Unsupervised registration method based on low-overlap three-dimensional point cloud map

Abstract

The invention provides an unsupervised registration method based on a low-overlap three-dimensional point cloud map. For the obtained source point cloud and target point cloud data of the three-dimensional point cloud map to be registered, carrying out global feature extraction on two groups of point clouds by adopting a feature extraction module; on the basis of the obtained global features, using a residual iteration Transform module to capture global information contact, and obtaining two groups of point cloud global fusion features; on the basis of the global fusion features, generating predicted overlapping weights of points of two groups of point clouds; adopting a key point screening method to generate two groups of point cloud overlapping region point cloud sets according to the predicted overlapping weight of each point; k neighbor cluster local feature fusion is carried out on the point cloud sets of the two point cloud overlapping areas, and an overlapping matching matrix is generated; on the basis of the overlapping matching matrix, the weight of the overlapping point pair to be registered is obtained, a rotation change matrix is solved according to the overlapping point pair, and finally a point cloud map registration result is generated; and based on the network loss function, iteratively training the network model until loss convergence, and obtaining an unsupervised registration model of the low-overlap three-dimensional point cloud map. The method can be used for high-precision registration work of a large-scene low-overlap three-dimensional point cloud map, and solves the problems that a model depends on labeled data, the global information association capture capability is weak, and the generalization of coping with a low-overlap three-dimensional point cloud map scene is poor.

Description

Technical Field

[0001] The present invention relates to an unsupervised registration method based on a low-overlap three-dimensional point cloud map, which can be used for high-precision registration of a low-overlap three-dimensional point cloud map and belongs to the technical field of three-dimensional point cloud registration. Background Art

[0002] 3D point cloud registration is the process of aligning two or more point cloud data to the same coordinate system. Low-overlap 3D point cloud map registration is crucial to improving the integrity of point cloud data, improving the quality of 3D reconstruction of target scenes, and enhancing visual effects.

[0003] Related technologies can generally be divided into two categories: traditional point cloud registration methods and deep learning-based point cloud registration methods. Traditional point cloud registration methods are mainly based on a combination of coarse registration (such as RANSAC, etc.) and fine registration (such as ICP, etc.), and gradually optimize the registration accuracy. However, traditional point cloud registration algorithms are sensitive to noise and outliers in complex point cloud data, and are often affected by the initial pose, resulting in unstable results. In recent years, deep learning-based three-dimensional point cloud registration methods have utilized the powerful representation capabilities of neural networks, automatically learning features and optimizing transformations during registration, and improving registration accuracy and robustness. While supervised point cloud registration networks rely too much on data annotation, unsupervised point cloud registration networks are more flexible in practical applications. In the existing unsupervised registration networks, for example, CEMNet and RIENet have good effects in specific scenes, but they do not fully capture the connection between global contexts, and the model robustness is poor; existing inventions such as "A method for robust registration of unsupervised point clouds based on virtual anchor guidance" (CN118229745A) have improved the robustness of the model, but there are certain drawbacks in terms of network complexity and coping with large-scene low-overlap 3D point cloud maps. Therefore, the current existing 3D point cloud map registration network model needs to be improved urgently.

[0004] The present invention discloses an unsupervised registration method based on a low-overlap three-dimensional point cloud map, proposes a residual Transformer module of multi-head self-attention and cross-attention iteration, improves the ability to capture global information connections in the registration problem of unknown point cloud map models; adopts global fusion features to predict the overlapping weights of each point, improves the accuracy of the problem of locking the overlapping area of ​​low-overlap point clouds; calculates the transformation matrix through the overlapping cluster feature matching matrix, improves the accuracy of the registration result; designs the overlapping cluster geometric alignment loss and the overlapping point pair feature consistency loss, improves the generalization of the low-overlap point cloud registration network model. Summary of the invention

[0005] The purpose of the present invention is to provide an unsupervised registration method based on low-overlap 3D point cloud maps, which can autonomously learn the registration and splicing of high-quality point cloud maps under unlabeled data conditions, and has a good ability to capture global information connections when registering unknown point cloud map models, and can accurately lock the overlapping areas of low-overlap point clouds. This is to solve the problems of excessive reliance on labeled data, weak ability to capture global information associations, and poor generalization of large-scale low-overlap scenes in related technologies.

[0006] To achieve the above purpose, the present invention designs an unsupervised registration method based on a low-overlap three-dimensional point cloud map, which includes the following contents:

[0007] The unsupervised registration method processes the input source point cloud and the target point cloud through seven modules: global feature extraction, global feature fusion, point overlap weight prediction, overlap area locking, overlap matching, rotation transformation matrix solution, and network loss calculation, and finally realizes a high-precision unsupervised registration model for low-overlap three-dimensional point cloud maps. The specific implementation process of the unsupervised registration network model includes: obtaining the source point cloud and target point cloud data of the three-dimensional point cloud map to be registered, and using the feature extraction module to extract global features of the two groups of point clouds; based on the obtained global features, using the residual iterative Transformer module to capture the global information connection and obtain the global fusion features of the two groups of point clouds; based on the global fusion features, generating the predicted overlapping weights of each point in the two groups of point clouds; using the key point screening method, according to the predicted overlapping weights of each point, respectively generating two groups of point cloud overlapping area point cloud sets; by performing k-nearest neighbor cluster local feature fusion on the two groups of point cloud overlapping area point cloud sets, and generating an overlapping matching matrix; based on the overlapping matching matrix, obtaining the weights of the overlapping point pairs to be registered, solving the rotation change matrix according to the overlapping point pairs, and finally generating the point cloud map registration result; based on the network loss function, iteratively training the network model until the loss converges, and obtaining an unsupervised registration model for low-overlap three-dimensional point cloud maps.

[0008] Optionally, based on the acquisition of the source point cloud and the target point cloud data of the three-dimensional point cloud map to be registered, using a feature extraction module to perform global feature extraction of two groups of point clouds, comprises:

[0009] Based on the source point cloud and target point cloud data of the three-dimensional point cloud map to be registered, the farthest point sampling method (FPS) is used to downsample the two groups of point clouds to obtain the downsampled source point cloud and target point cloud;

[0010] Based on the downsampled source point cloud and target point cloud obtained, feature extraction is performed using a feature extraction module based on DGCNN (Dynamic Graph Convolutional Neural Networks) to obtain global features of each point cloud.

[0011] Optionally, based on the obtained global features, a residual iterative Transformer module is used to capture global information connections to obtain two sets of point cloud global fusion features, including:

[0012] Based on the two sets of global features obtained as input, a Transformer module including multi-head self-attention and multi-head cross-attention is first used to perform feature connection learning. The multi-head self-attention mechanism is used to learn the feature connection between the points of a single point cloud itself, and obtain the self-connection features of the source point cloud and the target point cloud;

[0013] Based on the obtained source point cloud and target point cloud self-connection features, a multi-head cross attention mechanism is adopted to learn the feature connection between the source point cloud and the target point cloud and the target point cloud and the source point cloud, and obtain the fusion features of the source point cloud to the target point cloud and the fusion features of the target point cloud to the source point cloud;

[0014] Based on the initial global fusion feature, the two groups of fusion features are filled with features, and the residual module is used to reduce the disappearance of the own features during the learning process to obtain two groups of filled point cloud fusion features, and feature layer normalization and feature maximum pooling are used to generate two groups of global fusion features of the current iteration;

[0015] Based on the two sets of global fusion feature results generated in the current iteration, the residual iterative Transformer module is used to cyclically learn and capture the global connection between the two sets of point clouds to obtain the final global fusion feature.

[0016] Optionally, generating predicted overlap weights of each point in two groups of point clouds based on the obtained global fusion features includes:

[0017] Based on the global fusion features, a decoding output module including four layers of MLP is adopted, and the last layer uses a fully connected layer with a sigmoid activation function to reduce the impact of inaccurate prediction of corresponding contact points. Finally, the predicted overlapping weights of each point in the two sets of point clouds are obtained.

[0018] Optionally, the key point screening method is used to generate two groups of point cloud overlapping area point cloud sets according to the predicted overlapping weights of each point, including:

[0019] Based on the predicted overlapping weights of each point in the two groups of point clouds, the TOP-K key point screening method is used to select the top K points with the highest predicted overlapping weights of the two groups of point clouds as key points;

[0020] Key points of the two groups of point clouds are extracted to obtain a point cloud set in the overlapping area of ​​the two groups of point clouds.

[0021] Optionally, the step of fusing local features of k nearest neighbor clusters of point cloud sets in overlapping areas of two groups of point clouds and generating an overlapping matching matrix includes:

[0022] Based on the points in the point cloud set in the overlapping area of ​​the two groups of point clouds, k nearest neighbor clusters are indexed on the downsampled source point cloud and the target point cloud, and the features of each key point and its k nearest neighbor cluster are spliced ​​and fused;

[0023] Based on the features of the k nearest neighbor clusters of each key point, the maximum pooling is used to obtain the local fusion features of the key points;

[0024] Based on the local fusion features of the key points, the negative feature distance method is used to perform fusion feature matching to generate an overlapping matching matrix.

[0025] Optionally, the method of obtaining weights of overlapping point pairs to be registered based on the overlapping matching matrix, solving the rotation change matrix according to the overlapping point pairs, and finally generating a point cloud map registration result includes:

[0026] Performing softmax normalization based on the generated overlapping matching matrix to obtain the matching weight of each point pair;

[0027] According to the matching weights of each point pair, a weighted quaternion algorithm is used to solve the point cloud rotation transformation matrix;

[0028] Based on the obtained point cloud rotation transformation matrix, the source point cloud is subjected to posture rotation transformation, and finally a point cloud map registration result is obtained.

[0029] Optionally, the iterative training of the network model until the loss converges based on the network loss function to obtain an unsupervised registration model of a low-overlap three-dimensional point cloud map includes:

[0030] Based on the obtained point cloud rotation transformation matrix, the source point cloud overlapping area point cloud set and its k nearest neighbor points are used as the source point cloud overlapping cluster for rotation transformation, and based on the target point cloud overlapping area point cloud set and its k nearest neighbor points as the target point cloud overlapping cluster, the Euclidean distance from the source point cloud overlapping cluster to the target point cloud overlapping cluster after the rotation transformation is calculated to obtain the overlapping cluster geometric alignment loss;

[0031] According to the weight matrix of the overlapping point pairs to be registered, k point pairs with the largest weights are selected, the feature values ​​of each point in the point pair in the global feature, the global fusion feature and the local fusion feature are obtained, and the feature distances between the corresponding points' own features are calculated to obtain the feature consistency loss of the overlapping point pairs;

[0032] Based on the overlapping cluster geometric alignment loss and the overlapping point pair feature consistency loss, the registration network loss is fused and calculated to obtain the total network loss;

[0033] Based on the total network loss, back propagation guides the network model learning and training, and the network model is iteratively trained until the total network loss converges to obtain the final unsupervised registration model of the low-overlap three-dimensional point cloud map.

[0034] The present invention provides an unsupervised registration method based on a low-overlap three-dimensional point cloud map. The source point cloud and target point cloud data of the three-dimensional point cloud map to be registered are obtained, and a feature extraction module is used to extract global features of two groups of point clouds; based on the obtained global features, a residual iterative Transformer module is used to capture global information connections to obtain two groups of point cloud global fusion features; based on the global fusion features, the predicted overlap weights of each point in the two groups of point clouds are generated; a key point screening method is used to generate two groups of point cloud overlapping area point cloud sets according to the predicted overlap weights of each point; the k-nearest neighbor cluster local feature fusion is performed on the two groups of point cloud overlapping area point cloud sets, and an overlapping matching matrix is ​​generated; based on the overlapping matching matrix, the weights of the overlapping point pairs to be registered are obtained, and the rotation change matrix is ​​solved according to the overlapping point pairs, and finally the point cloud map registration result is generated; based on the network loss function, the network model is iteratively trained until the loss converges, and an unsupervised registration model of a low-overlap three-dimensional point cloud map is obtained. The present invention makes full use of the global information connection capture capability of the attention mechanism, and proposes a residual Transformer module of multi-head self-attention and cross-attention iteration. The global fusion feature is used to predict the overlap weight of each point, and the point cloud rotation transformation matrix is ​​calculated through the overlapping cluster feature matching matrix. The overlapping cluster geometric alignment loss and the overlapping point pair feature consistency loss are designed. The shortcomings of the existing point cloud registration model, such as poor generalization in low-overlap 3D point cloud map scenarios, insufficient global feature connection capture capability, and excessive reliance on label data, are overcome to achieve high-precision 3D point cloud map registration. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0036] Figure 1 It is a flowchart of the steps of an unsupervised registration method based on a low-overlap three-dimensional point cloud map provided by the present invention.

[0037] Figure 2 It is a schematic diagram of the overall structure of the unsupervised low-overlap three-dimensional point cloud map registration network provided by the present invention.

[0038] Figure 3 It is a comparison diagram before and after the three-dimensional point cloud map registration in the test example provided by the present invention. DETAILED DESCRIPTION

[0039] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0040] In the description of the present invention, “plurality” means two or more than two, unless otherwise clearly and specifically defined.

[0041] See also Figure 1 The present invention provides an unsupervised registration method based on a low-overlap three-dimensional point cloud map, comprising the following steps:

[0042] S101, obtaining source point cloud and target point cloud data of the three-dimensional point cloud map to be registered, and using a feature extraction module to perform global feature extraction on the two groups of point clouds.

[0043] Specifically, this model training uses the public outdoor point cloud kitti dataset. Based on the source point cloud and target point cloud data of the three-dimensional point cloud map to be registered, FPS is used to downsample the two groups of point clouds. The number of sampling points is set to 2048, and the downsampled source point cloud X∈R is obtained. 2048*3 and the target point cloud Y∈R 2048*3 ;

[0044] Based on the downsampled source point cloud and target point cloud, feature extraction is performed using a feature extraction module based on DGCNN, and the output feature dimension is set to 256 to obtain the global features ΦX∈R of each point cloud. 2048*256 ΦY∈R 2048 *256 .

[0045] S102, based on the obtained global features, using the residual iterative Transformer module to capture global information connections, and obtaining two sets of point cloud global fusion features.

[0046] Specifically, based on the two sets of global features ΦX and ΦY obtained as input, a Transformer module including multi-head self-attention and multi-head cross-attention is first used to perform feature connection learning, and the number of attention heads is 12;

[0047] A self-attention mechanism with 12 attention heads is used to learn the feature connections between points in a single point cloud, and obtain the self-connection features of the source point cloud and the target point cloud;

[0048] Based on the obtained source point cloud and target point cloud self-connection features, a multi-head cross attention mechanism with 12 attention heads is used to learn the feature connection between the source point cloud to the target point cloud and the target point cloud to the source point cloud, and obtain the fusion feature ΦX from the source point cloud to the target point cloud. i ∈R 2048*256 And the fusion feature ΦY from the target point cloud to the source point cloud j ∈R 2048 *256 ;

[0049] Based on the initial global fusion feature, the two sets of fusion features are filled with features, and the residual module is used to reduce the disappearance of the own features during the learning process. The filling formula is as follows:

[0050] ΦX i '=M(L N (ΦX+ΦX i )),ΦY j '=M(L N (ΦY+ΦY j ))(1)

[0051] Among them, + indicates splicing on the feature dimension. We get two sets of filled point cloud fusion features ΦX i '∈R 2048*512 ,ΦY' j ∈R 2048*512 , and use feature layer normalization L N And feature maximum pooling M, the output feature dimension is 2048*1, generating two sets of global fusion features for the current iteration;

[0052] Based on the two sets of global fusion feature results generated in the current iteration, the residual iterative Transformer module is used to cyclically learn and capture the global connection between the two sets of point clouds. The number of iterations is I=6. The last iteration does not perform feature layer normalization and feature maximum pooling to obtain the final global fusion feature.

[0053] S103: Based on the global fusion features, generate predicted overlap weights for each point in two groups of point clouds.

[0054] Specifically, based on the global fusion feature, a decoding output module including four layers of MLP is used, and the last layer uses a fully connected layer with a sigmoid activation function to reduce the impact of inaccurate prediction of corresponding contact points. Finally, the prediction overlap weights of each point in the two sets of point clouds are obtained. X ∈R 2048*1 , O Y ∈R 2048*1 .

[0055] S104, using a key point screening method, and generating two groups of point cloud overlapping area point cloud sets according to the predicted overlapping weights of the points.

[0056] Specifically, based on the predicted overlap weights of each point in the two groups of point clouds, the TOP-K key point screening method is used to screen the top K=20 points with the highest predicted overlap weights of the two groups of point clouds as key points;

[0057] Based on the key points of the two groups of point clouds, the point cloud set X'∈R in the overlapping area of ​​the two groups of point clouds is obtained. 20*3 , Y'∈R 20*3 .

[0058] S105, performing k-nearest-neighbor cluster local feature fusion on the point cloud sets in the overlapping area of ​​the two groups of point clouds, and generating an overlapping matching matrix.

[0059] Specifically, based on the points in the point cloud sets in the overlapping area of ​​the two groups of point clouds, k=5 nearest neighbor clusters are respectively indexed on the downsampled source point cloud X and the target point cloud Y, and the features of each key point and its k nearest neighbor clusters are spliced ​​and fused;

[0060] Based on the k nearest neighbor clusters of each key point, the maximum pooling is used to obtain the key point local fusion feature ΦX i ”、ΦY j ”;

[0061] Based on the local fusion features of the key points, the negative feature distance method is used for fusion feature matching. The negative feature distance formula is as follows:

[0062] D i,j = -||ΦX i ”-ΦY j ”||2(2)

[0063] Where X i ∈X',Y i ∈Y', and finally generate the overlapping matching matrix.

[0064] S106: Based on the overlapping matching matrix, the weights of the overlapping point pairs to be registered are obtained, and the rotation change matrix is ​​solved according to the overlapping point pairs, so as to finally generate a point cloud map registration result.

[0065] Specifically, normalization is performed based on the generated overlapping matching matrix, and the formula is as follows:

[0066] W i,j = S m ( D i,j )(3)

[0067] Among them, S mis the softmax function, and the matching weight W of each point pair is obtained i,j ;

[0068] Based on the matching weights of each point pair, a weighted quaternion algorithm is used to solve the point cloud rotation transformation matrix T = {R, t}. The weighted quaternion formula is as follows:

[0069]

[0070]

[0071] Among them, formula (4) is used to calculate the weighted centroid Weight O i , O j Use each point to predict the overlap weight O X , O Y , and X i ∈X',Y i ∈Y', formula (5) is used to calculate the weighted covariance matrix S, For X i , Y i The coordinates after removing the weighted center of mass. Based on S, the quaternion-related symmetric matrix N is constructed to obtain the rotation quaternion q, and q is converted into the rotation matrix R. The solution formula for the translation matrix t is as follows:

[0072] t = X'R - Y'(6)

[0073] Where X'∈R 20*3 , Y'∈R 20*3 For overlapping point cloud sets, the source point cloud is subjected to posture rotation transformation based on the obtained point cloud rotation transformation matrix T, and finally the point cloud map registration result is obtained.

[0074] S106. Based on the network loss function, iteratively train the network model until the loss converges, and obtain an unsupervised registration model of a low-overlap three-dimensional point cloud map.

[0075] Specifically, based on the obtained point cloud rotation transformation matrix, the source point cloud overlapping cluster X all Perform a rotation transformation and overlap cluster Y based on the target point cloud all , calculate the Euclidean distance from the overlapping cluster of the source point cloud to the overlapping cluster of the target point cloud after the rotation transformation, and obtain the overlapping cluster geometric alignment loss L1, the formula is as follows:

[0076] L1 = || X al – Y al ||2(7)

[0077] Where X all is the point cloud set in the overlapping area of ​​the source point cloud and its k nearest neighbor points, Y allThe point cloud set in the overlapping area of ​​the target point cloud and its k nearest neighbor points;

[0078] Based on the weight matrix of the overlapping point pairs to be registered, select the k point pairs with the largest weights, obtain the feature values ​​of each point in the point pair in the global feature, global fusion feature and local fusion feature, calculate the feature distances between the corresponding points' own features, and obtain the overlapping point pair feature consistency loss L2, which is as follows:

[0079] L2 = ||M(ΦX) –M(ΦY)||2+ ||ΦX i ' – ΦY j '||2+||ΦX i ” – ΦY j ”||2(8)

[0080] Based on the overlapping cluster geometric alignment loss and the overlapping point pair feature consistency loss, the registration network loss is fused and calculated to obtain the total network loss L = αL1 + βL2;

[0081] Based on the total network loss, back propagation guides the network model learning and training, and the network model is iteratively trained until the total network loss converges. The maximum number of iterations is 80 epochs, and the final unsupervised registration model of the low-overlap three-dimensional point cloud map is obtained.

[0082] Beneficial Effects

[0083] The present invention proposes an unsupervised registration method for high-precision and low-overlap 3D point clouds, and conducts example tests based on real seabed 3D point cloud map data. The results of the registration network model processing of the present invention are as follows: Figure 3 As shown, Figure 3 (a) Figure 3 (b) are comparison views before and after the point cloud registration obtained by using the method of the present invention based on the acquisition of the seafloor map point cloud. Figure 3 It can be seen from the comparison chart that through the high-precision unsupervised registration method of the three-dimensional point cloud map in this paper, a high-quality low-overlap point cloud registration and stitching result is finally obtained.

[0084] The present invention discloses an unsupervised registration method based on a low-overlap three-dimensional point cloud map. The source point cloud and target point cloud data of the three-dimensional point cloud map to be registered are obtained, and a feature extraction module is used to extract global features of two groups of point clouds; based on the obtained global features, a residual iterative Transformer module is used to capture global information connections to obtain two groups of point cloud global fusion features; based on the global fusion features, the predicted overlap weights of each point in the two groups of point clouds are generated; a key point screening method is used to generate two groups of point cloud overlapping area point cloud sets according to the predicted overlap weights of each point; the k-nearest neighbor cluster local feature fusion is performed on the two groups of point cloud overlapping area point cloud sets, and an overlapping matching matrix is ​​generated; based on the overlapping matching matrix, the weights of the overlapping point pairs to be registered are obtained, and the rotation change matrix is ​​solved according to the overlapping point pairs, and finally the point cloud map registration result is generated; based on the network loss function, the network model is iteratively trained until the loss converges, and an unsupervised registration model of a low-overlap three-dimensional point cloud map is obtained. The present invention makes full use of the global information connection capture capability of the attention mechanism, and proposes a residual Transformer module of multi-head self-attention and cross-attention iteration. The global fusion feature is used to predict the overlap weight of each point, and the point cloud rotation transformation matrix is ​​calculated through the overlapping cluster feature matching matrix. The overlapping cluster geometric alignment loss and the overlapping point pair feature consistency loss are designed. The shortcomings of the existing point cloud registration model, such as poor generalization in low-overlap 3D point cloud map scenarios, insufficient global feature connection capture capability, and excessive reliance on label data, are overcome to achieve high-precision 3D point cloud map registration.

[0085] What is disclosed above is only a preferred embodiment of the present invention, and it certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. An unsupervised registration method based on low-overlap 3D point cloud map, characterized in that: Includes the following: The unsupervised registration method processes the input source point cloud and the target point cloud through seven modules: global feature extraction, global feature fusion, point overlap weight prediction, overlap area locking, overlap matching, rotation transformation matrix solution, and network loss calculation, and finally realizes a high-precision unsupervised registration model for low-overlap three-dimensional point cloud maps. The specific implementation process of the unsupervised registration network model includes: obtaining the source point cloud and target point cloud data of the three-dimensional point cloud map to be registered, and using the feature extraction module to extract global features of the two groups of point clouds; based on the obtained global features, using the residual iterative Transformer module to capture the global information connection and obtain the global fusion features of the two groups of point clouds; based on the global fusion features, generating the predicted overlapping weights of each point in the two groups of point clouds; using the key point screening method, according to the predicted overlapping weights of each point, respectively generating two groups of point cloud overlapping area point cloud sets; by performing k-nearest neighbor cluster local feature fusion on the two groups of point cloud overlapping area point cloud sets, and generating an overlapping matching matrix; based on the overlapping matching matrix, obtaining the weights of the overlapping point pairs to be registered, solving the rotation change matrix according to the overlapping point pairs, and finally generating the point cloud map registration result; based on the network loss function, iteratively training the network model until the loss converges, and obtaining an unsupervised registration model for low-overlap three-dimensional point cloud maps.

2. The unsupervised registration method based on low-overlap three-dimensional point cloud map according to claim 1, characterized in that: Based on the source point cloud and target point cloud data of the 3D point cloud map to be registered, the feature extraction module is used to extract global features of two sets of point clouds, including: Based on the obtained point cloud, perform down sampling of the farthest point and select feature points; Based on the downsampled feature point cloud, a feature extraction module based on DGCNN is adopted to perform global feature extraction on the downsampled source point cloud and target point cloud.

3. The unsupervised registration method based on low-overlap three-dimensional point cloud map according to claim 1, characterized in that: Based on the obtained global features, the residual iterative Transformer module is used to capture the global information connection, and two sets of point cloud global fusion features are obtained, including: Based on the obtained global features, a Transformer module including multi-head self-attention and multi-head cross-attention is used to perform feature connection learning to obtain two sets of fusion features; Based on the initial global fusion features, the two groups of fusion features are filled with features, and the residual module is used to reduce the disappearance of the own features during the learning process to obtain two groups of filled point cloud fusion features. Feature layer normalization and feature maximum pooling are used to iteratively process the generated two groups of global fusion features.

4. The unsupervised registration method based on low-overlap three-dimensional point cloud map according to claim 1, characterized in that: Based on the global fusion features, the predicted overlap weights of each point in two sets of point clouds are generated, including: Based on the two sets of global fusion features generated, a decoding output module including four layers of MLP is adopted, and the last layer uses a fully connected layer with a sigmoid activation function to reduce the impact of inaccurate prediction of corresponding contact points. Finally, the predicted overlapping weights of each point in the two sets of point clouds are obtained.

5. The unsupervised registration method based on low-overlap three-dimensional point cloud map according to claim 1, characterized in that: Using the key point screening method, two groups of point cloud overlapping area point cloud sets are generated respectively according to the predicted overlapping weights of the points, including: Based on the predicted overlapping weights of each point in the two groups of point clouds, the TOP-K key point screening method is used to select the top K points with the highest predicted overlapping weights of the two groups of point clouds as key points; Key points of the two groups of point clouds are extracted to obtain a point cloud set in the overlapping area of ​​the two groups of point clouds.

6. The unsupervised registration method based on low-overlap three-dimensional point cloud map according to claim 1, characterized in that: The k-nearest-neighbor cluster local feature fusion is performed on the point cloud sets in the overlapping area of ​​the two groups of point clouds, and an overlapping matching matrix is ​​generated, including: Based on the points in the point cloud set in the overlapping area of ​​the two groups of point clouds, k nearest neighbor clusters are indexed on the downsampled source point cloud and the target point cloud, and the features of each key point and its k nearest neighbor cluster are spliced ​​and fused to obtain the local fusion features of the key point; Based on the local fusion features of the key points, the negative feature distance method is used to perform fusion feature matching to generate an overlapping matching matrix.

7. The unsupervised registration method based on low-overlap 3D point cloud map according to claim 1, characterized in that: Based on the overlapping matching matrix, the weights of the overlapping point pairs to be registered are obtained, and the rotation change matrix is ​​solved according to the overlapping point pairs, and finally the point cloud map registration result is generated, including: Based on the generated overlapping matching matrix, normalization is performed to obtain the matching weight of each point pair, and a weighted quaternion algorithm is used to solve the point cloud rotation transformation matrix; Based on the obtained point cloud rotation transformation matrix, the source point cloud is subjected to posture rotation transformation, and finally a point cloud map registration result is obtained.

8. The unsupervised registration method based on low-overlap 3D point cloud map according to claim 1, characterized in that: Based on the network loss function, the network model is iteratively trained until the loss converges to obtain an unsupervised registration model for low-overlap 3D point cloud maps, including: Based on the obtained point cloud rotation transformation matrix, overlapping cluster geometric alignment loss and overlapping point pair feature consistency loss are designed to perform registration network loss fusion calculation to obtain the total network loss; Based on the total network loss, back propagation guides the network model learning and training, and the network model is iteratively trained until the total network loss converges to obtain the final unsupervised registration model of the low-overlap three-dimensional point cloud map.

9. The unsupervised registration method based on low-overlap three-dimensional point cloud map according to claim 8, characterized in that: Design overlapping cluster geometric alignment loss and overlapping point pair feature consistency loss to perform registration network loss fusion calculation, including: The point cloud set in the overlapping area of ​​the source point cloud and its k nearest neighbor points are used as the source point cloud overlapping cluster for rotation transformation, and the point cloud set in the overlapping area of ​​the target point cloud and its k nearest neighbor points are used as the target point cloud overlapping cluster, and the Euclidean distance from the source point cloud overlapping cluster to the target point cloud overlapping cluster after the rotation transformation is calculated to obtain the overlapping cluster geometric alignment loss; Based on the weight matrix of the overlapping point pairs to be registered, k point pairs with the largest weights are selected, the feature values ​​of each point in the point pair in the global feature, global fusion feature and local fusion feature are obtained, the feature distances between the corresponding points' own features are calculated, and the feature consistency loss of the overlapping point pairs is obtained.

Images