A point cloud registration method and system based on global topology-aware interaction

Abstract

This invention discloses a point cloud registration method and system based on global topology-aware interaction, belonging to the 3D application field, including a training phase and an application phase. The training phase includes: training a point cloud registration model using a dataset, where training samples include source and target point clouds; the point cloud registration model includes: a local embedding feature extraction module, a global structural feature extraction module, a topology-aware interaction module, and a transformation matrix prediction module; wherein, the topology-aware interaction module is used to calculate geometric similarity metrics and project each point in the point cloud to its relative position, fusing the relative position features into the corresponding global structural features to obtain the corresponding position-aware features; topology-aware self-attention and cross-attention operations are applied to the position-aware features to extract the contextual interaction features of the point cloud. This invention solves the challenges of local structural similarity and matching ambiguity in unsupervised point cloud registration, improving the accuracy of point cloud registration.

Description

Technical Field

[0001] This invention belongs to the field of 3D applications, and more specifically, relates to a point cloud registration method and system based on global topology-aware interaction. Background Technology

[0002] Point cloud registration is a transformation estimation problem between point clouds, and the key lies in estimating the transformation matrix between two scanned point clouds. In recent years, the development of deep learning and its application in point cloud registration have greatly improved the robustness and efficiency of registration. Simultaneously, with the rapid development of high-precision sensors such as LiDAR and Kinect, point clouds have become the primary data format for representing the 3D world. Since sensors can only scan and acquire point cloud data within a limited field of view, registration algorithms are needed to generate a complete 3D scene. Based on the transformation matrix, partial scanned point clouds of the same 3D scene or object are merged into a complete 3D point cloud.

[0003] Unsupervised point cloud registration aims to align two point clouds (source and target point clouds) without ground reality transformations. Based on the type of features extracted, existing point cloud registration methods are divided into global and local methods. Compared to global methods, which are prone to local convergence issues, local methods have advantages. However, existing local methods, such as those that extract local geometric features of the point cloud and treat point cloud registration as a reinforcement learning task, or those that filter outliers based on differences in local structure, lack the interaction information between features due to similar local structural information, leading to incorrect feature matching results. Furthermore, both global and local methods simply use existing rotational variance extractors for feature extraction, which cannot collect global interaction information, also resulting in fuzzy feature matching. Summary of the Invention

[0004] To address the shortcomings and improvement needs of existing technologies, this invention provides a point cloud registration method and system based on global topology-aware interaction, with the aim of improving the accuracy of point cloud registration.

[0005] To achieve the above objectives, according to a first aspect of the present invention, a point cloud registration method based on global topology-aware interaction is provided, comprising:

[0006] Training phase: The point cloud registration model is trained using a dataset, where training samples include source point cloud S and target point cloud T; ​​wherein, the point cloud registration model includes:

[0007] The local embedding feature extraction module is used to extract the local embedding features F corresponding to the source point cloud S and the target point cloud T. s and F t ;

[0008] The global structural feature extraction module is used to extract features based on the local embedding features F. s and F t Construct the corresponding global graph g respectively s and g t In order to obtain the corresponding global structural features and

[0009] The topology-aware interaction module is used to calculate the global structural features. and The geometric similarity measure Θ between the features is used, and each point in the source point cloud S and the target point cloud T is projected to its relative position based on the geometric similarity measure Θ. and and relative position and Features are fused into the corresponding global structural features and In this process, the corresponding location-aware features are obtained. and and location-aware features and Topology-aware self-attention and cross-attention operations are employed to extract the contextual interaction features Φ of each point in the source point cloud S and the target point cloud T. s and Φ t ;

[0010] The transformation matrix prediction module is used to employ context-interactive features Φ s and Φ t Predict the transformation matrix R and the translation vector p;

[0011] The application phase includes: inputting the source point cloud S and target point cloud T of the actual object to be measured into the trained point cloud registration model to obtain the trained transformation matrix R and translation vector p; and registering the source point cloud S with the trained transformation matrix R and translation vector p.

[0012] Furthermore, in the topology-aware interaction module, location-aware features are... and Topology-aware self-attention and cross-attention operations are employed to extract the contextual interaction features Φ of each point in the source point cloud S and the target point cloud T. s and Φ t ,include:

[0013] S31. Calculate the feature vector query Q, keyword K, and value V using multi-head attention operation: in, Take respectively or Take respectively or These are the corresponding projection weights;

[0014] S32. Based on the feature vector, query Q, keyword K, and value V to perceive the features at the corresponding locations. and Performing topology-aware self-attention operation yields the self-interaction features of each point in the source point cloud S. Self-interaction features of each point in the target point cloud T

[0015] S33. Self-interaction features of each point in the source point cloud S. Perform topology-aware cross-attention operations to obtain cross-interaction features.

[0016] S34, the cross-interaction features With global structural features Feature merging is performed to obtain the contextual interaction features Φ of each point in the source point cloud S. s ; The self-interaction features of each point in the target point cloud T With global structural features Feature merging is performed to obtain the contextual interaction features Φ of each point in the target point cloud T. t .

[0017] Furthermore, in S32, based on the feature vector query Q, keyword K, and value V, the corresponding location-aware features are obtained respectively. and When performing topology-aware self-attention operation, the method further includes reweighting the weights of the relationships between features using the geometric similarity measure Θ; wherein the reweighting value is:

[0018]

[0019] Where dhead = d / H, H represents the number of heads in multi-head attention, d represents the feature dimension, and φ represents a linear layer with ReLU activation.

[0020] Furthermore, in the topology-aware interaction module, the global structural features are calculated and represented. and The geometric similarity measure Θ between features includes:

[0021] Calculate the global structural features and The affinity matrix A between them, where, W is the weight matrix;

[0022] The Sinkhorn algorithm is used to iteratively normalize the rows and columns of the affinity matrix until the Sinkhorn algorithm converges, thus obtaining the geometric similarity measure Θ.

[0023] Furthermore, in the topology-aware interaction module, each point in the source point cloud S and the target point cloud T is projected to its relative position according to the geometric similarity metric Θ. and include:

[0024] Perform a softmax operation along the rows and columns of the geometric similarity metric Θ to obtain the corresponding projection matrix Θ. s and Θ t ;

[0025] Using the projection matrix Θ s Project each point in the source point cloud S to its corresponding position. Using the projection matrix Θ t Project each point in the target point cloud T to its corresponding position.

[0026] Furthermore, in the local embedding feature extraction module, the global graph g s and g t They are respectively:

[0027] g s =(F s E s ), g t =(F t E t )

[0028] E s =softmax{<(F s ) T F s >}

[0029] E t =softmax{<(F t ) T F t >}

[0030] Among them, g s To base on local embedding features F s The constructed global graph, g t To base on local embedding features F t Constructed global graph; local embedding features F s and F t These are the nodes in the corresponding global graph, E s and E tThese represent the edges of the corresponding global graph; <·,·> represent inner product operations.

[0031] Furthermore, according to the global graph g s and g t In order to obtain the corresponding global structural features and include:

[0032] S21. Use in-graph convolution operations to extract the global graph g respectively. s and g t Corresponding global relation features and

[0033] S22. Calculate global relational features and The point-to-point Euclidean distance D between them i，j The distance is then normalized to obtain a normalized distance matrix; softmax operations are performed along the rows and columns of the normalized distance matrix to obtain the projection matrix P of the target point cloud T. t The projection matrix P of the source point cloud S s ;

[0034] S23. Update the global graph g using the projection matrix and global relation features. s and g t Corresponding to and

[0035]

[0036]

[0037] in, and ε s For the updated global graph Nodes and edges in the data; and ε t For the updated global graph Nodes and edges in the data;

[0038] S24. Use cross-graph convolution operation to extract the updated global graph. and The corresponding global structural features and

[0039] Furthermore, in the local embedding feature extraction module, the local embedding features F corresponding to the source point cloud S and the target point cloud T are extracted. s and F t ,include:

[0040] S11. Obtain the k-NN point set of the i-th point in the point cloud;

[0041] S12. Calculate the relative positions between adjacent points in the point set to establish k-NN connections;

[0042] S13. Perform convolution and max pooling operations on the established k-NN connections to obtain local embedding features; wherein, the point cloud is the source point cloud S or the target point cloud T, and the local embedding features are the local embedding features F of each point in the source point cloud S. s Or the local embedding features F of each point in the target point cloud T t .

[0043] According to a second aspect of the present invention, a point cloud registration system based on global topology-aware interaction is provided, including a computer-readable storage medium and a processor;

[0044] The computer-readable storage medium is used to store executable instructions;

[0045] The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method described in any of the first aspects.

[0046] According to a third aspect of the invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the first aspects.

[0047] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects:

[0048] (1) The point cloud registration method based on global topology-aware interaction of the present invention takes into account that the structural relationship between points in the point cloud and other points is ordered, while the position coordinates of points in the point cloud are unordered. Therefore, if the position coordinate information of points is directly extracted during local feature extraction, the extraction of structural information will be destroyed. To address this, this invention extracts only the local structural features between points in the point cloud during local embedding feature extraction. Based on these local structural features, a corresponding global graph is then built to extract global structural features. Based on the extracted global structural features, a geometric similarity metric Θ is calculated between the global structural features of the source and target point clouds. Using this Θ, the original position coordinates of points in the source and target point clouds are mapped to their relative positions, ensuring that points at two completely different locations in the source and target point clouds are mapped to the same feature space. In this case, the extracted relative position features do not destroy the structural information of the point cloud. The relative position information is then fused into the corresponding global structural features, and self-attention and cross-attention operations are performed to obtain the semantic interaction features between each point in the point cloud and other points. These semantic features contain both global and local structural and positional information; that is, they include local and global features, as well as feature associations between them. This enables topological awareness of both local and global features, allowing for accurate extraction of local structural information and accurate matching of global features, thereby improving the accuracy of point cloud registration.

[0049] (2) Furthermore, since the position coordinate information of the point cloud only affects the calculation of the weight of the relationship between structural features (the product of query Q and keyword K), this invention uses position-aware features when calculating the feature vector query Q, keyword K, and value V. Calculate query Q and keyword K using global structural features. The calculated value V ensures that the value of the feature itself is not affected when topological self-aware attention is performed, thus guaranteeing the consistency of the feature before and after the introduction of position coordinate information.

[0050] (3) Further, when calculating the self-interaction features of point clouds, the role of geometric similarity measure Θ is introduced into the attention mechanism. Geometric similarity measure Θ is used to reweight the mutual relationship weights between global structural features. Geometric similarity measure reflects the geometric similarity between features, encouraging feature pairs with strong topological similarity to have high attention scores, thus making the matching more accurate.

[0051] (4) Further, when calculating global structural features, graph convolution is first performed inside the point cloud to obtain global relational features. and Based on global relationship characteristics and The relationship between the points is used to obtain the projection matrix. The global graph is updated using the projection matrix and global relationship features. The new global graph is then subjected to cross-graph convolution operation to achieve graph convolution between the points, so that the obtained features can fully reflect the global structural features of the points, further avoiding registration errors caused by local similarity and achieving higher registration accuracy.

[0052] In summary, this invention can improve the discriminative power of features, solve the challenges of local structural similarity and matching ambiguity in unsupervised point cloud registration, and achieve more advanced unsupervised point cloud registration. Attached Figure Description

[0053] Figure 1 This is a flowchart of the point cloud registration method based on global topology-aware interaction according to the present invention.

[0054] Figure 2 A flowchart illustrating the global structural relationships provided by this invention.

[0055] Figure 3 A flowchart illustrating the context topology interaction provided by this invention.

[0056] Figure 4 This refers to the topology-aware self-attention process in this invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0058] In this invention, the terms "first," "second," etc., used in the invention and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0059] like Figure 1 As shown, the point cloud registration method based on global topology-aware interaction of the present invention includes: a training phase and an application phase;

[0060] The training phase includes:

[0061] The point cloud registration model is trained using a dataset, in which the training samples include source point cloud S and target point cloud T. The point cloud registration model includes: a local embedding feature extraction module, a global structural feature extraction module, a topology-aware interaction module, and a transformation matrix prediction module.

[0062] The local embedding feature extraction module is used to extract the local embedding features F of each point in the source point cloud S. s And the local embedding features F of each point in the target point cloud T. t ;

[0063] The global structural feature extraction module is used to extract features based on local embedding features F. s and F t Construct the corresponding global graph g respectively s and g t To obtain the global structural features corresponding to each point in the source point cloud S and the target point cloud T. and

[0064] The topology-aware interaction module is used to compute and represent global structural features. and The geometric similarity measure Θ between the features is used; and based on the geometric similarity measure Θ, each point in the source point cloud S and the target point cloud T is projected to its relative position. and and relative position and Features are fused into the corresponding global structural features and In this process, the corresponding location-aware features are obtained. and and location-aware features and Topology-aware self-attention and cross-attention operations are used to extract the contextual interaction features Φ of each point in the source point cloud S and the target point cloud T. s and Φ t ; where Φ s Φ represents the semantic interaction features between each point in the source point cloud S and other points. t This represents the semantic interaction features between each point in the target point cloud T and other points;

[0065] The transformation matrix prediction module is used to employ context-interactive features Φ s and Φ t Predict the transformation matrix R and the translation vector p;

[0066] The application phase includes: inputting the source point cloud S and target point cloud T of the actual object to be tested into the trained point cloud registration model to obtain the trained transformation matrix R and translation vector p; and registering the source point cloud S with the trained transformation matrix R and translation vector p; wherein, the object to be tested in this invention can be an aircraft, robot, automobile, etc.

[0067] Specifically, the local embedding feature extraction module extracts the local embedding features of each point in the point cloud, including:

[0068] S11. Obtain the set of k nearest neighbors (k-NN) of the i-th point in the point cloud;

[0069] S12. Calculate the relative positions between adjacent points in the point set to establish k-NN connections;

[0070] S13. Perform convolution and max pooling operations on the established k-NN connections to obtain local embedding features; where the point cloud is the source point cloud S or the target point cloud T, and the obtained local embedding features are the local embedding features F of each point in the source point cloud S. s , or the local embedding features F of each point in the target point cloud T t In this embodiment of the invention, the established k-NN connections are input to convolutional and max-pooling layers for convolution and max-pooling operations to obtain local embedded features.

[0071] Specifically, in S13, the local embedding feature of the i-th point in the point cloud is:

[0072]

[0073] Among them, f i s For the i-th point s in the source point cloud i Local embedding features, f i t The local embedding feature of the i-th point in the target point cloud; For the i-th point s in the source point cloud i The set of k nearest neighbors (k-NN); For the i-th point t in the target point cloud i The set of k nearest neighbors (k-NN) points; h represents the convolution and max pooling operations, and in this embodiment of the invention, is the corresponding convolution and max pooling layer.

[0074] Then the local embedding features F of each point in the source point cloud S s ={f i s ∈R d |i=1,…N s}, N s The number of points in the source point cloud;

[0075] Local embedding features F of each point in the target point cloud T t ={f i t ∈R d |i=1,…N t}, N t The number of points in the target point cloud.

[0076] Specifically, in the global structural feature extraction module, based on the local embedding feature Fs and F t The corresponding global graphs g constructed respectively s and g t for:

[0077] g s =(F s E s ), g t =(F t E t )

[0078] E s =softmax{<(F s ) T F s >}

[0079] E t =softmax{<(F t ) T F t >}

[0080] Among them, g s and g t Based on the local embedding features F of each point in the source point cloud s The constructed global graph, and the local embedding features F of each point in the target point cloud T. t Constructed global graph; local embedding features F s and F t These are the nodes in the corresponding global graph, E s and E t These represent the edges of the corresponding global graph; <·,·> represent inner product operations.

[0081] Specifically, such as Figure 2 As shown, according to the global graph g s and g t This yields the global structural features corresponding to each point in the source point cloud S and the target point cloud T. and include:

[0082] S21. Use in-graph convolution operations to extract the global graph g respectively. s and g t Global relational features in

[0083]

[0084] in, and These represent the global graph g. s Global relation features and global graph gt Global relational features in the graph; GConv represents the graph convolution operation.

[0085] S22. Calculate global relational features and The point-to-point Euclidean distance D between them i，j And the Euclidean distance D i，j Instance normalization is performed to ensure non-negativity, resulting in a normalized distance matrix. Softmax operations are then performed along the rows and columns of this normalized distance matrix to obtain the projection matrix P of the target point cloud T. t The projection matrix P of the source point cloud S s ;

[0086] S23, using projection matrix P t and P s Update the corresponding global graph g respectively s and g t Edges in the data, using global relation features As a node in the updated global graph, the updated global graph is obtained. and Among them, the updated global graph and The edges are:

[0087]

[0088] Where, ε s Represents the updated global graph The edge, ε t Represents the updated global graph The edge;

[0089] Updated global graph and They are respectively:

[0090]

[0091] in, For global graph g s Updated global graph; For global graph g t The updated global graph.

[0092] S24. Use cross-graph convolution operation to extract the updated global graph. and Global structural features in and

[0093]

[0094] Where CrossConv represents the cross-graph convolution operation; To the updated global graph Structural features extracted from; To the updated global graph The structural features extracted from it.

[0095] Specifically, in the topology-aware interaction module, the global structural features are computed and represented. and The geometric similarity measure Θ between features includes:

[0096] global structural features and Input is fed into the affinity layer to calculate global structural features. and The affinity matrix A between them Where W is the weight matrix, W∈R d×d is a learnable parameter in the affinity layer, and d represents the feature dimension; express The i-th element in express The j-th element in;

[0097] The Sinkhorn algorithm is used to iteratively process the affinity matrix A. i，j The rows and columns are normalized until the Sinkhorn algorithm converges, and the geometric similarity measure Θ is obtained.

[0098] Specifically, each point in the source point cloud S and the target point cloud T is projected to its relative position based on the geometric similarity metric Θ. and include:

[0099] Perform a softmax operation along the rows and columns of the geometric similarity measure Θ to obtain the corresponding projection matrix Θ. s and Θ t ; and using the projection matrix Θ s Project each point in the source point cloud S to its corresponding position. And using the projection matrix Θ t Project each point in the target point cloud T to its corresponding position.

[0100] Specifically, relative position and The input is fed into a convolutional layer, which learns relative position information to obtain the relative position. and feature.

[0101] Specifically, such as Figure 3 and Figure 4 As shown, location-aware features and Using self-attention and cross-attention operations, the contextual interaction features Φ of each point in the source point cloud S and the target point cloud T are extracted accordingly. s and Φ t ,include:

[0102] S31. Based on location-aware features and global structural features, multi-head attention operation is used to calculate the feature vector query Q, keyword K, and value V;

[0103]

[0104] Where Q, K and V respectively take the value of Q s K s V s Or Q t K t V t ; Take respectively or Take respectively or That is:

[0105]

[0106]

[0107] Among them, Q s K s V s Based on location-aware features and global structural features Get the feature vector query, keywords, and values; Q t K t V t Based on location-aware features and global structural features Obtain the feature vector query, keywords, and values; These are the projection weights for the query, keyword, and value, respectively. dhead = d / H, where H represents the number of heads in the tansformer multi-head attention.

[0108] S32. Based on the feature vector queries Q, keyword K, and value V, perform topological awareness self-attention operations on the position-aware features to obtain the corresponding point cloud self-interactive features. and

[0109]

[0110]

[0111] In this embodiment of the invention, a topology-aware self-attention layer is used to obtain the corresponding point cloud self-interaction features:

[0112]

[0113] Where Q, K, V, Θ respectively take the values ​​of Q s K s V s Θ s Or Q t K t V t Θ t φ represents a linear layer with ReLU activation.

[0114] S33, Self-interaction features Perform topology-aware cross-attention operations to integrate self-interactive features. Enhanced to cross-interaction features At this time, the cross-interaction characteristics It contains correlation information between the source point cloud and the target point cloud;

[0115]

[0116] S34, Features of cross-interaction With global structural features Feature merging is performed to obtain the contextual interaction features Φ of each point in the source point cloud S. s ; Self-interaction features With global structural features Feature merging is performed to obtain the contextual interaction features Φ of each point in the target point cloud T. t .

[0117]

[0118]

[0119] Here, ReLU represents the ReLU activation function.

[0120] The method of this invention introduces the role of geometric similarity metric Θ when calculating the self-interaction features of point clouds. Geometric similarity metric Θ is used to reweight the mutual relationship weights between global structural features. Geometric similarity metric reflects the geometric similarity between features. The higher the geometric similarity, the greater the weight obtained, and the closer the mutual relationship between features, making the matching more accurate.

[0121] Furthermore, since the point cloud's positional coordinate information only affects the calculation of the weights of the relationships between structural features (the product of query Q and keyword K), position-aware features are employed. Calculate query Q and keyword K using global structural features. The calculated value V ensures that the value of the feature itself is not affected when topological self-aware attention is performed, thus guaranteeing the consistency of the feature before and after the introduction of position coordinate information.

[0122] Furthermore, this invention discovers that using the keyword K and value V corresponding to the target point cloud to represent the self-interactive features of points in the source point cloud. Enhancement can further improve matching accuracy.

[0123] Specifically, in the transformation matrix prediction module, the context interaction feature Φ is used. s and Φ t The predicted transformation matrix R and translation vector p include:

[0124] S41, Calculate context interaction features Φ s and Φ t Euclidean distance matrix between

[0125] S42, Based on the Euclidean distance matrix Predict the pseudo-correspondence t of each point in the source point cloud. i This leads to the pseudo-target point cloud T′ for each point in the source point cloud, t. i '∈T', and according to the Euclidean distance matrix Calculate the built-in confidence matrix M;

[0126] Among them, the pseudo-correspondence t i 'for:

[0127]

[0128] In the formula, t j This represents any point in the target point cloud T; The dot (·) in the figure represents all rows or all columns.

[0129] The built-in confidence matrix M represents the reliability of the correspondence between points in the source point cloud and the target point cloud, that is, the correlation score between the two. The higher the correlation score, the more reliable the correspondence. M is calculated as follows:

[0130]

[0131] S43. Based on the pseudo-target point cloud T′ and the built-in confidence matrix M, the transformation matrix R and translation vector p are estimated by solving the weighted singular value decomposition (SVD):

[0132]

[0133] Specifically, during the training process, the loss function L of the point cloud registration model of this invention is:

[0134] L = L l +λ g L g +λ s L s

[0135] Where, λ g and λ s It is a hyperparameter that controls the weights of the global alignment loss and spatial geometry loss; L l L represents local alignment loss. g L represents the global alignment loss. s This represents the spatial geometric loss.

[0136] Global alignment loss L g Transform the source point cloud using the predicted transformation matrix R and translation vector p to generate the transformed source point cloud. The transformed source point cloud is measured using the chamfer loss method. The global alignment error between the target point cloud T and the target point cloud T is given by the following formula:

[0137]

[0138] Where T represents the target point cloud, t represents a point in the source point cloud after transformation, and t represents a point in the target point cloud;

[0139] H σ (·) represents the Huber function, as shown below:

[0140]

[0141] Where e represents the alignment error, and σ is a hyperparameter that defines the internal range.

[0142] During the training process of the loss function, the parameters of the corresponding network layers in the local embedding feature extraction module, global structural feature extraction module, topology-aware interaction module and transformation matrix prediction module of the above model are adjusted in reverse, as well as the built-in confidence matrix M, transformation matrix R and translation vector p.

[0143] Local alignment loss L l Using the built-in confidence matrix M, select the m pairs of points with the highest relevant scores in the built-in confidence matrix M, denoted as S. c ∈R m×3 and T c ∈Rm×3 For each point s i ∈S c and its corresponding point t i ∈T c Obtain the corresponding set of k nearest neighbors. and Local alignment error L l The difference between local regions of the transformed source point cloud and target point cloud is calculated as follows:

[0144]

[0145] The logcosh function is used to reduce the impact of outliers on local alignment errors, and it is more robust than the MSE function.

[0146] During the training process of the loss, similar to the global alignment loss, the relevant network parameters, as well as the built-in confidence matrix M, transformation matrix R, and translation vector p, are adjusted in reverse.

[0147] Spatial geometric loss L s By utilizing the predicted transformation matrix R and translation vector p, and minimizing the geometric difference between the pseudo-target point cloud and the target point cloud, a matching matrix is ​​defined for the maximum projection probability of each point in the source point cloud S.

[0148]

[0149] For point S in the selected source point cloud c , obtain with S c The corresponding matching probability is expressed as Spatial geometric loss L s To improve the matching probability between pseudo-target point clouds and actual target point clouds:

[0150]

[0151] in, S represents c The matching probability of any point in the equation.

[0152] The method provided by this invention will be experimentally illustrated below to demonstrate that the method of this invention can achieve better registration results.

[0153] Example 1

[0154] Four metrics—mean absolute error (MAE) and root mean square error (RMSE)—are used on the ModelNet40 dataset (which contains information about airplanes, cars, chairs, stairs, etc.) to measure the registration performance of the proposed method compared to existing classical methods. These metrics are denoted as RMSE(R), MAE(R), RMSE(p), and MAE(p).

[0155] The dataset used initially contains different training objects, each belonging to the same category. Both the training and testing datasets include all 40 categories. Furthest point sampling (FPS) is performed to remove 25% of the points, generating a partially overlapping point cloud. The pre-trained model provided by this invention is compared with existing traditional methods, supervised methods, and unsupervised methods; the comparison results are shown in Table I.

[0156] Table 1. Experimental data corresponding to Example 1

[0157]

[0158]

[0159] As can be seen, the present invention improves RMSE(R) and MAE(R) by 18.75% and 15.63% respectively compared to RIENet, which is the best performing technology in the prior art. That is, the present invention can achieve accurate conversion based on the framework of unsupervised learning.

[0160] Example 2

[0161] Similarly, on the ModelNet40 dataset, different classes were used for training samples. In this experiment, 20 classes were used to train and test the remaining 20 classes, as shown in Table 2.

[0162] Table 2 shows the experimental data corresponding to Example 2.

[0163]

[0164] As can be seen, the method of the present invention achieves the best performance across all metrics, which demonstrates that the global topology software interaction framework of the present invention can provide a powerful generalization of unseen categories.

[0165] Example 3

[0166] Based on Example 1, noise was added to the training samples, using Gaussian noise to jitter the partially overlapping source and target point clouds. The noise was sampled from a Gaussian distribution with a mean of 0 and a standard deviation of 0.01 on each axis. The noise was then trimmed using [-0.05, 0.05]. Models from several existing methods were trained using the same noise, and the experimental results are shown in Table 3.

[0167] Table 3 shows the experimental data corresponding to Example 3.

[0168]

[0169] It can be seen that the method of this invention performs better than the supervised method; compared with CEMNet, which has the best performance among unsupervised methods, the method of this invention has better RMSE(p) and MAE(p) in terms of translation vectors. At the same time, the RMSE(R) and MAE(R) of this invention are also improved by 68.88% and 57.3% respectively compared with CEMNet.

[0170] Example 4

[0171] This embodiment uses the 7Scene dataset (the data samples consist of RGB-D images, real camera poses, and 3D models of indoor rooms). The metrics used are MAE(R) and MAE(p), as well as the average isotropic error (MIE) of the transformation matrix R and the translation vector p, denoted as MIE(R) and MIE(p).

[0172] The source point cloud was resampled to 2048 points and then transformed using a transformation matrix to generate the target point cloud. To obtain partially overlapping data, FPS was applied to both point clouds, and they were downsampled to 1536 points. The experimental results are shown in Table 4.

[0173] Table 4 shows the experimental data corresponding to Example 4.

[0174]

[0175] As can be seen, the method of this invention achieves the best performance across all metrics compared to unsupervised methods. RMSE(R) and MAE(R) are improved by 36.4% and 28.4% respectively compared to RIENet. This demonstrates the feasibility of this method in indoor scenarios.

[0176] Example 5

[0177] In this embodiment, the KITTI odometry dataset (containing information on cities, rural areas, highways, etc.) is used. The point cloud is voxelized to a size of 0.3 meters, and 2048 points are randomly sampled for use in the learning-based method. For traditional methods, the experiment was conducted using the original point cloud. The experimental results are shown in Table 5.

[0178] Table 5 shows the experimental data corresponding to Example 5.

[0179]

[0180]

[0181] As can be seen, the method of this invention exhibits more significant advantages in outdoor scenarios: compared with previous unsupervised methods, RMSE(p) and MAE(p) are improved by more than 41.5% and 46.9%, respectively. The results demonstrate that the method of this invention can achieve higher registration accuracy in challenging outdoor environments.

[0182] The present invention also provides a point cloud registration system based on global topology-aware interaction, including a computer-readable storage medium and a processor;

[0183] Computer-readable storage media are used to store executable instructions;

[0184] The processor is used to read executable instructions stored in a computer-readable storage medium and execute the point cloud registration method based on global topology-aware interaction in the above embodiments.

[0185] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the point cloud registration method based on global topology-aware interaction as described in the above embodiments.

[0186] The point cloud registration method based on global topology-aware interaction of the present invention takes into account that the structural relationship between points in the point cloud and other points is ordered, while the position coordinates of points in the point cloud are unordered. Therefore, when extracting local features, directly extracting the position coordinate information of points will destroy the extraction of structural information. To address this, this invention extracts only the local structural features between points in the point cloud during local embedding feature extraction. Based on these local structural features, a corresponding global graph is then built to extract global structural features. Based on the extracted global structural features, a geometric similarity metric Θ is calculated between the global structural features of the source and target point clouds. Using this Θ, the original position coordinates of points in the source and target point clouds are mapped to their relative positions, ensuring that points at two completely different locations in the source and target point clouds are mapped to the same feature space. In this case, the extracted relative position features do not destroy the structural information of the point cloud. The relative position information is then fused into the corresponding global structural features, and self-attention and cross-attention operations are performed to obtain the semantic interaction features between each point in the point cloud and other points. These semantic features contain both global and local structural and positional information; that is, they include local and global features, as well as feature associations between them. This enables topological awareness of both local and global features, allowing for accurate extraction of local structural information and accurate matching of global features, thereby improving the accuracy of point cloud registration.

[0187] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A point cloud registration method based on global topology-aware interaction, characterized in that, include: Training phase: The point cloud registration model is trained using a dataset, where training samples include source point cloud S and target point cloud T; ​​wherein, the point cloud registration model includes: The local embedding feature extraction module is used to extract the local embedding features F corresponding to the source point cloud S and the target point cloud T. s and F t ; The global structural feature extraction module is used to extract features based on the local embedding features F. s and F t Construct the corresponding global graph g respectively s and g t In order to obtain the corresponding global structural features and The topology-aware interaction module is used to calculate the global structural features. and The geometric similarity measure Θ between the features is used, and each point in the source point cloud S and the target point cloud T is projected to its relative position based on the geometric similarity measure Θ. and and relative position and Features are fused into the corresponding global structural features and In this process, the corresponding location-aware features are obtained. and and location-aware features and Topology-aware self-attention and cross-attention operations are employed to extract the contextual interaction features Φ of each point in the source point cloud S and the target point cloud T. s and Φ t ; The transformation matrix prediction module is used to employ context-interactive features Φ s and Φ t Predict the transformation matrix R and the translation vector p; The application phase includes: inputting the source point cloud S and target point cloud T of the actual object to be measured into the trained point cloud registration model to obtain the trained transformation matrix R and translation vector p; and registering the source point cloud S with the trained transformation matrix R and translation vector p.

2. The method according to claim 1, characterized in that, In the topology-aware interaction module, location-aware features are... and Topology-aware self-attention and cross-attention operations are employed to extract the contextual interaction features Φ of each point in the source point cloud S and the target point cloud T. s and Φ t ,include: S31. Calculate the feature vector query Q, keyword K, and value V using multi-head attention operation: in, Take respectively or Take respectively or These are the corresponding projection weights; S32. Based on the feature vector, query Q, keyword K, and value V to perceive the features at the corresponding locations. and Performing topology-aware self-attention operation yields the self-interaction features of each point in the source point cloud S. Self-interaction features of each point in the target point cloud T S33. Self-interaction features of each point in the source point cloud S. Perform topology-aware cross-attention operations to obtain cross-interaction features. S34, the cross-interaction features With global structural features Feature merging is performed to obtain the contextual interaction features Φ of each point in the source point cloud S. s ; The self-interaction features of each point in the target point cloud T With global structural features Feature merging is performed to obtain the contextual interaction features Φ of each point in the target point cloud T. t .

3. The method according to claim 2, characterized in that, In S32, based on the feature vector queries Q, keyword K, and value V, the corresponding location-aware features are obtained respectively. and When performing topology-aware self-attention operation, the method further includes reweighting the weights of the relationships between features using the geometric similarity measure Θ; wherein the reweighting value is: Where dhead = d / H, H represents the number of heads in multi-head attention, d represents the feature dimension, and φ represents a linear layer with ReLU activation.

4. The method according to any one of claims 1-3, characterized in that, In the topology-aware interaction module, the global structural features are calculated and represented. and The geometric similarity measure Θ between features includes: Calculate the global structural features and The affinity matrix A between them, where, W is the weight matrix; The Sinkhorn algorithm is used to iteratively normalize the rows and columns of the affinity matrix until the Sinkhorn algorithm converges, thus obtaining the geometric similarity measure Θ.

5. The method according to claim 4, characterized in that, In the topology-aware interaction module, each point in the source point cloud S and the target point cloud T is projected to a relative position according to the geometric similarity metric Θ. and include: Perform a softmax operation along the rows and columns of the geometric similarity metric Θ to obtain the corresponding projection matrix Θ. s and Θ t ; Using the projection matrix Θ s Project each point in the source point cloud S to its corresponding position. Using the projection matrix Θ t Project each point in the target point cloud T to its corresponding position.

6. The method according to claim 1, characterized in that, In the local embedding feature extraction module, the global graph g s and g t They are respectively: g s =(F s ,E s ),g t =(F t ,E t ) E s =softmax{<(F s ) T ,F s >} E t =softmax{<(F t ) T ,F t >} Among them, g s To base on local embedding features F s The constructed global graph, g t To base on local embedding features F t Constructed global graph; local embedding features F s and F t These are the nodes in the corresponding global graph, E s and E t These represent the edges of the corresponding global graph; <·,·> represent inner product operations.

7. The method according to claim 6, characterized in that, According to the global graph g s and g t In order to obtain the corresponding global structural features and include: S21. Use in-graph convolution operations to extract the global graph g respectively. s and g t Corresponding global relation features and S22. Calculate global relational features and The point-to-point Euclidean distance D between them i，j The distance is then normalized to obtain a normalized distance matrix; softmax operations are performed along the rows and columns of the normalized distance matrix to obtain the projection matrix P of the target point cloud T. t The projection matrix P of the source point cloud S s ; S23. Update the global graph g using the projection matrix and global relation features. s and g t Corresponding to and in, and ε s For the updated global graph Nodes and edges in the data; and ε t For the updated global graph Nodes and edges in the data; S24. Use cross-graph convolution operation to extract the updated global graph. and The corresponding global structural features and 8. The method according to claim 1, characterized in that, In the local embedding feature extraction module, local embedding features F corresponding to the source point cloud S and the target point cloud T are extracted. s and F t ,include: S11. Obtain the k-NN point set of the i-th point in the point cloud; S12. Calculate the relative positions between adjacent points in the point set to establish k-NN connections; S13. Perform convolution and max pooling operations on the established k-NN connections to obtain local embedding features; wherein, the point cloud is the source point cloud S or the target point cloud T, and the local embedding features are the local embedding features F of each point in the source point cloud S. s Or the local embedding features F of each point in the target point cloud T t .

9. A point cloud registration system based on global topology-aware interaction, characterized in that, Includes computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-8.

Images