Multi-source point cloud registration method based on adaptive deformable cross-attention progressive matching

Abstract

The present invention belongs to the field of computer three-dimensional vision technology, and specifically discloses a multi-source point cloud registration method based on adaptive deformation cross-attention progressive matching, including: extracting features and downsampling the pre-processed dense point cloud and dense point cloud respectively, to obtain the features of the dense point cloud and dense point cloud and sparse point cloud; establishing a dual-branch sparse to dense soft matching relationship based on the features of the dense point cloud and dense point cloud and sparse point cloud; refining local features according to the soft matching relationship and adaptive deformation cross-attention network to obtain a sparse matching relationship; filtering the sparse corresponding relationship to obtain a filtered sparse matching relationship; performing local dense matching according to the filtered sparse matching relationship, restoring the rigid changes of the point cloud, and realizing multi-source point cloud registration. The present invention solves the problem that the consistency of multi-source point clouds is poor, resulting in a decrease in registration accuracy, and realizes efficient and accurate registration fusion of large-scale multi-source laser point clouds.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer three-dimensional vision, and in particular relates to the design of a multi-source point cloud registration method based on adaptive deformation cross-attention progressive matching. Background Art

[0002] Point cloud registration is an important research direction in the field of computer vision and 3D reconstruction. Its purpose is to generate a unified 3D model by aligning point cloud data from different perspectives. Point cloud registration technology has a wide range of applications in autonomous driving, robot navigation, 3D measurement, cultural heritage protection and other fields.

[0003] Traditional point cloud registration methods are mainly divided into feature-based registration and iterative optimization-based registration. Feature-based registration methods achieve point cloud registration by extracting key feature points and their descriptors from the point cloud and matching these feature points. However, the matching accuracy and robustness of this type of method depend on the selection of feature points and the stability of the descriptor. When faced with complex situations such as sparse point cloud data, noise interference, and local occlusion, matching failures or reduced accuracy are prone to occur. Registration methods based on iterative optimization, such as iterative closest point ( ICP ), when processing large-scale point cloud data, the computational complexity is high and it is difficult to meet the real-time requirements.

[0004] In recent years, with the development of deep learning technology, point cloud registration methods based on deep learning have gradually emerged. Such methods use the powerful learning ability of neural networks to learn the best strategies for feature extraction and matching from a large amount of data, thereby improving the registration accuracy and robustness. However, multi-source point cloud data often come from different sensors or different perspectives, and there are significant differences in their data distribution and structure. Existing deep learning methods usually fail to fully consider these differences when processing such data, resulting in a decrease in registration accuracy. Especially for multi-source remote sensing lidar point clouds, the scale of points is usually huge, so the commonly used method of solving the correspondence between dense point pairs is very time-consuming, and sparse key point matching is prone to key point detection errors due to its limited repeatability.

[0005] For multi-source remote sensing application scenarios, proposing a registration method that can effectively process multi-source point cloud data and has high precision and robustness is an important direction of current point cloud registration. Summary of the invention

[0006] The purpose of the present invention is to solve the problem that the consistency of multi-source point clouds is poor, resulting in a decrease in registration accuracy. A multi-source point cloud registration method based on adaptive deformation cross-attention progressive matching is proposed.

[0007] The technical solution of the present invention is: a multi-source point cloud registration method based on adaptive deformation cross-attention progressive matching, the method comprising the following steps:

[0008] S1. Preprocess the multi-source point cloud to obtain the processed dense point cloud and dense point clouds ;

[0009] S2. For dense point cloud and dense point clouds Perform feature extraction and downsampling respectively to obtain dense point cloud Features , dense point cloud Sparse point cloud , dense point cloud Features and dense point clouds Sparse point cloud ;

[0010] S3. Based on dense point cloud Features , dense point cloud Sparse point cloud , dense point cloud Features and dense point clouds Sparse point cloud , establish a dual-branch sparse to dense soft matching relationship;

[0011] S4. Refine local features based on soft matching relationship and adaptive deformation cross attention network to obtain sparse matching relationship;

[0012] S5. Filter the sparse matching relationship to obtain the filtered sparse matching relationship;

[0013] S6. Perform local dense matching based on the screened sparse matching relationship, restore the rigid changes of the point cloud, and realize multi-source point cloud registration.

[0014] The beneficial effects of the present invention are:

[0015] 1. By performing dual-branch sparse-to-dense matching, the overhead of solving large-scale point cloud correspondences is reduced, and bidirectional matching mines more multi-source point cloud consistency.

[0016] 2. By introducing adaptive deformable cross-attention to further enhance the feature interaction between multi-source point clouds, more robust point cloud features and more accurate point cloud correspondences are obtained, achieving efficient and accurate registration and fusion of large-scale multi-source laser point clouds.

[0017] Preferably, the multi-source point cloud is preprocessed in step S1, specifically: voxel-downsampling the multi-source point cloud using different downsampling parameters according to the density of the multi-source point cloud.

[0018] The beneficial effects of the above preferred solution are:

[0019] By voxelizing and downsampling the multi-source point cloud, the point cloud can be input into the network module for processing.

[0020] Preferably, the step S2 specifically comprises the following steps:

[0021] S21. Dense point cloud through kernel point convolution network and dense point clouds Perform feature extraction to obtain dense point cloud Features , dense point cloud Features ;

[0022] S22. Use the farthest point sampling algorithm to sample the dense point cloud and dense point clouds Downsample to get a dense point cloud Sparse point cloud and dense point clouds Sparse point cloud .

[0023] Preferably, the soft matching relationship in step S3 is established by a dual-branch structure, where branch one performs sparse point cloud and dense point clouds Matching, branch 2 performs sparse point cloud and dense point clouds The matching includes the following steps:

[0024] In step S3, the soft matching relationship is established through a dual-branch structure, and branch one performs sparse point cloud and dense point clouds Matching, branch 2 performs sparse point cloud and dense point clouds The matching includes the following steps:

[0025] S31. For sparse point clouds The sampling points in , calculate the sampling points Features With features The Euclidean distance between them is used to obtain the feature matching similarity of branch one;

[0026] S32. Select the first branch with the highest feature matching similarity. dense points as soft matching results, and obtain the soft matching relationship matrix of branch one ;

[0027] S33. For sparse point clouds The sampling points in , calculate the sampling points Features With features The Euclidean distance between them is used to obtain the feature matching similarity of branch 2, and the top feature matching similarity in branch 2 is selected. The dense points are taken as the soft matching results, and the soft matching relationship matrix of branch 2 is obtained. ;

[0028] S34. According to the soft matching relationship matrix and soft matching relationship matrix , and obtain a double-branch sparse to dense soft matching relationship.

[0029] Preferably, the step S4 specifically comprises the following steps:

[0030] S41. Based on soft matching relationship matrix and soft matching relationship matrix , get the sparse point cloud in branch 1 The corresponding dense point cloud The soft matching candidate point set in and the sparse point cloud in branch 2 The corresponding dense point cloud The soft matching candidate point set in ;

[0031] S42. Calculate the two soft matching candidate point sets respectively to obtain the centroid and centroid features of the two soft matching candidate point sets, and convert the sampling points Features And sampling points Features As a query in the attention mechanism ;

[0032] S43. Query , the centroid and centroid features of each soft matching candidate point set are input into the adaptive deformable cross attention network, and the local features of branch one and branch two are output;

[0033] S44. According to the local features refined by branch one and the local features refined by branch two, the feature similarity of each sampling point is calculated, and the sampling point with the highest feature similarity is selected as the sparse matching result to obtain the sparse matching relationship matrix and sparse matching relationship matrix ;

[0034] S45. Based on the sparse matching relationship matrix and sparse matching relationship matrix , and obtain a sparse matching relationship.

[0035] Preferably, the centroids of the two soft matching candidate point sets are The calculation formula is:

[0036]

[0037] in, Indicates the total number of soft matching candidate points, Indicates soft matching candidate points, Indicates The source point corresponds to Soft matching candidate points Axis coordinates, Indicates The source point corresponds to Soft matching candidate points Axis coordinates, Indicates The source point corresponds to Soft matching candidate points Axis coordinates;

[0038] The centroid feature The calculation formula is:

[0039]

[0040] in, Indicates The source point corresponds to Features of soft matching candidate points;

[0041] The formula of the adaptive deformable cross attention network is:

[0042]

[0043] in, Represents the key in the attention mechanism, = , Represents a mapping to query values, Represents a mapping of key values. represents the first multi-layer perceptron, represents the second multi-layer perceptron, Represents a multi-layer perceptron The predicted coordinate offset value, represents the original coordinates, = , represents the actual sampling point coordinates, Indicates that a nearest neighbor search is performed on the point coordinates. Indicates Attention Head Next The weight distribution of sampling points, Represents the mapping of features obtained by each attention head, represents the feature output obtained by the deformable attention network, that is, the refined local features, represents the total number of attention heads, Represents the total number of sampling points, Indicates The actual sampling point coordinates.

[0044] Preferably, the step S5 specifically comprises the following steps:

[0045] S51. Based on the sparse matching relationship matrix and sparse matching relationship matrix , obtain the characteristics of sparse matching relations;

[0046] S52. Input the features of the sparse matching relationship into FilterNet In the network, output the confidence of each group of sparse matching relationships ;

[0047] S53. Select confidence level Greater than a given threshold sparse matching relationship to obtain the filtered sparse matching relationship.

[0048] Preferably, the FilterNet The network formula is:

[0049]

[0050] in, represents a linear neural network, represents the activation function, Features representing sparse matching relationships.

[0051] Preferably, the step S6 specifically comprises the following steps:

[0052] S61. For each filtered sparse matching relationship, construct a local space neighborhood cluster of each matching point by nearest neighbor search, and obtain adjacent features of the local space neighborhood cluster;

[0053] S62. For any given set of sparse matching relationships, construct a corresponding cost matrix according to the adjacent neighborhood point features of its two sets of local spatial neighborhood clusters;

[0054] S63. According to the cost matrix, use Sinkhorn The soft assignment matrix is ​​calculated by iterative algorithm;

[0055] S64. Using the soft assignment matrix as a confidence matrix of candidate matches to obtain a local dense matching set;

[0056] S65. Based on local dense matching sets, using LGR The algorithm calculates the rotation matrix and translation matrix from the target point cloud to the source point cloud to achieve point cloud registration.

[0057] Preferably, the cost matrix in step S62 is Elements in The calculation formula is:

[0058]

[0059] in, and represents two sets of adjacent features, It means to calculate the bi-norm between eigenvectors. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Shown is a flowchart of a multi-source point cloud registration method based on adaptive deformable cross-attention progressive matching.

[0061] Figure 2 Shown is a dual-branch point cloud registration flow chart provided in an embodiment of the present invention.

[0062] Figure 3 Shown is the corresponding visualization effect of dual-branch point cloud sparse matching.

[0063] Figure 4 Shown is a diagram of an adaptive deformable cross-attention structure provided in an embodiment of the present invention.

[0064] Figure 5 Shown is the corresponding visualization effect of point cloud dense matching.

[0065] Figure 6 Shown is the visualization effect after the final registration of the point cloud. DETAILED DESCRIPTION

[0066] Now, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the accompanying drawings are only exemplary and are intended to explain the principles and spirit of the present invention, rather than to limit the scope of the present invention.

[0067] Example:

[0068] like Figure 1 and Figure 2 As shown, a multi-source point cloud registration method based on adaptive deformation cross-attention progressive matching, the method comprises the following steps:

[0069] S1. Preprocess the multi-source point cloud to obtain the processed dense point cloud and dense point clouds ;

[0070] S2. For dense point cloud and dense point clouds Perform feature extraction and downsampling respectively to obtain dense point cloud Features , dense point cloud Sparse point cloud , dense point cloud Features and dense point clouds Sparse point cloud ;

[0071] S3. Based on dense point cloud Features , dense point cloud Sparse point cloud , dense point cloud Features and dense point clouds Sparse point cloud , establish a dual-branch sparse to dense soft matching relationship;

[0072] S4. Refine local features based on soft matching relationship and adaptive deformation cross attention network to obtain sparse matching relationship;

[0073] S5. Filter the sparse matching relationship to obtain the filtered sparse matching relationship;

[0074] S6. Perform local dense matching based on the screened sparse matching relationship, restore the rigid changes of the point cloud, and realize multi-source point cloud registration.

[0075] In this embodiment, the multi-source point cloud is preprocessed as described in step S1, specifically: according to the density of the multi-source point cloud, the multi-source point cloud is voxelized and downsampled using different downsampling parameters; the multi-source point cloud after voxelization and downsampling is randomly rotated, randomly translated, randomly scaled and randomly noisy to obtain a dense point cloud after data enhancement. and dense point clouds ; Random rotation will be from A random value is selected to rotate the multi-source point cloud; random translation will be from Randomly select a value from the translation point cloud; random scaling is from Randomly select a ratio value to scale the point cloud proportionally; random noise addition adds a random Gaussian noise to the point cloud coordinates to simulate the noise of point cloud collection.

[0076] In this embodiment, step S2 specifically includes the following steps:

[0077] S21. Through kernel point convolution ( KPConv ) Network for dense point cloud and dense point clouds Perform feature extraction to obtain dense point cloud Features , dense point cloud Features ;in, represents the set of real numbers, and Represents dense point clouds and dense point clouds The number of points, 3 represents a dense point cloud and dense point clouds The three-dimensional coordinates of ;

[0078] S22. Using the farthest point sampling ( FPS ) algorithm respectively for dense point clouds and dense point clouds Downsample to get a dense point cloud Sparse point cloud and dense point clouds Sparse point cloud ,in, represents the set of real numbers, and Respectively represent the source point cloud, that is, the dense point cloud And the target point cloud is a dense point cloud The number of points after downsampling.

[0079] In this embodiment, the soft matching relationship in step S3 is established through a dual-branch structure, and branch one performs sparse point cloud and dense point clouds Matching, branch 2 performs sparse point cloud and dense point clouds The matching includes the following steps:

[0080] In step S3, the soft matching relationship is established through a dual-branch structure, and branch one performs sparse point cloud and dense point clouds Matching, branch 2 performs sparse point cloud and dense point clouds The matching includes the following steps:

[0081] S31. For sparse point clouds The sampling points in , calculate the sampling points Features With features The Euclidean distance between them is used to obtain the feature matching similarity of branch one;

[0082] S32. Select the first branch with the highest feature matching similarity. dense points as soft matching results, and obtain the soft matching relationship matrix of branch one ;

[0083] S33. For sparse point clouds The sampling points in , calculate the sampling points Features With features The Euclidean distance between them is used to obtain the feature matching similarity of branch 2, and the top feature matching similarity in branch 2 is selected. The dense points are taken as the soft matching results, and the soft matching relationship matrix of branch 2 is obtained. ;

[0084] S34. According to the soft matching relationship matrix and soft matching relationship matrix , and obtain a double-branch sparse to dense soft matching relationship.

[0085] In this embodiment, step S4 specifically includes the following steps:

[0086] S41. Based on soft matching relationship matrix and soft matching relationship matrix , get the sparse point cloud in branch 1 The corresponding dense point cloud The soft matching candidate point set in and the sparse point cloud in branch 2 The corresponding dense point cloud The soft matching candidate point set in ;

[0087] S42. Calculate the two soft matching candidate point sets respectively to obtain the centroid and centroid features of the two soft matching candidate point sets, and convert the sampling points Features And sampling points Features As a query in the attention mechanism ;

[0088] S43. Query , the centroid and centroid features of each soft matching candidate point set are input into the adaptive deformable cross attention network, and the local features of branch one and branch two are output;

[0089] S44. According to the local features refined by branch one and the local features refined by branch two, the feature similarity of each sampling point is calculated, and the sampling point with the highest feature similarity is selected as the sparse matching result to obtain the sparse matching relationship matrix and sparse matching relationship matrix ;

[0090] S45. Based on the sparse matching relationship matrix and sparse matching relationship matrix , we get a sparse matching relationship, and its visualization effect is as follows Figure 3 shown.

[0091] In this embodiment, the centroids of the two soft matching candidate point sets are The calculation formula is:

[0092]

[0093] in, Indicates the total number of soft matching candidate points, Indicates soft matching candidate points, Indicates The source point corresponds to Soft matching candidate points Axis coordinates, Indicates The source point corresponds to Soft matching candidate points Axis coordinates, Indicates The source point corresponds to Soft matching candidate points Axis coordinates;

[0094] The centroid feature The calculation formula is:

[0095]

[0096] in, Indicates The source point corresponds to Features of soft matching candidate points;

[0097] like Figure 4As shown, the formula of the adaptive deformation cross attention network is:

[0098]

[0099] in, Represents the key in the attention mechanism, , Represents a mapping to query values, Represents a mapping of key values. represents the first multi-layer perceptron, represents the second multi-layer perceptron, Represents a multi-layer perceptron The predicted coordinate offset value, represents the original coordinates, = , represents the actual sampling point coordinates, Indicates that a nearest neighbor search is performed on the point coordinates. Indicates Attention Head Next The weight distribution of sampling points, Represents the mapping of features obtained by each attention head, represents the feature output obtained by the deformable attention network, that is, the refined local features, represents the total number of attention heads, Represents the total number of sampling points, Indicates The actual sampling point coordinates.

[0100] In this embodiment, step S5 specifically includes the following steps:

[0101] S51. Based on the sparse matching relationship matrix and sparse matching relationship matrix , obtain the characteristics of sparse matching relations;

[0102] S52. Input the features of the sparse matching relationship into FilterNet In the network, output the confidence of each group of sparse matching relationships ;

[0103] S53. Select confidence level Greater than a given threshold sparse matching relationship to obtain the filtered sparse matching relationship.

[0104] In this embodiment, the FilterNet The network formula is:

[0105]

[0106] in, represents a linear neural network, Represents the activation function, which is used to map the features obtained by the linear layer to a confidence level between (0 and 1) , The features representing sparse matching relationships are , and Represents a sparse matching relationship Corresponding features.

[0107] In this embodiment, step S6 specifically includes the following steps:

[0108] S61. For each filtered sparse matching relationship, construct a local spatial neighborhood cluster of each matching point by nearest neighbor search and , and obtain the adjacent features of the local spatial neighborhood cluster;

[0109] S62. For any given set of sparse matching relationships, according to the adjacent neighborhood point features of its two sets of local spatial neighborhood clusters and ,in, represents the set of real numbers, represents the number of neighborhood points, Represents the feature dimension and constructs the corresponding cost matrix ;

[0110] S63. According to the cost matrix ,use Sinkhorn The soft assignment matrix is ​​calculated by iterative algorithm ;

[0111] S64. The soft allocation matrix As the confidence matrix of the candidate match, the local dense match set is obtained, and the corresponding visualization effect of the point cloud dense match is as follows: Figure 5 As shown;

[0112] S65. Based on local dense matching sets, using LGR The algorithm calculates the rotation matrix and translation matrix from the target point cloud to the source point cloud to achieve point cloud registration. The visualization effect after the final registration of the point cloud is as follows: Figure 6 shown.

[0113] In this embodiment, the cost matrix in step S62 is Elements in The calculation formula is:

[0114]

[0115] in, and represents two sets of adjacent features, It means to calculate the bi-norm between eigenvectors.

[0116] In this embodiment, during training, Adam The optimizer and the maximum number of training rounds are 200, the initial learning rate is set to 0.0004, and it is reduced to half of the original every 5 rounds. batch The size is set to 1, with different LiDAR Experiments on the point cloud collected by the device show that the method provided by the present invention has good sparse matching accuracy, and can provide more precise matching results for subsequent pose estimation through progressive matching from sparse to dense.

[0117] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A multi-source point cloud registration method based on adaptive deformation cross-attention progressive matching, characterized in that: The method comprises the following steps: S1. Preprocess the multi-source point cloud to obtain the processed dense point cloud and dense point clouds ; S2. For dense point cloud and dense point clouds Perform feature extraction and downsampling respectively to obtain dense point cloud Features , dense point cloud Sparse point cloud , dense point cloud Features and dense point clouds Sparse point cloud , specifically including the following steps: S21. Dense point cloud through kernel point convolution network and dense point clouds Perform feature extraction to obtain dense point cloud Features , dense point cloud Features ; S22. Use the farthest point sampling algorithm to sample the dense point cloud and dense point clouds Downsample to get a dense point cloud Sparse point cloud and dense point clouds Sparse point cloud ; S3. Based on dense point cloud Features , dense point cloud Sparse point cloud , dense point cloud Features and dense point clouds Sparse point cloud , establish a dual-branch sparse to dense soft matching relationship, the soft matching relationship is established through a dual-branch structure, branch one performs sparse point cloud and dense point clouds Matching, branch 2 performs sparse point cloud and dense point clouds The matching includes the following steps: S31. For sparse point clouds The sampling points in , calculate the sampling points Features With features The Euclidean distance between them is used to obtain the feature matching similarity of branch one; S32. Select the first branch with the highest feature matching similarity. dense points as soft matching results, and obtain the soft matching relationship matrix of branch one ; S33. For sparse point clouds The sampling points in , calculate the sampling points Features With features The Euclidean distance between them is used to obtain the feature matching similarity of branch 2, and the top feature matching similarity in branch 2 is selected. The dense points are taken as the soft matching results, and the soft matching relationship matrix of branch 2 is obtained. ; S34. According to the soft matching relationship matrix and soft matching relationship matrix , we get the double-branch sparse to dense soft matching relationship; S4. Refine the local features according to the soft matching relationship and the adaptive deformation cross attention network to obtain a sparse matching relationship, which specifically includes the following steps: S41. Based on soft matching relationship matrix and soft matching relationship matrix , get the sparse point cloud in branch 1 The corresponding dense point cloud The soft matching candidate point set in and the sparse point cloud in branch 2 The corresponding dense point cloud The soft matching candidate point set in ; S42. Calculate the two soft matching candidate point sets respectively to obtain the centroid and centroid features of the two soft matching candidate point sets, and convert the sampling points Features And sampling points Features As a query in the attention mechanism ; S43. Query , the centroids and centroid features of the two soft matching candidate point sets are input into the adaptive deformable cross attention network, and the refined local features of branch one and branch two are output; S44. According to the local features refined by branch one and the local features refined by branch two, the feature similarity of each sampling point is calculated, and the sampling point with the highest feature similarity is selected as the sparse matching result to obtain the sparse matching relationship matrix and sparse matching relationship matrix ; S45. Based on the sparse matching relationship matrix and sparse matching relationship matrix , and obtain a sparse matching relationship; S5. Filter the sparse matching relationship to obtain the filtered sparse matching relationship; S6. Perform local dense matching based on the screened sparse matching relationship, restore the rigid changes of the point cloud, and realize multi-source point cloud registration.

2. The multi-source point cloud registration method based on adaptive deformation cross-attention progressive matching according to claim 1 is characterized in that: The multi-source point cloud is preprocessed in step S1, specifically: voxel-downsampling the multi-source point cloud using different downsampling parameters according to the density of the multi-source point cloud.

3. The multi-source point cloud registration method based on adaptive deformation cross-attention progressive matching according to claim 1 is characterized in that: The centroids of the two soft matching candidate point sets The calculation formula is: in, Indicates the total number of soft matching candidate points, Indicates soft matching candidate points, Indicates The source point corresponds to Soft matching candidate points Axis coordinates, Indicates The source point corresponds to Soft matching candidate points Axis coordinates, Indicates The source point corresponds to Soft matching candidate points Axis coordinates; The centroid feature The calculation formula is: in, Indicates The source point corresponds to Features of soft matching candidate points; The formula of the adaptive deformable cross attention network is: in, Represents the key in the attention mechanism, = , Represents a mapping to query values, Represents a mapping of key values. represents the first multi-layer perceptron, represents the second multi-layer perceptron, Represents a multi-layer perceptron The predicted coordinate offset value, represents the original coordinates, = , represents the actual sampling point coordinates, Indicates that a nearest neighbor search is performed on the point coordinates. Indicates Attention Head Next The weight distribution of sampling points, Represents the mapping of features obtained by each attention head, represents the feature output obtained by the deformable attention network, that is, the refined local features, represents the total number of attention heads, Represents the total number of sampling points, Indicates The actual sampling point coordinates.

4. The multi-source point cloud registration method based on adaptive deformation cross-attention progressive matching according to claim 1 is characterized in that: The step S5 specifically comprises the following steps: S51. Based on the sparse matching relationship matrix and sparse matching relationship matrix , obtain the characteristics of sparse matching relations; S52. Input the features of the sparse matching relationship into FilterNet In the network, output the confidence of each group of sparse matching relationships ; S53. Select confidence level Greater than a given threshold sparse matching relationship to obtain the filtered sparse matching relationship.

5. The multi-source point cloud registration method based on adaptive deformation cross-attention progressive matching according to claim 4 is characterized in that: Said FilterNet The network formula is: in, represents a linear neural network, represents the activation function, Features representing sparse matching relationships.

6. The multi-source point cloud registration method based on adaptive deformation cross-attention progressive matching according to claim 1 is characterized in that: The step S6 specifically comprises the following steps: S61. For each filtered sparse matching relationship, construct a local spatial neighborhood cluster of each matching point by nearest neighbor search, and obtain neighborhood point features; S62. For any given set of sparse matching relationships, construct a corresponding cost matrix according to the adjacent neighborhood point features of its two sets of local spatial neighborhood clusters; S63. According to the cost matrix, use Sinkhorn The soft assignment matrix is ​​calculated by iterative algorithm; S64. Using the soft assignment matrix as a confidence matrix of candidate matches to obtain a local dense matching set; S65. Based on local dense matching sets, using LGR The algorithm calculates the rotation matrix and translation matrix from the target point cloud to the source point cloud to achieve point cloud registration.

7. The multi-source point cloud registration method based on adaptive deformation cross-attention progressive matching according to claim 6 is characterized in that: The cost matrix in step S62 Elements in The calculation formula is: in, and represents two sets of adjacent features, It means to calculate the bi-norm between eigenvectors.

Images