A point cloud registration method and system based on multi-scale feature similarity constraints

Abstract

The present invention belongs to the field of computer vision graphics, and specifically relates to a point cloud registration method and system based on multi-scale feature similarity constraints. The matching credibility of points is actively evaluated, and a network with an attention fusion mechanism is used to extract discriminative features. A bidirectional alternating correspondence search mechanism and a flexible misalignment loss function are used for robust learning. A point pair weight distribution scheme under consistency constraints is introduced, and finally a transformation matrix is ​​obtained by weighted SVD, thereby improving the accuracy of registration. The present invention can establish a robust correspondence relationship between point clouds, improve the effect of deep models on point cloud registration, and promote the development of related technical fields.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision graphics, and in particular relates to a point cloud registration method and system based on multi-scale feature similarity constraints. Background Art

[0002] The goal of 3D point cloud registration is to find an optimal spatial transformation to align one point cloud with another. Point cloud registration is an important research topic in the field of computer vision, and many applications are built on it, such as augmented reality, pose estimation, and 3D scene reconstruction. The most widely used traditional registration method is the ICP (Iterative Closest Point) algorithm, which mainly assigns each point in a set to the closest point in another set continuously, and then finds a rigid transformation that conforms to the least squares through iterative calculation. However, ICP is heavily dependent on good initial conditions and is prone to fall into local optimality.

[0003] With the development of deep learning, data-driven point cloud registration methods are constantly being proposed. This type of method usually extracts features from the data and then establishes corresponding relationships based on feature similarity. Compared with traditional methods, it has better registration accuracy, computational efficiency and generalization ability. However, there are still some problems behind it. For example, the registration of point clouds that only share some overlapping areas is not ideal. On the one hand, there are a large number of outliers between point clouds that interfere with point cloud registration. On the other hand, after the corresponding relationship is established, the wrong point pairs participate in the reasoning. In order to solve the problem of local to local point cloud registration, it is usually necessary to screen key points and then establish corresponding point relationships for the key points. Although some researchers have also proposed related methods, such as iterative multi-stage point elimination methods, mask elimination of the influence of non-overlapping points on the global features of point clouds, etc., these methods have problems such as poor registration accuracy and easy overfitting for point clouds with low overlap rates. Summary of the invention

[0004] To solve the above problems, the present invention provides a point cloud registration method and system based on multi-scale feature similarity constraints.

[0005] In a first aspect, a point cloud registration method based on multi-scale feature similarity constraints is provided, wherein a point cloud registration model is constructed and trained, the data to be registered is input into the trained point cloud registration model, a multi-scale consistency constraint is introduced to obtain point pair weights, and then the registration result is calculated;

[0006] The training process of the point cloud registration model includes:

[0007] S1. Obtain source point cloud and target point cloud;

[0008] S2. Extract the original features of each point in the two groups of point clouds, exchange the original features of the two groups of point clouds and obtain the deep features of each point through MLP, perform maximum pooling on the deep features to obtain the global features of the two groups of point clouds, concatenate the global features of the two groups of point clouds with the deep features of each point, and output the confidence of each point through a fully connected network; here, set the source point cloud confidence loss function;

[0009] S3. Use GNN to extract local structural features of the two groups of point clouds, record the GNN output features as the first-scale feature representation, and use the attention mechanism to obtain the final feature representations of the two groups of point clouds after enhancement, record them as the second-scale feature representation;

[0010] S4. randomly select K points in the source point cloud in the overlapping area of ​​the two groups of point clouds to form a first point set, and index the second scale feature representation corresponding to the first point set;

[0011] S5. multiplying the second scale feature representation corresponding to the first point set with the second scale feature representation of the target point cloud to obtain a similarity matrix;

[0012] S6. Normalize the similarity matrix by row to obtain a corresponding distribution matrix between the first point set and the target point cloud, and multiply the corresponding distribution matrix by the target point cloud to obtain a first virtual point set matching the first point set; here, a corresponding loss function is set for the first virtual point set and the real corresponding point set;

[0013] S7. Select the maximum element in each row of the corresponding distribution matrix, normalize it, and obtain the weight of the matching point pair between the first point set and the first virtual point set;

[0014] S8. Based on the corresponding distribution matrix and the weights of the matching point pairs, the SVD algorithm is used to solve the rotation matrix and the translation matrix between the source point cloud and the target point cloud;

[0015] S9. Transform the source point cloud according to the rotation matrix and translation matrix, return to step S2, and iterate 3 times;

[0016] S10. Start the gradient back-propagation mechanism, optimize the loss function, update the network parameters, and save the model when the model converges or reaches the set number of epochs.

[0017] Furthermore, the consistency constraints introduced into the trained point cloud registration model include:

[0018] The similarity matrix from the source point cloud to the target point cloud is calculated based on the first scale feature representation, which is denoted as M1. The similarity matrix from the source point cloud to the target point cloud is calculated based on the second scale feature representation, which is denoted as M2.

[0019] Multiply M1 and M2 bit by bit to get a new similarity matrix and normalize it row by row to get the corresponding point set of the source point cloud in the target point cloud;

[0020] Select the maximum element of each row in the new similarity matrix, align and normalize, and obtain the weight of the matching point pair between the source point cloud and the corresponding point set;

[0021] Based on the new similarity matrix and the weights of the matching point pairs, the SVD algorithm is used to solve the rotation matrix and translation matrix between the source point cloud and the target point cloud; the final transformation matrix is ​​calculated.

[0022] Furthermore, a bidirectional loss function is used in the training phase to optimize both the registration process from the source point cloud to the target point cloud and the registration process from the target point cloud to the source point cloud.

[0023] Furthermore, the bidirectional loss function of the point cloud registration model training phase is expressed as:

[0024]

[0025] in, represents the confidence loss of the source point cloud, represents the confidence loss of the target point cloud, represents the loss between the virtual points and the real points in the target point cloud corresponding to the source point cloud, Represents the loss between the virtual points in the source point cloud and the real points corresponding to the target point cloud.

[0026] Furthermore, the confidence loss of the source point cloud is calculated using the typical cross entropy loss, and the formula is:

[0027]

[0028] Among them, GT(x i ) represents point x i The true label is used to represent the point x i Whether it belongs to the overlapping area of ​​the two sets of point clouds or the non-overlapping area, it is represented by 1 and 0 respectively. i ) represents the point x in the source point cloud i confidence level.

[0029] Furthermore, the loss function error between the virtual point and the real point in the target point cloud corresponding to the source point cloud is It is expressed as:

[0030]

[0031] Among them, x i represents a point in the source point cloud, Represented as x iThe generated virtual corresponding points (R XY x i +t XY ) represents the target point cloud with x i The corresponding real point.

[0032] In a second aspect, the present invention provides a point cloud registration system based on multi-scale feature similarity constraints, comprising:

[0033] A data collection module is used to obtain source point cloud and target point cloud data input into the model;

[0034] Downsampling module, used to filter out the point set involved in registration;

[0035] The feature exchange module is used to extract the features of each point in the source point cloud and the target point cloud, and then exchange the feature information of the two sets of point clouds;

[0036] The MLP module is used to extract deep features of two sets of point clouds using multiple MLPs based on the output of the feature exchange module;

[0037] The pooling module is used to perform maximum pooling on the deep features of the two groups of point clouds to obtain the global features of the source point cloud and the global features of the target point cloud;

[0038] The splicing module is used to fuse the feature information of points and splice the global features of the source point cloud and the target point cloud with the deep features of each point;

[0039] The confidence calculation module is used to input the concatenated features of each point into the fully connected layer to obtain the confidence of each point;

[0040] A local structure feature acquisition module is used to obtain the first scale feature representation of the source point cloud and the target point cloud;

[0041] The attention module is used to obtain the correlation information between point clouds and enhance the feature representation of point clouds, that is, the first-scale feature representation is processed according to the self-attention mechanism and the cross-attention mechanism to obtain the second-scale feature representation of the two groups of point clouds;

[0042] The candidate point set module is used to select K points with the highest confidence in the source point cloud to form the first point set according to the output result of the confidence calculation module;

[0043] An indexing module, used to filter the different scale feature representations of the first point set in the registration process;

[0044] A matching module, used for obtaining a corresponding point set matching the first point set in the target point cloud;

[0045] A consistency constraint module is used to introduce consistency constraints into the point cloud registration model to obtain a corresponding distribution matrix between the first point set and the target point cloud;

[0046] SVD solving module, used to calculate the rotation matrix and translation matrix between the source point cloud and the target point cloud according to the results of the virtual matching module and the weights obtained by the consistency constraint module;

[0047] The transformation module is used to transform the source point cloud according to the rotation matrix and the translation matrix.

[0048] Beneficial effects of the present invention:

[0049] The present invention proposes a point cloud registration method and system based on multi-scale feature similarity constraints. The matching credibility of points is actively evaluated through a multi-scale feature exchange network. A network with an attention fusion mechanism is used to extract discriminative features. A bidirectional alternating correspondence search mechanism and a flexible misalignment loss function are used for robust learning. A point pair weight allocation scheme based on feature similarity consistency constraints established at multiple scales is introduced in the inference stage. Finally, the transformation matrix is ​​obtained by weighted SVD to improve the accuracy of registration. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a point cloud registration structure diagram based on multi-scale feature similarity constraints of the present invention;

[0051] Figure 2 The registration results and high-scoring correspondences based on consistency constraints of the present invention;

[0052] Figure 3 This is a diagram of the registration result according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0054] This paper proposes a point cloud registration method based on multi-scale feature similarity constraints, such as Figure 1 As shown, including:

[0055] S1. Obtain source point cloud and target point cloud. In this embodiment, the maximum number of epochs is set to 100 and the initial learning rate is 0.0001;

[0056] S2. Extract the original features of each point in the two sets of point clouds, exchange the original features of the two sets of point clouds and finally obtain the deep features of each point through multiple MLPs, perform maximum pooling on the deep features of each point to obtain the global features of the source point cloud and the global features of the target point cloud, concatenate the global features of the two sets of point clouds with the deep features of each point, and output the confidence of each point through a fully connected network;

[0057] Confidence is used to represent the high reliability matching between point clouds. After the model is trained, the higher the confidence point, the more likely it is that the point is in the overlapping area of ​​the source point cloud and the target point cloud, and the higher the matching accuracy.

[0058] Specifically, in difficult situations such as noise, outliers or large changes, using potential corresponding points with higher matching credibility is conducive to the robustness of registration. However, in local-local point cloud registration, non-overlapping areas do not have corresponding relationships in the other point cloud. Therefore, a multi-scale feature exchange network is proposed to focus on the global semantics of the overlapping credible areas of the two sets of point clouds, and a Transformer-based mechanism is used to exchange the original features of the two point clouds. The formula can be expressed as:

[0059] f x =f x +ζ(f x ,f y );

[0060] f y =f y +ζ(f y ,f x );

[0061] Among them, ζ(·) is the residual term learned based on the Transformer model, and f x represents the original features of the source point cloud, f y Represents the original features of the target point cloud.

[0062] Specifically, the concatenated points are sent to the fully connected layer to obtain the semantic information of each concatenated point, that is, the confidence of each point, expressed as:

[0063]

[0064]

[0065] Among them, G X Represents the global features of the source point cloud, G Y Represents the global features of the target point cloud, S x Represents the confidence set of each point in the source point cloud, S y Represents the confidence set of each point in the target point cloud, ρs () indicates a fully connected network, and r() indicates repetition of global features.

[0066] S3. Use GNN to extract local structural features of the two groups of point clouds, record the GNN output features as the first-scale feature representation, and use the attention mechanism to obtain the final feature representations of the two groups of point clouds after enhancement, record them as the second-scale feature representation;

[0067] Specifically, the first-scale feature representations of the two groups of point clouds are input into the self-attention mechanism to obtain two groups of results, and the two groups of results are input into the cross-attention mechanism to output the second-scale feature representations of the two groups of point clouds, wherein the source point cloud passes the first-scale feature representation through the two-layer attention mechanism to obtain the second-scale feature representation process as follows:

[0068]

[0069]

[0070] Among them, ψ X The first scale feature of the source point cloud represents the result after the self-attention mechanism, Φ X Represents the second scale feature representation of the source point cloud, represents the self-attention mechanism, It represents the cross attention mechanism. The first-scale feature representation of the target point cloud goes through the two-layer attention mechanism in the same way as the source point cloud.

[0071] S4. randomly select K points in the source point cloud in the overlapping area of ​​the two groups of point clouds to form a first point set, and index the second scale feature representation corresponding to the first point set;

[0072] The training process randomly selects points in the overlapping area of ​​the source point cloud and the target point cloud for training, ensuring that the points of the source point cloud and the target point cloud can establish a corresponding relationship and can be coupled with the subsequent processes; after the training is completed, when using the model for alignment, the first K points with the highest confidence level in the source point cloud are selected to form the first point set.

[0073] S5. multiplying the second scale feature representation corresponding to the first point set with the second scale feature representation of the target point cloud to obtain a similarity matrix;

[0074] S6. Normalize the similarity matrix by row to obtain a corresponding distribution matrix between the first point set and the target point cloud, and multiply the corresponding distribution matrix by the target point cloud to obtain a first virtual point set matching the first point set; here, a corresponding loss function is set for the first virtual point set and the real corresponding point set;

[0075] S7. Select the maximum element in each row of the corresponding distribution matrix, normalize it, and obtain the weight of the matching point pair between the first point set and the first virtual point set;

[0076] S8. Based on the corresponding distribution matrix and the weights of the matching point pairs, the SVD algorithm is used to solve the rotation matrix and the translation matrix between the source point cloud and the target point cloud;

[0077] S9. Transform the source point cloud according to the rotation matrix and translation matrix, return to step S2, and iterate 3 times;

[0078] S10. Start the gradient back-propagation mechanism, optimize the loss function, update the network parameters, and save the model when the model converges or reaches the set number of epochs.

[0079] Specifically, the point cloud registration model is a bidirectional model, that is, when the source point cloud is registered to the target point cloud, the target point cloud is registered to the source point cloud using the same method.

[0080] In the process of point cloud registration, corresponding points are established based on the feature similarity of points to form a pair of point pairs. In this derivation process, it is inevitable that the wrong point pair relationship will be affected. Therefore, after the corresponding relationship is established for the K key points of the first point set, it cannot be guaranteed that they have all found the correct corresponding points. In order to alleviate this problem, the present invention further screens reliable point pair relationships based on the idea of ​​consistency, including:

[0081] The similarity matrix from the source point cloud to the target point cloud is calculated based on the first scale feature representation, which is denoted as M1. The similarity matrix from the source point cloud to the target point cloud is calculated based on the second scale feature representation, which is denoted as M2.

[0082] Multiply M1 and M2 bit by bit to get a new similarity matrix and normalize it row by row to get the corresponding point set of the source point cloud in the target point cloud;

[0083] Select the maximum element of each row in the new similarity matrix, align and normalize, and obtain the weight of the matching point pair between the source point cloud and the corresponding point set;

[0084] Based on the new similarity matrix and the weights of the matching point pairs, the SVD algorithm is used to solve the rotation matrix and translation matrix between the source point cloud and the target point cloud; the final transformation matrix is ​​calculated.

[0085] Specifically, the fusion of the new similarity matrix is ​​expressed as:

[0086]

[0087] in, represents the M1 matrix, represents the M2 matrix, ⊙ represents the bitwise multiplication of matrix elements, It represents the feature representation of the source point cloud obtained by indexing the first point set at the first scale feature representation. Represents the first-scale feature representation of N points in the target point cloud, Represents the feature representation obtained by indexing the first point set in the second scale feature representation of the source point cloud, Represents the second-scale feature representation of N points in the target point cloud. Based on the consistency of feature similarity, the inconsistent distribution of the two matrices will lead to a decrease in their normalized similarity scores. Therefore, in order to eliminate the impact of this unreliable correspondence on reasoning, reselect C XY The largest element in each row is renormalized and then calculated according to C XY The certainty of each row gives the weight of the contribution of the matching point pair to the calculation:

[0088]

[0089] Among them, X K represents the point set obtained after downsampling of the source point cloud, Y N represents the target point cloud, is X in SVD K To Y N Establish the corresponding point pair weights, is a function with a value of 0 or 1, using Y K represents the point set obtained after sampling the target point cloud, X M Represents the source point cloud, represented by Y K To X M The weights are calculated in the same way.

[0090] In one embodiment, four corresponding loss functions are set, and the sum of them constitutes the total loss function of the present invention, which is expressed as:

[0091]

[0092] in, represents the confidence loss of the source point cloud, represents the confidence loss of the target point cloud, Represents the error between the virtual point and the real point in the target point cloud corresponding to the source point cloud, Indicates the error between the virtual point in the source point cloud corresponding to the target point cloud and the real point. In this embodiment, this bidirectional loss function is used to train the point cloud registration model, which not only expands the training samples but also further improves the registration accuracy of the model.

[0093] Specifically, the confidence loss of the source point cloud is calculated using the typical cross entropy loss, and its formula is:

[0094]

[0095] Among them, GT(x i ) represents point x i The true label, P(x i ) represents the point x in the source point cloud i confidence level.

[0096] Specifically, the error between the virtual point in the target point cloud corresponding to the source point cloud and the real point It is expressed as:

[0097]

[0098] Among them, x i represents a point in the source point cloud, Represented as x i The generated virtual corresponding points (R XY x i +t XY ) represents the target point cloud with x i The corresponding real point.

[0099] The present invention proposes a point cloud registration system based on multi-scale feature similarity constraints, comprising:

[0100] A data collection module is used to obtain source point cloud and target point cloud data input into the model;

[0101] Downsampling module, used to filter out the point set involved in registration;

[0102] The feature exchange module is used to extract the features of each point in the source point cloud and the target point cloud, and then exchange the feature information of the two sets of point clouds;

[0103] The MLP module is used to extract deep features of two sets of point clouds using multiple MLPs based on the output of the feature exchange module;

[0104] The pooling module is used to perform maximum pooling on the deep features of the two groups of point clouds to obtain the global features of the source point cloud and the global features of the target point cloud;

[0105] The splicing module is used to fuse the feature information of points and splice the global features of the source point cloud and the target point cloud with the deep features of each point;

[0106] The confidence calculation module is used to input the concatenated features of each point into the fully connected layer to obtain the confidence of each point;

[0107] A local structure feature acquisition module is used to obtain the first scale feature representation of the source point cloud and the target point cloud;

[0108] The attention module is used to obtain the correlation information between point clouds and enhance the feature representation of point clouds, that is, the first-scale feature representation is processed according to the self-attention mechanism and the cross-attention mechanism to obtain the second-scale feature representation of the two groups of point clouds;

[0109] The candidate point set module is used to select K points with the highest confidence in the source point cloud to form the first point set according to the output result of the confidence calculation module;

[0110] An indexing module, used to filter the different scale feature representations of the first point set in the registration process;

[0111] A matching module, used for obtaining a corresponding point set matching the first point set in the target point cloud;

[0112] A consistency constraint module is used to introduce consistency constraints into the point cloud registration model to obtain a corresponding distribution matrix between the first point set and the target point cloud;

[0113] SVD solving module, used to calculate the rotation matrix and translation matrix between the source point cloud and the target point cloud according to the results of the matching module and the weights obtained by the consistency constraint module;

[0114] The transformation module is used to transform the source point cloud according to the rotation matrix and the translation matrix.

[0115] In one embodiment, a data set (ModelNet40) is used for experiments. The data set (ModelNet40) is a synthetic data set that is widely used for registration tasks. The data set contains 12311 CAD models, 9843 of which are usually used as training sets and 2468 as test sets. To obtain the two point clouds in the registration, it is necessary to downsample and simulate rotation and translation for sample preparation. Each original sample is recorded as a source point cloud. In order to obtain the target point cloud, three Euler angles are randomly generated along the X / Y / Z axes within [0,45°], and the translation parameters are set between [-0.5,0.5]. The rotation and translation are applied to the source point cloud to obtain the target point cloud. In order to obtain partial point cloud pairs, two random sampling points are set, and 768 points are retained respectively. The experiment is divided into three groups. The first group: trained on all categories and then tested in the test set; the second group: trained on the first 20 categories and tested in the last 20 categories; the third group: trained in noisy data and then tested in the test set. The root mean square error (RMSE) and mean absolute error (MAE) are used to measure the quality of registration. Table 1 shows the experimental results of the second set of tests using different methods.

[0116] Table 1 Test results on unseen categories

[0117] Method RMSE(R) MAE(R) RMSE(t) MAE(t) ICP 37.9513 26.7103 0.29395 0.2507 DCP 9.5585 6.9228 0.0461 0.0348 PRNet 6.0438 3.6295 0.0425 0.0291 GNN+IDAM 37.5242 17.9965 0.1028 0.0549

[0118] It can be seen from Table 1 that the errors of the rotation angle (R) and translation parameter (t) obtained by this patent are greatly reduced through testing with different algorithms. Figure 2 As shown in the figure, the first column is the two input point clouds, the second column is the registration result based on the consistency constraint of the present invention, the third column is the actual registration result, and the last column represents the high-weight correspondence relationship established under the consistency constraint. This result shows that the present invention can effectively screen out more reliable point pairs based on the feature similarity consistency constraint. Figure 3 is the visualization of the registration results in the three sets of experiments, Figure 2 , 3 The results show that the present invention is effective.

[0119] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A point cloud registration method based on multi-scale feature similarity constraints, characterized in that: Build and train a point cloud registration model, input the data to be registered into the trained point cloud registration model, introduce multi-scale consistency constraints to obtain point pair weights, and then calculate the registration results; The training process of the point cloud registration model includes: S1. Obtain source point cloud and target point cloud; S2. Extract the original features of each point in the two groups of point clouds, exchange the original features of the two groups of point clouds and obtain the deep features of each point through MLP, perform maximum pooling on the deep features to obtain the global features of the two groups of point clouds, concatenate the global features of the two groups of point clouds with the deep features of each point, and output the confidence of each point through a fully connected network; here, set the source point cloud confidence loss function; S3. Use GNN to extract the local structural features of the two groups of point clouds, record the GNN output features as the first-scale feature representation, and use the attention mechanism to obtain the final feature representations of the two groups of point clouds after enhancement, which are recorded as the second-scale feature representation; S4. randomly select K points in the source point cloud in the overlapping area of ​​the two groups of point clouds to form a first point set, and index the second scale feature representation corresponding to the first point set; S5. multiplying the second scale feature representation corresponding to the first point set with the second scale feature representation of the target point cloud to obtain a similarity matrix; S6. Normalize the similarity matrix by row to obtain a corresponding distribution matrix between the first point set and the target point cloud, and multiply the corresponding distribution matrix by the target point cloud to obtain a first virtual point set matching the first point set; here, a corresponding loss function is set for the first virtual point set and the real corresponding point set; S7. Select the maximum element in each row of the corresponding distribution matrix, normalize it, and obtain the weight of the matching point pair between the first point set and the first virtual point set; S8. Based on the corresponding distribution matrix and the weights of the matching point pairs, the SVD algorithm is used to solve the rotation matrix and the translation matrix between the source point cloud and the target point cloud; S9. Transform the source point cloud according to the rotation matrix and translation matrix, return to step S2, and iterate 3 times; S10. Start the gradient back-propagation mechanism, optimize the loss function, update the network parameters, and save the model when the model converges or reaches the set number of epochs.

2. The point cloud registration method based on multi-scale feature similarity constraint according to claim 1, characterized in that: The consistency constraints introduced into the trained point cloud registration model include: The similarity matrix from the source point cloud to the target point cloud is calculated based on the first scale feature representation, which is denoted as M1. The similarity matrix from the source point cloud to the target point cloud is calculated based on the second scale feature representation, which is denoted as M2. Multiply M1 and M2 bit by bit to get a new similarity matrix and normalize it row by row to get the corresponding point set of the source point cloud in the target point cloud; Select the maximum element of each row in the new similarity matrix, normalize it, and obtain the weight of the matching point pair between the source point cloud and the corresponding point set; Based on the new similarity matrix and the weights of the matching point pairs, the SVD algorithm is used to solve the rotation matrix and translation matrix between the source point cloud and the target point cloud, and the final transformation matrix is ​​calculated.

3. The point cloud registration method based on multi-scale feature similarity constraint according to claim 1, characterized in that: A bidirectional loss function is used in the training phase to optimize both the registration process from the source point cloud to the target point cloud and the registration process from the target point cloud to the source point cloud.

4. The point cloud registration method based on multi-scale feature similarity constraint according to claim 3, characterized in that: The bidirectional loss function of the point cloud registration model training phase is expressed as: in, represents the confidence loss of the source point cloud, represents the confidence loss of the target point cloud, represents the loss between the virtual points and the real points in the target point cloud corresponding to the source point cloud, Represents the loss between the virtual points in the source point cloud and the real points corresponding to the target point cloud.

5. The point cloud registration method based on multi-scale feature similarity constraint according to claim 4, characterized in that: The confidence loss of the source point cloud is calculated using the typical cross entropy loss, and its formula is: Among them, GT(x i ) represents point x i The true label, P(x i ) represents the point x in the source point cloud i confidence level.

6. The point cloud registration method based on multi-scale feature similarity constraint according to claim 4, characterized in that: The loss between the virtual points and the real points in the target point cloud corresponding to the source point cloud It is expressed as: Among them, x i represents a point in the source point cloud, Represented as x i The generated virtual corresponding points (R XY x i +t XY ) represents the target point cloud with x i The corresponding real point.

7. A point cloud registration system based on multi-scale feature similarity constraints, characterized in that: include: A data collection module is used to obtain source point cloud and target point cloud data input into the model; Downsampling module, used to filter out the point set involved in registration; The feature exchange module is used to extract the features of each point in the source point cloud and the target point cloud, and then exchange the feature information of the two sets of point clouds; The MLP module is used to extract deep features of two sets of point clouds using multiple MLPs based on the output of the feature exchange module; The pooling module is used to perform maximum pooling on the deep features of the two groups of point clouds to obtain the global features of the source point cloud and the global features of the target point cloud; The splicing module is used to fuse the feature information of points and splice the global features of the source point cloud and the target point cloud with the deep features of each point; The confidence calculation module is used to input the concatenated features of each point into the fully connected layer to obtain the confidence of each point; A local structure feature acquisition module is used to obtain the first scale feature representation of the source point cloud and the target point cloud; The attention module is used to obtain the correlation information between point clouds and enhance the feature representation of point clouds, that is, the first-scale feature representation is processed according to the self-attention mechanism and the cross-attention mechanism to obtain the second-scale feature representation of the two groups of point clouds; The candidate point set module is used to select K points with the highest confidence in the source point cloud to form the first point set according to the output result of the confidence calculation module; An indexing module, used to filter the different scale feature representations of the first point set in the registration process; A matching module, used for obtaining a corresponding point set matching the first point set in the target point cloud; A consistency constraint module is used to introduce consistency constraints into the point cloud registration model to obtain a corresponding distribution matrix between the first point set and the target point cloud; SVD solving module, used to calculate the rotation matrix and translation matrix between the source point cloud and the target point cloud according to the results of the matching module and the weights obtained by the consistency constraint module; The transformation module is used to transform the source point cloud according to the rotation matrix and the translation matrix.

Images