A partial point cloud registration method based on consistency learning
Through point consistency learning and correspondence consistency learning modules, the non-overlapping area impact and error correspondence relationship problems of point cloud registration in some overlapping scenarios are solved, and the accurate point cloud registration effect is achieved.
Patent Information
- Application Number
- CN202310095301.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-02-10
- Publication Date
- 2025-08-19
- Estimated Expiration
- 2043-02-10
AI Technical Summary
In point cloud registration in partial overlapping scenarios, the negative impact of non-overlapping areas leads to ineffective feature learning, many error correspondences, resulting in large deviations in transformation parameters and failed registration.
The point consistency learning module predicts overlap fractions and removes non-overlapping area points. The correspondence consistency learning module is used to explore the local to global context information identification and removes the error correspondence relationship. The estimation rigid transformation module is used to solve the transformation parameters.
Effectively removes the negative effects of non-overlapping areas, achieves accurate point cloud registration, reduces task difficulty, and converts partial-to-partial registration into approximately full-to-full registration.
Smart Images
Figure CN115994933B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer vision and relates to a partial point cloud registration method based on consistency learning. Background Art
[0002] Point cloud registration is a fundamental task in computer vision and is widely used in areas such as simultaneous localization and mapping (SLAM), autonomous driving, and 3D reconstruction. The goal of point cloud registration is to find a rigid transformation (rotation matrix and translation vector) between two 3D point clouds (source and target) to align them. Unfortunately, point cloud registration in partially overlapping scenes becomes more challenging due to the negative impact of non-overlapping regions of the point clouds.
[0003] In recent years, numerous methods have been proposed to address the point cloud registration problem. Iterative Closest Point (ICP) is one of the most classic algorithms. It iteratively performs two steps: finding the closest point in the other point cloud and using SVD to find the optimal transformation and update the source point cloud state until convergence. However, ICP is prone to getting stuck in local optima and is sensitive to noise, limiting its applicability. To address these challenges, Go-ICP, symmetric-ICP, and Generalized-ICP have been proposed. In recent years, deep learning-based methods have made significant breakthroughs. Deep Closest Point (DCP) uses the nearest neighbors in feature space instead of Euclidean space. IDAM uses a convolutional neural network to directly predict the mapping matrix. DeepGMR uses a Gaussian mixture model to estimate the optimal transformation. However, these methods do not assume registration under partially overlapping scenes, resulting in poor performance. PRNet solves part-to-part point cloud registration by extracting keypoints from point clouds and using them to construct a mapping matrix. RPMNet uses Sinkhorn regularization to compute the mapping matrix. DeepBBS improves the registration performance in partially overlapping scenarios by using soft mutual neighbor constraints.
[0004] Most current deep learning-based methods divide point cloud registration into three steps: extracting point features, building correspondences based on feature matching, and estimating transformation parameters. However, the negative impact of non-overlapping regions in partially overlapping scenes can prevent the learning of an effective feature space.
[0005] Most current methods are unable to effectively identify and remove incorrect correspondences. When there are too many incorrect correspondences, the estimated transformation parameters will differ significantly from the true transformation, leading to registration failure. Summary of the Invention
[0006] In view of this, the purpose of the present invention is to provide a partial point cloud registration method based on consistency learning. In order to address the negative impact of non-overlapping areas in partially overlapping scenes, the consistency constraint at the point level is used to predict an overlap score for each point, and the influence of points in non-overlapping areas is removed according to the overlap score. In order to address the problem of too many erroneous correspondences, the erroneous correspondences are identified and eliminated by exploring the local to global contextual information between the correspondences.
[0007] In order to achieve the above object, the present invention provides the following technical solutions:
[0008] A partial point cloud registration method based on consistency learning includes the following steps:
[0009] S1: Collect source point cloud P, target point cloud Q and real transformation from the actual scene, construct a dataset, and divide it into training set and test set after preprocessing;
[0010] S2: Constructing a partial point cloud registration model, the partial point cloud registration model includes a point consistency learning module, a correspondence consistency learning module, and an estimated rigid transformation module; the point consistency learning module first extracts the local features of each point, then estimates the corresponding overlap score, and finally removes points in the non-overlapping area according to the overlap score of each point;
[0011] S3: using the training set to train the partial point cloud registration model, and using the test set to test it;
[0012] S4: Initialize the source point cloud and target point cloud to be registered and input them into the trained partial point cloud registration model, and output the rigid transformation between the source point cloud and the target point cloud.
[0013] Furthermore, the dataset is preprocessed, including removing outliers in the point cloud, performing voxel downsampling and regularization.
[0014] Furthermore, the specific calculation steps of the point consistency learning module are as follows:
[0015] A1: Use graph neural network to extract point p from the source point cloud i ∈P and point q in the target point cloud j Extract local features of points in ∈Q;
[0016] A2: Predict the overlap score for each point;
[0017] A3: Points with an overlap score less than 70% of the point cloud size are considered as points in the non-overlapping area and the points in the non-overlapping area are removed.
[0018] Furthermore, the step A1 specifically includes the following steps:
[0019] A11: For each point p i ∈P, first perform k-nearest neighbor search to obtain its k-nearest neighbor point set
[0020] A12: For each point p i The difference between the 3D coordinates of ∈P and the 3D coordinates of its neighborhood points constitutes the initial input features:
[0021]
[0022] A13: Using a graph neural network f g Will Mapped to a high-dimensional feature space, the obtained features are recorded as Among them, the graph neural network f g (·) The interior is divided into four layers, each layer is expressed as:
[0023]
[0024] Where, is the output feature and also the input feature of the next layer; MLP is a multi-layer perceptron; MaxPooling represents the maximum pooling operation; for The k-nearest neighbor set of [·,·] indicates the splicing along the channel direction of the feature;
[0025] A14: Use the same method to calculate the point q in the target point cloud j ∈Q is used for feature extraction, and the features obtained are in The difference between the 3D coordinates of each point in the target point cloud Q and the 3D coordinates of its neighboring points constitutes the initial input feature;
[0026] A15: Using self-attention and cross-attention modules for local features of points and Perform feature enhancement and record the enhanced features as and
[0027] Furthermore, the step A2 specifically includes the following steps:
[0028] A21: For the extracted and Apply the maximum pooling operation to obtain the global feature G of the entire point cloud P =maxpooling(C P ) and G Q =maxpooling(C Q );
[0029] A22: Combine the global features of the two point clouds and the features of each point into a multi-layer perceptron to predict the overlap score of each point and Where MLP is a multi-layer perceptron with an output dimension of N×1, where N is the number of points in the point cloud; δ(·) is a repeated operation along the feature channel direction; [x, y, z] represents the concatenation of features x, y, and z in the channel dimension; O P and O Q Represent the overlapping score sets of point cloud P and point cloud Q respectively.
[0030] Furthermore, the specific calculation steps of the correspondence consistency learning module are as follows:
[0031] B1: Define a mapping matrix M:
[0032]
[0033]
[0034] Where M ij is the element in row i and column j in the mapping matrix M; and Point p i and dot q j Features enhanced by self-attention and cross-attention modules; It represents the inner product of the eigenvector after being amplified by the exponential function and can be regarded as an intermediate variable in the calculation.
[0035] For a point p in the source point cloud P r , its corresponding point in the target point cloud Q is
[0036]
[0037] Thus, we obtain a set of initial correspondences
[0038] B2: Exploring local to global contextual information between correspondences to identify incorrect correspondences;
[0039] B3: Remove incorrect correspondences.
[0040] Furthermore, the step B2 specifically includes:
[0041] B21: First calculate the initial correspondence feature set H C :
[0042]
[0043]
[0044] Where, Represents the corresponding relationship c r =(p r ,q r )∈C; [·,·,...,·] represents the concatenation of features in the channel dimension; Represents the corresponding relationship characteristics of distance perception, express The jth nearest neighbor of for The k-nearest neighbor set of , ‖·‖ is the modulus of the vector;
[0045] B22: Input the initial correspondence features into the multi-scale attention module to explore local to global context information; the multi-scale attention module contains three layers of identical operations, each layer operates as follows:
[0046] The features Input into a graph neural network to explore local consistency, and the output feature is recorded as Y r (1) ;
[0047] The local feature Y r (1) Input into an inner product attention module and output feature F r (1) , then calculate the offset feature OF r (1) =Y r (1) -F r (1) , then offset the feature OF r (1) Input into a multi-layer perceptron to obtain The final feature is
[0048] B23: Apply maximum pooling to the features of the three layers, combine the obtained features together and input them into a multi-layer perceptron to predict the weight of each corresponding relationship W = MLP ([F (1) ,F (2) ,F (3) ]); where MLP is a multi-layer perceptron, and its final output dimension is |C|×1, where |C| is the size of the corresponding relationship set C.
[0049] Furthermore, in step B3, the weight W is used to determine whether a corresponding relationship should be retained. The final corresponding relationship set is recorded as The corresponding weights are
[0050] Furthermore, the calculation steps of the rigid transformation estimation module are as follows:
[0051] According to the obtained correspondence set and weights The final transformation is obtained by solving the following objective function:
[0052]
[0053] Where, (R e ,t e ) is the final transformation parameter, where R e is the rotation matrix, t e is the translation vector; For the corresponding relationship (p i ,q i ) corresponding weight;
[0054] The objective function is calculated by a weighted singular value decomposition method.
[0055] Furthermore, in step S3, when using the training set to train the partial point cloud registration model, the loss function used is divided into three parts: mapping matrix loss, point prediction loss and correspondence prediction loss, wherein the point prediction loss and correspondence prediction loss are binary cross entropy loss;
[0056] The mapping matrix loss L1 is:
[0057]
[0058] j * =argmin j ||R * p i +t * -q j || 2
[0059] Where M is the mapping matrix; a i Is a parameter that determines whether the current item is calculated. If it satisfies Then a i =1; otherwise a i =0; R * and t * are the real rotation matrix and translation vector respectively; p i and q j j is a point in the source point cloud P and the target point cloud Q respectively; * Represents point p iThe number of the actual corresponding point in the target point cloud Q;
[0060] Point prediction loss L 21 and L 22 for:
[0061]
[0062]
[0063] Where N is the size of the point cloud; and are the predicted points p i and dot q j The overlap fraction of For point p i The true overlap fraction, if p i In the overlapping area, otherwise For point q i The true overlap fraction, if q i In the overlapping area, otherwise
[0064] The correspondence prediction loss L3 is:
[0065]
[0066] Where w i ∈W is the weight of the predicted i-th correspondence, is the weight of the true i-th correspondence;
[0067] The final loss is
[0068] L=L1+L 21 +L 22 +L3
[0069] Where L represents the final loss.
[0070] The beneficial effects of the present invention are as follows: During the point consistency learning phase, the present invention fundamentally addresses the negative impact of non-overlapping areas by predicting the overlapping areas of the source and target point clouds, transforming partial-to-partial point cloud registration into approximately complete-to-complete point cloud registration, greatly reducing the difficulty of the task. During the correspondence consistency learning phase, the present invention obtains initial correspondences through feature similarity, then identifies mismatches by exploring the local-to-global contextual information between the initial correspondences, achieving accurate point cloud registration.
[0071] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0072] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:
[0073] Figure 1 Flowchart of the partial point cloud registration method based on consistency learning;
[0074] Figure 2 This is the registration result image. DETAILED DESCRIPTION
[0075] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.
[0076] Among them, the accompanying drawings are only for illustrative purposes and represent only schematic diagrams rather than actual pictures, and should not be understood as limiting the present invention. In order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the dimensions of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions may be omitted in the accompanying drawings.
[0077] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "back", etc. indicating directions or positional relationships, they are based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.
[0078] like Figure 1As shown, the present invention provides a partial point cloud registration method based on consistency learning. Given a source point cloud P and a target point cloud Q, the rigid transformations R and t between the two point clouds are found through three stages: point consistency learning, correspondence consistency learning, and estimation of the rigid transformations R and t.
[0079] Point consistency learning phase:
[0080] Point consistency learning first extracts the local features of each point, then estimates the corresponding overlap score, and finally removes points in the non-overlapping area based on the overlap score of each point.
[0081] First, we use the Graph Neural Network to extract the i ∈P and q j ∈Q, the local features of the points can capture the local information of the points and focus on the local fine-grained structure within their neighborhood. Specifically, for each point p i ∈P, first perform k-nearest neighbor search to obtain its k-nearest neighbor point set Then the difference between the 3D coordinates of each point and the 3D coordinates of its neighboring points constitutes the initial input feature Then use a graph neural network f g Will Mapped to a high-dimensional feature space, the obtained features are recorded as Point q in the target point cloud j ∈Q also uses similar operations to extract features, and the features obtained are Local features of points and The similarity between points p i and q j However, in order to better predict the overlap score of each point and build a reliable correspondence later, it is necessary to enhance the extracted local features. Therefore, we use self-attention and cross-attention modules to enhance the features. The enhanced features are respectively denoted as and
[0082] In order to predict the overlap score of each point, that is, the probability that each point is located in the overlapping area, the extracted and Apply the maximum pooling operation to obtain the global feature G of the entire point cloud P =maxpooling(C P ) and G Q =maxpooling(C Q). The global features of the two point clouds and the features of each point are then combined and input into a multi-layer perceptron (MLP) to predict the overlap score of each point. and
[0083] Points with an overlap score less than 70% of the point cloud size are considered to be points in the non-overlapping area. Finally, these points are removed and the remaining points participate in subsequent operations.
[0084] Correspondence consistency learning stage:
[0085] After the point consistency learning phase, the points in the non-overlapping area have been removed. The remaining point clouds are still recorded as P and Q, and the corresponding features are still recorded as C P and C Q In order to obtain the mapping relationship between two point clouds, the inner product of the feature vectors is used to measure the similarity of the features. Define a mapping matrix M, where For a point p in the source point cloud P r , its corresponding point in the target point cloud Q is In this way, we obtain a set of initial correspondences However, there are some false matches in the initial correspondence C. If they cannot be removed, the final registration result will be seriously affected. To this end, the present invention identifies false correspondences by exploring the local to global context information between correspondences, and then removes them, and finally uses the remaining correct correspondences to solve the rigid transformation. For each correspondence, the initial correspondence feature is first calculated. The feature is then input into a multi-scale attention module to explore local to global context information. The multi-scale attention module has three layers. Since the operations between different layers are exactly the same, we will only describe the operation of the first layer below. Input into a graph neural network to explore local consistency, and the output feature is recorded as Y r (1) . Then Y r (1) Input to an OffsetNonlocal module to explore global consistency. Specifically, the local feature Y r (1) Input into an inner product attention module and output feature F r (1) . Then calculate the offset feature OF r (1) =Y r (1) -F r (1), then offset the feature OF r (1) Input into a multi-layer perceptron to obtain The final feature is Finally, we apply the maximum pooling to the features of the three layers, combine the features together and input them into a multi-layer perceptron to predict the weight of each corresponding relationship W = MLPs ([F (1) ,F (2) ,F (3) ]). The weight W is used to determine whether a certain correspondence should be retained. In the implementation, the first 50% of the correspondences are retained for subsequent solution transformations based on the size of the weight W. The final correspondence set is recorded as The corresponding weights are
[0086] Estimate a rigid transformation:
[0087] According to the obtained correspondence set and weights The final transformation can be obtained by solving the following objective function: This objective function can be solved by weighted singular value decomposition (Weighted SVD).
[0088] Training part:
[0089] Prepare training data according to the actual scenario, namely the source point cloud, target point cloud and real transformation: If there are a large number of outliers in the point cloud, it is necessary to remove the outliers to improve the quality of the point cloud. Use the Open3D tool's radius outlier removal (remove_radius_outlier) function to delete points with less than 30 points in a given sphere within a radius of 0.05m. Use Open3D for voxel downsampling (voxel_down_sample), and set the voxel size to 0.025. Then randomly sample 1024 points from the downsampled source point cloud and target point cloud respectively, and record the sampled points as P∈R 1024×3 and Q∈R 1024×3 . Finally, regularize P and Q: In 3D space, a random direction transformation T is generated (including a rotation from 0° to 360° and a translation within a distance of ±0.05 units), and the random transformation is applied to the target point cloud Q.
[0090] Input the point clouds P and Q into the partial point cloud registration model of the present invention, and record the generated mapping matrix M, overlap score O P and O Q The weight W of the corresponding relationship is calculated by comparing this data with the real data, and backpropagation is performed to update the parameters. The above describes the operations performed in one epoch.
[0091] In this paper, the total number of epochs for training was set to 1000, the Adam optimizer was used, the initial learning rate was 0.0001, and the batch size was set to 8. The loss function used was divided into three parts: mapping matrix loss, point prediction loss, and correspondence prediction loss. Both the point prediction loss and the correspondence prediction loss were binary cross-entropy losses.
[0092] (1) The mapping matrix loss L1 is:
[0093]
[0094] j * =argmin j ||R * p i +t * -q j || 2
[0095] Where M is the mapping matrix; a i Is a parameter that determines whether the current item is calculated. If it satisfies Then a i =1; otherwise a i =0; R * and t * are the real rotation matrix and translation vector respectively; p i and q j j is a point in the source point cloud P and the target point cloud Q respectively. * Represents point p i The number of the actual corresponding point in the target point cloud Q.
[0096] (2) Point prediction loss L 21 and L 22 for:
[0097]
[0098]
[0099] Where N is the size of the point cloud; and are the predicted points p i and dot q j The overlap fraction of For point p i The true overlap fraction, if p i In the overlapping area, otherwise For point q i The true overlap fraction, if qi In the overlapping area, otherwise
[0100] (3) The correspondence prediction loss L3 is:
[0101]
[0102] Where w i ∈W is the weight of the predicted i-th correspondence, is the weight of the true i-th correspondence.
[0103] The final loss L is
[0104] L=L1+L 21 +L 22 +L3
[0105] Testing and practical application:
[0106] 1. Prepare the source point cloud and target point cloud to be registered and perform preprocessing operations on them: remove outliers, downsample voxels and regularize.
[0107] 2. Load the trained model for estimating the transformation and record the model as f θ , where θ represents the pre-training parameters.
[0108] 3. Input the preprocessed point clouds P and Q into the trained model to solve the transformation (R, t) = f θ (P,Q).
[0109] 4. The previous step obtains the transformation (R, t) of the regularized space, so it is necessary to map it back to the original unnormalized space to obtain the final result:
[0110]
[0111]
[0112] like Figure 2 The following figure shows the registration results of this example. The first three rows correspond to the results of unseen shapes, unseen categories, and unseen shapes with noise, respectively. Inputs are the input source point cloud and target point cloud. GT is the ground-truth registration result.
[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.
Claims
1. A partial point cloud registration method based on consistency learning, characterized by: Including the following step: S1: Collect source point cloud P, target point cloud Q and real transformation from the actual scene, construct a dataset, and divide it into training set and test set after preprocessing; S2: Constructing a partial point cloud registration model, the partial point cloud registration model includes a point consistency learning module, a correspondence consistency learning module, and an estimated rigid transformation module; the point consistency learning module first extracts the local features of each point, then estimates the corresponding overlap score, and finally removes points in the non-overlapping area according to the overlap score of each point; S3: using the training set to train the partial point cloud registration model, and using the test set to test it; S4: Initialize the source point cloud and target point cloud to be registered and input them into the trained partial point cloud registration model, and output the rigid transformation between the source point cloud and the target point cloud; The specific calculation steps of the point consistency learning module are as follows: A1: Use graph neural network to extract point p from the source point cloud i ∈P and point q in the target point cloud j Extract local features of points in ∈Q; A2: Predict the overlap score for each point; A3: Points with an overlap score less than 70% of the point cloud size are considered as points in the non-overlapping area and are removed; The specific calculation steps of the correspondence consistency learning module are as follows: B1: Define a mapping matrix M: Where M ij is the element in row i and column j in the mapping matrix M; and Point p i and dot q j Features enhanced by self-attention and cross-attention modules; Represents the inner product of the eigenvector after exponential function amplification, which is regarded as the calculation of M ij An intermediate variable of For a point p in the source point cloud P r , its corresponding point in the target point cloud Q is Thus, we obtain a set of initial correspondences B2: Exploring local to global contextual information between correspondences to identify incorrect correspondences; B3: Remove incorrect correspondences.
2. The method for partial point cloud registration based on consistency learning according to claim 1, characterized in that: The dataset is preprocessed, including removing outliers in the point cloud, performing voxel downsampling and regularization.
3. The partial point cloud registration method based on consistency learning according to claim 1, characterized in that: The step A1 specifically includes the following steps: A11: For each point p i ∈P, first perform k-nearest neighbor search to obtain its k-nearest neighbor point set A12: For each point p i The difference between the 3D coordinates of ∈P and the 3D coordinates of its neighborhood points constitutes the initial input features: A13: Using a graph neural network f g Will Mapped to a high-dimensional feature space, the obtained features are recorded as Among them, the graph neural network f g (·) The interior is divided into four layers, each layer is expressed as: Where, is the output feature and also the input feature of the next layer; MLP is a multi-layer perceptron; MaxPooling represents the maximum pooling operation; for The k-nearest neighbor set of [·,·] indicates splicing along the channel direction of the feature; A14: Use the same method to calculate the point q in the target point cloud j ∈Q is used for feature extraction, and the features obtained are in The difference between the 3D coordinates of each point in the target point cloud Q and the 3D coordinates of its neighboring points constitutes the initial input feature; A15: Using self-attention and cross-attention modules for local features of points and Perform feature enhancement and record the enhanced features as and 4. The method for partial point cloud registration based on consistency learning according to claim 1, characterized in that: The step A2 specifically includes the following steps: A21: For the extracted and Apply the maximum pooling operation to obtain the global feature G of the entire point cloud P =maxpooling(C P ) and G Q =maxpooling(C Q ); A22: Combine the global features of the two point clouds and the features of each point into a multi-layer perceptron to predict the overlap score of each point and Where MLP is a multi-layer perceptron with an output dimension of N×1, where N is the number of points in the point cloud; δ(·) is a repeated operation along the feature channel direction; [x, y, z] represents the concatenation of features x, y, and z in the channel dimension; O P and O Q Represent the overlapping score sets of point cloud P and point cloud Q respectively.
5. The method for partial point cloud registration based on consistency learning according to claim 1, characterized in that: The step B2 specifically includes: B21: First calculate the initial correspondence feature set H C : Where, Represents the corresponding relationship c r =(p r ,q r )∈C; [·,·,...,·] represents the concatenation of features in the channel dimension; Represents the corresponding relationship characteristics of distance perception, express The jth nearest neighbor of for The k-nearest neighbor set of , ||·|| is the modulus of the vector; B22: Input the initial correspondence features into the multi-scale attention module to explore local to global context information; the multi-scale attention module contains three layers of identical operations, each layer operates as follows: The features Input into a graph neural network to explore local consistency, and the output feature is recorded as Y r (1) ; The local feature Y r (1) Input into an inner product attention module and output feature F r (1) , then calculate the offset feature OF r (1) =Y r (1) -F r (1) , then offset the feature OF r (1) Input into a multi-layer perceptron to obtain The final feature is B23: Apply maximum pooling to the features of the three layers, combine the obtained features together and input them into a multi-layer perceptron to predict the weight of each corresponding relationship W = MLP ([F (1) , F (2) , F (3) ]); where MLP is a multi-layer perceptron, and its final output dimension is |C|×1, where |C| is the size of the corresponding relationship set C.
6. The partial point cloud registration method based on consistency learning according to claim 1, characterized in that: In step B3, the weight W is used to determine whether a corresponding relationship should be retained. The final corresponding relationship set is recorded as The corresponding weights are 7. The method for partial point cloud registration based on consistency learning according to claim 1, characterized in that: The calculation steps of the estimated rigid transformation module are as follows: According to the obtained correspondence set and weights The final transformation is obtained by solving the following objective function: Where, (R e , t e ) is the final transformation parameter, where R e is the rotation matrix, t e is the translation vector; For the corresponding relationship (p i ,q i ) corresponding weight; The objective function is calculated by a weighted singular value decomposition method.
8. The method for partial point cloud registration based on consistency learning according to claim 1, characterized in that: In step S3, when using the training set to train the partial point cloud registration model, the loss function used is divided into three parts: mapping matrix loss, point prediction loss and correspondence prediction loss, wherein the point prediction loss and the correspondence prediction loss are binary cross entropy losses; The mapping matrix loss L1 is: j * =argmin j ||R*p i +t*-q j || 2 Where M is the mapping matrix; a i Is a parameter that determines whether the current item is calculated. If it satisfies Then a i =1; otherwise a i =t;R * and t * are the real rotation matrix and translation vector respectively; p i and q j j is a point in the source point cloud P and the target point cloud Q respectively; * Represents point p i The number of the actual corresponding point in the target point cloud Q; Point prediction loss L 21 and L 22 for: Where N is the size of the point cloud; and are the predicted points p i and dot q j The overlap fraction of For point p i The true overlap fraction, if p i In the overlapping area, otherwise For point q i The true overlap fraction, if q i In the overlapping area, otherwise The correspondence prediction loss L3 is: Where w i ∈W is the weight of the predicted i-th correspondence, is the weight of the true i-th correspondence; The final loss is L = L1 + L 21 +L 22 +L3 Where L represents the final loss.