GPCL-Net grouping point corresponding learning method for 6D pose estimation in complex environment

Abstract

The invention provides a GPCL-Net grouping point corresponding learning method for 6D pose estimation in a complex environment, and the method is based on a GPCL-Net grouping point corresponding learning network cooperatively driven by an association normalization feature extraction module and a double-FA feature alignment module. The GPCL-Net grouping point corresponding learning network comprises an associated normalized feature extraction module, a double-FA feature alignment module, a pose estimation module and a pose iteration module; according to the method, point corresponding relations of'part-to-part 'and'all-to-all' are established in a feature space at the same time, advantage complementation is utilized to achieve the purpose that the network solves the pose by establishing the point corresponding relation between two coordinates, and a pose iteration refining network matched with the mode of the GPCL-Net network is designed according to the structural characteristics of the GPCL-Net network so as to iteratively improve the pose estimation precision; a double-FA feature alignment module is designed, point features from different coordinate systems are decomposed into pose and matching features, two feature alignment is carried out along channel dimension grouping, pose-matching feature pairs are generated, and a point corresponding relation of two point sets is established.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sensors, in particular to a GPCL-Net grouping point correspondence learning method for 6D pose estimation in complex environments. BACKGROUND

[0002] 6D object pose estimation is widely used in robot grasping, augmented reality, and autonomous driving. However, in complex environments with varying lighting and severe occlusions, achieving high-precision 6D pose estimation remains challenging. The rotational dimension of 6D pose is described by SO(3), and its inherent nonlinearity makes it difficult to directly regress 6D pose from camera observation data. Therefore, many existing methods achieve indirect estimation of pose by learning the point correspondence between the camera and object coordinate systems, providing an effective technical approach to solving this nonlinearity problem.

[0003] At the same time, with the continuous development of sensors, sensor technology is no longer limited to the capture of planar images, but can accurately obtain the depth image and complete spatial geometric features of the target, providing key data support for three-dimensional scene analysis and target positioning. The technical characteristics of different sensors differ, directly leading to three major categories of 6D pose estimation technology: RGB-based methods, RGB-D-based methods, and point cloud-based methods. Since RGB-D sensors capture both RGB and depth images, they provide complementary color and geometric information. Therefore, RGB-D-based methods are generally superior to RGB-based and point cloud-based methods. Some of these methods directly concatenate RGB and depth values and input them into a multi-stage model to estimate pose. Unlike these methods, some recent methods convert depth images into point cloud data using camera intrinsic matrices to extract geometric information, which is then fused with color features and input into a model to estimate pose.

[0004] Based on this, drawing on the core idea of establishing point correspondence in DCL-Net, combining RGB-D sensor data and geometric information of prior CAD models, by converting RGB-D to point cloud (referred to as target point set) and matching it with CAD model sampling point cloud (referred to as source point set), the point correspondence between camera coordinate system and object coordinate system is established. For example, Figure 1As shown in the prior art, taking the observation data of the target object in the camera coordinate system and the CAD model in the object coordinate system as inputs, two methods for establishing point correspondence are given: (1) predicting the "part-to-part" correspondence of the target point set in the source point set; (2) predicting the "complete-to-complete" correspondence of the source point set in the target point set. These two correspondence methods have their own advantages and limitations: the "part-to-part" correspondence does not need to perform shape completion operation, so it has more advantages in correspondence accuracy; but when facing objects with severe occlusion, it is usually difficult to effectively establish point correspondence. In contrast, the "complete-to-complete" correspondence can more stably cope with severe occlusion scenes and achieve more stable pose estimation, but it needs to solve the technical problem of shape completion.

[0005] Therefore, how to effectively combine these two point correspondence establishment methods to achieve complementary advantages and avoid their limitations to meet the object pose estimation needs in different scenarios has become a key problem to be solved in the field of 6D pose estimation. SUMMARY

[0006] The present application aims to solve the above-mentioned problems in the prior art, and the main purpose is to provide a GPCL-Net grouping point correspondence learning method for 6D pose estimation in complex environments, which can simultaneously establish "part-to-part" and "all-to-all" point correspondence in the feature space, utilize complementary advantages, and achieve the purpose of the network solving the pose by establishing point correspondence between the two coordinates. In view of the characteristics of the GPCL-Net network structure, a pose iterative refinement network is designed to adapt to its mode, which is used to iteratively improve the pose estimation accuracy. A scale parameter shared by the source point set and the target point set is introduced in feature extraction, which has strong robustness to incomplete observation values and outliers, and can improve the similarity of the feature distribution of the two point sets. The T-S FA feature alignment module and the S-T FA feature alignment module are designed to decompose the point features from different coordinate systems into pose and matching features, and then group them along the channel dimension for alignment of the two features. Finally, the pose-matching feature pair is generated, and the point correspondence of the two point sets can be established based on the aligned feature pair. It performs well on the challenging Occlusion-LineMOD and HB-Vision datasets, achieving an accuracy of 73.7% and 79.4%, respectively.

[0007] To achieve the purpose of the present application, the technical scheme adopted is: A GPCL-Net grouped point correspondence learning method for 6D pose estimation in complex environments is proposed. The GPCL-Net grouped point correspondence learning method is based on a GPCL-Net grouped point correspondence learning network driven by a correlation normalized feature extraction module and a dual FA feature alignment module. The GPCL-Net grouped point correspondence learning network includes a correlation normalized feature extraction module, a dual FA feature alignment module, a pose estimation module, and a pose iteration module.

[0008] A GPCL-Net grouped point correspondence learning method for 6D pose estimation in complex environments, specifically including the following steps: S1. Correlation Normalization Feature Extraction: Using the target point set Source point set As input, the association-normalized feature extraction module performs a set of target points... Source point set After centering, the target point sets are extracted by the output channel convolutional layer. Point-by-point features Source point set Point-by-point features For each point, the feature is processed point by point. , After normalization, the data is correlated and spliced ​​along the channel dimension to obtain the target point set. Correlation normalized pointwise features Source point set Correlation normalized pointwise features ; S2, using correlation-normalized point-by-point features , As input, the target point set is generated through processing by the dual FA feature alignment module. Matching features Postural characteristics Source point set Matching features Postural characteristics Matching features , and posture characteristics , The system is divided into n channel groups, each group is executed independently, and attention graphs are generated for each group. and Then use the attention diagram. Matching features and posture characteristics Obtain the target point set by performing matrix multiplication. Matching features in object coordinate system and posture characteristics Use attention diagram Matching features and posture characteristics Matrix multiplication to obtain the source point set Matching features in object coordinate system and posture characteristics And correlated with normalized point-by-point features in the camera coordinate system , Align and generate matching feature pairs respectively ( and ) and pose features ( and ); S3, using matching feature pairs ( and ) and pose features ( and As input, the pose estimation module constructs a matching feature pair matrix. Pose feature pair matrix The matching feature pair matrix is ​​processed by MLP (Multilayer Perceptron). Generate confidence Pose features to matrix The data is transmitted to the MLP multilayer perceptron and processed by confidence level. Weighted regression yields 6D pose ; S4. Using 6D pose Update target point set , source set and the updated target point set The pose iteration module takes the source point set as input again. and the updated target point set The data is transmitted to the correlation normalization feature extraction module, the dual FA feature alignment module, and the pose estimation module to obtain pose variables. The pose iteration module completes multiple iterations of refinement, and the refined 6D pose is used as the final output target 6D pose.

[0009] Furthermore, the correlation normalization feature extraction module in step S1 of the GPCL-Net group point correspondence learning method includes a four-layer Block_CB structure, and each Block_CB structure is equipped with a Correlation Normalization module and a Batch Normalization module.

[0010] Furthermore, the specific operation of step S1 in the GPCL-Net group point correspondence learning method is as follows: S11. Input target point set and source point set , respectively, are extracted by the convolutional layers with output channels wherein, is the size of the point-wise feature , is the number of points in the target point set , is the size of the point-wise feature , is the number of points in the source point set , is the size of the point-wise feature , is the number of points in the source point set , is the number of features contained in each point of the point-wise feature , is the number of features contained in each point of the point-wise feature , is the one-dimensional convolution operation, is a point set with is a point set with is the point-wise feature of the target point set , is the point-wise feature of the source point set , is the feature of the th point in , is the feature of the th point in i , is the target point set , is the source point set i is the point cloud data obtained by converting the RGB-D data collected by the depth camera through the camera intrinsic parameters; the source point set is the point cloud data obtained by uniformly sampling the CAD model; The Correlation Normalization module normalizes the point-wise features of each point in the target point set and the source point set : wherein, is the scale parameter shared in the first layer Block_CB structure, is the normalized feature , is the normalized feature , and and​​​​​​​​​​​​​​​​ Used to calculate point-by-point features and These are all process variables and have no practical significance; Scale parameters shared in the first-level Block_CB structure Depend on The calculation yielded: , in, Indicates the first i The first point j One feature, this parameter is derived from the target point set. Source point set shared, For the source point set The number of midpoints For point-by-point features and The number of features contained in each point; S12. Point-by-point features obtained after normalization by the Correlation Normalization module. and Then use the Batch Normalization module to perform batch normalization processing and Activate the normalized pointwise feature extraction for the first Block_CB structure. The normalized pointwise feature extraction for other Block_CB structures follows the same operation. , , , in, for Point-by-point features after normalization for Point-by-point features after normalization For quantity The set of points, For quantity The set of points, for Features after normalization for Features after normalization For activation function, This is an abbreviation for Batch Normalization, which stands for Batch Normalization. For the target point set After batch normalization of the output of the first-level Block_CB structure and Activated pointwise features, For the source point set After batch normalization of the output of the first-level Block_CB structure and Activated pointwise features; S13. Normalize the output batch of the four-layer Block_CB structure and... Activated pointwise features and By concatenating along the channel dimension, we obtain the extracted, finally normalized, point-by-point features. and : , in, For the target point set The correlation normalized pointwise features, For the source point set The correlation normalized pointwise features, These are the batch normalizations of the outputs of the Block_CB structures in the first, second, third, and fourth layers, respectively. Activated pointwise features, "Correlation Normalization Feature Extraction" is an abbreviation for the entire correlation normalization feature extraction module, specifically referring to the correlation normalization operation. This is a splicing operation along the channel dimension.

[0011] Furthermore, the dual FA feature alignment module in step S2 of the GPCL-Net grouping point correspondence learning method includes a TS FA feature alignment module and an ST FA feature alignment module.

[0012] Furthermore, the specific operation of step S2 in the GPCL-Net group point correspondence learning method is as follows: S21, with target point set Source point set Correlation normalized pointwise features and As input, the target point set is generated through the T-SFA feature alignment module and processed by the MLP multilayer perceptron. Matching features and posture characteristics The source point set is generated through the ST FA feature alignment module. Matching features and posture characteristics : , in, For the target point set The correlation normalized pointwise features, is a set of source points , is a set of target points , is a set of source points , is a set of target points , is a set of source points , represents a subnetwork of MLP (Multi-Layer Perceptron); S22, in the double FA feature alignment module, the matching feature , and the pose feature , are divided into n channel groups according to the channel, and each group is independently executed, respectively, in the T-S FA feature alignment module to generate the attention map , in the S-T FA feature alignment module to generate the attention map : , , , wherein, is the matching feature of the i-th channel group of the set of target points , i is the matching feature of the i-th channel group of the set of source points , is the pose feature of the i-th channel group of the set of source points i , is an activation function, represents a matrix transposition operation; i S23, in the T-S FA feature alignment module, the attention map is respectively multiplied with the matching feature , the pose feature to infer the matching feature and the pose feature of the set of target points in the object coordinate system, and align with the associated normalized point-by-point feature of the set of target points in the camera coordinate system: , , wherein, is the associated normalized point-by-point feature of the set of target points , ​​​​It is a formula operation process variable, which has no practical significance, and is only convenient for formula writing and operation, is a target point set is a matching feature in the object coordinate system, is a target point set is a pose feature in the object coordinate system, is a target point set is a matching feature, is a target point set is a pose feature, is a concatenation operation along the channel dimension, represents a subnetwork of MLP (Multi-Layer Perceptron); In the S-T FA feature alignment module, the attention graph is respectively multiplied by the matching feature and the pose feature to infer the source point set is a matching feature in the object coordinate system and a pose feature , and is aligned with the associated normalized point-wise feature of the source point set in the camera coordinate system: , , , wherein, is the associated normalized point-wise feature of the source point set , is a formula operation process variable, which has no practical significance, and is only convenient for formula writing and operation, is a matching feature in the object coordinate system of the source point set is a pose feature in the object coordinate system of the source point set , is a matching feature of the source point set in the object coordinate system, is a pose feature of the source point set , is a pose feature of the source point set ; Through the T-S FA feature alignment module and the S-T FA feature alignment module, partial-to-partial and full-to-full correspondence relationships are realized, and then through these two correspondence relationships, a matching feature pair ( and ) and a pose feature pair ( and ) are generated in the feature space.

[0013] Further, the specific operation of step S3 of the multi-channel cross-fusion method is: S31, a matching feature pair ( and ) and pose feature pairs and ) as input, the matching feature pair matrix and pose feature pair matrix : , wherein, is the target point set matching features in the object coordinate system, is the target point set pose features in the object coordinate system, is the target point set matching features, is the target point set pose features, is the source point set matching features in the object coordinate system, is the source point set pose features in the object coordinate system, is the source point set matching features, is the source point set pose features. S32, the matching feature pair matrix is processed by the MLP multi-layer perceptron to generate the confidence , the pose feature pair matrix is transmitted to the MLP multi-layer perceptron, and the 6D pose is obtained by weighted regression according to the confidence : , , , , wherein, is the value of the i confidence, is the number of points in the target point set , is the number of points in the source point set , is the sum along the edge, is the activation function, represents a subnetwork of the MLP multi-layer perceptron, is an intermediate variable in the pose estimation process, and is used for simplifying the formula for easy understanding. Further, the specific operation of the multi-channel cross-fusion method step S4 is as follows:

[0014] Further, the specific operation of the multi-channel cross-fusion method step S4 is as follows: ​S41. The 6D pose obtained is processed using the correlation normalization feature extraction module, the dual FA feature alignment module, and the pose estimation module. Update target point set iteration -1 target point set for: , in, For the target point set The number of points, For the target point set The Middle i One point, Iterating for the pose iteration module -1 times in 6D pose, when When =1, 6D pose This is the first time that a 6D pose has been obtained through the correlation normalization feature extraction module, the dual FA feature alignment module, and the pose estimation module. S42, Source Set and the updated target point set As input, it is transmitted to the correlation normalization feature extraction module to obtain iterative... -1 new association normalized pointwise features , : , in, "Correlation Normalization Feature Extraction" is an abbreviation for the entire correlation normalization feature extraction module, specifically referring to the correlation normalization operation. S43, Iteration -1 new association normalized pointwise features , The data is transmitted to the dual FA feature alignment module, where the TS FA feature alignment module and the ST FA feature alignment module establish partial-to-partial and full-to-full correspondences. These two correspondences are then used to generate iterative features in the feature space. -1 matching feature pairs ( and ) and pose features ( and Then it is transmitted to the pose estimation module to construct the iterative... -1 matching feature pair matrix and pose feature pair matrix : , in, For the updated target point set iterating in the object coordinate system -1st order matching features, for the updated target point set iterating -1st order matching features, for the updated target point set iterating in the object coordinate system -1st order pose features, for the updated target point set iterating -1st order pose features, for the source point set iterating in the object coordinate system -1st order matching features, for the source point set iterating -1st order matching features, for the source point set iterating in the object coordinate system -1st order pose features, for the source point set iterating -1st order pose features; S44, processing the iteration -1st order matching feature pair matrix generating the iteration -1st order confidence , transferring the iteration -1st order pose feature pair matrix into the MLP, weighted regression by confidence to get the iteration -1st order 6D pose variable : , , , , , where, is the sum along the edge, is the activation function, denotes a subnetwork of the MLP, is an intermediate variable in the iteration -1st order pose estimation process, used for simplifying the formula for better understanding, is the iteration -1st order 6D pose variable is obtained by the pose iteration module : , wherein, is the 6D pose variable obtained by iterative -1 order weighted regression, is the 6D pose obtained by iterative -1 order weighted regression, is the 6D pose obtained by iterative order weighted regression.

[0015] The application provides a GPCL-Net grouping point correspondence learning method for 6D pose estimation in a complex environment. The GPCL-Net grouping point correspondence learning method for 6D pose estimation in a complex environment can simultaneously establish point correspondence relationships of 'part-to-part' and 'all-to-all' in a feature space, utilize the complementary advantages, achieve the purpose of solving the pose by establishing the point correspondence relationship between two coordinates, and design a pose iterative refinement network suitable for the mode of the GPCL-Net network structure characteristics, so as to iteratively further improve the pose estimation accuracy; a scale parameter shared by the source point set and the target point set is introduced in feature extraction, the method has strong robustness to incomplete observation values and outlier data, and can improve the similarity of the feature distribution of the two point sets; a T-S FA feature alignment module and an S-T FA feature alignment module are designed, point features from different coordinate systems are decomposed into pose and matching features, then the two features are aligned along the channel dimension in groups, and finally a pose-matching feature pair is generated, the point correspondence relationship of the two point sets can be established based on the aligned feature pair; the method performs well on the challenging Occlusion-LineMOD and HB-Vision data sets, and achieves an accuracy of 73.7% and 79.4% respectively, which has significant superiority and progressiveness. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 It is two point correspondence relationships established between the target point set and the source point set in the background art of the application.

[0017] Figure 2 It is a general structure framework diagram of the GPCL-Net grouping point correspondence learning method for 6D pose estimation in a complex environment.

[0018] Figure 3 It is a pose iterative network schematic diagram of the GPCL-Net grouping point correspondence learning method for 6D pose estimation in a complex environment.

[0019] Figure 4is a comparison diagram of three feature extraction modules of the GPCL-Net grouping point correspondence learning method for 6D pose estimation in a complex environment according to the present application.

[0020] Figure 5 is a feature distribution extracted in point cloud data by different network layers of the GPCL-Net grouping point correspondence learning method for 6D pose estimation in a complex environment according to the present application; wherein, the first row: the feature distribution extracted from the source point set and the target point set is obviously different using batch normalization; the second row: using our method, the similarity of the feature distribution can be more accurately represented.

[0021] Figure 6 is a visualization result of different methods of the GPCL-Net grouping point correspondence learning method for 6D pose estimation in a complex environment on the YCB-Video dataset.

[0022] Figure 7 is a visualization result of different methods of the GPCL-Net grouping point correspondence learning method for 6D pose estimation in a complex environment on the HB-Vision dataset. DETAILED DESCRIPTION

[0023] The present application will be further described and illustrated below in conjunction with specific embodiments and the accompanying drawings.

[0024] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0025] In the description of the present application, it should be understood that the terms "upper", "lower", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0026] As Figure 2As shown, a GPCL-Net grouped point correspondence learning method for 6D pose estimation in complex environments is presented. The GPCL-Net grouped point correspondence learning method is based on a GPCL-Net grouped point correspondence learning network driven by a correlation normalized feature extraction module and a dual FA feature alignment module. The GPCL-Net grouped point correspondence learning network includes a correlation normalized feature extraction module, a dual FA feature alignment module, a pose estimation module, and a pose iteration module.

[0027] like Figure 2 As shown, a GPCL-Net grouped point correspondence learning method for 6D pose estimation in complex environments is presented. The GPCL-Net grouped point correspondence learning method specifically includes the following steps: S1. Correlation Normalization Feature Extraction: Based on the target point set Source point set As input, the association-normalized feature extraction module performs a set of target points... Source point set After centering, the target point sets are extracted by the output channel convolutional layer. Point-by-point features Source point set Point-by-point features For each point, the feature is processed point by point. , After normalization, the data is correlated and spliced ​​along the channel dimension to obtain the target point set. Correlation normalized pointwise features Source point set Correlation normalized pointwise features ; S2, using correlation-normalized point-by-point features , As input, the target point set is generated through the dual FA feature alignment module. Matching features Postural characteristics Source point set Matching features Postural characteristics Matching features , and posture characteristics , The system is divided into n channel groups, each group is executed independently, and attention graphs are generated for each group. and Then use the attention diagram. Matching features and posture characteristics Obtain the target point set by performing matrix multiplication. Matching features in object coordinate system and posture characteristics Use attention diagram Matching features and posture characteristics Matrix multiplication to obtain the source point set Matching features in object coordinate system and posture characteristics And correlated with normalized point-by-point features in the camera coordinate system , Align and generate matching feature pairs respectively ( and ) and pose features ( and ); S3, using matching feature pairs ( and ) and pose features ( and As input, the pose estimation module constructs a matching feature pair matrix. Pose feature pair matrix The matching feature pair matrix is ​​processed by MLP (Multilayer Perceptron). Generate confidence Pose features to matrix The data is transmitted to the MLP multilayer perceptron and then processed using confidence levels. Weighted regression yields 6D pose ; S4. Using 6D pose Update target point set , source set and the updated target point set The pose iteration module takes the source point set as input again. and the updated target point set The data is transmitted to the correlation normalization feature extraction module, the dual FA feature alignment module, and the pose estimation module to obtain pose variables. The pose iteration module completes multiple iterations of refinement, and the refined 6D pose is used as the final output target 6D pose.

[0028] Furthermore, such as Figure 2 As shown, a GPCL-Net grouped point correspondence learning method for 6D pose estimation in complex environments is presented. The correlation normalization feature extraction module in step S1 of the GPCL-Net grouped point correspondence learning method includes a four-layer Block_CB structure, and each Block_CB structure is equipped with a Correlation Normalization module and a BatchNormalization module.

[0029] Further, as shown in Figure 2 GPCL-Net grouping point correspondence learning method for 6D pose estimation in complex environment, the specific operation of step S1 of GPCL-Net grouping point correspondence learning method is: S11, the input target point set and the source point set , respectively, through the convolution layer with output channel to extract the point-by-point feature and : , , wherein, is the size of the point-by-point feature , is the number of points in the target point set , is the size of the point-by-point feature , is the number of points in the source point set , is the number of features contained in each point of the point-by-point feature and , is a one-dimensional convolution operation, is a point set with number, is a point set with number, is the point-by-point feature of the target point set , is the point-by-point feature of the source point set , is the feature of the th point in i , is the feature of the th point in i , the target point set is the point cloud data obtained by converting the RGB-D data collected by the depth camera through the camera intrinsic parameter; the source point set is the point cloud data obtained by uniformly sampling the CAD model; Correlation Normalization module normalizes the point-by-point feature and of each point of the target point set and the source point set : , , These are scale parameters shared within the first-level Block_CB structure. for Features after normalization for Features after normalization and Used to calculate point-by-point features and These are all process variables and have no practical significance; Scale parameters shared in the first-level Block_CB structure Depend on The calculation yielded: , in, Indicates the first i The first point j One feature, this parameter is derived from the target point set. Source point set shared, For the source point set The number of midpoints For point-by-point features and The number of features contained in each point; S12. Point-by-point features obtained after normalization by the Correlation Normalization module. and Then use the Batch Normalization module to perform batch normalization processing and Activate the normalized pointwise feature extraction for the first Block_CB structure. The normalized pointwise feature extraction for other Block_CB structures follows the same operation. , , , in, for Point-by-point features after normalization for Point-by-point features after normalization For quantity The set of points, For quantity The set of points, for Features after normalization for Features after normalization For activation function, BN is the abbreviation of Batch Normalzation, which means batch normalization, is the target point set is the batch normalization and activated point-wise feature output by the first layer Block_CB structure is the source point set is the batch normalization and activated point-wise feature output by the first layer Block_CB structure S13, batch normalizing the output of the four-layer Block_CB structure and activated point-wise feature and concatenating along the channel dimension to obtain the final extracted correlation normalized point-wise feature and : , wherein, is the correlation normalized point-wise feature of the target point set is the correlation normalized point-wise feature of the source point set is the batch normalization and activated point-wise feature output by the first layer, second layer, third layer and fourth layer Block_CB structure respectively, is the abbreviation of Correlation Normalization FeatureExtraction, which means the entire correlation normalization feature extraction module, specifically the correlation normalization operation, is the concatenation operation along the channel dimension. Further, as shown in , a GPCL-Net grouping point correspondence learning method for 6D pose estimation in a complex environment, the double FA feature alignment module in step S2 of the GPCL-Net grouping point correspondence learning method includes a T-S FA feature alignment module and an S-T FA feature alignment module.

[0030] Figure 2 Further, as shown in , a GPCL-Net grouping point correspondence learning method for 6D pose estimation in a complex environment, the specific operation of step S2 of the GPCL-Net grouping point correspondence learning method is:

[0031] S21, taking the correlation normalized point-wise feature of the target point set Figure 2 and the source point set and ​​​As input, the matching features and pose features of the target point set are generated by the S-T FA feature alignment module , and pose features of the source point set : , are the associated normalized per-point features of the target point set , are the associated normalized per-point features of the source point set , are the matching features of the target point set , are the matching features of the source point set , are the pose features of the target point set , are the pose features of the source point set , denotes a subnetwork of the MLP multi-layer perceptron; S22, in the double-FA feature alignment module, the matching features , and pose features , are divided into n channel groups according to the channel, and each group is independently executed, respectively, to generate the attention map in the T-S FA feature alignment module, and the attention map is generated in the S-T FA feature alignment module: , , , wherein, are the matching features of the th channel group of the target point set i , are the matching features of the th channel group of the source point set i , are the pose features of the th channel group of the source point set i , is an activation function, denotes a matrix transposition operation; S23, in the T-S FA feature alignment module, the attention map is respectively multiplied with the matching features and pose features to infer the target point set Matching features in object coordinate system and pose features , and associated normalized per-point features of target point set in camera coordinate system align: , , , where, is the associated normalized per-point features of target point set , is a formula operation process variable, which has no actual meaning, only for formula writing and operation, is the matching features of target point set in object coordinate system, is the pose features of target point set in object coordinate system, is the matching features of target point set , is the pose features of target point set , is the concatenation operation along the channel dimension, denotes the subnetwork of MLP multi-layer perceptron; In the S-T FA feature alignment module, the attention graph is respectively multiplied with the matching features and the pose features to infer the matching features and pose features of target point set in object coordinate system, and associated normalized per-point features of source point set in camera coordinate system align: , , , where, is the associated normalized per-point features of source point set , is a formula operation process variable, which has no actual meaning, only for formula writing and operation, is the matching features of source point set in object coordinate system, is the pose features of source point set in object coordinate system, is the matching features of source point set , is the pose features of source point set Part-to-part and all-to-all correspondence is realized through the T-S FA feature alignment module and the S-T FA feature alignment module, and then the two kinds of correspondence are used to generate matching feature pairs ( and ) and pose feature pairs ( and ) in the feature space.

[0032] Further, as shown in Figure 2 , a GPCL-Net grouping point correspondence learning method for 6D pose estimation in a complex environment, the specific operation of the multi-channel cross fusion method step S3 is: S31, taking the matching feature pairs ( and ) and the pose feature pairs ( and ) as inputs, the matching feature pair matrix and the pose feature pair matrix are constructed through the pose estimation module: , wherein, is the matching feature of the target point set in the object coordinate system, is the pose feature of the target point set in the object coordinate system, is the matching feature of the target point set , is the pose feature of the target point set , is the matching feature of the source point set in the object coordinate system, is the pose feature of the source point set in the object coordinate system, is the matching feature of the source point set , is the pose feature of the source point set ; S32, the matching feature pair matrix is processed through the MLP multi-layer perceptron to generate the confidence , the pose feature pair matrix is transmitted to the MLP multi-layer perceptron, and the 6D pose is obtained through weighted regression of the confidence : , , , , wherein, To indicate the first i A confidence level value. For the target point set The number of midpoints For the source point set The number of midpoints To achieve a summation along the border, For activation function, This represents a subnetwork of an MLP (Multilayer Perceptron). This is an intermediate variable in the pose estimation process, used to simplify the formula for easier understanding.

[0033] Furthermore, such as Figure 2 , 3 As shown, a GPCL-Net grouped point correspondence learning method for 6D pose estimation in complex environments, the specific operation of step S4 of the multi-channel cross-fusion method is as follows: S41. The 6D pose obtained is processed using the correlation normalization feature extraction module, the dual FA feature alignment module, and the pose estimation module. Update target point set iteration -1 target point set for: , in, For the target point set The number of points, For the target point set The Middle i One point, Iterating for the pose iteration module -1 times in 6D pose, when When =1, 6D pose This is the first time that a 6D pose has been obtained through the correlation normalization feature extraction module, the dual FA feature alignment module, and the pose estimation module. S42, Source Set and the updated target point set As input, it is transmitted to the correlation normalization feature extraction module to obtain iterative... -1 new association normalized pointwise features , : , in, "Correlation Normalization Feature Extraction" is an abbreviation for the entire correlation normalization feature extraction module, specifically referring to the correlation normalization operation. S43, Iteration -1 new association normalized pointwise features , The data is transmitted to the dual FA feature alignment module, where the TS FA feature alignment module and the ST FA feature alignment module establish partial-to-partial and full-to-full correspondences. These two correspondences are then used to generate iterative features in the feature space. -1 matching feature pairs ( and ) and pose features ( and Then it is transmitted to the pose estimation module to construct the iterative... -1 matching feature pair matrix Pose feature pair matrix : , in, For the updated target point set Iteration in the object coordinate system -1 matching features, For the updated target point set Iteration -1 matching features, For the updated target point set Iteration in the object coordinate system -1 posture characteristics, For the updated target point set Iteration -1 posture characteristics, For the source point set Iteration in the object coordinate system -1 matching features, For the source point set Iteration -1 matching features, For the source point set Iteration in the object coordinate system -1 posture characteristics, For the source point set Iteration -1 posture characteristics; S44. Iteration is processed through MLP multilayer perceptron. -1 matching feature pair matrix Generation Iteration -1 confidence level , will iterate -1 order pose feature pair matrix The data is transmitted to the MLP multilayer perceptron and then processed using confidence levels. Weighted regression is iterated -1 6D pose variables : , , , , , in, To achieve a summation along the border, For activation function, This represents a subnetwork of an MLP (Multilayer Perceptron). For iteration -1 is an intermediate variable in the pose estimation process, used to simplify the formula for easier understanding. For the 6D pose variables after one iteration; The pose iteration module is used to refine the iteration and obtain the iterations. The next iteration of 6D pose : , in, For iteration -1 weighted regression yielded 6D pose variables. For iteration -1 weighted regression yields the 6D pose. For iteration The weighted regression yields the 6D pose.

[0034] Experimental environment This experiment was conducted on a computer configured with an RTX A6000 GPU, and the training environment was set up using PyTorch. For the two input point sets, we set the number of point samples to 1024. During training, the parameters were... , , and The hyperparameters were set to 4.5, 1.5, 1.0, and 1.0 respectively to balance the hyperparameters. Then set it to 0.001; set the pose iteration count to 3; evaluate our method on three public datasets: YCB-Video, LineMOD, and Occlusion LineMOD, as well as the custom dataset HB-Vision: The YCB-Video dataset covers 21 object classes, each with a unique shape and texture. It includes 92 RGB-D videos that present subsets of the 21 object classes in different indoor scenes. In terms of data splitting, 16,189 frames were selected from 80 videos and used as the training set along with another 80,000 synthetic frames. The remaining 2,949 keyframes from the 12 videos were used as the test set. LineMOD dataset: consists of 13 videos, involving 13 low-texture objects, totaling 15780 frames. Each frame is a group, and each group contains the corresponding RGB image, depth image, object CAD model, and 6D pose annotation; the test set and training set are consistent with DenseFusion; Occlusion LineMOD dataset: based on the extension of the LineMOD dataset, it introduces different occlusion levels of interference on its basis, which contains 8 objects, and each scene is composed of multiple annotated objects that are severely occluded; HB-Vision dataset: uses Realsense D435i depth camera to shoot a video of about 40 seconds, with a shooting speed of 30 frames per second, a resolution of 1280x720, and saves picture information frame by frame; this dataset contains 6 kinds of household items, a total of 7050 groups, each group has corresponding RGB image, depth image, object 3D model and 6DOF pose annotation; the training set and test set are divided by 1:4, that is, 1410 groups with group number multiples of 5 are extracted as the training set, and the remaining 5640 groups are used as the test set; this dataset uses two modes: 4 to 5 different objects are selected, different placements are made under different light intensities, and the angle is adjusted during shooting, which simulates 30%~60% object occlusion and light changes to verify the performance of the proposed network in complex environments.

[0035] Evaluation index Since there is a significant difference between the pose estimation of asymmetric objects and symmetric objects, we use ADD and ADD-S to evaluate our network; first, for asymmetric objects, we use the average distance of model points (ADD), that is, the distance between the points sampled in the predicted pose transformed model and the corresponding points in the real pose model; the definition of ADD is as follows: , Second, for symmetric objects, the average distance of the nearest points (ADD-S) is used, that is, the distance between each point in the estimated model direction and the nearest point on the ground real model is minimized; the definition of ADD-S is as follows: , Wherein, represents the i-th point in the randomly sampled three-dimensional points from the three-dimensional CAD model of the object, is the ground truth real pose, is the pose after the network predicts the pose transformation, represents the predicted pose of the i-th pixel point after the pose transformation.

[0036] ​​The evaluation metrics used can vary slightly for different datasets; in the YCB-Video dataset, the average distance of the closest point (ADD-S) is used to measure the pose error, and two metrics are reported: the area under the ADD-S curve AUC, with a maximum threshold of 0.1 m; the percentage of ADD-S less than 2 cm (<2 cm); in the LineMOD, Occlusion LineMOD, and HB-Vision datasets, the percentage of distances less than 10% of the object diameter (ADD-0.1d) is reported.

[0037] Ablation experiments and analysis To verify the effectiveness of each component in the GPCL-Net of the present application, and to ensure fairness, the same evaluation metrics are used on the YCB-Video dataset in line with the method. The correlation point correspondence learning network proposed in the present application has certain advantages compared to some existing methods, mainly in the design of the three components: (1) CNFE effectively improves the similarity of the feature distributions of the target point set and the source point set by sharing parameters; (2) DFGA can simultaneously establish two types of point correspondence and fully utilize the detailed features by using a grouping and synchronous approach to establish high-quality point correspondence; (3) the designed iterative refinement network greatly improves the pose estimation effect. Therefore, we conducted experimental evaluation on the above three components, and the ablation experiment results are shown in Table 1.

[0038] Table 1 Ablation experiment on the YCB-Video dataset for CNFE, DFGA, and pose iterative network

[0039] Based on the effectiveness analysis of CNFE. We compared CNFE with batch normalization-based feature extraction module (BNFE) and feature extraction module from Backbone (Backbone) to prove the effectiveness of increasing the similarity of the feature distributions of the two point sets. The structures of the three feature extraction modules are shown in Figure 4 To ensure the simplicity and fairness of the experiment, we did not perform the pose iterative process. As shown in Table 1, CNFE improved the evaluation metrics compared to BNFE (B0 vs. D0) and Backbone (B1 vs. D0). This indicates that CNFE can more effectively extract features and provide higher-quality feature inputs for subsequent pose estimation tasks, thereby effectively improving the algorithm performance.

[0040] Effectiveness analysis of DFGA. To explore the effectiveness of DFGA, we conduct a series of ablation experiments. Also, to ensure simplicity and fairness, we do not perform the pose iteration process. First, we compare DFGA with T-S FGA (D0 vs. C0) and S-T FGA (D0 vs. C1), and the evaluation index AUC is increased by 1% and 1.4%, respectively, and the evaluation index of ADD-S less than 2 cm (<2 cm) is increased by 0.4% and 5%. Thus, compared with only establishing one point correspondence, simultaneously establishing "partial-to-partial" and "complete-to-complete" point correspondences has a great advantage in improving the accuracy of 6D pose estimation. In addition, under the condition of simultaneously establishing two kinds of point correspondences, DFGA further demonstrates the advantage of DFGA in detail feature processing and point correspondence quality compared with FDA (D0 vs. C2).

[0041] Effectiveness analysis based on pose iteration refinement. To compare the performance of the pose iteration network proposed in the present application, we introduce the iteration network in DCL-Net and design three groups of comparative experiments: (1) without introducing any iteration network; (2) introducing the iteration refinement network of DCL-Net; (3) introducing the pose iteration network proposed in the present application. The iteration refinement in DCL-Net is: directly using the feature of the source point set and the target point set after updating to perform feature stitching, and then using the first N elements of the confidence to weight them and then regress the feature after iteration. However, the iteration refinement designed in the present application (as shown in Figure 5 ) is different, and the feature of the source point set and the target point set is not used, but the correlation normalization module is selected again in the new iteration, so that the source point set and the target point set again have the same feature space distribution, and then the feature alignment and pose estimation steps are performed again to obtain the pose feature pair , and then the confidence parameter is used to weight them, and finally the pose after iteration is regressed. As can be seen from Table 1, compared with not introducing an iteration network (A0 vs. D0 and D1 vs. D0) and the iteration network from DCL-Net (A0 vs. D1), both iteration modes can adapt to the network of the present application and achieve the effect of pose iteration refinement, but the pose iteration network proposed in the present application has a higher degree of adaptation to CLAP-Net and can achieve higher pose estimation accuracy.

[0042] Comparison with existing methods To further demonstrate the performance advantage of the proposed holistic network, we compare our method with some existing RGB-D based pose estimation methods on YCB-Video, LineMOD, Occlousion LineMOD and HB-Vision datasets.

[0043] The comparative results on YCB-Video dataset. We compare our method with some existing methods. The results are shown in Table 2. Compared with DenseFusion using iterative refinement, our network improves the ADD-S(AUC) and ADD-S(<2 cm) by 3.1% and 2.3% respectively without using iterative refinement, and by 3.9% and 2.7% respectively with using iterative refinement. Compared with some recent methods PVN3D, DCL-Net and FFB6D, our method also shows higher ADD-S(AUC) and ADD-S(<2 cm) values. Meanwhile, we also show the qualitative visualization results of different methods on YCB-Video dataset as shown in Fig. 2, where the sampling points of CAD model are transformed by the estimated pose and projected to 2D image. From the figure, we can also see the significant advantage of our method in handling occlusion cases. Figure 6

[0044] Table 2. Results of different methods on YCB-Video dataset. The evaluation metrics are ADD(AUC) and ADD(<2 cm). The class names in bold represent symmetric objects, and the values in bold represent the maximum value in each row

[0045] The comparative results on LineMOD dataset. As can be seen in Table 3, compared with some existing methods, our method achieves the best results in most categories of the LineMOD dataset, and also reaches 99.7% in the ADD-0.1d metric.

[0046] Table 3. Results of different methods on LineMOD dataset in ADD-0.1d. The class names in bold represent symmetric objects, and the values in bold represent the maximum value in each row

[0047] ​Comparative experiment results on Occlousion LineMOD dataset. To verify the performance of our method in the occlusion state, comparative experiments were conducted on this dataset. The results are shown in Table 4. It can be seen that without pose iteration processing, compared with the previous methods PoseCNN and HybridPose, we improved by 43.5% and 20.9% respectively on ADD-0.1d. After using pose iteration, our method is also superior to the recent method, reaching 73.7%.

[0048] Comparative experiment results on HB-Vision dataset. To verify the performance of our method in the light change and occlusion state, comparative experiments were conducted on the HB-Vision dataset, and the results are shown in Table 5. Without using iterative refinement, we reached 72% on the ADD-0.1d index, which improved by 5.1%, 3% and 0.8% compared with PVN3D, FFB6D and GCPose respectively. After using iterative refinement, we improved to 79.4% on the index, which improved by 7.4% compared with the non-iterative case, and improved by 5.4% compared with DCL-Net. In addition, Figure 7 Some visualization effect figures of some methods on this dataset are also shown in the HB-Vision dataset. From the comparison in the first row, it can be seen that in the dark light environment, the pose estimation effect of our method is obviously better than that of other methods. In the second row, the excellent performance of our method in the occlusion state is also shown.

[0049] Table 4 ADD-0.1d results of different methods on Occlousion LineMOD dataset. The class name in bold represents the symmetric object, and the numerical value in bold represents the maximum value in each row

[0050] Table 5 ADD-0.1d results of different methods on HB-Vision dataset. The bold sign represents the maximum value in each row

[0051] The corresponding relationship is established between the target point cloud converted from the RGB-D image and the source point cloud uniformly sampled from the CAD model, which provides a new idea for 6D pose estimation. Based on this idea, a new grouping point correspondence learning network (GPCL-Net) is proposed. The method first introduces a correlation normalization feature extraction strategy, effectively reduces the feature distribution difference between the source and target point sets by sharing a scale parameter, and ensures a certain feature similarity between the two. Then, the T-S FA feature alignment module and the S-T FA feature alignment module are designed to realize the construction of the "part-to-part" and "all-to-all" correspondence between the two point sets, while ensuring the correspondence quality and improving the robustness of the network in complex environments. At the same time, due to the high adaptability of the designed pose iterative network and GPCL-Net, the 6D pose estimation accuracy can be significantly improved. We evaluate the performance of the proposed network on four datasets: YCB-Video, LineMOD, Occlusion-LineMOD and HB-Vision. The results show that the performance of GPCL-Net on each dataset is better than that of existing methods, verifying its effectiveness and robustness in complex scenarios.

[0052] In the present application, the GPCL-Net grouping point correspondence learning method for 6D pose estimation in complex environments can simultaneously establish "part-to-part" and "all-to-all" point correspondence in the feature space, utilize the complementary advantages to achieve the purpose of the network solving the pose by establishing the point correspondence between the two coordinates, and design a pose iterative refinement network that adapts to the mode of the GPCL-Net network structure to further improve the pose estimation accuracy. In feature extraction, a scale parameter shared by the source point set and the target point set is introduced. This method has strong robustness to incomplete observations and outliers, and can improve the similarity of the feature distribution of the two point sets. The T-S FA feature alignment module and the S-T FA feature alignment module are designed to decompose the point features from different coordinate systems into pose and matching features, then group them along the channel dimension for alignment of the two features, and finally generate pose-matching feature pairs. Based on the aligned feature pairs, the point correspondence of the two point sets can be established. The method performs well on the highly challenging Occlusion-LineMOD and HB-Vision datasets, achieving an accuracy of 73.7% and 79.4% respectively, with significant superiority and progressiveness.

[0053] The technical solutions disclosed by the embodiments of the present application are described in detail above, the principles and implementation manners of the embodiments of the present application are described by applying the specific embodiments, the above embodiment descriptions are only applicable to help understand the principles of the embodiments of the present application; meanwhile, for the general technical personnel in the art, according to the embodiments of the present application, the specific implementation manners and application scopes will be changed, and the above descriptions should not be understood as the limitation of the present application.

Claims

1. A GPCL-Net grouped point correspondence learning method for 6D pose estimation in complex environments, characterized in that, The GPCL-Net group point correspondence learning method is based on the GPCL-Net group point correspondence learning network driven by the association normalized feature extraction module and the dual FA feature alignment module. The GPCL-Net group point correspondence learning network includes an association normalized feature extraction module, a dual FA feature alignment module, a pose estimation module, and a pose iteration module. The GPCL-Net group point correspondence learning method specifically includes the following steps: S1. Correlation Normalization Feature Extraction: Using the target point set Source point set As input, the association-normalized feature extraction module performs a set of target points... Source point set After centering, the target point sets are extracted by the output channel convolutional layer. Point-by-point features Source point set Point-by-point features For each point, the feature is processed point by point. , After normalization, the data is correlated and spliced ​​along the channel dimension to obtain the target point set. Correlation normalized pointwise features Source point set Correlation normalized pointwise features ; S2, using correlation-normalized point-by-point features , As input, the target point set is generated through processing by the dual FA feature alignment module. Matching features Postural characteristics Source point set Matching features Postural characteristics Matching features , and posture characteristics , The system is divided into n channel groups, each group is executed independently, and attention graphs are generated for each group. and Then use the attention diagram. Matching features and posture characteristics Obtain the target point set by performing matrix multiplication. Matching features in object coordinate system and posture characteristics Use attention diagram Matching features and posture characteristics Matrix multiplication to obtain the source point set Matching features in object coordinate system and posture characteristics And correlated with normalized point-by-point features in the camera coordinate system , Align and generate matching feature pairs respectively ( and ) and pose features ( and ); S3, using matching feature pairs ( and ) and pose features ( and As input, the pose estimation module constructs a matching feature pair matrix. Pose feature pair matrix The matching feature pair matrix is ​​processed by MLP (Multilayer Perceptron). Generate confidence Pose features to matrix The data is transmitted to the MLP multilayer perceptron and then processed using confidence levels. Weighted regression yields 6D pose ; S4. Using 6D pose Update target point set , source set and the updated target point set The pose iteration module takes the source point set as input again. and the updated target point set The data is transmitted to the correlation normalization feature extraction module, the dual FA feature alignment module, and the pose estimation module to obtain pose variables. The pose iteration module completes multiple iterations of refinement, and the refined 6D pose is used as the final output target 6D pose.

2. The GPCL-Net grouped point correspondence learning method for 6D pose estimation in complex environments according to claim 1, characterized in that, The correlation normalization feature extraction module in step S1 of the GPCL-Net group point correspondence learning method includes a four-layer Block_CB structure, and each Block_CB structure is equipped with a Correlation Normalization module and a Batch Normalization module.

3. The GPCL-Net grouped point correspondence learning method for 6D pose estimation in complex environments according to claim 2, characterized in that, The specific operation of step S1 in the GPCL-Net group point correspondence learning method is as follows: S11. Input target point set Source point set Each is output through a separate channel. Point-by-point features are extracted from convolutional layers. and : ， ， in, For point-by-point features Size, For the target point set The number of midpoints For point-by-point features Size, For the source point set The number of midpoints For point-by-point features and The number of features contained in each point. This is a one-dimensional convolution operation. For quantity The set of points, For quantity The set of points, For the target point set Point-by-point features, For the source point set Point-by-point features, for The Middle i Features of each point for The Middle i Features of each point, target point set This refers to point cloud data derived from RGB-D data acquired by a depth camera through camera intrinsic parameter conversion; source point set Point cloud data obtained by uniform sampling from a CAD model; The Correlation Normalization module applies the target point set Source point set Point-by-point features and Normalization is performed: ， ， in, These are scale parameters shared within the first-level Block_CB structure. for Features after normalization for Features after normalization and Used to calculate point-by-point features and These are all process variables and have no practical significance; Scale parameters shared in the first-level Block_CB structure Depend on The calculation yielded: ， in, Indicates the first i The first point j One feature, this parameter is derived from the target point set. Source point set shared, For the source set The number of midpoints For point-by-point features and The number of features contained in each point; S12. Point-by-point features obtained after normalization by the Correlation Normalization module. and Then use the Batch Normalization module to perform batch normalization processing and Activate the normalized pointwise feature extraction for the first Block_CB structure. The normalized pointwise feature extraction for other Block_CB structures follows the same operation. ， ， ， in, for Point-by-point features after normalization for Point-by-point features after normalization For quantity The set of points, For quantity The set of points, for Features after normalization for Features after normalization For activation function, This is an abbreviation for Batch Normalization, which stands for Batch Normalization. For the target point set After batch normalization of the output of the first-level Block_CB structure and Activated pointwise features, For the source set After batch normalization of the output of the first-level Block_CB structure and Activated pointwise features; S13. Normalize the output batch of the four-layer Block_CB structure and... Activated pointwise features and By concatenating along the channel dimension, we obtain the extracted, finally normalized, point-by-point features. and : ， in, For the target point set The correlation normalized pointwise features, For the source set The correlation normalized pointwise features, These are the batch normalizations of the outputs of the Block_CB structures in the first, second, third, and fourth layers, respectively. Activated pointwise features, "Correlation Normalization Feature Extraction" is an abbreviation for the entire correlation normalization feature extraction module, specifically referring to the correlation normalization operation. This is a splicing operation along the channel dimension.

4. The GPCL-Net grouped point correspondence learning method for 6D pose estimation in complex environments according to claim 1, characterized in that, The dual FA feature alignment module in step S2 of the GPCL-Net grouping point correspondence learning method includes a TS FA feature alignment module and an ST FA feature alignment module.

5. The GPCL-Net grouped point correspondence learning method for 6D pose estimation in complex environments according to claim 4, characterized in that, The specific operation of step S2 in the GPCL-Net group point correspondence learning method is as follows: S21, with target point set Source point set Correlation normalized pointwise features and As input, the target point set is generated through the TS FA feature alignment module and processed by the MLP multilayer perceptron. Matching features and posture characteristics The source point set is generated through the ST FA feature alignment module. Matching features and posture characteristics : , in, For the target point set The correlation normalized pointwise features, For the source set The correlation normalized pointwise features, For the target point set Matching features For the source set Matching features For the target point set Postural characteristics, For the source set Postural characteristics, This represents a subnetwork of an MLP (Multilayer Perceptron). S22. In the dual FA feature alignment module, the matching features are... , and posture characteristics , The channel is divided into n channel groups, and each group is executed independently, generating attention maps in the TS FA feature alignment module. Attention maps are generated in the ST FA feature alignment module. : , , , in, For the target point set The i Matching features of each channel group For the source set The i Matching features of each channel group For the source set The i The pose characteristics of each channel group For activation function, This represents the matrix transpose operation; S23. In the TS FA feature alignment module, use attention diagrams. respectively with matching features Postural characteristics Inferring the target point set by performing matrix multiplication Matching features in object coordinate system and posture characteristics and the target point set in the camera coordinate system Correlation normalized pointwise features Alignment: , , ， in, For the target point set The correlation normalized pointwise features, These are variables used in the formula calculation process; they have no practical meaning and are only used to facilitate formula writing and calculation. For the target point set Matching features in the object coordinate system For the target point set Pose characteristics in the object coordinate system For the target point set Matching features For the target point set Postural characteristics, This is a splicing operation along the channel dimension. This represents a subnetwork of an MLP (Multilayer Perceptron). In the ST FA feature alignment module, attention diagrams are used. respectively with matching features and posture characteristics Inferring the source point set by performing matrix multiplication Matching features in object coordinate system and posture characteristics and the source point set in the camera coordinate system Correlation normalized pointwise features Alignment: , , , in, For the source set The correlation normalized pointwise features, These are variables used in the formula calculation process; they have no practical meaning and are only used to facilitate formula writing and calculation. For the source set Matching features in the object coordinate system For the source set Pose characteristics in the object coordinate system For the source set Matching features For the source set Postural characteristics; The TS FA feature alignment module and the ST FA feature alignment module are used to achieve partial-to-partial and full-to-full correspondences, and then matching feature pairs are generated in the feature space based on these two correspondences. and ) and pose features ( and ).

6. The GPCL-Net grouped point correspondence learning method for 6D pose estimation in complex environments according to claim 1, characterized in that, The specific operation of step S3 in the multi-channel cross-fusion method is as follows: S31, using matching feature pairs ( and ) and pose features ( and As input, the pose estimation module constructs a matching feature pair matrix. Pose feature pair matrix : ， in, For the target point set Matching features in the object coordinate system For the target point set Pose characteristics in the object coordinate system For the target point set Matching features For the target point set Postural characteristics, For the source set Matching features in the object coordinate system For the source set Pose characteristics in the object coordinate system For the source set Matching features For the source set Postural characteristics; S32. Processing the matching feature pair matrix using an MLP (Multilayer Perceptron) Generate confidence Pose features to matrix The data is transmitted to the MLP multilayer perceptron and then processed using confidence levels. Weighted regression yields 6D pose : ， ， ， ， in, To indicate the first i A confidence level value. For the target point set The number of midpoints For the source set The number of midpoints To achieve a summation along the border, For activation function, This represents a subnetwork of an MLP (Multilayer Perceptron). This is an intermediate variable in the pose estimation process, used to simplify the formula for easier understanding.

7. The GPCL-Net grouped point correspondence learning method for 6D pose estimation in complex environments according to claim 1, characterized in that, The specific operation of step S4 in the multi-channel cross-fusion method is as follows: S41. The 6D pose obtained is processed using the correlation normalization feature extraction module, the dual FA feature alignment module, and the pose estimation module. Update target point set iteration -1 target point set for: ， in, For the target point set The number of points, For the target point set The Middle i One point, Iterating for the pose iteration module -1 times in 6D pose, when When =1, 6D pose This is the first time that a 6D pose has been obtained through the correlation normalization feature extraction module, the dual FA feature alignment module, and the pose estimation module. S42, Source Set and the updated target point set As input, it is transmitted to the correlation normalization feature extraction module to obtain iterative... -1 new association normalized pointwise features , : ， in, "Correlation Normalization Feature Extraction" is an abbreviation for the entire correlation normalization feature extraction module, specifically referring to the correlation normalization operation. S43, Iteration -1 new association normalized pointwise features , The data is transmitted to the dual FA feature alignment module, where the TS FA feature alignment module and the ST FA feature alignment module establish partial-to-partial and full-to-full correspondences. These two correspondences are then used to generate iterative features in the feature space. -1 matching feature pairs ( and ) and pose features ( and Then it is transmitted to the pose estimation module to construct the iterative... -1 matching feature pair matrix Pose feature pair matrix : ， in, For the updated target point set Iteration in the object coordinate system -1 matching features, For the updated target point set Iteration -1 matching features, For the updated target point set Iteration in the object coordinate system -1 posture characteristics, For the updated target point set Iteration -1 posture characteristics, For the source set Iteration in the object coordinate system -1 matching features, For the source set Iteration -1 matching features, For the source set Iteration in the object coordinate system -1 posture characteristics, For the source set Iteration -1 posture characteristics; S44. Iteration is processed through MLP multilayer perceptron. -1 matching feature pair matrix Generation Iteration -1 confidence level , will iterate -1 order pose feature pair matrix The data is transmitted to the MLP multilayer perceptron and then processed using confidence levels. Weighted regression is iterated -1 6D pose variables : , ， ， ， ， in, To achieve a summation along the border, For activation function, This represents a subnetwork of an MLP (Multilayer Perceptron). For iteration -1 is an intermediate variable in the pose estimation process, used to simplify the formula for easier understanding. For 6D pose variables in one iteration; The pose iteration module is used to refine the iteration and obtain the iterations. The next iteration of 6D pose : , in, For iteration -1 weighted regression yielded 6D pose variables. For iteration -1 weighted regression yields the 6D pose. For iteration The weighted regression yields the 6D pose.

Images