Non-rigid 3D point cloud registration method and system based on attention mechanism

Abstract

The present invention relates to a non-rigid three-dimensional point cloud registration method and system based on an attention mechanism. The three-dimensional point cloud registration method first uses an adaptive instance normalization module as the normalization layer in a standard Transformer network, and then obtains the original point cloud data of a non-rigid target under two different actions, respectively as the source point cloud and the target point cloud. Then, the high-dimensional point-by-point features of the source point cloud and the target point cloud are extracted, and superimposed with the position encoding to obtain the initial feature embeddings of the source point cloud and the target point cloud. Then, a multi-head cross-attention module is used to perform linear matrix operations, and multiple attention scores are spliced to obtain an intermediate feature embedding of the source point cloud that maps the target point cloud information. Then, the adaptive instance normalization module is used to normalize the intermediate feature embedding, so that the deformed point cloud and the source point cloud are as consistent as possible in terms of identity features, and the surface is smoother, further improving the registration effect.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a non-rigid three-dimensional point cloud registration method and system based on an attention mechanism. Background Art

[0002] Due to the limited scanning range of 3D scanning equipment, the diverse shapes of objects to be scanned, and the influence of the environment, it is often impossible to obtain a complete data model of the object with a single scan. Multiple point clouds must usually be captured from different angles, each associated with a different coordinate system, and may be subject to translational and rotational misalignments. By calculating an appropriate coordinate transformation, the point cloud data scanned from different perspectives can be unified into the same coordinate system to form a complete data model. Alternatively, a transformation matrix can be calculated to unify two point clouds of exactly the same shape and size into the same coordinate system. The key technology to achieve this is point cloud registration.

[0003] Point cloud registration can be classified into rigid registration and non-rigid registration according to the object. The transformation parameters of rigid registration are usually represented by a low-dimensional transformation matrix, which performs global rotation and translation coordinate transformation. The form of non-rigid registration transformation is more complex, involving local rotation and translation and non-rigid deformation, and cannot be described using simple transformation parameters. In practical applications, non-rigid point cloud registration faces many difficulties: First, due to the movement of the scanned target or the difference in scanning angles, the overlapping part of the two point clouds to be registered is unknown. In large-scale deformation, the difference between the two point clouds is greater and the overlapping part is less. Second, unlike images, point clouds have no color features, only spatial coordinates. Due to these reasons, the existing non-rigid point cloud registration methods have problems such as mismatching, insufficient registration effect, and high algorithm overhead in large-scale deformation, which greatly limits their use in industrial automation.

[0004] Current non-rigid point cloud registration methods based on traditional methods typically use traditional feature descriptors to calculate the features of the input point cloud. Reliable point-to-point correspondences are established by comparing the feature similarities between the original and target point clouds, resulting in a point-by-point rotation and translation transformation matrix. However, these registration methods are prone to generating erroneous point pair correspondences when faced with large deformations, thus affecting the registration results. Furthermore, in recent years, deep learning-based non-rigid point cloud registration methods have emerged, such as the Coherent Point Drift Network (CPD-Net) and the Recurrent Multi-view Alignment Network (RMAN). These methods share a similar overall concept to traditional methods, typically comprising a feature extraction component and a correspondence search component. However, the greater degrees of freedom in non-rigid registration increases the difficulty of network training, and the lack of labeled data further limits network training, making current deep learning-based non-rigid point cloud registration methods applicable only to small-scale non-rigid deformations. Furthermore, due to the disordered nature of point clouds, initial deep learning methods addressed the point cloud registration problem by converting point clouds into voxels. This type of method divides the point cloud into multiple cube grids, but the gridded voxel data has great redundancy, which makes this method computationally intensive and occupies a lot of space. Summary of the Invention

[0005] Based on this, it is necessary to address the technical problem in the existing technology that the target three-dimensional point cloud registration effect under large-scale non-rigid deformation is poor. The present invention provides a non-rigid three-dimensional point cloud registration method and system based on the attention mechanism.

[0006] The present invention discloses a non-rigid three-dimensional point cloud registration method based on an attention mechanism, comprising the following steps:

[0007] S1. Use at least one set of adaptive instance normalization modules as normalization layers in a standard Transformer network to obtain a modified Transformer network for performing non-rigid 3D point cloud registration tasks. The modified Transformer network also includes a multi-head crisscross attention module.

[0008] S2. Obtain the original point cloud data of a non-rigid target under two different actions, as the source point cloud S and target point cloud T respectively.

[0009] S3. Extract high-dimensional point-by-point features of the source point cloud and the target point cloud, and superimpose the output features with the position encoding in the improved Transformer network to obtain the initial feature embedding X of the source point cloud and the target point cloud S and X T .

[0010] S4. Using the linear transformation matrix W in the multi-head crisscross attention module Q 、W K 、W V Perform matrix operations on the preliminary feature embeddings of the source point cloud and the target point cloud to obtain the linear transformation result matrices Q, K, V and calculate the attention scores.

[0011] S5. By concatenating multiple attention score matrices, we obtain the intermediate feature embedding Z of the source point cloud that maps the target point cloud information.

[0012] S6. Use the adaptive instance normalization module to normalize the intermediate feature embedding Z of the source point cloud, thereby keeping the posture of the source point cloud close to the target point cloud without changing the original identity information, and obtaining the normalized feature embedding Z′ of the source point cloud.

[0013] S7. Set the loss function and use S4 to S6 as the decoder part of the improved Transformer network to embed the initial features of the source point cloud into X s The normalized feature embedding Z′ is used as the input of the encoder, and after a preset number of cycles, the output optimized feature embedding Z″ is obtained.

[0014] S8. Use a multi-layer perceptron structure to convolve and activate the optimized feature embedding Z″, and use the following formula to calculate the coordinates of the resulting point cloud to complete the point cloud registration. The calculation formula is as follows:

[0015] R = 2*tanh[Conv(Z″)]

[0016] Where R represents the resulting point cloud and tanh is the hyperbolic tangent activation function.

[0017] As a further improvement of the above scheme, in S2, the original point cloud data is sampled using an equal downsampling method so that the number of points in the source point cloud and the target point cloud is the same.

[0018] As a further improvement of the above scheme, in S3, a dynamic graph convolutional neural network is used to extract high-dimensional point-by-point features of the source point cloud and the target point cloud.

[0019] As a further improvement to the above solution, in S4, the calculation formula of the attention score is:

[0020]

[0021] Among them, d k is the number of columns of matrix Q.

[0022] As a further improvement of the above scheme, the number of adaptive instance normalization modules is set to 3 groups.

[0023] As a further improvement of the above scheme, in S6, residual connection is introduced to normalize the source point cloud feature embedding.

[0024] As a further improvement to the above scheme, in S6, the expression formula of the normalized feature embedding Z′ of the source point cloud is:

[0025] Z′=[Conv(AdaIN(Conv(AdaIN(Z))))+Conv(AdaIN(Z))]

[0026] Where Conv(·) represents the convolution processing of ·. AdaIN(·) represents the adaptive instance normalization processing of ·.

[0027] As a further improvement of the above scheme, in S7, the point-by-point grid Euclidean distance PMD is used as the loss function:

[0028]

[0029] Where M i and M′ i They represent the three-dimensional coordinates of the i-th point of the target point cloud and the result point cloud respectively, i = 1, 2, 3...n, and n is the total number of points in the target point cloud.

[0030] As a further improvement of the above scheme, in S2, a lidar scanner or a Kinect camera is used to capture different actions of the same non-rigid target at different time points.

[0031] The present invention also discloses a non-rigid 3D point cloud registration system based on an attention mechanism, which adopts any of the above-mentioned non-rigid 3D point cloud registration methods based on an attention mechanism. The registration system includes: a network construction module, a point cloud acquisition module, and a point cloud feature extraction module.

[0032] A network building module is configured to use at least one set of adaptive instance normalization modules as normalization layers in a standard Transformer network to obtain a modified Transformer network for performing non-rigid 3D point cloud registration tasks. The modified Transformer network also includes a multi-head crisscross attention module.

[0033] The point cloud acquisition module is used to obtain the original point cloud data of a non-rigid target under two different actions, which are used as the source point cloud S and the target point cloud T respectively.

[0034] The point cloud feature extraction module is used to extract high-dimensional point-by-point features of the source point cloud and the target point cloud, and superimpose the output features with the position encoding in the improved Transformer network to obtain the initial feature embedding X of the source point cloud and the target point cloud. S and X T .

[0035] Among them, the linear transformation matrix W in the multi-head cross attention module Q 、W K 、W V Matrix operations are performed on the preliminary feature embeddings of the source and target point clouds to obtain the linear transformation result matrices Q, K, and V, and the attention scores are calculated. By splicing multiple attention score matrices, the intermediate feature embedding Z of the source point cloud that maps the target point cloud information is obtained. The adaptive instance normalization module is used to normalize the intermediate feature embedding Z of the source point cloud, thereby keeping the posture of the source point cloud close to the target point cloud while not changing the original identity information, and obtaining the normalized feature embedding Z′ of the source point cloud. The improved Transformer network also uses its own encoder and decoder to obtain the output optimized feature embedding after a preset number of cycles. The optimized feature embedding is then convolved and activated using a multi-layer perceptron structure to realize the coordinate calculation of the resulting point cloud, completing the point cloud registration.

[0036] Compared with the prior art, the technical solution disclosed in the present invention has the following beneficial effects:

[0037] The present invention directly uses the point cloud structure for registration, which can more accurately reflect the actual size and shape structure of the original object compared to the traditional voxel-based representation. Directly learning the set features from the point cloud improves the registration performance to a certain extent. At the same time, the present invention encodes the potential relationship between the input source point cloud and the target point cloud through a multi-head cross-attention mechanism, giving higher weights to the parts that are considered more relevant. While capturing the dependency between point clouds, it also maintains the original point cloud local structure, thereby improving the degree of point cloud registration under large-scale non-rigid deformation. In addition, the present invention also introduces an adaptive instance normalization module in the image field, so that the deformed point cloud and the source point cloud are as consistent as possible in identity features, and the surface is smoother, further improving the registration effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flowchart of the non-rigid three-dimensional point cloud registration method based on the attention mechanism in Example 1 of the present invention;

[0039] Figure 2 This is a diagram of the architecture for registering an input three-dimensional point cloud of a human body in Example 1 of the present invention;

[0040] Figure 3 A comparison diagram of the source point cloud and the target point cloud in the normal viewing direction in Example 1 of the present invention;

[0041] Figure 4 A comparison diagram of the result point cloud and the target point cloud in the front view direction in Example 1 of the present invention;

[0042] Figure 5 A comparison diagram of the source point cloud and the target point cloud in the side view direction in Example 1 of the present invention;

[0043] Figure 6 This is a comparison diagram of the result point cloud and the target point cloud in the side view direction in Example 1 of the present invention. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used herein in the specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.

[0046] Example 1

[0047] This embodiment provides a non-rigid three-dimensional point cloud registration method based on the attention mechanism. This method combines adaptive instance normalization in the image field and effectively solves the problem of target three-dimensional point cloud registration under large-scale non-rigid deformation, such as the three-dimensional point cloud registration of the human body. The registration speed and registration effect have been improved to a certain extent.

[0048] See also Figure 1 , the registration method may include the following steps:

[0049] S1. Use at least one set of adaptive instance normalization modules as normalization layers in a standard Transformer network to obtain a modified Transformer network for performing non-rigid 3D point cloud registration tasks. The modified Transformer network also includes a multi-head cross-attention module. In this embodiment, the number of adaptive instance normalization modules can be set to three.

[0050] S2. Obtain the original point cloud data of a non-rigid target under two different actions, as the source point cloud S and target point cloud T respectively.

[0051] In this embodiment, a laser radar scanner (LIDAR) or a Microsoft Kinect camera can be used to capture different movements of the same person at different time points, resulting in two original point clouds with approximately 20,000 points. The original point clouds are sampled using an equal downsampling method with certain sampling rules. While ensuring that the overall geometric features remain unchanged, the density of the point cloud is reduced to 6,000 points, thereby reducing the amount of data and algorithm complexity required for related processing. The two sampled point clouds are denoted as the source point cloud S and the target point cloud T. The two point clouds have large non-rigid deformations and low overlap.

[0052] S3. Extract high-dimensional point-by-point features of the source point cloud and the target point cloud, and superimpose the output features with the position encoding in the improved Transformer network to obtain the initial feature embedding X of the source point cloud and the target point cloud S and X T .

[0053] In this embodiment, a dynamic graph convolutional neural network (DGCNN) can be used to extract the features of the source point cloud and the target point cloud, wherein the stacked EdgeConv layer can effectively extract the features of the local shape of the point cloud while maintaining the permutation invariance. The output features are superimposed with the Transformer position encoding to obtain the global feature embedding X S and X T , established topological connections between points, and further enhanced the representation ability of features.

[0054] S4. Using the linear transformation matrix W in the multi-head crisscross attention module Q 、W K 、W V Perform matrix operations on the preliminary feature embeddings of the source point cloud and the target point cloud to obtain the linear transformation result matrices Q, K, V and calculate the attention scores.

[0055] Among them, in the multi-head cross attention module, the linear transformation matrix W is used Q 、W K 、W V Perform matrix operations on the features of the input source point cloud and target point cloud to obtain the linear transformation result matrices Q, K, and V, and calculate the attention score:

[0056]

[0057] Where, d k is the number of columns of matrix Q.

[0058] S5. By concatenating multiple attention score matrices, we obtain the intermediate feature embedding Z of the source point cloud that maps the target point cloud information. Among them, multi-head attention helps the network capture richer information.

[0059] S6. Use the adaptive instance normalization module to normalize the intermediate feature embedding Z of the source point cloud, thereby keeping the posture of the source point cloud close to the target point cloud without changing the original identity information. At the same time, residual connection can also be introduced to obtain the normalized feature embedding Z′ of the source point cloud.

[0060] In this embodiment, the Adaptive Instance Normalization (AdaIN) in image style transfer is introduced into the normalization layer to make the pose close to the target point cloud without changing the original identity feature information, which further accelerates the convergence of the network. The residual layer uses residual connections to allow the network to focus on the current difference, which is conducive to solving the gradient disappearance and network degradation problems that occur during multi-layer network training. The expression formula of the normalized feature embedding Z′ of the source point cloud is:

[0061] Z′=[Conv(AdaIN(Conv(AdaIN(z))))+Conv(AdaIN(z))]

[0062] Where Conv(·) represents the convolution processing of ·. AdaIN(·) represents the adaptive instance normalization processing of ·.

[0063] S7. Set the loss function and use S4 to S6 as the decoder part of the improved Transformer network to embed the initial features of the source point cloud into X S The normalized feature embedding Z′ is used as the input of the encoder, and after a preset number of cycles, the output optimized feature embedding Z″ is obtained.

[0064] S8. Use a multi-layer perceptron (MLP) structure to perform convolution and activation on the optimized feature embedding Z″ to obtain the deformed human point cloud. In this training process, the point-wise mesh Euclidean distance (PMD) is used as the loss function:

[0065]

[0066] Where M i and M′ i They represent the three-dimensional coordinates of the i-th point of the target point cloud and the result point cloud respectively, i = 1, 2, 3...n, and n is the total number of points in the target point cloud.

[0067] In S8, the coordinates of the resulting point cloud are calculated using the following formula to complete the point cloud registration. The calculation formula is as follows:

[0068] R = 2*tanh[Conv(Z″)]

[0069] Where R represents the resulting point cloud and tanh is the hyperbolic tangent activation function.

[0070] In order to verify the registration performance of the non-rigid three-dimensional point cloud registration method based on the attention mechanism of the present invention, this embodiment also provides a set of qualitative experiments. The experimental environment is: Intel (R) Xeon (R) CPU E5-2609 v4 @ 1.70GHz, Ubuntu 18.04 operating system, 32G memory, Nvidia TITAN X graphics card, programming environment is Pycharm, deep learning framework is Pytorch 1.5.0, and the data sets used include MPI FAUST, MPI Dynamic FAUST, etc.

[0071] See also Figures 3 to 6 , where, in the normal direction, Figure 3 is the comparison diagram of source point cloud and target point cloud, Figure 4 This is a comparison diagram of the result point cloud and the target point cloud. In the side view direction, Figure 5 is the comparison diagram of source point cloud and target point cloud, Figure 6 The comparison between the result point cloud and the target point cloud shows that the proposed method has a good registration effect on large-scale deformed point clouds. The result point cloud fits the target point cloud very well at the elbow and knee joints, and the fitting effect at the extremities such as the hands and feet is also relatively ideal.

[0072] The present invention directly uses the point cloud structure for registration, which can more accurately reflect the actual size and shape structure of the original object compared to the traditional voxel-based representation. Directly learning the set features from the point cloud improves the registration performance to a certain extent. At the same time, the present invention encodes the potential relationship between the input source point cloud and the target point cloud through a multi-head cross-attention mechanism, giving higher weights to the parts that are considered more relevant. While capturing the dependency between point clouds, it also maintains the original point cloud local structure, thereby improving the degree of point cloud registration under large-scale non-rigid deformation. In addition, the present invention also introduces an adaptive instance normalization module in the image field, so that the deformed point cloud and the source point cloud are as consistent as possible in identity features, and the surface is smoother, further improving the registration effect.

[0073] Example 2

[0074] This embodiment provides a non-rigid 3D point cloud registration system based on an attention mechanism, which adopts the non-rigid 3D point cloud registration method based on an attention mechanism in Example 1. The registration system includes: a network construction module, a point cloud acquisition module, and a point cloud feature extraction module.

[0075] A network building module is configured to use at least one set of adaptive instance normalization modules as normalization layers in a standard Transformer network to obtain a modified Transformer network for performing non-rigid 3D point cloud registration tasks. The modified Transformer network also includes a multi-head crisscross attention module.

[0076] The point cloud acquisition module is used to obtain the original point cloud data of a non-rigid target under two different actions, which are used as the source point cloud S and the target point cloud T respectively.

[0077] The point cloud feature extraction module is used to extract high-dimensional point-by-point features of the source point cloud and the target point cloud, and superimpose the output features with the position encoding in the improved Transformer network to obtain the initial feature embedding X of the source point cloud and the target point cloud. S and X T .

[0078] Among them, the linear transformation matrix W in the multi-head cross attention module Q 、W K 、W V Matrix operations are performed on the preliminary feature embeddings of the source and target point clouds to obtain the linear transformation result matrices Q, K, and V, and the attention scores are calculated. By splicing multiple attention score matrices, the intermediate feature embedding Z of the source point cloud that maps the target point cloud information is obtained. The adaptive instance normalization module is used to normalize the intermediate feature embedding Z of the source point cloud, thereby keeping the posture of the source point cloud close to the target point cloud while not changing the original identity information, and obtaining the normalized feature embedding Z′ of the source point cloud. The improved Transformer network also uses its own encoder and decoder to obtain the output optimized feature embedding after a preset number of cycles. The optimized feature embedding is then convolved and activated using a multi-layer perceptron structure to realize the coordinate calculation of the resulting point cloud, completing the point cloud registration.

[0079] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0080] The above-described embodiments merely illustrate several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that a person skilled in the art would be able to make numerous modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A non-rigid 3D point cloud registration method based on attention mechanism, characterized in that: The following steps are involved: S1. Using at least one set of adaptive instance normalization modules as normalization layers in a standard Transformer network to obtain a modified Transformer network for performing non-rigid 3D point cloud registration tasks; the modified Transformer network also includes a multi-head cross attention module; S2. Obtain the original point cloud data of a non-rigid target under two different actions, as the source point cloud S and target point cloud T respectively; S3. Extract high-dimensional point-by-point features of the source point cloud and the target point cloud, and superimpose the output features with the position encoding in the improved Transformer network to obtain the initial feature embedding X of the source point cloud and the target point cloud S and X T ; S4. Using the linear transformation matrix W in the multi-head cross attention module Q 、W K 、W V Perform matrix operations on the preliminary feature embeddings of the source point cloud and the target point cloud to obtain linear transformation result matrices Q, K, and V and calculate attention scores; S5. Obtain the intermediate feature embedding Z of the source point cloud that maps the target point cloud information by concatenating multiple attention score matrices; S6. Using the adaptive instance normalization module to normalize the intermediate feature embedding Z of the source point cloud, thereby keeping the pose of the source point cloud close to the target point cloud while not changing the original identity information, to obtain the normalized feature embedding Z′ of the source point cloud; S7. Set the loss function and use S4 to S6 as the decoder part of the improved Transformer network to embed the initial features of the source point cloud into X s The normalized feature embedding Z′ is used as the input of the encoder, and after a preset number of cycles, the output optimized feature embedding Z″ is obtained; S8. Use a multi-layer perceptron structure to convolve and activate the optimized feature embedding Z″, and calculate the coordinates of the resulting point cloud using the following formula to complete point cloud registration; the calculation formula is as follows: R = 2*tanh[Conv(Z″)] Wherein, R represents the result point cloud; tanh is the hyperbolic tangent activation function.

2. The non-rigid 3D point cloud registration method based on the attention mechanism according to claim 1, characterized in that: In S2, the original point cloud data is sampled using an equal downsampling method so that the number of points in the source point cloud and the target point cloud is the same.

3. The non-rigid 3D point cloud registration method based on attention mechanism according to claim 1, characterized in that: In S3, a dynamic graph convolutional neural network is used to extract high-dimensional point-by-point features of the source point cloud and the target point cloud.

4. The non-rigid 3D point cloud registration method based on attention mechanism according to claim 1, characterized in that: In S4, the calculation formula of the attention score is: Among them, d k is the number of columns of matrix Q.

5. The non-rigid 3D point cloud registration method based on attention mechanism according to claim 1, characterized in that: The number of the adaptive instance normalization modules is set to 3 groups.

6. The non-rigid 3D point cloud registration method based on attention mechanism according to claim 5, characterized in that: In S6, a residual connection is introduced to normalize the source point cloud feature embedding.

7. The non-rigid 3D point cloud registration method based on attention mechanism according to claim 6, characterized in that: In S6, the expression formula of the normalized feature embedding Z′ of the source point cloud is: Z'=[Conv(AdaIN(Conv(AdaIW(Z))))+Conv(AdaIW(Z))] Where Conv(·) represents the convolution processing of ·; AdaIN(·) represents the adaptive instance normalization processing of ·.

8. The non-rigid 3D point cloud registration method based on attention mechanism according to claim 1, characterized in that: In S7, the point-by-point grid Euclidean distance PMD is used as the loss function: Where M i and M′ i Represent the three-dimensional coordinates of the i-th point of the target point cloud and the result point cloud respectively, i=1,2,3...n, and n is the total number of points in the target point cloud.

9. The non-rigid 3D point cloud registration method based on attention mechanism according to claim 1, characterized in that: In S2, a lidar scanner or a Kinect camera is used to capture different actions of the same non-rigid target at different time points.

10. A non-rigid 3D point cloud registration system based on attention mechanism, characterized in that: It adopts the non-rigid three-dimensional point cloud registration method based on the attention mechanism as described in any one of claims 1 to 9; the registration system includes: A network construction module, configured to use at least one set of adaptive instance normalization modules as normalization layers in a standard Transformer network to obtain a modified Transformer network for performing non-rigid 3D point cloud registration tasks; the modified Transformer network also includes a multi-head crisscross attention module; A point cloud acquisition module is used to acquire the original point cloud data of a non-rigid target under two different actions, as the source point cloud S and the target point cloud T respectively; and The point cloud feature extraction module is used to extract the high-dimensional point-by-point features of the source point cloud and the target point cloud, and superimpose the output features with the position encoding in the improved Transformer network to obtain the initial feature embedding X of the source point cloud and the target point cloud. S and X T ; Among them, the linear transformation matrix W in the multi-head cross attention module Q 、W K 、W V Matrix operations are performed on the preliminary feature embeddings of the source point cloud and the target point cloud to obtain the linear transformation result matrices Q, K, and V and calculate the attention scores. The intermediate feature embedding Z of the source point cloud that maps the target point cloud information is obtained by splicing multiple attention score matrices; the adaptive instance normalization module is used to normalize the intermediate feature embedding Z of the source point cloud, thereby keeping the posture of the source point cloud close to the target point cloud without changing the original identity information, and obtaining the normalized feature embedding Z′ of the source point cloud; the improved Transformer network is also used to obtain the output optimized feature embedding after a preset number of cycles of its own encoder and decoder, and use a multi-layer perceptron structure to convolve and activate the optimized feature embedding to realize the coordinate calculation of the result point cloud, thereby completing point cloud alignment.

Images