A point cloud registration method based on position-enhanced attention mechanism

Abstract

The application discloses a point cloud registration method based on a position enhanced attention mechanism, which comprises the following steps: firstly, multi-scale features of a source point cloud and a target point cloud are extracted; secondly, position information of the source point cloud and the target point cloud is respectively extracted, and context feature information of the source point cloud and the target point cloud is learned from the multi-scale features; the context feature information is fused with the position information to obtain fused feature information; the fused feature information of the source point cloud and the target point cloud is subjected to feature information interaction to generate mixed features; then, an outlier parameter is generated according to an alignment state of the source point cloud and the target point cloud; the outlier parameter and the mixed features are used to obtain a corresponding relationship of all points in the source point cloud and the target point cloud; finally, a singular value decomposition method is used to obtain a rotation matrix and a translation matrix between the source point cloud and the target point cloud, and iterative solving is performed to obtain a final rotation matrix and a final translation matrix, so that the registration process is completed.

Description

Technical Field

[0001] This invention relates to the field of 3D point cloud registration in deep learning and computer vision, and particularly to a 3D point cloud registration method based on a position-enhanced attention mechanism. Technical Background

[0002] Point cloud registration is the process of transforming point clouds acquired by a point cloud scanning device from different viewpoints to the same coordinate system through rotation and translation. It is widely used in pose estimation, 3D reconstruction, and mobile robotics. However, point cloud registration faces certain challenges in practical applications, mainly because: 1) point clouds scanned from different viewpoints contain noise and partially invisible elements; 2) point clouds are disordered and sparse. Therefore, improving the accuracy and robustness of the algorithm is indispensable in practical point cloud registration tasks.

[0003] Based on the data transformation method, existing point cloud registration methods can be divided into voxel-based, multi-view, and point cloud-based methods. Since the first two methods result in information loss, point cloud-based methods are currently the most widely used. The PointNet algorithm solved the problems of point cloud disorder and rotation invariance. Subsequently, the processing of point cloud registration problems was extended to deep learning-based methods. Compared with the traditional iterative nearest-point algorithm, deep learning-based methods solve the problem of getting trapped in local optima. However, when the point cloud contains noise or some points are missing, deep learning-based methods still fail to maintain good registration results.

[0004] Traditional point cloud registration methods, such as the Iterative Closest Point (ICP) algorithm, are prone to getting trapped in local optima. Deep learning-based methods, such as Robust Point Matching using Learned Features (Rpm-Net), can achieve efficient registration even with some noise and missing points. However, when the missing portion of the point cloud reaches 30%, the registration performance deteriorates. This is because current deep learning-based registration methods only focus on local geometric features of the point cloud, lacking a global understanding and failing to consider the feature interactions between the source and target point clouds, as well as the positional information of corresponding points. This makes the learned features less discriminative, resulting in numerous erroneous point correspondences.

[0005] Therefore, in scenarios where point clouds are noisy and the missing portion of the point cloud reaches 30% or even higher, how to improve the registration effect is a problem that urgently needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a point cloud registration method based on a location-enhanced attention mechanism, which enables the network to learn the contextual information between each point cloud while combining it with the geometric structure of the point cloud to obtain more geometrically relevant features and improve registration performance.

[0007] To achieve the above objectives, the present invention employs the following technical solution:

[0008] A point cloud registration method based on a location-enhanced attention mechanism includes:

[0009] First, the source point cloud and target point cloud data are input into the adaptive graph convolution feature extraction module to extract multi-scale features of the source point cloud and target point cloud respectively.

[0010] Secondly, the multi-scale features of the source point cloud and the target point cloud are input into the position enhancement attention mechanism module to extract the position information of the source point cloud and the target point cloud respectively, and the contextual feature information of the source point cloud and the target point cloud is learned from the multi-scale features. The contextual feature information is fused with the position information to obtain fused feature information. The fused feature information of the source point cloud and the target point cloud is then interacted to generate hybrid features.

[0011] Then, the alignment state of the source point cloud and the target point cloud is obtained based on their three-dimensional coordinates. The alignment state of the source point cloud and the target point cloud is input into the outlier parameter module to generate outlier parameters. The outlier parameters and the hybrid features are then input into the similarity matching module to obtain the correspondence between all points in the source point cloud and the target point cloud.

[0012] Finally, the singular value decomposition method is used to obtain the rotation and translation matrices between the source and target point clouds, and the process is iteratively solved to obtain the final rotation and translation matrices, thus completing the registration process.

[0013] Furthermore, the iterative solution to obtain the final rotation and translation matrices includes:

[0014] The loss function between the obtained rotation and translation matrices and the true rotation and translation matrices provided by the dataset itself is calculated. If the loss function does not converge, the source point cloud is multiplied by the rotation matrix and summed with the translation matrix to obtain a new source point cloud, and a new round of iterative registration is continued. If the loss function converges, the rotation and translation matrices are output.

[0015] Furthermore, the processing procedure of the location-enhanced attention mechanism module is as follows:

[0016] Contextual information is extracted from the multi-scale features of the source point cloud. Simultaneously, the location information of the source point cloud is directly extracted and concatenated with the contextual information. The concatenated features and location information are then input into the self-attention module in the location-enhanced attention module to interact with the contextual features and location information, thus obtaining the fused feature information of the source point cloud. The same process is repeated for the target point cloud. Finally, the fused feature information of the source point cloud and the target point cloud is input into the cross-attention module to perform feature interaction between the source point cloud and the target point cloud, ultimately outputting the hybrid features.

[0017] Furthermore, the location information is encoded by the distance and normal vector coordinates between points; any two points in the source point cloud The spatial distance between them is calculated as follows: For normal vector information, a linear layer is directly used to encode the normal vectors of points in the source point cloud in space; the spatial distance information of the point is concatenated with the encoded normal vector information to obtain the position information of the point.

[0018] Furthermore, the calculation process of the location-enhanced attention mechanism module is as follows:

[0019] Q = F·W Q K = F·W K V = F·W V

[0020]

[0021]

[0022] Among them, the multi-scale characteristics of the source point cloud are The fusion feature information of the source point cloud is J represents the number of points in the source point cloud, and d represents the number of feature dimensions. Let S represent the set of real numbers. j The attention weights are represented by the three projection matrices Q, K, and V of the source point cloud input features; j represents the index of a point, and the superscript T indicates the matrix transpose. It is the location information of the point, W Q W K ,and These are the learnable parameters, MLP represents multilayer perceptron, softmax(.) represents row softmax, and cat[·,·] represents concatenation.

[0023] Furthermore, the step of inputting the alignment status of the source point cloud and the target point cloud into the outlier parameter module to generate outlier parameters includes:

[0024] Using a parameter prediction network, the unaligned point cloud is used as input to predict the parameters of the current iteration. First, the source point cloud and the target point cloud are concatenated into a matrix (B,3,J+K), where B is the batch size, J and K are the number of source point clouds and target point clouds, respectively, and 3 represents 3D coordinates.

[0025] To characterize which point cloud a point comes from, a fourth feature column is added, where 0 indicates that the point comes from the source point cloud and 1 indicates that the point comes from the target point cloud. Therefore, the input data of the parameter prediction module is (B, 4, J+K) dimensional. After passing through a multilayer perceptron and a max pooling layer, the outlier parameters α and β are finally obtained.

[0026] Furthermore, the step of inputting the outlier parameters and the mixed features together into the similarity matching module to obtain the correspondence between all points in the source point cloud and the target point cloud includes:

[0027] Outlier parameters α and β, along with the mixed features, are input into the similarity matching module to construct a matching matrix M; each element m of the matching matrix... jk ∈M is initialized as follows:

[0028]

[0029] Where Fx j ,Fy k These are the mixed features of the source point cloud and the target point cloud, respectively.

[0030] Then, alternating row and column normalization is performed on the matching matrix. By repeatedly applying alternating row and column normalization, a double random matrix can be obtained from any square matrix with all positive terms, thus obtaining the correspondence between all points in the source point cloud and the target point cloud.

[0031] Furthermore, for each point p in the source point cloud j It can calculate the corresponding points in the target point cloud.

[0032]

[0033] Where k represents the index of a point in the target point cloud, K represents the number of points in the target point cloud, j represents the index of a point in the source point cloud, and q k This represents the k-th point in the target point cloud.

[0034] Furthermore, the singular value decomposition method is used to obtain the rotation and translation matrices between the source and target point clouds, and the results are iteratively solved to obtain the final rotation and translation matrices, including:

[0035] The corresponding points in the target point cloud Substituting into the following equation and using singular value decomposition to solve for the rotation matrix R in the nth iteration(n) Translation matrix t (n) :

[0036]

[0037] Where n represents the number of iterations, J represents the number of points in the source point cloud, and R and t represent the rotation and translation matrices obtained in the current iteration;

[0038] Obtain the rotation matrix R (n) Translation matrix t (n) Then, iterate over the current source point cloud to obtain a new source point cloud, and repeat the iteration process until the loss function converges and the final rotation matrix R is obtained. (*) Translation matrix t (*) .

[0039] Furthermore, the loss function is defined as the true transformation matrix {R} of the source point cloud. gt ,t gt The prediction transformation matrix {R} of the source point cloud (*) ,t (*) The L1 distance between} is calculated using the loss function as follows:

[0040]

[0041] Where J is the number of points in the source point cloud, j is the index of a point in the source point cloud, and x j This represents the j-th point in the source point cloud.

[0042] Compared with the prior art, the present invention has the following technical features:

[0043] In scenarios where point clouds are noisy and some point clouds are missing, traditional algorithms and learning-based algorithms are unable to register points effectively. This invention proposes an end-to-end point cloud registration scheme: local geometric features are extracted from a pair of original point clouds and concatenated into multi-scale features. These features, along with point distance and normal information, are input into a position-enhanced attention mechanism to obtain hybrid features. Point correspondences are then calculated based on these hybrid features, and finally, singular value decomposition is used to obtain the final rotation and translation matrices. Compared to traditional registration methods and deep learning-based registration methods, this invention achieves better registration results in scenarios with noisy point clouds and missing point clouds. The proposed position-enhanced attention mechanism improves the registration network's understanding of positional information and the discriminative power of learned features, thereby enhancing registration performance. Attached Figure Description

[0044] Figure 1 This is a network diagram illustrating the method of the present invention;

[0045] Figure 2A schematic diagram of the location-enhanced attention module;

[0046] Figure 3 This is a flowchart illustrating the method of the present invention.

[0047] Figure 4 This is a diagram showing the registration effect of the method of the present invention in an embodiment. Detailed Implementation

[0048] This invention proposes a point cloud registration method based on a location-enhanced attention mechanism, the structure of which is as follows: Figure 1 As shown:

[0049] First, the source point cloud and target point cloud data are input into the adaptive graph convolution feature extraction module to extract multi-scale features of the source point cloud and target point cloud respectively.

[0050] Secondly, the multi-scale features of the source point cloud and the target point cloud are input into the position enhancement attention mechanism module to extract the position information of the source point cloud and the target point cloud respectively, and the contextual feature information of the source point cloud and the target point cloud is learned from the multi-scale features. The contextual feature information is fused with the position information to obtain fused feature information. The fused feature information of the source point cloud and the target point cloud is then interacted to generate hybrid features.

[0051] Then, the alignment state of the source point cloud and the target point cloud is obtained based on their three-dimensional coordinates. The alignment state of the source point cloud and the target point cloud is input into the outlier parameter module to generate outlier parameters. The outlier parameters and the hybrid features are then input into the similarity matching module to obtain the correspondence between all points in the source point cloud and the target point cloud.

[0052] Finally, the singular value decomposition method is used to obtain the rotation and translation matrices between the source and target point clouds, and the process is iteratively solved to obtain the final rotation and translation matrices, thus completing the registration process.

[0053] The process involves iterative solving to obtain the final rotation and translation matrices, specifically:

[0054] The loss function between the obtained rotation and translation matrices and the true rotation and translation matrices provided by the dataset itself is calculated. If the loss function does not converge, the source point cloud is multiplied by the rotation matrix and summed with the translation matrix to obtain a new source point cloud, and a new round of iterative registration is continued. If the loss function converges, the rotation and translation matrices are output.

[0055] The adaptive graph convolution proposed by Wei M et al. overcomes the shortcomings of the fixed kernel in standard graph convolution by adaptively establishing relationships between pairs of points based on feature attributes, thereby generating an adaptive kernel that can more effectively extract different relationships between points in different parts of the point cloud. However, it lacks a global understanding. Therefore, this invention inputs the point cloud into an adaptive graph convolution feature extraction module consisting of four adaptive graph convolution layers (64, 64, 128, 256) for multi-level feature extraction, and then performs a concatenation operation to generate multi-scale features.

[0056] For the multi-scale features of the source point cloud and the target point cloud, the processing procedure of the location-enhanced attention mechanism module is as follows: First, contextual information is extracted from the multi-scale features of the source point cloud. Simultaneously, the location information of the source point cloud is directly extracted and concatenated with the contextual information. The concatenated features and location information are then input into the self-attention module within the location-enhanced attention module for interaction between the contextual features and location information, thus obtaining the fused feature information of the source point cloud. The same process is repeated for the target point cloud. Finally, the fused feature information of the source and target point clouds is input into the cross-attention module for feature interaction between the source and target point clouds, ultimately outputting a hybrid feature, the structure of which is as follows: Figure 2 As shown. The position-enhanced attention mechanism in this invention mainly embeds the position information of the point cloud into the calculation of the attention module. This helps the model learn the spatial structure between points in the point cloud and allows the model to pay more attention to the information of specific regions when processing the point cloud, thereby improving the ability to perceive key points and reducing erroneous point correspondences between the source point cloud and the target point cloud.

[0057] Location information is primarily encoded by the distance and normal vector coordinates between points. This invention calculates the Euclidean distance between points, i.e., the distance between any two points in the source point cloud. The spatial distance between them is calculated as follows: For normal vector information, a linear layer is directly used to encode the spatial normal vectors of points in the source point cloud. The spatial distance information of the point is concatenated with the encoded normal vector information to obtain the point's position information, thus encoding the point's position information.

[0058] The following section describes the computation process of the source point cloud in the self-attention module.

[0059] Given an input feature matrix (multi-scale features of the source point cloud). Output feature matrix (fused feature information of source point clouds) (J represents the number of points in the source point cloud, and d represents the number of feature dimensions) S represents the set of real numbers, which is the weighted sum of projections of all input features F. j To represent the attention weights, first calculate Q,K,V (the projection matrix of the source point cloud input features):

[0060] Q = F·W Q K = F·W K V = F·W V

[0061]

[0062]

[0063] Here, J represents the number of points in the source point cloud, j represents the index of the point, and the superscript T indicates matrix transpose. It is the location information of the point, W Q W K ,and These are the learnable parameters, obtained through training on the dataset. MLP stands for Multilayer Perceptron, softmax(.) represents row softmax, and cat[·,·] represents concatenation.

[0064] A cross-attention module was added after the self-attention module, which can effectively exchange feature information between the source point cloud and the target point cloud, and finally obtain hybrid features.

[0065] The parameter prediction network proposed by Yew ZJ et al. can select appropriate outlier parameters based on the alignment status of the current point cloud. Therefore, using the parameter prediction network, the misaligned point cloud is taken as input to predict the parameters of the current iteration. First, the source and target point clouds are concatenated into a (B, 3, J+K) matrix (B is the batch size, J and K are the number of source and target point clouds respectively, and 3 represents 3D coordinates). To characterize which point cloud a point originates from, a fourth feature column is added: 0 indicates that the point comes from the source point cloud, and 1 indicates that the point comes from the target point cloud. Therefore, the input data of the parameter prediction module is (B, 4, J+K) dimensional. After passing through a multilayer perceptron and a max pooling layer, the final data obtained is (B, 2), which are the outlier parameters α and β.

[0066] The outlier parameters α and β, along with the mixed features, are then input into the similarity matching module to construct a matching matrix M; each element m of the matching matrix... jk ∈M is initialized as follows:

[0067]

[0068] Where Fx j ,Fy k These are the mixed features of the source point cloud and the target point cloud, respectively.

[0069] Then, alternating row and column normalization is performed on the matching matrix. By repeatedly applying alternating row and column normalization, a double random matrix can be obtained from any square matrix with all positive terms, thus obtaining the correspondence between all points in the source point cloud and the target point cloud:

[0070] For each point p in the source point cloud j It can calculate the corresponding points in the target point cloud.

[0071]

[0072] Where k represents the index of a point in the target point cloud, K represents the number of points in the target point cloud, j represents the index of a point in the source point cloud, and q k This represents the k-th point in the target point cloud.

[0073] Finally, singular value decomposition is used to solve for the rigid transformation:

[0074] The corresponding points in the target point cloud Substituting into the following equation and using singular value decomposition to solve for the rotation matrix R in the nth iteration (n) Translation matrix t (n) :

[0075]

[0076] Where n represents the number of iterations, J represents the number of points in the source point cloud, and R and t represent the rotation and translation matrices obtained in the current iteration.

[0077] Obtain the rotation matrix R (n) Translation matrix t (n) Then, iterate over the current source point cloud to obtain a new source point cloud, and repeat the above iterative process until the loss function converges and the final rotation matrix R is obtained. (*) Translation matrix t (*) .

[0078] Take the first three rows and first three columns of the true transformation matrix from the dataset used for network training; this is represented as the true rotation matrix R. gt Accordingly, the first three rows and last column of the true transformation matrix in the dataset are taken, and it is represented as the true translation matrix t. gt The loss function is defined as the true transformation matrix {R} of the source point cloud. gt ,t gt The prediction transformation matrix {R} of the source point cloud (*) ,t (*) The L1 distance between} is the Manhattan distance. The loss function is calculated as follows:

[0079]

[0080] Where J is the number of points in the source point cloud, j is the index of a point in the source point cloud, and x j This represents the j-th point in the source point cloud.

[0081] The point cloud registration results of this invention are as follows: The visualized registration results are as follows. Figure 4 (The left image shows the initial point cloud, and the right image shows the registration result.) It can be seen that the registration method proposed in this invention can achieve accurate registration in scenarios where there is noise in the point cloud and some point clouds are missing.

[0082] The computer used in this experiment was an Intel(R) Xeon(R) Bronze 3204 CPU@1.90GHz, an RTX3090 GPU, PyTorch version 1.13.0, and the Python programming language.

[0083] The method of this invention is compared with the traditional algorithm ICP (Iterative Closest Point) and the deep learning-based algorithm Rpm-Net (Robust Point Matching using LearnedFeatures), and the results are shown in Table 1.

[0084] All methods were tested on the ModelNet40 dataset. Noise was added and 30% of the point cloud was missing. Anisotropic mean absolute errors of rotation and translation matrices, MAE(r) and MAE(t), were used, as well as isotropic mean absolute errors of rotation and translation matrices, Error(r) and Error(t), respectively.

[0085] Table 1. Experimental registration results using different methods

[0086]

[0087] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for point cloud registration based on location-enhanced attention mechanism, characterized in that, Comprise: First, the source point cloud and target point cloud data are input into the adaptive graph convolution feature extraction module respectively, and the multi-scale features of the source point cloud and the target point cloud are extracted; Second, the multi-scale features of the source point cloud and the target point cloud are input into the position-enhanced attention mechanism module, the position information of the source point cloud and the target point cloud is extracted respectively, and the context feature information of the source point cloud and the target point cloud is learned from the multi-scale features, the context feature information and the position information are fused to obtain the fused feature information; The fused feature information of the source point cloud and the target point cloud is interacted to generate a mixed feature; Then, the alignment state of the source point cloud and the target point cloud is obtained according to their three-dimensional coordinates, the alignment state of the source point cloud and the target point cloud is input into the outlier parameter module to generate an outlier parameter, and the outlier parameter and the mixed feature are input into a similarity matching module to obtain the corresponding relationship of all points in the source point cloud and the target point cloud; Finally, the singular value decomposition method is used to obtain the rotation matrix and the translation matrix between the source point cloud and the target point cloud, and iterative solution is performed to obtain the final rotation matrix and translation matrix, and the registration process is completed. The calculation process of the position-enhanced attention mechanism module is: Q = F · W Q , K = F · W K , V = F · W V wherein the multi-scale feature of the source point cloud is the fusion feature information of the source point cloud is J is the number of points in the source point cloud, and d is the number of feature dimensions, denotes a real set, S j denotes an attention weight, three projection matrices of the source point cloud input feature are Q, K, and V; j denotes an index of a point in the source point cloud, and a superscript T denotes matrix transposition, is position information of the point, W Q , W K , and are learnable parameters, respectively, MLP denotes a multi-layer perceptron, softmax(.) denotes a row softmax, and cat[·,·] denotes concatenation; The outlier parameter and the mixed feature are input into the similarity matching module to obtain the corresponding relationship of all points in the source point cloud and the target point cloud, comprising: The abnormal value parameters α and β are input into the similarity matching module together with the mixed features to construct a matching matrix M; each element m of the matching matrix is initialized as follows: jk ∈Minitialized as follows: wherein Fx j Fy k are the mixed features of the source point cloud and the target point cloud, respectively; Then, the alternating row and column normalization is performed on the matching matrix, and by repeatedly applying the alternating row and column normalization, a doubly stochastic matrix can be obtained from any square matrix with all positive terms, thereby obtaining the corresponding relationship of all points in the source point cloud and the target point cloud.

2. The point cloud registration method based on location enhanced attention mechanism according to claim 1, characterized in that, The iterative solution is performed to obtain the final rotation matrix and translation matrix, comprising: The loss function between the obtained rotation matrix and translation matrix and the real rotation matrix and translation matrix provided by the data set itself is calculated, if the loss function does not converge, the source point cloud is multiplied by the rotation matrix and summed with the translation matrix to obtain a new source point cloud, and a new round of iterative registration is continued; if the loss function converges, the rotation matrix and the translation matrix are output. 3.The point cloud registration method based on location enhanced attention mechanism according to claim 1, characterized in that, The processing process of the position-enhanced attention mechanism module is: The context information is extracted from the multi-scale features of the source point cloud, at the same time, the position information of the source point cloud is directly extracted and spliced with the context information, then the spliced features and the position information are input into the self-attention module in the position-enhanced attention module to interact the context features and the position information, thereby obtaining the fused feature information of the source point cloud, and the target point cloud is the same; finally, the fused feature information of the source point cloud and the target point cloud is input into the cross-attention module to interact the features between the source point cloud and the target point cloud, and the mixed feature is finally output.

4. The point cloud registration method based on location enhanced attention mechanism according to claim 1, characterized in that, The location information is encoded by the distance between points and their normal vector coordinates; any two points in the source point cloud The spatial distance between them is calculated as follows: For normal vector information, a linear layer is directly used to encode the normal vectors of points in the source point cloud in space; the spatial distance information of the point is concatenated with the encoded normal vector information to obtain the position information of the point.

5. The point cloud registration method based on location enhanced attention mechanism according to claim 1, characterized in that, The alignment state of the source point cloud and the target point cloud is input into the outlier parameter module to generate an outlier parameter, comprising: Using a parameter prediction network, the misaligned point cloud is input to predict the parameters of the current iteration; first, the source point cloud and the target point cloud are spliced into a (B, 3, J+K) matrix, where B is the batch size, J and K are the number of source point clouds and target point clouds respectively, and 3 represents 3-dimensional coordinates; In order to characterize which point cloud a point comes from, the fourth column feature is added, 0 represents that a point comes from the source point cloud, and 1 represents that a point comes from the target point cloud; therefore, the input data of the parameter prediction module is (B, 4, J+K) dimensions, and finally the abnormal value parameters a and b are obtained through a plurality of multilayer perceptron and maximum pooling layers.

6. The method of claim 1, wherein, For each point p in the source point cloud j a corresponding point p' in the target point cloud can be computed wherein k represents an index of a point in the target point cloud, K represents a number of points in the target point cloud, j represents an index of a point in the source point cloud, q k represents the kth point in the target point cloud.

7. The method of claim 1, wherein, The singular value decomposition method is used to obtain the rotation matrix and the translation matrix between the source point cloud and the target point cloud, and iterative solving is performed to obtain the final rotation matrix and the translation matrix, including: The corresponding points in the target point cloud Substituting into the following equation and using singular value decomposition to solve for the rotation matrix R in the nth iteration (n) Translation matrix t (n) : Wherein, n represents the number of iterations, J represents the number of points of the source point cloud, R and t represent the rotation matrix and the translation matrix obtained in the current iteration; get rotation matrix R (n) and translation matrix t (n) After the current source point cloud iteration, repeat the iteration process to get the new source point cloud until the loss function converges to get the final rotation matrix R (*) and translation matrix t (*) .

8. The method of claim 1, wherein, The loss function is defined as the l1 distance between the true transformation matrix {R gt ,t gt} of the source point cloud and the predicted transformation matrix {R (*) ,t (*)} of the source point cloud, and the loss function is calculated as: where J is the number of points in the source point cloud, j is the index of a point in the source point cloud, x j denotes the jthpoint in the source point cloud.

Images