud registration of the ground point cloud according to the screened corresponding point set, and completing vertical registration based on a horizontal point cloud registration result and the airborne point cloud. According to the method, the density difference and the scale difference of the point cloud data are comprehensively considered, and the point cloud registration accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud registration, and more particularly to a ground-to-air point cloud registration method in an urban scene. Background Art

[0002] With the acceleration of urbanization, fields such as urban 3D modeling, topographic mapping, and infrastructure management have placed higher demands on the accuracy and efficiency of spatial data acquisition. Terrestrial Laser Scanning (TLS) and Airborne Laser Scanning (ALS) technologies have been widely used in the collection and updating of urban geographic spatial data due to their ability to efficiently acquire high-precision 3D spatial information. TLS can acquire high-density, high-quality facade point clouds, but due to limitations such as observation angle, top point cloud information is easily lost. ALS can acquire top information of objects in a short period of time, but facade details are blurred and missing. Therefore, aligning the two point cloud data can obtain more comprehensive and accurate 3D scene point cloud data. However, point cloud data acquired from different platforms often have problems such as density differences, scale differences, and large rotation angles, making cross-source point cloud registration more difficult.

[0003] The registration method for ground-based lidar point clouds and airborne lidar point clouds includes two key steps: feature point extraction and matching, and geometric transformation parameter calculation. Feature point extraction and matching are used to establish spatial correspondences between point cloud data from different sources, while geometric transformation parameter calculation is used to determine the rotation and translation matrices to achieve spatial alignment of the point cloud data. Existing automatic registration methods have poor feature extraction performance for two sets of data with large density differences, thus affecting the reliability of the registration results. Furthermore, existing registration methods are generally rigid registration methods that do not fully consider the scale differences between cross-source point clouds, which can easily lead to low registration accuracy or even registration failure.

[0004] Therefore, it is necessary to propose a ground-to-air point cloud registration method that comprehensively considers density differences and scale differences. Summary of the Invention

[0005] In view of this, the present invention provides a ground-to-air point cloud registration method in an urban scene, so as to at least solve some of the technical problems in the background technology.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A ground-to-air point cloud registration method in an urban scene includes the following steps:

[0008] Based on point density projection and Hough transform, the two-dimensional building corner points of the ground point cloud in the target area are obtained, and a topological relationship atlas of the ground corner points is constructed;

[0009] Based on morphological operations and Hough transform, the two-dimensional building corner points of the airborne point cloud in the target area are obtained, and a topological relationship atlas of the airborne point cloud corner points is constructed;

[0010] Calculate the similarity between each ground corner point topology relationship graph in the ground corner point topology relationship graph set and each airborne point cloud corner point topology relationship graph in the airborne point cloud corner point topology relationship graph set, and determine the initial corresponding points;

[0011] The random sample consistency algorithm is used to screen and optimize the initial corresponding points to obtain the ground corner point corresponding point set and the airborne corner point corresponding point set;

[0012] Perform horizontal point cloud registration based on the ground corner point corresponding point set and the airborne corner point corresponding point set to obtain the horizontal registration result of converting the ground point cloud data into the airborne point cloud data;

[0013] The vertical point cloud registration is completed based on the horizontal registration results of the ground point cloud data and the original airborne point cloud data.

[0014] Furthermore, the two-dimensional building corner points of the ground point cloud of the target area are obtained based on point density projection and Hough transform, which specifically includes the following steps:

[0015] Use the cloth simulation algorithm to filter out the ground plane points in the ground point cloud data of the target area;

[0016] The point cloud after filtering the ground plane points is projected onto the horizontal grid, and the two-dimensional building contour points are extracted using the point density projection method;

[0017] Hough transform is used to perform straight line fitting on the two-dimensional contour points, and the intersection of the two-dimensional contour segments is used as the two-dimensional building corner points of the ground point cloud in the target area.

[0018] Furthermore, constructing a ground corner point topological relationship atlas specifically includes the following steps:

[0019] For each 2D building corner point of the ground point cloud, take it as the origin O T Construct a local coordinate system and determine the distance O T The farthest corner point A, all corner points around O T Rotate until O T A points to the positive direction of the X axis of the local coordinate system, and then the other corner points are aligned with O T Connect to form the ground corner point topology relationship graph m of the corner point;

[0020] The topological relationship graphs of all ground corner points form a set M = {m1,m2,m3,...,m i}, where i is the number of ground corner points.

[0021] Furthermore, the 2D building corner points of the airborne point cloud in the target area are obtained based on morphological operations and Hough transform, specifically including:

[0022] Filter the point clouds with a specified threshold distance from the ground in the airborne point cloud data of the target area;

[0023] The building roof extraction method based on seed point set is used to extract the filtered airborne point cloud data of the target area to obtain the building top surface point cloud of the target area;

[0024] Project the building top surface point cloud onto the horizontal grid to generate a binary image, and perform image contour expansion on the binary image to fill the gaps in the point cloud;

[0025] Hough transform is used to perform linear fitting on the binary image after filling the gaps in the point cloud, and the intersection of the image contour segments is used as the initial airborne building corner point;

[0026] The initial airborne building corner points are corrected along the opposite direction of the image contour expansion to obtain the final airborne point cloud two-dimensional building corner points.

[0027] Furthermore, a topological relationship atlas of the airborne point cloud corners is constructed, specifically including:

[0028] For each 2D airborne building corner point, take it as the origin O A Construct a local coordinate system, determine the other corner points as point A in turn, and let all corner points revolve around O A Rotate until O A A points to the positive direction of the X axis, and then the other corner points are aligned with O A Connect, forming the airborne corner point topological relationship graph n of the corner point, and the topological relationship graphs of all airborne corner points form a set N = {{n 11 ,n 12 ,...,n 1j-1},{n 21 ,n 22 ,...,n 2j-1},...,{n j1 ,n j2 ,...,n jj-1}}, where j is the number of airborne corner points.

[0029] Furthermore, the similarity between each ground corner point topology relationship graph in the ground corner point topology relationship graph set and each airborne point cloud corner point topology relationship graph in the airborne point cloud corner point topology relationship graph set is calculated, and the initial corresponding points are determined, specifically including:

[0030] Compare the angle similarity and side length similarity of each ground corner point topology relationship graph m in the ground corner point topology relationship atlas M and the airborne point cloud corner point topology relationship graph n in the airborne point cloud corner topology relationship atlas N;

[0031] Set a ratio threshold num. When the angle consistency ratio of the ground corner point topology relationship map m and the airborne point cloud corner point topology relationship map n exceeds num, the side length ratios that meet the angle conditions are counted. If the standard deviation of the side length ratios is less than 0.04, these point pairs that meet the conditions are used as initial corresponding points.

[0032] Furthermore, horizontal point cloud registration is performed based on the ground corner point corresponding point set and the airborne corner point corresponding point set, specifically including:

[0033] Calculate the center points of the ground corner point corresponding point set P and the airborne corner point corresponding point set Q, and use them as the geometric center C of the ground corner point corresponding point set P The geometric center C of the point set corresponding to the airborne corner point Q ;

[0034] The ground corner point corresponding point set P and the airborne corner point corresponding point set Q are processed to remove the geometric center respectively;

[0035] Calculate the covariance matrix of the ground corner point corresponding point set and the airborne corner point corresponding point set after removing the geometric center to quantify the correlation between the two sets of point sets;

[0036] Perform singular value decomposition on the covariance matrix to obtain matrix U, singular value ∑ and matrix V T ;

[0037] According to the singular value decomposition results, calculate the rotation matrix R:

[0038] R=VU T ,

[0039] If det(R) < 0, then invert the last row of the matrix V and recalculate the rotation R;

[0040] Solve the optimal scale factor u so that the scaled point set is as close as possible to the target point set:

[0041]

[0042] Where S is the number of corresponding points, P i ' represents the i-th point in the ground corner corresponding point set after removing the geometric center, Q i ' represents the i-th point in the corresponding point set of the airborne corner point after removing the geometric center;

[0043] According to the center point relationship and translation invariance, the translation vector T is calculated:

[0044] T=C Q -uRC P

[0045] Apply the scale u, rotation matrix R and translation matrix T to the ground point cloud data to be registered to complete the horizontal registration.

[0046] Furthermore, vertical point cloud registration is completed based on the horizontal registration results of the ground point cloud data and the original airborne point cloud data, specifically including:

[0047] Obtain the horizontal registration result of the ground point cloud data converted to the airborne point cloud data, and use the cloth simulation algorithm to obtain the ground plane point cloud in the horizontal registration result, which is recorded as the first ground plane point cloud;

[0048] The cloth simulation algorithm is used to obtain the ground plane point cloud in the original airborne point cloud data of the target area, which is recorded as the second ground plane point cloud;

[0049] The first ground plane point cloud and the second ground plane point cloud are set to the same space and filtered to remove the non-common area point clouds between the two;

[0050] The two filtered ground plane point clouds are mapped to the horizontal grid respectively, retaining only the Z coordinate value; the average elevation value of the point cloud falling into the grid cell is used to update the grid initial value NAN;

[0051] The elevation difference between the corresponding grids is calculated and the average value is obtained to obtain the vertical displacement Δz. Each time the average value is calculated, the random sample consistency algorithm is first used to eliminate outliers;

[0052] The vertical displacement Δz is applied to the z coordinate of the ground point cloud to be vertically registered: Z = uz + Δz to complete the registration, where Z is the z coordinate of the target airborne point cloud and u is the optimal scale factor.

[0053] It can be seen from the above technical solutions that, compared with the prior art, the present invention provides a ground-to-air point cloud registration method in urban scenarios, which has the following beneficial effects:

[0054] To address density discrepancies, this paper effectively extracts high-precision 2D building corners through point density projection, Hough transform, mathematical morphological dilation, and corner correction strategies. To address significant scale discrepancies, a topological relationship graph is constructed based on the 2D corners, and a scale-adaptive matching strategy is employed to obtain robust corresponding point pairs. Subsequently, through singular value decomposition and the introduction of a scale factor, the rotation matrix, translation matrix, and scale parameter are jointly solved to achieve horizontal registration. Finally, vertical translation correction is performed using the elevation characteristics of the terrain surface, thereby achieving overall 3D registration. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0056] Figure 1 This is an overall schematic diagram of the registration method provided in Example 1 of the present invention;

[0057] Figure 2 The ground LiDAR point cloud data provided in Example 1 of the present invention;

[0058] Figure 3 The airborne LiDAR point cloud data provided in Example 1 of the present invention;

[0059] Figure 4 This is the ground LiDAR point cloud corner point extraction result diagram provided in Example 1 of the present invention;

[0060] Figure 5 This is the result of corner point extraction from the airborne LiDAR point cloud provided in Example 1 of the present invention;

[0061] Figure 6 Schematic diagram of the airborne corner correction method provided in Example 1 of the present invention;

[0062] Figure 7 This is the horizontal registration result diagram provided in Example 1 of the present invention;

[0063] Figure 8 This is the overall registration result diagram provided in Example 1 of the present invention;

[0064] Figure 9 Schematic diagram of the method for constructing a topological relationship graph of corner points provided in Example 1 of the present invention, wherein (a) is a schematic diagram of the method for constructing a topological relationship graph of corner points of a ground point cloud, and (b) is a schematic diagram of the method for constructing a topological relationship graph of corner points of an airborne point cloud.

[0065] Figure 10 Schematic diagram of the vertical registration process provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0066] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0067] The embodiment of the present invention discloses a ground-to-air point cloud registration method in an urban scene, comprising the following steps:

[0068] Based on point density projection and Hough transform, the two-dimensional building corner points of the ground point cloud in the target area are obtained, and a topological relationship atlas of the ground corner points is constructed;

[0069] Based on morphological operations and Hough transform, the two-dimensional building corner points of the airborne point cloud in the target area are obtained, and a topological relationship atlas of the airborne point cloud corner points is constructed;

[0070] Calculate the similarity between each ground corner point topology relationship graph in the ground corner point topology relationship graph set and each airborne point cloud corner point topology relationship graph in the airborne point cloud corner point topology relationship graph set, and determine the initial corresponding points;

[0071] The random sample consistency algorithm is used to screen and optimize the initial corresponding points to obtain the ground corner point corresponding point set and the airborne corner point corresponding point set;

[0072] Perform horizontal point cloud registration based on the ground corner point corresponding point set and the airborne corner point corresponding point set to obtain the horizontal registration result of converting the ground point cloud data into the airborne point cloud data;

[0073] The vertical point cloud registration is completed based on the horizontal registration results of the ground point cloud data and the original airborne point cloud data.

[0074] The specific process of the present invention is further described below through specific embodiments.

[0075] Example 1

[0076] The following will be combined with the Figure 1 , further introduces the technical solution in Example 1 of the present invention:

[0077] The specific implementation method disclosed in Example 1 of the present invention is based on the buildings on a certain city street. The ground LiDAR point cloud is obtained by a Huace AU20 LiDAR scanner with an average point density of 330 points / square meter and a total of 19,152,760 points (the ground point cloud is as follows Figure 2 The airborne LiDAR point cloud was acquired by scanning with a UAV equipped with a Huace AU20 LiDAR, with an average point density of 90 points / m2 and a total of 4,931,411 points (the airborne point cloud is shown in Figure 3 shown).

[0078] Example 1 discloses a ground-to-air point cloud registration method in an urban scene, comprising the following steps:

[0079] S1. Preprocess the two point cloud data, specifically including the following steps:

[0080] Remove outliers and noise points from ground LiDAR data and clean up redundant points inside buildings;

[0081] Remove redundant point clouds and noise points from airborne LiDAR data;

[0082] The ground-based LiDAR data and airborne LiDAR data are downsampled to a size that is convenient for processing.

[0083] S2. Extracting 2D building corner points from ground LiDAR data, specifically including the following steps:

[0084] Use the Cloth Simulation Filter (CSF) algorithm to filter out ground plane points from terrestrial LiDAR data to reduce noise and improve outline clarity.

[0085] The point cloud after filtering ground points is projected onto a horizontal grid, and the density of point projection (DoPP) method is used to extract the two-dimensional building contour points;

[0086] Hough transform is used to fit the two-dimensional contour points into straight lines, and the intersection of the two-dimensional contour segments is used as the ground building corner point (the ground building corner point is as follows Figure 4 As shown in Figure 3, 9 two-dimensional ground building corner points are extracted.

[0087] S3. Extracting 2D building corner points from airborne LiDAR data, specifically including the following steps:

[0088] Filter out the point clouds below 3 meters from the ground in the airborne LiDAR data to remove interference from vehicles and low vegetation;

[0089] The building roof point cloud is extracted using the building roof extraction method based on seed point set;

[0090] The building top surface point cloud is projected onto a horizontal grid with a unit grid of 0.2 meters to generate a binary image, and a 7×7 dilated convolution kernel is used to dilate the binary image to fill the gaps in the point cloud;

[0091] Hough transform is used to perform linear fitting on the binary image after filling the gaps in the point cloud, and the intersection of the image contour segments is used as the initial airborne building corner point (the airborne building corner point is as follows Figure 5 As shown in Figure 3, 32 two-dimensional airborne building corner points are extracted.

[0092] The initial airborne building corner points are corrected along the opposite direction of the image contour expansion, and the following is obtained: Figure 6The final airborne building corner point shown is: S = (k / 2) grid_size, where S is the size of the reverse-corrected contour, k is the size of the dilated convolution kernel used in the dilation operation, and grid_size is the size of the horizontal grid, that is, the corrected length of the line segment here is 0.7 meters.

[0093] S4. Construct a topological relationship graph based on the corner points and combine it with a scale-adaptive matching strategy to achieve matching of corresponding points. Specifically, the following steps are included:

[0094] refer to Figure 9 As shown in (a), for each two-dimensional ground building corner point, take it as the origin O T Construct a local coordinate system and determine the distance O T The farthest corner point A, all corner points around O T Rotate until O T A points to the positive direction of the X axis, and then the other corner points are aligned with O T Connect them to form the ground corner point topological relationship graph m of the corner point. The topological relationship graphs of all ground corner points form a set M = {m1,m2,m3,...,m i}, where i is the number of ground corner points. Each topological graph stores the angles (α1,α1,...,α n ) and side lengths (L1, L2, ..., L n ) and other information.

[0095] refer to Figure 9 As shown in (b), for each 2D airborne building corner point, take it as the origin O A Construct a local coordinate system, determine the other corner points as point A in turn, and let all corner points revolve around O A Rotate until O A A points to the positive direction of the X axis, and then the other corner points are aligned with O A Connect, forming the airborne corner point topological relationship graph n of the corner point, and the topological relationship graphs of all airborne corner points form a set N = {{n 11 ,n 12 ,...,n 1j-1},{n 21 ,n 22 ,...,n 2j-1},...,{n j1 ,n 92 ,...,n jj-1}}, where j is the number of airborne corner points. Since the extracted airborne building corner points are usually more than the ground building corner points, when constructing the airborne corner point cloud topology relationship diagram, the ground and airborne data differ only in the selection of point A. For ground LiDAR two-dimensional building corner point data, point A is the distance from the ground building corner point O TThe farthest point means that each ground building corner point has only one topological relationship graph m, that is, 9 ground corner point topological relationship graphs are constructed. For the airborne LiDAR 2D building corner point data, point A is the point except the airborne building corner point O. A All points except , which means that each ground building corner point has j-1 topological relationship graphs, that is, 992 airborne corner point topological relationship graphs are constructed, where j is the number of airborne corner points.

[0096] Since the relative positions of the corner points remain unchanged, ideally, the topological relationship diagrams of the corner points at the same location in the target area in the airborne and ground coordinate systems are exactly the same. That is, by comparing the similarity of the topological relationship diagrams of the ground corner points and the airborne corner points (consistency of angles and side lengths), the corresponding matching point pairs can be found. Taking into account the scale differences of cross-source data, the consistency of side lengths is determined by the corresponding side length ratio that meets the angle condition. The specific steps are as follows:

[0097] Compare the similarity (including angle and side length) of each topological relationship graph m in the set M with each topological relationship graph n in the set N to determine the corresponding point pairs; since there is a certain deviation in the extraction of building corner points, when the angle between the two points and the origin is less than 1°, it is considered to meet the angle consistency condition; set a ratio threshold num, such as 50%, and when the angle consistency ratio of m and n exceeds num, count the side length ratios that meet the angle condition. If the standard deviation of the side length ratio is less than 0.04, these point pairs that meet the condition are used as initial corresponding points.

[0098] Then the random sample consensus (RANSAC) algorithm is used to screen and optimize the initial corresponding points to obtain the final corresponding points P and Q. P is the point set of the ground corner point corresponding points, and Q is the point set of the airborne corner point corresponding points. The corresponding two-dimensional coordinates are shown below.

[0099] Table 1 Coordinates of corresponding points of two-dimensional building corners on the ground

[0100]

[0101] Table 2 Coordinates of corresponding points of airborne two-dimensional building corners

[0102]

[0103] S5. Combining singular value decomposition and introducing a scale factor to restore the two-dimensional transformation parameters to complete horizontal registration, specifically including the following steps:

[0104] Calculate the center points of two sets of corresponding points P and Q to determine the geometric center of each set of points. The calculation formula for the center point is: Where S is the number of corresponding points, P i and Q iRespectively represent the i-th point in the two sets of point sets; in the present invention, the number of the airborne corner point corresponding point set Q = (q1, q2, ... qs) and the ground corner point corresponding point set P = (p1, p2, ..., ps) are consistent.

[0105] Decentralize the two sets of corresponding points P and Q: P i '=P i -C P ,Q i '=Q i -C Q , and calculate P i and Q i The covariance matrix H is used to quantify the correlation between two sets of points: H = P' T Q', where P' T Represents the transpose of the point set P'.

[0106] Then perform singular value decomposition on the covariance matrix H to obtain matrix U, singular value ∑ and matrix V T :H=U∑V T ;

[0107] According to the result of singular value decomposition, calculate the rotation matrix R: R = VU T To ensure the determinant of the rotation matrix is ​​positive, if det(R) < 0, negate the last row of V and recalculate R. In this step, det() represents the determinant-finding function. To ensure that R is a pure rotation matrix, det(R) = 1. In practical applications, the rotation matrix R obtained by singular value decomposition may not satisfy the properties of an orthogonal matrix. In particular, its determinant det(R) may be -1. A matrix with a determinant of -1 indicates a reflection transformation, not a pure rotation. This is done to avoid reflection transformations.

[0108] Solve the optimal scale factor u so that the scale-adjusted point set is as close as possible to the target point set. The expression of the scale factor is expressed as:

[0109] Then, based on the center point relationship and translation invariance, the translation vector T is calculated: T = C Q -uRC P ;

[0110] The coordinate system of the ground laser point cloud TLS system is o-xyz, and the coordinate system of the airborne laser point cloud ALS system is O-XYZ. The scale u, rotation matrix R and translation matrix T are applied to the ground point cloud to be registered to complete the horizontal registration, that is, the ground point cloud is converted to the coordinate system of the airborne point cloud (the horizontal registration result is shown in the figure). Figure 7 shown): in is the X,Y coordinate of the airborne point cloud, are the x, y coordinates of the ground point cloud to be registered; after horizontal registration, the ground laser point cloud TLS and the airborne laser point cloud ALS data are in the same O-XYZ coordinate system, and there is only a vertical translation Δz between the two.

[0111] S6. Finally, vertical point cloud registration is performed based on the horizontal registration result converted from the ground point cloud to the airborne point cloud coordinate system and the original airborne point cloud data of the target area. Specifically, the following steps are included:

[0112] Obtain horizontally registered ground point cloud data and airborne point cloud data, use the cloth simulation algorithm (CSF) to extract the ground plane point clouds from each, convert the two ground plane point clouds into the same space, and filter out the point clouds in non-common areas;

[0113] The two ground plane point clouds of the public area machine-ground data are mapped to a horizontal grid with a unit grid of 2 meters, retaining only the Z coordinate value; the average elevation value of the point cloud falling within the grid unit is used to update the grid initial value NAN;

[0114] The elevation difference between the corresponding grids is calculated and the average value is obtained to obtain the vertical displacement Δz. Each time the average value is calculated, RANSAC is used to remove outliers.

[0115] Apply the vertical displacement Δz to the z coordinate of the ground LiDAR point cloud to be vertically registered: Z = uz + Δz to complete the registration (the overall registration result is as follows Figure 8 As shown), the vertical registration process can refer to Figure 10 shown.

[0116] according to Figure 2 , Figure 3 As can be seen, for the building in this example, the ground LiDAR point cloud model has a large amount of missing tops, while the airborne LiDAR point cloud model is missing more than the facade point cloud, and the size difference between the two is quite large. The resulting scaling ratio is 2.4939. In this example, extracting ground corner points takes 13.1374 seconds, while extracting airborne corner points from the roof point cloud takes only 0.1605 seconds. The time to query corresponding points is 7.6578 seconds, with a very low root mean square error, meeting the registration requirements. This ground-to-air point cloud registration method for urban scenes can successfully align point clouds from two different sources.

[0117] The above analysis found that:

[0118] 1) The registration method of the present invention can register point clouds of any angle, density and size.

[0119] 2) The registration method of the present invention is highly efficient and its accuracy meets the registration requirements.

[0120] In summary, the present invention can meet the requirements of point cloud registration with large angles, large density differences, and different scales while ensuring computational efficiency.

[0121] The present invention uses the airborne laser point cloud as a reference system to align the ground three-dimensional laser point cloud to the airborne point cloud model, thereby constructing more complete and high-precision three-dimensional scene point cloud data. To address the problem of density differences, the present invention effectively extracts high-precision two-dimensional building corner points through point density projection, Hough transform, mathematical morphological dilation operation, and corner correction strategy. To address the problem of significant scale differences, a topological relationship diagram is further constructed based on the two-dimensional corner points, and a scale-adaptive matching strategy is adopted to obtain robust corresponding point pairs. Subsequently, through the singular value decomposition method and the introduction of the scale factor, the rotation matrix, translation matrix, and scale parameter are jointly solved to achieve the purpose of horizontal registration. Finally, the height characteristics of the terrain surface are used to perform vertical translation correction, thereby achieving overall registration in three-dimensional space.

[0122] Experiments show that the ground-to-air point cloud registration method in this paper is used to register this example with high registration efficiency and low root mean square error, which meets the registration requirements. A ground-to-air point cloud registration method in an urban scene can complete the registration of point clouds from two different sources.

[0123] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0124] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A ground-to-air point cloud registration method in an urban scene, characterized by: The following steps are involved: Based on point density projection and Hough transform, the two-dimensional building corner points of the ground point cloud in the target area are obtained, and a topological relationship atlas of the ground corner points is constructed; Based on morphological operations and Hough transform, the two-dimensional building corner points of the airborne point cloud in the target area are obtained, and a topological relationship atlas of the airborne point cloud corner points is constructed; Calculate the similarity between each ground corner point topology relationship graph in the ground corner point topology relationship graph set and each airborne point cloud corner point topology relationship graph in the airborne point cloud corner point topology relationship graph set, and determine the initial corresponding points; The random sample consistency algorithm is used to screen and optimize the initial corresponding points to obtain the ground corner point corresponding point set and the airborne corner point corresponding point set; Perform horizontal point cloud registration based on the ground corner point corresponding point set and the airborne corner point corresponding point set to obtain the horizontal registration result of converting the ground point cloud data into the airborne point cloud data; The vertical point cloud registration is completed based on the horizontal registration results of the ground point cloud data and the original airborne point cloud data.

2. The ground-to-air point cloud registration method in an urban scene according to claim 1, characterized in that: The two-dimensional building corner points of the ground point cloud in the target area are obtained based on point density projection and Hough transform, which specifically includes the following steps: Use the cloth simulation algorithm to filter out the ground plane points in the ground point cloud data of the target area; The point cloud after filtering the ground plane points is projected onto the horizontal grid, and the two-dimensional building contour points are extracted using the point density projection method; Hough transform is used to perform straight line fitting on the two-dimensional contour points, and the intersection of the two-dimensional contour segments is used as the two-dimensional building corner points of the ground point cloud in the target area.

3. The ground-to-air point cloud registration method in an urban scene according to claim 1, characterized in that: Constructing a topological atlas of ground corner points includes the following steps: For each 2D building corner point of the ground point cloud, take it as the origin O T Construct a local coordinate system and determine the distance O T The farthest corner point A, all corner points around O T Rotate until O T A points to the positive direction of the X axis of the local coordinate system, and then the other corner points are aligned with O T Connect to form the ground corner point topology relationship graph m of the corner point; The topological relationship graphs of all ground corner points form a set M = {m1,m2,m3,...,m i }, where i is the number of ground corner points.

4. The ground-to-air point cloud registration method in an urban scene according to claim 1, characterized in that: The 2D building corner points of the airborne point cloud in the target area are obtained based on morphological operations and Hough transform, including: Filter the point clouds with a specified threshold distance from the ground in the airborne point cloud data of the target area; The building roof extraction method based on seed point set is used to extract the filtered airborne point cloud data of the target area to obtain the building top surface point cloud of the target area; Project the building top surface point cloud onto the horizontal grid to generate a binary image, and perform image contour expansion on the binary image to fill the gaps in the point cloud; Hough transform is used to perform linear fitting on the binary image after filling the gaps in the point cloud, and the intersection of the image contour segments is used as the initial airborne building corner point; The initial airborne building corner points are corrected along the opposite direction of the image contour expansion to obtain the final airborne point cloud two-dimensional building corner points.

5. The ground-to-air point cloud registration method in an urban scene according to claim 1, characterized in that: Construct an airborne point cloud corner point topology atlas, including: For each 2D airborne building corner point, take it as the origin O A Construct a local coordinate system, determine the other corner points as point A in turn, and let all corner points revolve around O A Rotate until O A A points to the positive direction of the X axis, and then the other corner points are aligned with O A Connect, forming the airborne corner point topological relationship graph n of the corner point, and the topological relationship graphs of all airborne corner points form a set N = {{n 11 ,n 12 ,...,n 1j-1 },{n 21 ,n 22 ,...,n 2j-1 },...,{n j1 ,n j2 ,...,n jj-1 }}, where j is the number of airborne corner points.

6. The ground-to-air point cloud registration method in an urban scene according to claim 1, characterized in that: Calculate the similarity between each ground corner point topology relationship graph in the ground corner point topology relationship graph set and each airborne point cloud corner point topology relationship graph in the airborne point cloud corner point topology relationship graph set, and determine the initial corresponding points, specifically including: Compare the angle similarity and side length similarity of each ground corner point topology relationship graph m in the ground corner point topology relationship atlas M and the airborne point cloud corner point topology relationship graph n in the airborne point cloud corner topology relationship atlas N; Set a ratio threshold num. When the angle consistency ratio of the ground corner point topology relationship map m and the airborne point cloud corner point topology relationship map n exceeds num, the side length ratios that meet the angle conditions are counted. If the standard deviation of the side length ratios is less than 0.04, these point pairs that meet the conditions are used as initial corresponding points.

7. The ground-to-air point cloud registration method in an urban scene according to claim 1, characterized in that: Horizontal point cloud registration is performed based on the ground corner point correspondence point set and the airborne corner point correspondence point set, specifically including: Calculate the center points of the ground corner point corresponding point set P and the airborne corner point corresponding point set Q, and use them as the geometric center C of the ground corner point corresponding point set P The geometric center C of the point set corresponding to the airborne corner point Q ; The ground corner point corresponding point set P and the airborne corner point corresponding point set Q are processed to remove the geometric center respectively; Calculate the covariance matrix of the ground corner point corresponding point set and the airborne corner point corresponding point set after removing the geometric center to quantify the correlation between the two sets of point sets; Perform singular value decomposition on the covariance matrix to obtain matrix U, singular value ∑ and matrix V T ; According to the singular value decomposition results, calculate the rotation matrix R: R=VU T , If det(R) < 0, then invert the last row of the matrix V and recalculate the rotation R; Solve the optimal scale factor u so that the scaled point set is as close as possible to the target point set: Among them, S is the number of corresponding points, P i ' represents the i-th point in the ground corner corresponding point set after removing the geometric center, Q i ' represents the i-th point in the corresponding point set of the airborne corner point after removing the geometric center; According to the center point relationship and translation invariance, the translation vector T is calculated: T=C Q -uRC P Apply the scale u, rotation matrix R and translation matrix T to the ground point cloud data to be registered to complete the horizontal registration.

8. The ground-to-air point cloud registration method in an urban scene according to claim 7, characterized in that: Based on the horizontal registration results of the ground point cloud data and the original airborne point cloud data, the vertical point cloud registration is completed, specifically including: Obtain the horizontal registration result of the ground point cloud data converted to the airborne point cloud data, and use the cloth simulation algorithm to obtain the ground plane point cloud in the horizontal registration result, which is recorded as the first ground plane point cloud; The cloth simulation algorithm is used to obtain the ground plane point cloud in the original airborne point cloud data of the target area, which is recorded as the second ground plane point cloud; The first ground plane point cloud and the second ground plane point cloud are set to the same space and filtered to remove the non-common area point clouds between the two; The two filtered ground plane point clouds are mapped to the horizontal grid respectively, retaining only the Z coordinate value; the average elevation value of the point cloud falling into the grid cell is used to update the grid initial value NAN; The elevation difference between the corresponding grids is calculated and the average value is obtained to obtain the vertical displacement Δz. Each time the average value is calculated, the random sample consistency algorithm is first used to eliminate outliers; The vertical displacement Δz is applied to the z coordinate of the ground point cloud to be vertically registered: Z = uz + Δz to complete the registration, where Z is the z coordinate of the target airborne point cloud and u is the optimal scale factor.

Images