Multi-modal visual data fusion curved antenna key feature reconstruction method

Abstract

The invention provides a curved antenna key feature reconstruction method based on multi-mode visual data fusion, and relates to the field of feature extraction based on vision. Comprising the following steps: designing an artificially enhanced feature pattern and projecting the artificially enhanced feature pattern to the surface of a to-be-measured curved antenna; a structured light camera is adopted to collect surface images and three-dimensional point clouds of the to-be-detected curved surface antenna containing the artificially enhanced feature patterns at different viewpoints; establishing feature vectors of all assembly hole circle center points in the viewpoint based on the surface image; obtaining a matching relationship of the assembly holes in the three-dimensional point cloud, mapping a matching result of the assembly holes to the three-dimensional point cloud, and splicing and fusing the assembly holes under a plurality of different viewpoints to obtain complete point cloud data of the to-be-detected curved antenna; semantic segmentation and feature fitting are carried out on the complete point cloud data, and an assembly hole circle center point set and a bonding pad circle center point set are output. According to the invention, through non-contact measurement of multi-modal visual data fusion, high-precision measurement of the size of the curved antenna and feature reconstruction are realized on the premise that the resolution is not reduced.

Description

Technical Field

[0001] The present invention relates to the field of feature extraction based on vision, and in particular to a method for reconstructing key features of curved antennas by fusion of multi-modal visual data. Background Art

[0002] Existing mechanical parts measurement methods are mainly implemented by contact and non-contact methods. The contact measurement method uses a high-precision probe to contact the workpiece surface to obtain the workpiece coordinate data, and realize the measurement of three-dimensional size, shape and position. This method usually has a slow measurement speed, and the measurement of smaller features is unchanged. In addition, the contact of the probe may cause surface damage to some precision workpieces, which is not applicable to the measurement of curved conformal antenna parts studied in the present invention. The non-contact error measurement method is mainly implemented through photoelectric imaging technology, such as structured light scanning, visible light or infrared ultraviolet camera photography, etc. This method does not require contact with the object being measured, and is suitable for measuring delicate or complex surface objects. At the same time, the data acquisition speed is fast and the use flexibility is high. It is suitable for the measurement object of the present invention, but the photoelectric imaging technology has a fundamental contradiction between the field of view and the resolution size, which makes it difficult to measure large-scale parts with high precision.

[0003] There is an existing technology: the turbine blade is clamped on a cylindrical fixture, and the secondary search method of the marker point is used to improve the stability of the point cloud coarse registration, thereby achieving high-quality multi-point cloud stitching. However, this method is highly dependent on the marker points, and the number of target points in the overlapping area between two adjacent images must reach about 20, which has certain limitations for 3D reconstruction applications with small overlapping areas.

[0004] The Chinese invention patent with publication number CN118298110A discloses a 3D point cloud image reconstruction method and system for the parts manufacturing process. The method uses a quantitatively measurable rotating fixture to clamp the object to be tested, and integrates the rotation matrix of the rotating fixture into the fast ICP process to achieve the reconstruction of the object to be tested. This method needs to rely on the geometric information provided by the external motion mechanism and does not make full use of the image's own information.

[0005] The above method realizes the stitching of multi-viewpoint three-dimensional point clouds, but it puts forward higher requirements on the inherent characteristics of the object to be measured. When the inherent characteristics of the object to be measured are insufficient, additional features such as marker points are added or an external motion mechanism is used to obtain the conversion relationship between images from different viewpoints, which increases the complexity of the measurement process. Summary of the invention

[0006] Purpose of the invention: To propose a method for reconstructing key features of curved antennas based on multimodal visual data fusion to solve the above-mentioned problems existing in the prior art.

[0007] The present invention proposes a method for reconstructing key features of curved antennas by fusion of multimodal visual data, comprising the following steps:

[0008] Design an artificially enhanced feature pattern and project it onto the surface of the curved antenna to be tested;

[0009] Using a structured light camera to capture the surface image of the curved antenna to be tested containing the artificially enhanced characteristic pattern at different viewpoints as a two-dimensional image, and simultaneously collecting a three-dimensional point cloud of the curved antenna to be tested;

[0010] Based on the two-dimensional image, establishing feature vectors of the center points of all assembly holes in the viewpoint;

[0011] Obtain the matching relationship of the assembly holes in the 3D point cloud, map the matching results of the assembly holes to the 3D point cloud, and merge them under multiple different viewpoints to obtain the complete point cloud data of the curved antenna to be tested;

[0012] Semantic segmentation and feature fitting are performed on the complete point cloud data, and an assembly hole center point set and a pad center point set are output.

[0013] In a further embodiment, the projection area A of the artificially enhanced characteristic pattern on the surface of the curved antenna to be tested is a rectangular area, and the length of the projection area A is , width ;

[0014] The aspect ratio of the projection area A is the same as the aspect ratio of the minimum envelope rectangle of the curved antenna to be measured on the imaging plane of the two-dimensional camera;

[0015] The length x and width y of the artificially enhanced characteristic pattern satisfy , , k is the safety factor.

[0016] In a further embodiment, projecting the artificially enhanced characteristic pattern onto the surface of the curved antenna to be tested specifically includes:

[0017] The projection area A is meshed as , divide rectangular cells, the number of cells satisfies , , d is the diameter of the end face of the assembly hole of the curved antenna to be tested;

[0018] Generate a random point in each cell obtained ;

[0019] Random Points Draw a filled circular colored area with the center as the circle. The diameter of the circular colored area is satisfy , MP is the pixel size of the camera used to acquire the two-dimensional image.

[0020] In a further embodiment, the random points generated Coordinates satisfy:

[0021]

[0022]

[0023] In the formula, i and j are the horizontal and vertical indexes of the grid respectively; , is the minimum value of the horizontal and vertical coordinates of the projection area A in the coordinate system; , is the length and width of the cell, which satisfies , ; is a random coefficient.

[0024] In a further embodiment, when the structured light camera captures the surface image of the curved antenna to be tested containing the artificially enhanced characteristic pattern at different viewpoints:

[0025] The overlapping area of ​​two adjacent viewpoint images contains at least two assembly hole features.

[0026] In a further embodiment, based on the two-dimensional image, a feature vector of the center points of all assembly holes in the viewpoint is established, specifically including:

[0027] Fit the circle in the two-dimensional image under all viewpoints and obtain its center to obtain the center point set of the assembly hole and the center point set of artificial feature patterns ;

[0028] The points in the assembly hole center point set P As the target point, search for the point in the center point set Q of the artificial feature pattern The a points with the closest Euclidean distance are used to establish an a-dimensional distance vector in ascending order of Euclidean distance. The feature vectors of the mounting holes of the curved antenna.

[0029] In a further embodiment, obtaining the matching relationship of the assembly holes in the three-dimensional point cloud specifically includes:

[0030] Fit the center of the assembly hole in the 3D point cloud and project it on the XY plane to obtain the projection point set of the center of the assembly hole , projection point set of the center of the assembly hole The corresponding relationship between the feature centers is obtained by aligning with the center point set P of the assembly hole, so as to establish the matching relationship between the assembly holes in the overlapping area of ​​the 3D point cloud of adjacent viewpoints and obtain the center point pair set of the assembly hole ,in is the center point of a certain assembly hole in the 3D point cloud of a certain viewpoint, The 3D point cloud of its adjacent viewpoints is The center points of the mounting holes with the same name in physical space.

[0031] In a further embodiment, the matching results of the assembly holes are mapped to a three-dimensional point cloud and spliced ​​and fused under multiple different viewpoints, specifically including:

[0032] According to the matching relationship of the assembly holes in the 3D point cloud, select the point cloud data in the cube area with the center of the assembly hole as the geometric center and the edge length a in the 3D point cloud of a certain viewpoint and its adjacent viewpoints, and use it as the point cloud to be registered and ;

[0033] Point cloud to be registered is the source point cloud, the point cloud to be registered For the target point cloud, search for the nearest neighbor point in the target point cloud for each point in the source point cloud to obtain a paired point set ;

[0034] For each pair of points in the paired point set Define the weights of the composite Gaussian form :

[0035]

[0036] In the formula, and Points and The minimum value of the Euclidean distance to the center of all assembly holes in the respective point cloud; Parameters that control the speed of weight decay;

[0037] Compute the weighted centroid:

[0038]

[0039]

[0040] Where N is the pairing point set The number of elements, M is the collocation pair set The number of elements in ; Assign weights to the point pairs of the assembly hole centers;

[0041] Calculate the weighted covariance matrix H:

[0042]

[0043] Perform SVD decomposition on the covariance matrix H, calculate the rotation matrix R and translation vector t, and obtain the registration matrix T according to the kinematic principle;

[0044] Iteratively update the transformation matrix T until the convergence condition is met;

[0045] The point clouds of all viewpoints are registered one by one in turn to obtain several transformation matrices;

[0046] Take the point cloud data of a certain viewpoint as the overall target point cloud and perform rigid transformation on the remaining point clouds:

[0047]

[0048] Where X represents the number of pairwise registrations required to transform to the overall target point cloud coordinate system. The size of X depends on the number of pairwise registrations required to transform the point cloud C to the overall target point cloud coordinate system. C represents the point cloud before transformation. Represents the transformed point cloud; is the transformation matrix under the current viewpoint;

[0049] The overall target point cloud is fused with all the point clouds after rigid transformation to obtain the complete point cloud data of the curved antenna to be tested.

[0050] In a further embodiment, semantic segmentation and feature fitting are performed on the complete point cloud data, specifically including:

[0051] The voxel grid size is set according to the desired accuracy requirement, and the complete point cloud data of the curved antenna to be measured is subjected to through-filtering, voxelization, statistical filtering and smoothing to obtain a pre-processed point cloud;

[0052] Calculate the normal vector of the pre-processed point cloud;

[0053] The boundary points of the point cloud are segmented according to the normal vector difference, and the boundary points are segmented again to obtain the boundary points of the assembly holes;

[0054] Calculate the point cloud curvature based on the normal vector and identify the pad plane;

[0055] Verify and eliminate the incorrect recognition results, and finally obtain the assembly hole center point set and the pad center point set.

[0056] Compared with the prior art, the present invention has the following significant advantages: non-contact measurement through multimodal visual data fusion can overcome the contradiction between field of view and resolution, and increase the measurement range without reducing the resolution, thereby achieving high-precision measurement of large-size parts represented by curved antennas and reconstruction of tiny features such as assembly holes and pads, providing methods and technical support for the measurement of manufacturing errors and adaptive assembly processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a schematic diagram of the process of the present invention.

[0058] Figure 2 It is a schematic diagram of the point cloud semantic segmentation and assembly hole and pad feature fitting process of the present invention. DETAILED DESCRIPTION

[0059] In the following description, a large number of specific details are provided to provide a more thorough understanding of the present invention. However, it is apparent to those skilled in the art that the present invention can be implemented without one or more of these details. In other examples, in order to avoid confusion with the present invention, some technical features known in the art are not described.

[0060] The embodiment of the present invention provides a method for reconstructing key features of curved antennas by fusion of multi-modal visual data. The method uses a structured light camera based on a monocular telecentric lens to capture the surface image of the curved antenna part to be tested containing an artificially enhanced feature pattern to obtain a two-dimensional image, and based on the principle of multi-light source encoding and phase resolution, obtains a three-dimensional point cloud of the corresponding surface. The fusion of the two-dimensional image and the three-dimensional point cloud realizes the splicing of the multi-viewpoint three-dimensional point cloud, thereby realizing the reconstruction of the key features of the curved antenna. Figure 1 , its technical solution is divided into the following four steps:

[0061] Step S1: Artificially enhance the feature pattern design:

[0062] Step S11: In order to make the method proposed by the present invention have better environmental robustness, the present invention uses a laser projector with adjustable brightness and focal length to project the artificially enhanced characteristic pattern onto the surface of the curved antenna to be tested.

[0063] Step S12: The artificially enhanced feature pattern projection area A is a rectangular area with an aspect ratio of The length x and width y of the artificially enhanced feature pattern should be the same as the minimum envelope rectangle length-to-width ratio of the curved antenna to be tested on the two-dimensional camera imaging plane. , , where k is the safety factor, which is determined based on engineering experience and is usually greater than 1. x and y are the length and width of the minimum envelope rectangle of the curved antenna to be tested in the two-dimensional camera imaging plane.

[0064] Step S13: grid the projection area A into , divide rectangular cells, the number of cells should satisfy , , where d is the diameter of the end face of the assembly hole of the curved antenna to be tested.

[0065] Step S14: the A point is randomly generated in each cell of , whose coordinates satisfy:

[0066]

[0067]

[0068] in and is the grid index;

[0069] , is the minimum value of the projection area A in a certain coordinate system;

[0070] , is the length and width of the rectangular cell, which satisfies , ;

[0071] is a random coefficient.

[0072] Step S15: using the random points generated in step S14 Draw a filled circular colored area with the center of the circle. Should meet , where MP is the pixel size of the camera used to acquire the two-dimensional image.

[0073] Step S16: To ensure that there is no overlap in the circular areas of the artificially enhanced feature pattern, the random coefficient Should meet:

[0074]

[0075]

[0076] Step S2: Unique coding of the curved antenna assembly hole based on the artificially enhanced characteristic pattern:

[0077] In order to achieve the unique coding of the part's own features, the circular light spot and the center of the hole feature of the curved antenna itself are fitted in the two-dimensional image, and the feature vector is constructed according to the relative position relationship between the two in space. Step S2 specifically includes:

[0078] Step S21, using a monocular structured light camera based on a telecentric lens to obtain two-dimensional images and three-dimensional point clouds from different viewpoints, ensuring that the spatial position of the telecentric lens does not change when obtaining the two-dimensional image and three-dimensional point cloud from the same viewpoint. Reasonable viewpoint planning is used to ensure that the overlapping area of ​​two adjacent viewpoint images contains at least two assembly hole features.

[0079] Step S22: Use the Hough transform method to fit the circles in all viewpoint two-dimensional images and obtain the center of the circle to obtain the assembly hole center point set. and the center point set of artificial feature patterns .

[0080] Step S23: Points in P As the target point, search for the point in the set Q The points with the closest Euclidean distance are a, and an a-dimensional distance vector is established in ascending order of Euclidean distance. The feature vectors of the mounting holes of the curved antenna.

[0081] Step S24, repeat step S23 to establish the feature vectors of the center points of the assembly holes in all viewpoints, design a fuzzy threshold, regard the feature vector elements whose difference is not greater than the fuzzy threshold as the same element, perform fuzzy matching on the hole feature vectors in adjacent viewpoints, and achieve matching of the assembly holes in the overlapping area of ​​adjacent viewpoints.

[0082] Step S25: Use the RANSAC algorithm to fit the center of the assembly hole feature in the three-dimensional point cloud, and project it on the XY plane to obtain the assembly hole center projection point set. , using ICP algorithm to point set The corresponding relationship between the feature circle centers is obtained by registering with the point set P, so as to establish the matching relationship between the assembly holes in the overlapping area of ​​the 3D point clouds of adjacent viewpoints and obtain the matching point pair set of the assembly hole center ,in is the center point of a certain assembly hole in the 3D point cloud of a certain viewpoint, The 3D point cloud of its adjacent viewpoints is The center points of the mounting holes with the same name in physical space.

[0083] Step S3: Multi-view 3D point cloud registration based on improved ICP (Iterative Closest Point) algorithm:

[0084] In order to realize the splicing of multi-viewpoint 3D point clouds, an improved ICP algorithm based on the spatial geometric information weighted algorithm is designed, and the center of the hole feature of the curved antenna itself is added to the objective function of the improved ICP algorithm to realize the splicing of 3D point clouds from different viewpoints. Step S3 specifically includes:

[0085] Step S31, the assembly hole matching relationship established in step S25, selects the point cloud data in a cube area with the center of the assembly hole as the geometric center and the edge length a in a certain viewpoint 3D point cloud and its adjacent viewpoint 3D point cloud, and uses it as the point cloud to be registered and .

[0086] Step S32: point cloud is the source point cloud, point cloud For the target point cloud, KD-tree is used to accelerate the nearest neighbor search, and the nearest neighbor point in the target point cloud is found for each point in the source point cloud to obtain a paired point set .

[0087] Step S33: for each point pair Define the weights of the composite Gaussian form :

[0088]

[0089] in, and Points and The minimum value of the Euclidean distance to the center of all assembly holes in the respective point cloud; The parameter for controlling the weight decay rate has a base range determined by engineering experience, usually 0.1-0.3;

[0090] Step S34, calculate the weighted centroid:

[0091]

[0092]

[0093] Wherein, N is the number of elements in the pairing set described in step S32, and M is the number of elements in the assembly hole center point pair set described in step S25. is the weight of the assembly hole center point pair set, and its specific size is determined by engineering experience and is related to N.

[0094] Step S35, calculate the weighted covariance matrix:

[0095]

[0096] Perform SVD decomposition on the covariance matrix H, calculate the rotation matrix R and translation vector t, and obtain the registration matrix T according to the kinematic principle.

[0097] Step S36, repeat steps S32 to S35, iteratively update the transformation matrix T until a convergence condition is met, and the convergence condition is usually the number of iterations or the error change threshold.

[0098] Step S37: perform pairwise registration on all viewpoints to obtain several transformation matrices. Take the point cloud data of a certain viewpoint as the overall target point cloud, and perform rigid transformation on the remaining point clouds. The size of X depends on the number of pairwise registrations that need to be performed to transform the point cloud C into the overall target point cloud coordinate system.

[0099] Step S38: Fusing the overall target point cloud with all the point clouds after rigid transformation to obtain complete point cloud data of the curved antenna to be tested.

[0100] Step S4: 3D point cloud semantic segmentation and feature fitting:

[0101] In order to reconstruct the key features of the curved antenna, a curved antenna assembly hole and pad position fitting algorithm based on attribute clustering is designed to reconstruct the key features of the curved antenna. Figure 2 , step S4 specifically includes:

[0102] Step S41, point cloud pre-processing: set the voxel grid size according to the accuracy requirement, and perform through filtering, voxelization, statistical filtering and smoothing on the acquired complete point cloud data of the curved antenna. The through filtering, voxelization, statistical filtering and smoothing in this process are known methods, which can be directly used by those skilled in the art and will not be described in detail here.

[0103] Step S42, based on PCA (principal component analysis), calculate the normal vector of the pre-processed point cloud. According to the difference in normal vectors, the boundary points of the point cloud are segmented, and the boundary points are segmented twice by the region growing method to obtain the boundary points of the assembly hole; the curvature of the point cloud is calculated according to the normal vector to realize the recognition of the pad plane, and the RANSAC method is used for shape fitting. Finally, feature verification and pseudo-feature removal are performed to obtain the center point set of the assembly hole and the center point set of the pad. In this process, PCA (principal component analysis), region growing method, and RANSAC method are existing algorithms, and no specific improvements are made. Those skilled in the art can use them directly, or use other similar technologies as substitutes. This patent does not make special limitations on this, and there is no need to elaborate.

[0104] The measurement accuracy of the present invention is measured by using a three-dimensional coordinate measuring machine as the true value. The measurement results are shown in Table 1 below:

[0105] Table 1: Comparison between three-coordinate measurement values ​​and measurement values ​​of the present invention

[0106] Assembly hole number X Y Z 1 0.000158 0.006166 0.003988 2 0.010763 0.014723 0.019025 3 0.003842 0.000247 0.011533 4 0.009104 0.014592 0.014971 5 0.011877 0.010277 0.004582 6 0.012532 0.012617 0.002821 7 0.005480 0.000533 0.011419 8 0.004652 0.017415 0.000250 9 0.018239 0.009327 0.012900 Error mean 0.008516 0.009544 0.009054 Maximum error 0.018239 0.017415 0.019025

[0107] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes may be made to it in form and detail without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A method for reconstructing key features of curved antennas based on multimodal visual data fusion, characterized in that: The steps include: Design an artificially enhanced feature pattern and project it onto the surface of the curved antenna to be tested; Using a structured light camera to capture the surface image of the curved antenna to be tested containing the artificially enhanced characteristic pattern at different viewpoints as a two-dimensional image, and simultaneously collecting a three-dimensional point cloud of the curved antenna to be tested; Based on the two-dimensional image, establishing feature vectors of the center points of all assembly holes in the viewpoint; Obtain the matching relationship of the assembly holes in the 3D point cloud, map the matching results of the assembly holes to the 3D point cloud, and merge them under multiple different viewpoints to obtain the complete point cloud data of the curved antenna to be tested; Semantic segmentation and feature fitting are performed on the complete point cloud data, and an assembly hole center point set and a pad center point set are output.

2. The method for reconstructing key features of curved antennas by multimodal visual data fusion according to claim 1, characterized in that: The projection area A of the artificially enhanced characteristic pattern on the surface of the curved antenna to be tested is a rectangular area, and the length of the projection area A is , width ; The aspect ratio of the projection area A is the same as the aspect ratio of the minimum envelope rectangle of the curved antenna to be measured on the imaging plane of the two-dimensional camera; The length x and width y of the artificially enhanced characteristic pattern satisfy , , k is the safety factor.

3. The method for reconstructing key features of curved antennas by multimodal visual data fusion according to claim 2, characterized in that: Projecting the artificial enhanced characteristic pattern onto the surface of the curved antenna to be tested includes: The projection area A is meshed as , divide rectangular cells, the number of cells satisfies , , d is the diameter of the end face of the assembly hole of the curved antenna to be tested; Generate a random point in each cell obtained ; Random Points Draw a filled circular colored area with the center as the circle. The diameter of the circular colored area is satisfy , MP is the pixel size of the camera used to acquire the two-dimensional image.

4. The method for reconstructing key features of curved antennas by multimodal visual data fusion according to claim 3, characterized in that: The random points generated Coordinates satisfy: ； ； In the formula, i and j are the horizontal and vertical indexes of the grid respectively; , is the minimum value of the horizontal and vertical coordinates of the projection area A in the coordinate system; , is the length and width of the cell, which satisfies , ; is a random coefficient.

5. The method for reconstructing key features of curved antennas by multimodal visual data fusion according to claim 1, characterized in that: When the structured light camera captures the surface image of the curved antenna to be tested containing the artificially enhanced characteristic pattern at different viewpoints: The overlapping area of ​​two adjacent viewpoint images contains at least two assembly hole features.

6. The method for reconstructing key features of curved antennas by multimodal visual data fusion according to claim 1, characterized in that: Based on the two-dimensional image, the feature vectors of the center points of all assembly holes in the viewpoint are established, specifically including: Fit the circle in the two-dimensional image under all viewpoints and obtain its center to obtain the center point set of the assembly hole and the center point set of artificial feature patterns ; The points in the assembly hole center point set P As the target point, search for the point in the center point set Q of the artificial feature pattern The a points with the closest Euclidean distance are used to establish an a-dimensional distance vector in ascending order of Euclidean distance. The feature vectors of the mounting holes of the curved antenna.

7. The method for reconstructing key features of curved antennas by multimodal visual data fusion according to claim 6, characterized in that: The obtaining of the matching relationship of the assembly holes in the three-dimensional point cloud specifically includes: Fit the center of the assembly hole in the 3D point cloud and project it on the XY plane to obtain the projection point set of the center of the assembly hole , projection point set of the center of the assembly hole The corresponding relationship between the feature centers is obtained by aligning with the center point set P of the assembly hole, so as to establish the matching relationship between the assembly holes in the overlapping area of ​​the 3D point cloud of adjacent viewpoints and obtain the center point pair set of the assembly hole ,in is the center point of a certain assembly hole in the 3D point cloud of a certain viewpoint, The 3D point cloud of its adjacent viewpoints is The center points of the mounting holes with the same name in physical space.

8. The method for reconstructing key features of curved antennas by multimodal visual data fusion according to claim 7, characterized in that: Map the matching results of the assembly holes to the 3D point cloud and merge them from multiple viewpoints, including: According to the matching relationship of the assembly holes in the 3D point cloud, select the point cloud data in the cube area with the center of the assembly hole as the geometric center and the edge length a in the 3D point cloud of a certain viewpoint and its adjacent viewpoints, and use it as the point cloud to be registered and ; Point cloud to be registered is the source point cloud, the point cloud to be registered For the target point cloud, search for the nearest neighbor point in the target point cloud for each point in the source point cloud to obtain a paired point set ; For each pair of points in the paired point set Define the weights of the composite Gaussian form : ； In the formula, and Points and The minimum value of the Euclidean distance to the center of all assembly holes in the respective point cloud; Parameters that control the speed of weight decay; Compute the weighted centroid: ； ； Where N is the pairing point set The number of elements, M is the collocation pair set The number of elements in ; Assign weights to the point pairs of the assembly hole centers; Calculate the weighted covariance matrix H: ； Perform SVD decomposition on the covariance matrix H, calculate the rotation matrix R and translation vector t, and obtain the registration matrix T according to the kinematic principle; The transformation matrix T is updated iteratively until the convergence condition is met.

9. The method for reconstructing key features of curved antennas by multimodal visual data fusion according to claim 8, characterized in that: Map the matching results of the assembly holes to the 3D point cloud and merge them from multiple viewpoints. It also includes: The point clouds of all viewpoints are registered one by one in turn to obtain several transformation matrices; Take the point cloud data of a certain viewpoint as the overall target point cloud and perform rigid transformation on the remaining point clouds: ； Where X represents the number of pairwise registrations required to transform to the overall target point cloud coordinate system; C represents the point cloud before transformation, Represents the transformed point cloud; is the transformation matrix under the current viewpoint; The overall target point cloud is fused with all the point clouds after rigid transformation to obtain the complete point cloud data of the curved antenna to be tested.

10. The method for reconstructing key features of curved antennas by multimodal visual data fusion according to claim 9, characterized in that: The complete point cloud data is subjected to semantic segmentation and feature fitting, specifically including: The voxel grid size is set according to the desired accuracy requirement, and the complete point cloud data of the curved antenna to be measured is subjected to through-filtering, voxelization, statistical filtering and smoothing to obtain a pre-processed point cloud; Calculate the normal vector of the pre-processed point cloud; The boundary points of the point cloud are segmented according to the normal vector difference, and the boundary points are segmented again to obtain the boundary points of the assembly holes; Calculate the point cloud curvature based on the normal vector and identify the pad plane; Verify and eliminate the incorrect recognition results, and finally obtain the assembly hole center point set and the pad center point set.

Images