Category-level object pose estimation method based on shape enhancement and position embedding coding

Abstract

The present invention relates to a category-level target pose estimation method based on shape enhancement and position embedding coding, comprising the following steps: step S1: performing data enhancement on a public category-level pose estimation dataset by using a nonlinear shape enhancement method, thereby enriching the diversity of the dataset and increasing the shape diversity during network training; step S2: performing position embedding coding on input features by using an encoder, perceiving point-by-point geometric features from a local to a global perspective, and alleviating the problem of complex geometric structure of the target; step S3: reducing the sparse effect of point clouds caused by the data enhancement in step S1 by using a skip connection module, avoiding information loss in the process of cross-scale feature propagation, and further fusing multi-scale geometric features by using a decoder; step S4: using a symmetric perception reconstruction loss and a distance loss function to jointly improve the efficiency of training supervision, and finally completing category-level object 6D pose estimation and recognition; the method can effectively perform category-level pose estimation on the target.

Description

Technical Field

[0001] The present invention relates to the fields of pattern recognition and computer vision, and in particular to a category-level target pose estimation method based on shape enhancement and position embedding coding. Background Art

[0002] In recent years, 3D graph convolutional networks (GCNNs) have achieved tremendous success in computer vision. Deep learning-based category-level pose estimation methods, leveraging the GCNNs' strengths in feature extraction, have consistently set new records on public datasets and achieved significant progress. Category-level object pose estimation predicts the 6D pose and 3D size of previously unseen instances from a given set of object categories, freeing itself from the limitations of instance-level pose estimation. It has been widely used in fields such as robotics, augmented reality, and autonomous driving, and holds enormous commercial value. Academia and industry are constantly exploring new category-level pose estimation techniques, and a growing number of researchers are publishing papers related to this topic at various computer vision conferences. In recent years, leveraging significant breakthroughs in deep learning and neural networks in computer vision, category-level pose estimation algorithms have consistently set new records on various public benchmark datasets and have produced numerous commercial applications in industry.

[0003] Although category-level pose estimation technology has made great progress, it still faces many challenges in real environments, such as intra-class variations caused by color and shape differences between instance targets in a category, misidentification problems caused by similarities between object instances from different categories, and objects with complex geometric structures. Many factors will affect the performance of category-level pose estimation methods. Summary of the Invention

[0004] The object of the present invention is to provide a category-level target pose estimation method based on shape enhancement and position embedding coding, which can effectively perform category-level pose estimation on the target.

[0005] To achieve the above object, the technical solution of the present invention is: a category-level target pose estimation method based on shape enhancement and position embedding coding, comprising the following steps:

[0006] Step S1: Data enhancement is performed on the public category-level pose estimation dataset using a nonlinear shape enhancement method to enrich the diversity of the dataset and increase the shape diversity during network training;

[0007] Step S2: The input features are positionally embedded and encoded through an encoder to perceive point-by-point geometric features from a local to a global perspective, alleviating the problem of complex target geometric structures.

[0008] Step S3: The skip connection module is used to reduce the point cloud sparsity caused by the data enhancement in step S1, avoid information loss in the process of cross-scale feature propagation, and further fuse multi-scale geometric features through the decoder;

[0009] Step S4: Use the symmetry-aware reconstruction loss and the distance loss function to jointly improve the efficiency of training supervision, and finally complete the category-level object 6D pose estimation and recognition.

[0010] Furthermore, the step S1 specifically includes the following steps:

[0011] Step S11: Obtain a public category-level pose estimation dataset from the Internet and obtain relevant annotations of the training data;

[0012] Step S12: Generate different unseen instance targets through nonlinear shape enhancement methods while retaining the representative shape features of each category. For symmetric targets, a nonlinear scaling transformation based on the symmetry axis is performed to deform the target shape by adjusting the scale along the symmetry axis.

[0013] For each point cloud P in the regular object space, its deformation scale Sym(P) is given by Sym(P) = ξ(P * ) is obtained, where ξ(·) is a random nonlinear function, a parabolic function; P * is the projection of P on the axis of symmetry, * represents any axis of symmetry, where P x is the projection of P on the x-axis, P y is the projection of P on the y-axis, P z is the projection of P on the z-axis; each point cloud P generates a point cloud O after deformation operation, and obtains point cloud data O' for subsequent calculations. The deformation scale calculation formula is as follows:

[0014] Sym(P)=ξ(P * )=γ min +4(γ max -γ min )(P*) 2

[0015] Among them, γ max , γ min is a uniformly sampled random variable that controls the upper and lower bounds of Sym(P);

[0016] Furthermore, when the x-axis is the symmetric axis of the object, the x-axis is selected as the enhancement direction, the y-axis and the z-axis are the asymmetric axes, and the transformation function Defined as:

[0017]

[0018] where θ1 is a random variable that controls the scaling transformation along the asymmetric axes y and z.

[0019] Furthermore, when the y-axis is the symmetric axis of the object, the y-axis is selected as the enhancement direction, the x-axis and the z-axis are the asymmetric axes, and the transformation function Defined as:

[0020]

[0021] where θ2 is a random variable that controls the scaling transformation along the asymmetric x-axis and z-axis.

[0022] Furthermore, when the z-axis is the symmetric axis of the object, the z-axis is selected as the enhancement direction, the x-axis and the y-axis are the asymmetric axes, and the transformation function Defined as:

[0023]

[0024] where θ3 is a random variable that controls the scaling transformation along the asymmetric x-axis and y-axis.

[0025] Furthermore, the step S2 specifically includes the following steps:

[0026] Step S21: compose a 3D graph convolution encoder through a 3D graph convolution layer to transform the point cloud data As the input data of the encoder, N o Represents the number of all observed point clouds of the encoder; use the KNN algorithm to search each observed point cloud The M nearest neighbors of p m Represents the M nearest neighbor point clouds of point cloud O, m represents the neighbor point cloud number, and M represents the number of nearest neighbors that need to be calculated by the KNN algorithm;

[0027] Step S22: Calculate the direction vector dis of each point cloud O, perceive the relative position and direction of each point cloud in its local neighborhood, and initialize the learnable support kernel vector from the support kernel point to the central kernel point through uniform distribution S' represents the number of learnable support kernel vectors, s represents the number of learnable support kernel vectors; the calculation formula of the direction vector dis is:

[0028] dis=p m —O

[0029] Step S23: embed the position information of each point cloud O into a C through a 3D graph convolution encoder o The feature vector of dimension is passed to the ReLU activation function to obtain the position embedding code Repeat the above process and finally convert the original xyz coordinates of all input point cloud data O' into high-dimensional feature embedding code The encoder process calculation formula is as follows:

[0030]

[0031] Where ROI represents the receptive field range of the three-dimensional graph convolution operation, ReLU(·) represents the ReLU activation function, and ||·|| represents the norm;

[0032] Step S24: embedding encoding in high-dimensional features The graph attention mechanism is divided into four stages. The point-by-point feature extraction is performed through the point self-attention layer. Three different shared MLP layers are used to convert the local feature F of the i-th stage into i Project to produce the query vector Q i , key vector V i Sum value vector Key i , extracting point-by-point geometric features from the input point features in a local to global manner to solve the problem of complex geometric shapes; the calculation formula of the point self-attention layer is as follows:

[0033]

[0034] Among them, i represents the number of stages, ranging from [1,4], A i Indicates that the i-th stage is composed of local features F i The global geometric features generated after the point attention layer, Represents the value vector Key i The transpose of C i Represents the local feature F i The number of channels, Softmax(·) represents the Softmax activation function;

[0035] Finally, A i With F i Perform residual connection to obtain the global shape features of different scales obtained in the four stages, which are Among them, C1, C2, C3 and C4 are the number of feature channels in the first, second, third and fourth stages respectively.

[0036] Furthermore, the step S3 specifically includes the following steps:

[0037] Step S31: Fusing multi-scale information to reduce the sparse effect of the point cloud caused by the data enhancement operation in step S1; adding a skip connection module at the junction of the encoder and decoder modules, including a 1×1 Conv-BN-ReLU layer, three 3×3 AtrousConv-BN-ReLU layers, and a skip connection operation; connecting the 1×1 Conv-BN-ReLU layer and the three 3×3 AtrousConv-BN-ReLU layers, and sending them to another 1×1 BN-ReLU layer to generate the spatial dimension required by the decoder; setting the output spatial dimension to one-quarter of the input to ensure the dominance of the main channel, and then connecting the fused feature with the main channel, E4 becomes the skip feature SE4 after the skip connection module operation, and sends it to the decoder for training;

[0038] Step S32: Use the KNN algorithm to determine the features of the point cloud in the four stages of the decoder, and perform alignment operations with the four global shape features of different scales obtained in step S24, and use the connection operation and high-dimensional feature embedding coding The global shape features E1, E2, E3, and E4 are aggregated to generate the final geometric feature H, avoiding information loss during cross-scale feature propagation without the need for additional learnable parameters.

[0039] Furthermore, step S4 specifically includes the following steps:

[0040] Step S41: Use reflection symmetry and rotational symmetry to extract effective point-by-point features and predict the input symmetric point cloud Z of the input point cloud O: for the reflection symmetry category, predict the symmetric point cloud of the point cloud O about the reflection plane; for the rotational symmetry category, predict the symmetric point cloud of the point cloud O about the symmetry axis; for the remaining categories, use the point cloud O as the predicted symmetric point cloud; the loss function of symmetry-aware reconstruction is:

[0041] L sym =λ sym ||Z-ε(O,R pre , t pre , R gt , t gt )||1

[0042] Among them, λ sym represents the weight parameter of symmetric-aware reconstruction, ε(·) represents the symmetric projection function, R pre represents the predicted rotation matrix, R ge represents the true value rotation matrix, t pre represents the predicted translation vector, t gt represents the true translation vector, ||·||1 represents the L1 norm;

[0043] Step S42: Generate a predicted point cloud Q that matches the point cloud O through shape prior adaptation based on the geometric features H obtained in step S32 pre , through the Chamfer distance L CD Perform loss supervision to minimize the predicted point cloud Q pre The distance between the nearest neighbor point in point cloud O, L CD The calculation formula is as follows:

[0044]

[0045] Among them, λ CD Represents the Chamfer distance weight parameter, qp represents the predicted point cloud Q pre All points on, q represents all points on the point cloud O, ||·||2 represents the L2 norm; the total loss function L loss The calculation formula is as follows:

[0046] L loss =L sym +L CD

[0047] Step S43: The symmetry-aware reconstruction loss and the distance loss function are combined to optimize the performance of the model at different levels, and the recognition ability of symmetrical objects is improved in the case of posture ambiguity caused by symmetrical objects. The distance loss function helps the model to make fine adjustments in spatial alignment, improve the accuracy of geometric shape alignment, and ensure the consistency of the predicted posture with the true posture in geometry. The Umeyama algorithm and the RANSAC algorithm are used to calculate the 6D object posture and 3D size; the Umeyama algorithm is used to estimate the point cloud O and the predicted point cloud Q pre The best similarity transformation parameters between them include rotation, translation and scale parameters, where rotation and translation parameters correspond to 6D object posture and scale parameter corresponds to object size; the RANSAC algorithm is used to remove outliers and achieve robust estimation.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. It can effectively expand the data of existing public category-level pose estimation datasets, increase the diversity of shapes during training, reduce overfitting, better learn the intra-class changes of targets, and enhance the generalization ability of the network.

[0050] 2. Through position embedding coding, the local geometric features of the target are captured, and the graph attention mechanism is used to improve the network's perception of global information, perceiving point-by-point geometric features from local to global perspectives, alleviating the problem of complex target geometric structures.

[0051] 3. The skip connection module is used to reduce the sparse point cloud effect caused by the shape enhancement operation in step S1. The decoder further fuses multi-scale geometric features, propagating point-by-point geometric features from finer scales to coarser scales to avoid information loss during cross-scale feature propagation.

[0052] 4. The efficiency of network training supervision is improved through the symmetric perception reconstruction loss and distance loss function, and finally the 6D pose estimation and recognition of category-level objects are completed. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 Flowchart of the method implementation of the embodiment of the present invention DETAILED DESCRIPTION

[0054] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0055] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0056] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0057] like Figure 1 As shown, the present invention provides a category-level target pose estimation method based on shape enhancement and position embedding coding, comprising the following steps:

[0058] Step S1: Perform data augmentation on the public category-level pose estimation dataset using a nonlinear shape augmentation method to enrich the diversity of the dataset and increase the shape diversity during network training. This specifically includes the following steps:

[0059] Step S11: Obtain a public category-level pose estimation training set from the Internet and obtain relevant annotations of the training data;

[0060] Step S12: Generate different unseen instance targets through nonlinear shape enhancement methods while retaining the representative shape features of each category. For symmetric targets, a nonlinear scaling transformation based on the symmetry axis is performed to deform the target shape by adjusting the scale along the symmetry axis.

[0061] For each point cloud P in the regular object space, its deformation scale Sym(P) is given by Sym(P) = ξ(P * ) is obtained, where ξ(·) is a random nonlinear function, a parabolic function. P * is the projection of P on the axis of symmetry, * represents any axis of symmetry, for example, P x is the projection of P on the x-axis, P y is the projection of P on the y-axis, P z is the projection of P on the z-axis. Each point cloud P generates a point cloud O after deformation operation, and obtains point cloud data O' for subsequent calculations. The deformation scale calculation formula is as follows:

[0062] Sym(P)=ξ(P * )=γ min +4(γ max -γ min )(P * ) 2

[0063] Among them, γ max , γ min are uniformly sampled random variables that control the upper and lower bounds of Sym(P).

[0064] When the x-axis is the symmetric axis of the object, the x-axis is selected as the enhancement direction, the y-axis and the z-axis are the asymmetric axes, and the transformation function Defined as:

[0065]

[0066] where θ1 is a random variable that controls the scaling transformation along the asymmetric axes y and z.

[0067] When the y-axis is the symmetric axis of the object, the y-axis is selected as the enhancement direction, the x-axis and the z-axis are the asymmetric axes, and the transformation function Defined as:

[0068]

[0069] where θ2 is a random variable that controls the scaling transformation along the asymmetric x-axis and z-axis.

[0070] When the z-axis is the symmetric axis of the object, the z-axis is selected as the enhancement direction, the x-axis and y-axis are the asymmetric axes, and the transformation function Defined as:

[0071]

[0072] where θ3 is a random variable that controls the scaling transformation along the asymmetric x-axis and y-axis.

[0073] Step S2: The encoder embeds the input features into positional encoding, perceives point-by-point geometric features from local to global perspectives, and alleviates the problem of complex target geometric structures. This specifically includes the following steps:

[0074] Step S21: compose a 3D graph convolution encoder through a 3D graph convolution layer to transform the point cloud data As the input data of the encoder, N o Represents the number of all observed point clouds of the encoder. Use the KNN algorithm to search each observed point cloud The M nearest neighbors of p m Represents the M nearest neighbor point clouds of point cloud O, m represents the neighbor point cloud number, and M represents the number of nearest neighbors that need to be calculated by the KNN algorithm.

[0075] Step S22: Calculate the direction vector dis of each point cloud O, perceive the relative position and direction of each point cloud in its local neighborhood, and initialize the learnable support kernel vector from the support kernel point to the central kernel point through uniform distribution S' represents the number of learnable support kernel vectors, and s represents the number of learnable support kernel vectors. The calculation formula of the direction vector dis is:

[0076] dis=Pm-0

[0077] Step S23: embed the position information of each point cloud O into a C through a 3D graph convolution operation o The feature vector of dimension is passed to the ReLU activation function to obtain the position embedding code Repeat the above process and finally convert the original xyz coordinates of all input point cloud data O' into high-dimensional feature embedding code The encoder process calculation formula is as follows:

[0078]

[0079] Among them, ROI represents the receptive field range of the three-dimensional graph convolution operation, ReLU(·) represents the ReLU activation function, and ||·|| represents the norm.

[0080] Step S24: embedding encoding in high-dimensional features The graph attention mechanism is divided into four stages. The point-by-point feature extraction is performed through the point self-attention layer. Three different shared MLP layers are used to convert the local feature F of the i-th stage into i Project to produce the query vector Q i , key vector V i Sum value vector Key i, extracting point-by-point geometric features from the input point features in a local to global manner to solve problems with complex geometric shapes. The calculation formula of the point self-attention layer is as follows:

[0081]

[0082] Among them, i represents the number of stages, ranging from [1,4], A i Indicates that the i-th stage is composed of local features F i The global geometric features generated after the point attention layer, Represents the value vector Key i The transpose of C i Represents the local feature F i The number of channels, Softmax(·) represents the Softmax activation function.

[0083] Finally, A i With F i Perform residual connection to obtain the global shape features of different scales obtained in the four stages, which are Among them, C1, C2, C3 and C4 are the number of feature channels in the first, second, third and fourth stages respectively.

[0084] Step S3: The skip connection module is used to reduce the point cloud sparsity caused by the data enhancement in step S1, avoid information loss during cross-scale feature propagation, and further fuse multi-scale geometric features through the decoder. Specifically, it includes the following steps:

[0085] Step S31: Fuse multi-scale information to reduce the point cloud sparsity caused by the data augmentation operation in step S1. Add a skip connection module at the junction of the encoder and decoder modules, including a 1×1 Conv-BN-ReLU layer, three 3×3 AtrousConv-BN-ReLU layers, and a skip connection operation. Connect the 1×1 Conv-BN-ReLU layer and the three 3×3 AtrousConv-BN-ReLU layers and send them to another 1×1 BN-ReLU layer to generate the spatial dimensions required by the decoder. Set the output spatial dimension to one-quarter of the input to ensure the dominance of the main channel, then connect the fused feature with the main channel. E4 is converted into the skip feature SE4 through the skip connection module operation and sent to the decoder for training.

[0086] Step S32: Use the KNN algorithm to determine the features of the point cloud in the four stages of the decoder, and perform alignment operations with the four global shape features of different scales obtained in step S24, and use the connection operation and high-dimensional feature embedding coding The global shape features E1, E2, E3, and E4 are aggregated to generate the final geometric feature H, avoiding information loss during cross-scale feature propagation without the need for additional learnable parameters.

[0087] Step S4: Use the symmetry-aware reconstruction loss and the distance loss function to improve the efficiency of training supervision, and finally complete the 6D pose estimation and recognition of category-level objects. Specifically, it includes the following steps:

[0088] Step S41: Use reflection symmetry and rotational symmetry to extract effective point-by-point features and predict the input symmetric point cloud Z of the input point cloud O: For the reflection symmetry category, predict the symmetric point cloud of point cloud O about the reflection plane; for the rotational symmetry category, predict the symmetric point cloud of point cloud O about the symmetry axis; for the remaining categories, use point cloud O as the predicted symmetric point cloud. The loss function of symmetry-aware reconstruction is:

[0089] L sym =λ sym ||Z-ε(O,R pre , t pre , R gt , t gt )||1

[0090] Among them, λ sym represents the weight parameter of symmetric-aware reconstruction, ε(·) represents the symmetric projection function, R pre represents the predicted rotation matrix, R gt represents the true value rotation matrix, t pre represents the predicted translation vector, t gt represents the true translation vector, and ||·||1 represents the L1 norm.

[0091] Step S42: Generate a predicted point cloud Q that matches the point cloud O through shape prior adaptation based on the geometric features H obtained in step S32 pre , through the Chamfer distance L CD Perform loss supervision to minimize the predicted point cloud Q pre The distance between the nearest neighbor point in point cloud O, L CD The calculation formula is as follows:

[0092]

[0093] Among them, λ CD represents the Chamfer distance weight parameter, qp represents the predicted point cloud Q pre All points on, q represents all points on the point cloud O, ||·||2 represents the L2 norm. The total loss function L loss The calculation formula is as follows:

[0094] Lloss =L sym ++L CD

[0095] Step S43: The symmetry-aware reconstruction loss and the distance loss function are combined to optimize the performance of the model at different levels. The recognition ability of symmetrical objects is improved in the case of symmetrical objects that may cause posture ambiguity. The distance loss function helps the model to make fine adjustments in spatial alignment, improve the accuracy of geometric shape alignment, and ensure the consistency of the predicted posture with the true posture in geometry. The Umeyama algorithm and RANSAC algorithm are used to calculate the 6D object posture and 3D size. The Umeyama algorithm is used to estimate the point cloud O and the predicted point cloud Q pre The best similarity between the transformation parameters (rotation, translation and scale), where the rotation and translation parameters correspond to the 6D object pose and the scale parameter corresponds to the object size. The RANSAC algorithm is used to remove outliers and achieve robust estimation.

[0096] The above are preferred embodiments of the present invention. Any changes made according to the technical solution of the present invention, as long as the resulting functions and effects do not exceed the scope of the technical solution of the present invention, shall fall within the scope of protection of the present invention.

Claims

1. A category-level target pose estimation method based on shape enhancement and position embedding coding, characterized in that: The following steps are involved: Step S1: Data enhancement is performed on the public category-level pose estimation dataset using a nonlinear shape enhancement method to enrich the diversity of the dataset and increase the shape diversity during network training; Step S2: The input features are positionally embedded and encoded through an encoder to perceive point-by-point geometric features from a local to a global perspective, alleviating the problem of complex target geometric structures. Step S3: The skip connection module is used to reduce the point cloud sparsity caused by the data enhancement in step S1, avoid information loss in the process of cross-scale feature propagation, and further fuse multi-scale geometric features through the decoder; Step S4: Use symmetric perception reconstruction loss and distance loss function to jointly improve the efficiency of training supervision and finally complete the 6D pose estimation and recognition of category-level objects; Step S1 specifically includes the following steps: Step S11: Obtain a public category-level pose estimation dataset from the Internet and obtain relevant annotations of the training data; Step S12: Generate different unseen instance targets through nonlinear shape enhancement methods while retaining the representative shape features of each category. For symmetric targets, a nonlinear scaling transformation based on the symmetry axis is performed to deform the target shape by adjusting the scale along the symmetry axis. For each point cloud P in the regular object space, its deformation scale Sym(P) is given by Sym(P) = ξ(P * ) is obtained, where ξ(·) is a random nonlinear function, a parabolic function; P * is the projection of P on the axis of symmetry, * represents any axis of symmetry, where P x is the projection of P on the x-axis, P y is the projection of P on the y-axis, P z is the projection of P on the z-axis; each point cloud P generates a point cloud O after deformation operation, and obtains point cloud data O' for subsequent calculations. The deformation scale calculation formula is as follows: Sym(P)=ξ(P * )=γ min +4(c max -c min )(P * ) 2 Among them, γ max , γ min is a uniformly sampled random variable that controls the upper and lower bounds of Sym(P); Step S4 specifically includes the following steps: Step S41: Use reflection symmetry and rotational symmetry to extract effective point-by-point features and predict the input symmetric point cloud Z of the input point cloud O; the loss function of symmetry-aware reconstruction is: L sym =λ sym ||Z-ε(O,R pre ,t pre ,R gt ,t gt )||1 Among them, λ sym represents the weight parameter of symmetric-aware reconstruction, ε(·) represents the symmetric projection function, R pre represents the predicted rotation matrix, R gt represents the true value rotation matrix, t pre represents the predicted translation vector, t gt represents the true translation vector, ||·||1 represents the L1 norm; Step S42: Generate a predicted point cloud Q that matches the point cloud O through shape prior adaptation based on the geometric features H obtained in step S32 pre , through the Chamfer distance L CD Perform loss supervision to minimize the predicted point cloud Q pre The distance between the nearest neighbor point in point cloud O, L CD The calculation formula is as follows: Among them, λ CD represents the Chamfer distance weight parameter, qp represents the predicted point cloud Q pre All points on, q represents all points on the point cloud O, ||·||2 represents the L2 norm; the total loss function L loss The calculation formula is as follows: L loss =L sym +L CD Step S43: The symmetry-aware reconstruction loss and the distance loss function are combined to optimize the performance of the model at different levels. The recognition ability of symmetrical objects is improved in the case of posture ambiguity caused by symmetrical objects. The distance loss function helps the model to make fine adjustments in spatial alignment, improve the accuracy of geometric shape alignment, and ensure the consistency of the predicted pose and the true pose in geometry. The Umeyama algorithm is used to estimate the point cloud O and the predicted point cloud Q. pre The best similarity transformation parameters between them include rotation, translation and scale parameters, where rotation and translation parameters correspond to 6D object posture and scale parameter corresponds to object size; the RANSAC algorithm is used to remove outliers and achieve robust estimation.

2. The method for class-level target pose estimation based on shape enhancement and position embedding coding according to claim 1, characterized in that When the x-axis is the symmetric axis of the object, the x-axis is selected as the enhancement direction, the y-axis and the z-axis are the asymmetric axes, and the transformation function Defined as: where θ1 is a random variable that controls the scaling transformation along the asymmetric axes y and z.

3. The method for class-level target pose estimation based on shape enhancement and position embedding coding according to claim 1, characterized in that When the y-axis is the symmetric axis of the object, the y-axis is selected as the enhancement direction, the x-axis and the z-axis are the asymmetric axes, and the transformation function Defined as: where θ2 is a random variable that controls the scaling transformation along the asymmetric x-axis and z-axis.

4. The method for class-level target pose estimation based on shape enhancement and position embedding coding according to claim 1, wherein: When the z-axis is the symmetric axis of the object, the z-axis is selected as the enhancement direction, the x-axis and y-axis are the asymmetric axes, and the transformation function Defined as: where θ3 is a random variable that controls the scaling transformation along the asymmetric x-axis and y-axis.

5. The method for class-level target pose estimation based on shape enhancement and position embedding coding according to claim 1, wherein: Step S2 includes the following steps: Step S21: compose a 3D graph convolution encoder through a 3D graph convolution layer to transform the point cloud data As the input data of the encoder, N o Represents the number of all observed point clouds of the encoder; use the KNN algorithm to search each observed point cloud The M nearest neighbors of p m Represents the M nearest neighbor point clouds of point cloud O, m represents the neighbor point cloud number, and M represents the number of nearest neighbors that need to be calculated by the KNN algorithm; Step S22: Calculate the direction vector dis of each point cloud O, perceive the relative position and direction of each point cloud in its local neighborhood, and initialize the learnable support kernel vector from the support kernel point to the central kernel point through uniform distribution S' represents the number of learnable support kernel vectors, s represents the number of learnable support kernel vectors; the calculation formula of the direction vector dis is: dis=p m -O Step S23: embed the position information of each point cloud O into a C through a 3D graph convolution encoder o The feature vector of dimension is passed to the ReLU activation function to obtain the position embedding code Repeat the above process and finally convert the original xyz coordinates of all input point cloud data O' into high-dimensional feature embedding code Step S24: embedding encoding in high-dimensional features The graph attention mechanism is divided into four stages. The point-by-point feature extraction is performed through the point self-attention layer. Three different shared MLP layers are used to convert the local feature F of the i-th stage into i Project to produce the query vector Q i , key vector V i Sum value vector Key i , extracting point-by-point geometric features from the input point features in a local to global manner; the calculation formula is as follows: Among them, i represents the number of stages, ranging from [1,4], A i Indicates that the i-th stage is composed of local features F i The global geometric features generated after the point attention layer, Represents the value vector Key i The transpose of C i Represents the local feature F i The number of channels, Softmax(·) represents the Softmax activation function; Finally, A i With F i Perform residual connection to obtain the global shape features of different scales obtained in the four stages, which are Among them, C1, C2, C3 and C4 are the number of feature channels in the first, second, third and fourth stages respectively.

6. The method for class-level target pose estimation based on shape enhancement and position embedding coding according to claim 5, characterized in that: The encoder process calculation formula is as follows: Among them, ROI represents the receptive field range of the three-dimensional graph convolution operation, ReLU(·) represents the ReLU activation function, and ||·|| represents the norm.

7. The method for class-level target pose estimation based on shape enhancement and position embedding coding according to claim 6, characterized in that: Step S3 specifically includes the following steps: Step S31: Fusing multi-scale information to reduce the sparse effect of the point cloud caused by the data enhancement operation in step S1; adding a skip connection module at the junction of the encoder and decoder modules, including a 1×1 Conv-BN-ReLU layer, three 3×3 AtrousConv-BN-ReLU layers, and a skip connection operation; connecting the 1×1 Conv-BN-ReLU layer and the three 3×3 AtrousConv-BN-ReLU layers, and sending them to another 1×1 BN-ReLU layer to generate the spatial dimension required by the decoder; setting the output spatial dimension to one-quarter of the input to ensure the dominance of the main channel, and then connecting the fused feature with the main channel, E4 becomes the skip feature SE4 after the skip connection module operation, and sends it to the decoder for training; Step S32: Use the KNN algorithm to determine the features of the point cloud in the four stages of the decoder, and perform alignment operations with the four global shape features of different scales obtained in step S24, and use the connection operation and high-dimensional feature embedding coding Aggregate with the global shape features E1, E2, E3, and E4 to generate the final geometric feature H.

8. The method for class-level target pose estimation based on shape enhancement and position embedding coding according to claim 1, wherein: The prediction of the symmetric point cloud Z is as follows: for the reflection symmetry category, predict the symmetric point cloud Z of point cloud O about the reflection plane; for the rotational symmetry category, predict the symmetric point cloud Z of point cloud O about the symmetry axis; for the remaining categories, use point cloud O as the predicted symmetric point cloud Z.

Images