A robust point cloud registration method and device based on common view query embedding

Abstract

The present invention belongs to the technical field of point cloud data processing, and more specifically, relates to a robust point cloud registration method and device based on common view query embedding. The method includes using KPConv‑FPN as a feature encoder to perform multi-scale sampling on the input point cloud pairs and extract point-by-point features; then the coarsest sampling points and their features are input into the inter-block common view feature interaction module to perform common view feature interaction and common view point segmentation; then based on the enhanced features and common view points, they are input into the filtered block matching module to extract the corresponding relationship at the coarse scale; based on the coarse matching results, the upsampled points and their features are input into the point matching module to generate a dense corresponding point set; using these corresponding point sets, a set of transformation candidates are calculated; finally, registration is performed using a local to global registration scheme. The present invention solves the problems of high training cost and additional computational overhead brought about by the traditional two-stage design, and the low overall efficiency of point cloud registration.

Description

Technical Field

[0001] The present invention belongs to the technical field of point cloud data processing, and more specifically, relates to a robust point cloud registration method and device based on common view query embedding. Background Art

[0002] Rigid point cloud registration is a fundamental task in computer vision and robotics. Its core lies in estimating the rigid transformation that aligns two point clouds. In recent years, learning-based methods have dominated this field, primarily employing feature extractors such as PointNet and KPConv. Researchers have introduced attention mechanisms into point cloud registration. The self-attention mechanism enables global feature interaction within a single point cloud, significantly improving feature representation. The cross-attention mechanism specifically strengthens feature interaction across point clouds, effectively enhancing feature relevance.

[0003] Chinese patent document CN109767464A discloses a point cloud registration method with low overlap rate, which includes the following steps: Step 1: Obtain source point cloud and target point cloud; Step 2: Downsample the obtained source point cloud and target point cloud to obtain point cloud A and point cloud B respectively; Step 3: Divide point cloud A and point cloud B into multiple local areas respectively; Step 4: Perform preliminary matching and select the corresponding four-point congruent set through the 4PCS algorithm; Step 5: Select the optimal transformation of the local area; Step 6: Construct the matching confidence coefficient of the local area, enumerate and calculate all local area matching situations, sort them, and select the local area with the highest confidence coefficient; Step 7: Use the LMedS sampling consistency algorithm to perform alignment to obtain the final rigid transformation matrix to complete the point cloud.

[0004] In summary, accurately distinguishing between overlapping and non-overlapping areas between two point clouds is crucial for establishing the correct correspondence. Existing registration methods that distinguish overlapping areas can be divided into two categories: implicit methods and explicit methods. Implicit methods indirectly drive the neural network to learn regional distinction capabilities through overlapping perception loss functions, such as Prodator and Geo Transformer, but this approach cannot guarantee that the network truly masters the distinction ability. In contrast, explicit methods adopt a two-stage strategy: first, the overlapping area between the two point clouds is estimated through a segmentation model, and then it is used as prior information to guide feature interaction. Although this method explicitly uses overlapping information to enhance feature discriminability, the two-stage method requires step-by-step training, and the independent region estimation module also increases training and computational costs, reducing the overall efficiency of point cloud registration. Summary of the Invention

[0005] The present invention aims to overcome at least one of the above-mentioned defects of the prior art and provide a robust point cloud registration method and device based on common view query embedding to solve the problems of high training cost and additional computational overhead brought about by the traditional two-stage design and low overall efficiency of point cloud registration.

[0006] The detailed technical solutions of the present invention are as follows:

[0007] A robust point cloud registration method based on common view query embedding, the method comprising:

[0008] S1. Obtain a 3D point cloud pair, i.e., a source point cloud and a target point cloud, generated by scanning the same scene at different angles. The sampled points, i.e., super points and dense points, are obtained through encoder-point sampling. At the same time, point-by-point local features are extracted to obtain the sampled points and the corresponding local features, i.e., super point features and dense point features.

[0009] S2, perform inter-block matching between the sampling points and the corresponding local features to obtain the super-point correspondence set of the source point cloud and the target point cloud;

[0010] S3. Assign the dense points and dense features of the source point cloud and the target point cloud to the superpoint closest to them in the geometric space to form a patch corresponding to each superpoint; then, calculate the similarity matrix within each patch in the source point cloud and the target point cloud to obtain the local dense point correspondence set: use the local to global registration method to filter the final rigid transformation matrix from the local dense point correspondence set, so that the source point cloud and the target point cloud are successfully registered.

[0011] Furthermore, the S1 specifically includes:

[0012] S11: Use KPConv-FPN as the backbone network to perform downsampling and local feature extraction on the input source point cloud P and target point cloud Q at different scales. The output downsampled points are called super points, and { , }, the corresponding local feature, that is, the super-point feature is expressed as , where d represents the dimension of local features, Represents a feature set;

[0013] Superpoint features are calculated by aggregating the features of the neighboring points of each superpoint and using a distance-weighted approach;

[0014] S12: After the backbone network completes downsampling of the source point cloud and the target point cloud, it gradually upsamples the downsampled points. During the upsampling process, local point features are extracted for each upsampled point. After upsampling, the output upsampled points are dense points with the resolution of the original 3D point cloud. and 1 / 2 of the corresponding local point features, that is, dense point features, are recorded as and , where d represents the dimension of the feature, Represents a feature set.

[0015] Furthermore, the S2 specifically includes:

[0016] S21. First stage: Perform geometric attention transformation on super-points and their corresponding super-point features, encoding the relative position information of each super-point in the point cloud into the super-point features of each point cloud; at the same time, encode the geometric and contextual information within and across point clouds, capture local and global feature relationships, and output super-point geometric features that are transformation invariant;

[0017] S22, the second stage: performing common view feature transformation and common view point segmentation on the super-point geometric features, and outputting the common view features and common view probability maps of the source point cloud and the target point cloud;

[0018] S23, the third stage: based on the obtained common view features, a secondary geometric attention transformation is performed again to achieve feature enhancement to obtain global geometric features;

[0019] S24. Finally, the global geometric features and their common view probability maps are matched between blocks, and the super-point correspondence between the source point cloud and the target point cloud is finally output.

[0020] Furthermore, the geometric attention transformation is used to encode the geometric structure information of the point cloud and realize the interaction of geometric features within and across point clouds, specifically including: a geometric structure encoding layer, a self-attention layer, and a cross-attention layer;

[0021] The geometric structure encoding layer is responsible for encoding the relative position information between superpoints in the point cloud. The self-attention layer is used to learn the global correlation of superpoints in the feature space and geometric space within a single point cloud. The cross-attention layer, as a typical module for point cloud registration tasks, focuses on feature exchange and geometric consistency encoding between two input point clouds.

[0022] The processing of geometric attention transformation is as follows:

[0023] First, the superpoints of the source point cloud and the target point cloud and their corresponding superpoint features are taken as input, and the relative position information between the superpoints is processed by the geometric structure encoding layer. Then, the global superpoint features of each point cloud are encoded using the self-attention layer. Finally, these global features are input into the cross-attention layer to complete the feature interaction of the two point clouds, and finally the hybrid features with transformation invariance, namely the superpoint geometric features, are output, which are expressed as { },in and represents the super points after downsampling of the source point cloud and the target point cloud, Indicates the first stage.

[0024] Furthermore, the common view feature transformation and common viewpoint segmentation include two parts: common view attention transformation and common viewpoint segmentation:

[0025] The common view attention transformation performs common view transformation based on the extracted super-point geometric features to achieve the interaction and alignment of dual point cloud features, and outputs the common view features of the source point cloud and the target point cloud;

[0026] Common viewpoint segmentation further calculates the overlap fraction of the superpoints of the two point clouds based on the common view features, and extracts the common view area of ​​the two point cloud pairs. The overlap fraction effectively reflects the distribution of the common view area between the two point clouds, that is, a common view probability map with clear geometric meaning is then generated based on the transformed common view features.

[0027] The co-viewing attention transformation specifically includes:

[0028] First, the common view information of the super-point geometric features is decoded to obtain the common view information of the two point clouds; specifically, the query vector Q is initialized and the common view information of the two point clouds is obtained by decoding using the Transformer decoder. , expressed as { }:

[0029] (1);

[0030] In formula (1), Represents the super point after downsampling of the i-th source point cloud and target point cloud, and outputs { } represents the geometric features from the super point { The common view information decoded in} represents the overlapping information of two point cloud pairs, Q represents the initialized query vector, { } represents the output super-point geometric features, Indicates the first stage, and Represents the super points after downsampling of the source point cloud and the target point cloud, q, k, v represent The receiving variable of the input;

[0031] Then, the decoded common view information { }As a priori information and the superpoint geometric features{ } is input into the cross attention layer to obtain the common view feature of the two point clouds represented as { },in Indicates the second stage;

[0032] Finally, in order to make the common view feature { } is more discriminative, and the obtained common view features { } is used as input and a secondary geometric attention transformation is performed to achieve deep encoding and optimization of features, and the final output is the enhanced global geometric features and ,in Indicates the third stage.

[0033] Furthermore, the common viewpoint segmentation is specifically as follows:

[0034] First, the decoded common view information is transformed into Projection, constructing a weight map, which is used to enhance the common view feature and Next, a 1×1 convolution operation is applied to the obtained weight map and processed by the Sigmoid activation function to output the overlap score of the source point cloud and the target point cloud superpoints, as follows:

[0035] (2);

[0036] (3);

[0037] In formulas (2)~(3), represents the superpoint after downsampling of the i-th source point cloud and target point cloud, and Represents the common view feature of two point clouds, { } represents the common view information of two point clouds, represents a weight map, which is used to represent the probability that each superpoint is located in the overlapping area, represents element-wise multiplication, represents the Sigmoid activation function, express and The common view probability map is specifically expressed as { , }, Represents a 1×1 convolution operation.

[0038] Furthermore, the inter-block matching specifically includes:

[0039] Patch matching with filtering using common view probability map { , Construct an overlap matrix and fuse it with the feature similarity matrix to filter out false matches in non-overlapping regions;

[0040] Given global geometric features and , calculate their similarity matrix and overlap matrix :

[0041] (4);

[0042] In formula (4), and They are super points { , } midpoint index, and Represents the inner product operation.

[0043] Next, these two matrices are fused by element-wise multiplication to obtain the feature matching matrix :

[0044] (5);

[0045] In formula (5), The operation is to convert the similarity matrix and overlap matrix Multiply the elements of the corresponding positions to filter out invalid matches outside the common view area. Finally, from the feature matching matrix Select the first k matching pairs as the super-point correspondence set:

[0046] (6);

[0047] In formula (6), is the feature matching matrix between the source point cloud and the target point cloud, n and m represent the super point indexes of the two respectively, represents the superpoint matching set of the source point cloud filtered from the matching matrix, is a collection The elements in is the corresponding superpoint matching set in the target point cloud, is a collection Elements in

[0048] Finally, these matching pairs will be point matched for further refined matching. Represents the set of superpoint correspondences.

[0049] Furthermore, the S3 is specifically:

[0050] S31. First, the dense points and dense features of the source point cloud and the target point cloud are assigned to the nearest superpoints in the geometric space to form a patch corresponding to each superpoint;

[0051] Given a superpoint , the patch assigned to it is denoted as .and The associated feature group is defined as ,in Similarly, calculate and express Patches in and feature groups , for each superpoint correspondence , first calculate the similarity matrix :

[0052] (7);

[0053] In formula (7), d is the feature dimension, and Represents the source point cloud and target point cloud Corresponding superpoint index

[0054] S32, calculate the similarity matrix in each patch of the source point cloud and the target point cloud, and obtain the local dense point correspondence set: use the Sinkhorn algorithm to calculate the similarity matrix Optimize and get the allocation matrix ; Then, from Extract the first k correspondences and construct a local dense point correspondence set

[0055] S33, then aggregate the local dense point correspondences of all patches, all local dense point correspondences are set Merge into the final global dense point correspondence set ,in is the number of superpoint correspondences;

[0056] S34. In order to avoid the slow convergence and instability of RANSAC iteration, a local to global registration scheme is used to estimate the transformation of the final output, using weighted SVD from the local dense point correspondence set. Compute the local candidate rigid transformation matrix Set, then calculate the inlier rate of each transformation candidate, and select the transformation matrix with the largest inlier rate as the final rigid transformation matrix ;

[0057] The calculation can be expressed as follows:

[0058] (8);

[0059] In formula (8), τ represents the acceptance radius, and [⋅] is the Iverson bracket;

[0060] The source point cloud and target point cloud are successfully registered through the rigid transformation matrix T.

[0061] In another aspect of the present invention, a device for a robust point cloud registration method based on common view query embedding is provided, the device comprising: a data acquisition module, a feature extraction module, an inter-block matching module, a point matching module, and a pose screening and refinement module;

[0062] The data acquisition module is used to acquire three-dimensional point clouds generated by scanning the same scene at different angles;

[0063] The feature extraction module is used to extract point-by-point local features of the three-dimensional point cloud;

[0064] The inter-block matching module performs inter-block matching on the sampling points and the corresponding features to obtain corresponding point sets of the source point cloud and the target point cloud;

[0065] The inter-block matching module includes a geometric attention transformation module, a common view attention transformation module, a common viewpoint segmentation module and a block matching module with filtering;

[0066] The geometric attention transform is used to encode the geometric structure information of the point cloud and realize the interaction of geometric features within and across point clouds. It specifically includes: a geometric structure encoding layer, a self-attention layer, and a cross-attention layer;

[0067] Geometric Attention Transformation Module: The geometric structure encoding layer is responsible for encoding the relative position information between superpoints in the point cloud. The self-attention layer is used to learn the global correlation of superpoints in the feature space and geometric space within a single point cloud. The cross-attention layer, as a typical module for point cloud registration tasks, focuses on feature exchange and geometric consistency encoding between two input point clouds.

[0068] The common view attention transformation module performs common view transformation based on the extracted super point geometric features to achieve interaction and alignment of dual point cloud features, and outputs the common view features of the source point cloud and the target point cloud;

[0069] The common viewpoint segmentation module further calculates the overlap score of the superpoints of the two point clouds based on the common view feature, and extracts the common view area of ​​the two point cloud pairs and represents it as a common view probability map;

[0070] The block matching module with filtering constructs an overlap matrix through the common view probability map and fuses it with the feature similarity matrix to filter out false matches in non-overlapping areas and obtain a super point correspondence set;

[0071] The point matching module assigns the dense points and their dense features of the source point cloud and the target point cloud to the superpoint closest to them in the geometric space, forming a patch corresponding to each superpoint; then, the similarity matrix within each patch in the source point cloud and the target point cloud is calculated to obtain a local dense point correspondence set;

[0072] The pose screening and refinement module is used to obtain a series of local candidate rigid transformation matrices from the local dense point correspondence set using weighted SVD, calculate the inlier rate of each candidate transformation, and select the local candidate rigid transformation matrix with the largest inlier rate as the final rigid transformation matrix, so that the source point cloud and the target point cloud are successfully aligned.

[0073] In another aspect of the present invention, a computer-readable storage medium is provided, which stores executable instructions. When the instructions are executed, the machine performs the robust point cloud registration method based on common view query embedding as described above.

[0074] Compared with the prior art, the present invention has the following beneficial effects:

[0075] (1) The present invention provides a robust point cloud registration method and device based on common view query embedding, and proposes a common view feature interaction method for partial point cloud registration. This method couples common view point segmentation with feature interaction, which can effectively reduce unnecessary computational overhead while enhancing feature interaction in the common view area.

[0076] (2) The present invention provides a robust point cloud registration method and device based on common view query embedding, and proposes a common view region dual constraint strategy to improve the matching quality of corresponding points. At the feature level, the common view query embedding mechanism is introduced to effectively eliminate the interference of non-overlapping regions; at the matching level, the common view segmentation mask is used to filter out erroneous correspondences.

[0077] (3) The present invention provides a robust point cloud registration method and device based on common view query embedding, which cascades a feature interaction module based on common view attention after the backbone network, couples the feature interaction with the common view feature decoding, and enables the common view feature to guide the global feature interaction, making the encoded feature more discriminative. By embedding the query vector into the two point cloud pairs to obtain the common view area, the introduction of an additional estimation module is avoided. Compared with the scheme of introducing an independent estimation network to obtain the overlapping area, it has fewer parameters and lower time complexity. At the same time, a lightweight segmentation module is used to obtain the common view area to further filter out unmatched features. By explicitly utilizing the overlapping area of ​​the point cloud, it can effectively focus on the feature interaction of the overlapping area, significantly reduce the interference of the non-overlapping area, and achieve a more reliable rough matching through the common view feature interaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 It is a flowchart of the robust point cloud registration method based on common view query embedding described in the present invention.

[0079] Figure 2 It is a schematic diagram of the overlapping area in Example 1 of the present invention.

[0080] Figure 3 This is a co-image diagram of the segmentation result in Example 1 of the present invention.

[0081] Figure 4 It is a schematic diagram of the effects in partial overlap and low overlap scenarios in Example 1 of the present invention. DETAILED DESCRIPTION

[0082] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0083] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0084] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0085] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0086] Example 1

[0087] Ginseng Figure 1 ,This embodiment provides a robust point cloud registration method based on ,common-view query embedding, which is built on GeoTrans and also follows the ,coarse-to-fine registration paradigm.

[0088] Point cloud registration is a fundamental task in computer vision and robotics. It involves using an airborne LiDAR to scan two point clouds from different angles of the same scene, referred to as the source point cloud and the target point cloud. The goal of point cloud registration is to predict the correspondence between the two point clouds and then estimate a rigid transformation matrix to transform the source point cloud into the target point cloud's coordinate system while aligning their overlapping regions. The method specifically involves:

[0089] S1, encoder - point sampling and local feature extraction;

[0090] The airborne laser radar obtains three-dimensional point cloud pairs, namely source point cloud and target point cloud, generated by scanning the same scene at different angles. The sampling points are obtained through encoder-point sampling, and point-by-point local features are extracted to obtain sampling points and corresponding local features. The encoder processes the input point cloud pairs through multi-scale sampling, thereby obtaining sampling points at different scales and extracting point-by-point local features.

[0091] Said S1 comprises the following steps:

[0092] S11: Use KPConv-FPN as the backbone network to perform downsampling and local feature extraction on the input source point cloud P and target point cloud Q at different scales. The output downsampled points are called super points, and { ,}, the corresponding local feature, that is, the super-point feature is expressed as , where d represents the dimension of local features, Represents a feature set;

[0093] Superpoint features are calculated by aggregating the features of the neighboring points of each superpoint and using a distance-weighted approach;

[0094] The local features are obtained by aggregating the features of the neighboring points of each superpoint and calculating them in a distance-weighted manner. In order to distinguish them from other features below, these superpoints and their superpoint features are then input into the inter-block matching module for common view feature interaction.

[0095] S12: After the backbone network completes downsampling of the source point cloud and the target point cloud, it gradually upsamples the downsampled points. During the upsampling process, local point features are extracted for each upsampled point. After upsampling, the output upsampled points are dense points with the resolution of the original 3D point cloud. and 1 / 2 of the corresponding local point features, that is, dense point features, are recorded as and , where d represents the dimension of the feature, Represents a feature set;

[0096] The method of obtaining local point features is consistent with the method of obtaining local super-point features in S11, that is, by aggregating the features of neighboring points and using distance weighted calculation.

[0097] S2. Perform inter-block matching between the sampling points and the corresponding local features to obtain a set of super-point correspondences between the source point cloud and the target point cloud.

[0098] S2 specifically includes a cross-patch common view feature interaction part and a block matching part with filtering, wherein the cross-patch common view feature interaction part includes three stages:

[0099] S21. In the first stage, geometric attention transformation is performed on super-points and their corresponding super-point features, and the relative position information of each super-point in the point cloud is encoded into the super-point features of each point cloud; at the same time, the geometric and contextual information within and across point clouds is encoded to capture local and global feature relationships, and output super-point geometric features with transformation invariance.

[0100] S22. In the second stage, the super-point geometric features are transformed into common view features and common view points are segmented, and the common view features and common view probability maps of the source point cloud and the target point cloud are output.

[0101] S23, the third stage, based on the acquired common view features, a secondary geometric attention transformation is performed again to achieve feature enhancement to obtain global geometric features.

[0102] S24. Finally, the global geometric features and their common view probability maps are matched between blocks, and the super-point correspondence between the source point cloud and the target point cloud is finally output.

[0103] Furthermore, the geometric attention transformation is used to encode the geometric structure information of the point cloud and realize the interaction of geometric features within and across point clouds, specifically including: a geometric structure encoding layer, a self-attention layer, and a cross-attention layer;

[0104] The geometric structure encoding layer is responsible for encoding the relative position information between superpoints in the point cloud. The self-attention layer is used to learn the global correlation of superpoints in the feature space and geometric space within a single point cloud. The cross-attention layer, as a typical module for point cloud registration tasks, focuses on feature exchange and geometric consistency encoding between the two input point clouds.

[0105] The processing of geometric attention transformation is as follows:

[0106] First, the superpoints of the source point cloud and the target point cloud and their corresponding superpoint features are taken as input, and the relative position information between the superpoints is processed by the geometric structure encoding layer. Then, the global superpoint features of each point cloud are encoded using the self-attention layer. Finally, these global features are input into the cross-attention layer to complete the feature interaction of the two point clouds, and finally the hybrid features with transformation invariance, namely the superpoint geometric features, are output, which are expressed as { },in and represents the super points after downsampling of the source point cloud and the target point cloud, Indicates the first stage.

[0107] Furthermore, the common view feature transformation and common viewpoint segmentation include two parts: common view attention transformation and common viewpoint segmentation:

[0108] The common view attention transformation performs common view transformation based on the extracted super-point geometric features to achieve the interaction and alignment of dual point cloud features, and outputs the common view features of the source point cloud and the target point cloud; the common view point segmentation further calculates the overlap score of the super-points of the two point clouds based on the common view features, and extracts the common view area of ​​the two point cloud pairs. The overlap score effectively reflects the distribution of the common view area between the two point clouds, that is, a common view probability map with clear geometric meaning is then generated based on the transformed common view features.

[0109] Since the global feature interaction mechanism based on Transformer can effectively encode the global structural information of point clouds, in registration tasks with low overlap or partial overlap, such as Figure 1As shown in the figure, the feature interaction in the non-overlapping area may interfere with the feature representation in the overlapping area. This interference will reduce the distinguishability of the features and thus produce more erroneous matching results. To solve this problem, the present invention proposes a method for explicitly utilizing the overlapping information of point clouds. By designing a common view feature interaction module, it aims to focus on the overlapping area of ​​two point clouds and obtain more discriminative common view features, which helps in the subsequent feature transformation process to make the transformed features more discriminative. The existing methods mostly establish correspondences in non-overlapping areas, such as Figure 2 As shown in (a) and (b) in FIG, the method of the present invention is established within the overlapping area, as shown in FIG. Figure 2 As shown in (c):

[0110] The co-viewing attention transformation specifically includes:

[0111] First, the common view information of the super-point geometric features is decoded to obtain the common view information of the two point clouds. Specifically, the query vector Q is initialized and the common view information of the two point clouds is obtained by decoding using the Transformer decoder. , expressed as { }:

[0112] (1);

[0113] In formula (1), Represents the super point after downsampling of the i-th source point cloud and target point cloud, and outputs { } represents the geometric features from the super point { The common view information decoded in} represents the overlapping information of two point cloud pairs, Q represents the initialized query vector, { } represents the output super-point geometric features, Indicates the first stage, and Represents the super points after downsampling of the source point cloud and the target point cloud, q, k, v represent The receiving variable of the input.

[0114] Next, the decoded common view information { } as prior information together with the superpoint geometric features output by S21 { } is input into the cross attention layer to obtain the common view feature of the two point clouds represented as { },in Indicates the second stage.

[0115] Finally, in order to make the common view feature { } is more discriminative, and the obtained common view features { } is used as input and a secondary geometric attention transformation is performed to achieve deep encoding and optimization of features, and the final output is the enhanced global geometric features and ,in Indicates the third stage.

[0116] After decoding the common view features, we design a lightweight common view point segmentation module to extract the common view area of ​​two point cloud pairs. The common view area segmentation task is regarded as a prediction problem of logarithmic probability map, where each element represents the probability that the point is located in the overlapping area. The common view map of the segmentation result is as follows: Figure 3 As shown, it is visualized through two different scenes. The first and third columns show the input point clouds (point cloud pairs), where the green dots represent the predicted common viewpoints between the two point clouds; the second and fourth columns show the true transformation, i.e., the rigid transformation, and the black box represents the true overlapping area between the two point clouds. The overlap rate of the first rigid change is 11.4%, and the overlap rate of the second rigid change is 28.9%. The rigid transformation is: the source point cloud is transformed to the target point cloud through a rigid transformation matrix to achieve alignment of the overlapping areas.

[0117] The common viewpoint segmentation specifically includes:

[0118] Specifically, first, the decoded common view information is converted into Projection, constructing a weight map, which is used to enhance the common view feature and The common view context information in the source point cloud is obtained; next, a 1×1 convolution operation is applied to the obtained weight map and processed by the Sigmoid activation function. The specific calculation process is as shown in the formula below, and the overlap score of the source point cloud and the target point cloud super point is output, that is, the formula probability map of the two point clouds:

[0119] (2);

[0120] (3);

[0121] In formulas (2)~(3), represents the super points after downsampling of the source point cloud and the target point cloud, and Represents the common view feature of two point clouds, { } represents the common view information of two point clouds, represents a weight map, which is used to represent the probability that each superpoint is located in the overlapping area, represents element-wise multiplication, represents the Sigmoid activation function, express and The common view probability map is specifically expressed as { , }, Represents a 1×1 convolution operation.

[0122] Furthermore, the inter-block matching specifically includes:

[0123] Patch matching with filtering via common view probability map { , Construct an overlap matrix and fuse it with the feature similarity matrix to filter out false matches in non-overlapping regions;

[0124] Specifically, given the global geometric features and , calculate their similarity matrix and overlap matrix :

[0125] (4);

[0126] In formula (4), and Super point and The index of the midpoint, and Represents the inner product operation.

[0127] Next, these two matrices are fused by element-wise multiplication to obtain the feature matching matrix That is, the correlation matrix:

[0128] (5);

[0129] In formula (5), The operation is to convert the similarity matrix and overlap matrix Multiply the elements of the corresponding positions to filter out invalid matches outside the common view area. Finally, from the feature matching matrix Select the first k matching pairs as the super point correspondence set:

[0130] (6);

[0131] In formula (6), is the feature matching matrix between the source point cloud and the target point cloud, n and m represent the super point indexes of the two respectively, represents the superpoint matching set of the source point cloud filtered from the matching matrix, is a collection The elements in is the corresponding superpoint matching set in the target point cloud, is a collection Elements in .

[0132] Finally, these matching pairs will be point matched for further refined matching. Represents the set of superpoint correspondences.

[0133] Step S3: point matching and pose estimation;

[0134] After obtaining the superpoint correspondence set Afterwards, a point matching module is used to extract point correspondences, such as the point matching module in GeoTrans. The method of the present invention adopts a point-to-node strategy:

[0135] The dense points and dense features of the source point cloud and the target point cloud are assigned to the superpoint closest to them in the geometric space to form a patch corresponding to each superpoint; then, the similarity matrix within each patch in the source point cloud and the target point cloud is calculated to obtain the local dense point correspondence set: using the local to global registration method, the final rigid transformation matrix is ​​screened from the local dense point correspondence set to successfully register the source point cloud and the target point cloud.

[0136] Specifically, the S3 is:

[0137] S31. Given a superpoint , the patch assigned to it is denoted as ,and The associated feature group is defined as ,in ; Similarly, calculate and express Patches in and feature groups , for each superpoint correspondence , first calculate the similarity matrix :

[0138] (7);

[0139] In formula (7), d is the feature dimension. and Represents the source point cloud and target point cloud The corresponding superpoint index.

[0140] S32, then, the present invention uses the Sinkhorn algorithm to calculate the similarity matrix Optimize and get the allocation matrix ; Then, from Extract the first k correspondences and construct a local dense point correspondence set .

[0141] S33. Then, all local dense points correspond to the set Merge into the final global dense point correspondence set ,in is the number of superpoint correspondences.

[0142] S34. In order to avoid the slow convergence and instability of RANSAC iteration, a local to global registration scheme is used to estimate the transformation of the final output, using weighted SVD from the local dense point correspondence set. Compute the local candidate rigid transformation matrix Set, then calculate the inlier rate of each transformation candidate, and select the transformation matrix with the largest inlier rate as the final rigid transformation matrix ;

[0143] The calculation can be expressed as follows:

[0144] (8);

[0145] In formula (8), τ represents the acceptance radius, and [⋅] is the Iverson bracket;

[0146] The source point cloud and target point cloud are successfully registered through the rigid transformation matrix T.

[0147] Furthermore, the present invention conducts an in-depth analysis of the feature interaction mechanism of the coarse-to-fine point cloud registration method at the coarse scale through experiments: the previous global interaction method based on transformer often introduces interference from non-overlapping areas, which hinders the previous methods from achieving more robust performance. To this end, the present invention proposes that in partial overlap and low overlap scenarios, common view area estimation can effectively improve the registration performance, such as Figure 4 shown.

[0148] We compare the visualization effects of Geo Transformer and our method, and show that the common view feature interaction mechanism explicitly guides the model to focus on overlapping areas, while its learnable segmentation module effectively eliminates abnormal matches. Experimental results show that even in challenging scenes, such as walls that lack unique geometric features, Figure 4 As shown in (e), and the bedroom furniture arrangement with local symmetry, such as Figure 4 As shown in (f), the proposed method can also reliably establish accurate patch-level correspondences.

[0149] In order to verify the effectiveness of the present invention, the point clouds of the present invention in different scenes are visualized. Figure 4 The first line shows the visualization of the first scenario, and the second line shows the visualization of the second scenario. These are visualizations in different scenarios. Take the visualization of the first scenario as an example:

[0150] (a) represents the input point cloud, i.e., the source point cloud (orange) and the target point cloud (blue), without rigid transformation. It can be seen that the two point clouds are not aligned;

[0151] (b) shows the application of true rigid transformation to the source point cloud. It can be seen that the source point cloud and the target point cloud are aligned. The overlap ratio represents the ratio of the number of overlapping points in the source point cloud to the number of points in the source point cloud. The larger the ratio, the higher the overlap ratio. The overlap ratio is 26.1% in the first scenario and 18.4% in the second scenario.

[0152] (c) shows the rigid transformation moment predicted by the GeoTransformer method for the source point cloud. It is found that the transformation effect is not ideal and there is a large error. The root mean square error indicates the degree of alignment between the source point cloud and the target point cloud after the transformation. The lower the error, the higher the registration accuracy. The root mean square error in the first scenario is 1.193m, and the root mean square error in the second scenario is 1.031m.

[0153] (d) shows the application of the rigid transformation matrix predicted by the present invention to the source point cloud. It is found that the effect is significantly improved, and the root mean square error is significantly smaller than the error in (c). This shows that the registration method of the present invention has higher accuracy and smaller error. The root mean square error in the first scenario is 0.048m, and the root mean square error in the second scenario is 0.053m.

[0154] (e) represents the correspondence predicted by GeoTransformer, and the connecting lines represent the super-point correspondence established at a coarse scale:

[0155] Red indicates an incorrect prediction; green indicates a correct prediction. The number of correspondences indicates that the 256 pairs of coarse-scale super-point correspondences predicted by GeoTransformer were filtered. The number of corresponding points in the first scenario is 256, and the number of corresponding points in the second scenario is also 256.

[0156] The inlier rate indicates the proportion of correctly matched point pairs. The lower inlier rate of GeoTransformer indicates that the number of correctly predicted correspondences is small. The inlier rate is 1.33% in the first scenario and 11.6% in the second scenario.

[0157] (f) represents the number of correspondences predicted by the present invention. From the inlier rate, it can be seen that the super-point correspondences predicted by the present invention at a coarse scale are significantly higher than those of the GeoTransformer method. The number of corresponding points in the first scenario is 256, and the number of corresponding points in the second scenario is also 256; the inlier rate in the first scenario is 59.7%, and the inlier rate in the second scenario is 44.1%.

[0158] However, introducing a complete common view area estimation network will consume a lot of computing resources and time costs. To solve this problem, the present invention cascades a common view attention transformation module after the backbone network. This module couples feature interaction with common view feature decoding, so that the common view feature can guide the global feature interaction, making the encoded features more discriminative. By embedding the query vector into the two point cloud pairs to obtain the common view area, the introduction of an additional estimation module is avoided. Compared with the solution of introducing an independent estimation network to obtain the overlapping area, it has fewer parameters and lower time complexity. At the same time, a lightweight common view point segmentation module can be used to obtain the common view area to further filter out mismatched features. By explicitly utilizing the overlapping area of ​​the point cloud, the present invention can effectively focus on the feature interaction of the overlapping area and significantly reduce the interference of the non-overlapping area.

[0159] Through the interaction of common view features, a more reliable rough match can be achieved. Extensive experiments on the 3DMatch and KITTI datasets show that, especially in the indoor dataset registration task, the method of the present invention surpasses all previous methods with scores of 93.9 and 75.0, respectively, verifying its effectiveness in high and low overlap situations. The specific results are shown in Tables 1 and 2:

[0160] On outdoor datasets, our method achieved top performance across all metrics. By introducing a common-view feature interaction mechanism, we successfully improved the registration accuracy of Geo Transformer, validating the importance of rational feature interaction in point cloud registration. On the synthetic datasets ModelNet40 and ModelLoNet40, our method achieved state-of-the-art performance across nearly all metrics, demonstrating its strong universality.

[0161] Table 1 Experimental results of the present invention on 3DMatch and 3DLoMatch datasets;

[0162]

[0163] The effectiveness of our method was verified on the indoor datasets 3DMatch and 3DLoMatch. The overlap between the two point clouds in 3DMatch was greater than 30%, while the overlap in 3DLoMatch was greater than 10% and less than 30%. The specific evaluation metric remained the same as in previous methods: RR (registration success rate).

[0164] Specifically, the present invention first uses the RANSAC estimator to use 50k iterations to calculate the rigid transformation T. The present invention finds that under the standard of 5000 sampling points, the present invention's registration success rate achieves good performance on both 3DMatch and 3DLoMatch datasets. At the same time, the present invention also uses the method in Geo Transformer: LGR (local to global registration method) to calculate the rigid transformation T. It can be seen that the registration success rate is higher than GeoTransformer, achieving the optimal result. The above information shows the effectiveness of the method of the present invention.

[0165] Table 2 Experimental results of the present invention and other registration methods on the KITTI dataset under the same settings;

[0166]

[0167] To verify the robustness of our method, we tested it on the outdoor dataset KITTI. We used RANSAC (with 50k iterations) and LGR as estimators to solve the rigid transformation matrix T. The effectiveness of our method was evaluated using translation error, rotation error, and registration success rate. Translation error and rotation error provide a direct reflection of the accuracy of the predicted rigid transformation matrix T. The experimental results show that the RANSAC-based method achieves the best translation error and registration success rate, with the rotation error ranking second, only 0.04m lower than the optimal result. This fully demonstrates the effectiveness of our method on outdoor datasets. Furthermore, the LGR-based method also performs well, ultimately achieving the best performance. These experimental results not only validate the superiority of our method on indoor datasets but also demonstrate its applicability in outdoor scenarios, further confirming the broad effectiveness of the proposed method.

[0168] Table 3 Experimental results of the present invention on ModelNet40 and ModelLoNet40 datasets;

[0169]

[0170] For a comprehensive comparison, we also conducted experiments on synthetic datasets. Following the settings in the Predator method, we used 5112 CAD samples from ModelNet40 for training, 1202 CAD samples for validation, and 1266 CAD samples for testing. We also used ModelLoNet40 from Predator, which achieved an overlap of 53.6%, compared to 73.5% for ModelNet40. We compared our method with state-of-the-art learning-based methods. RPM-Net and RGM are specifically designed for registration on synthetic datasets, while the other methods are more general. All methods used their optimal settings. Following the settings in Predator, we used relative rotation error, relative translation error, and chamfer distance as evaluation metrics. In most cases, our method achieved comparable results. Compared to Geo Transformer, our method demonstrated superior registration performance under large rotation conditions, demonstrating its excellent generalization capability.

[0171] Table 4 Experimental results of ablation experiments on 3DMatch and 3DLoMatch datasets

[0172]

[0173] The present invention conducts a systematic ablation study on 3DMatch to verify the effectiveness of each component in the method of the present invention.

[0174] We gradually remove key modules to analyze their impact on registration performance. In practice, we find that the LGR registration estimator is more stable and faster than the RANSAC estimator while achieving comparable performance (see Table 1).

[0175] We perform an ablation study on 3DMatch and 3DLoMatch by gradually adding key components, as shown in Table 4. Since the common view point segmentation (CPS) and the block matching module with filtering (PMF) rely on the common view information extracted by feature interaction with common view attention (FICA), we first replace the geometric transformer (GeometricTransformer) in Geo Transformer with the FICA module, defined as Var1.

[0176] Experimental results show that Var1’s RR on 3DMatch is 0.2pp higher than Geo Transformer, but it decreases by 2.2pp on 3DMatch and 3.1pp on 3DLoMatch compared to the complete method of the present invention. Experimental results show that Var1’s RR on 3DMatch is 0.2pp higher than Geo Transformer, but it decreases by 2.2pp on 3DMatch and 3.1pp on 3DLoMatch compared to the complete method of the present invention. Next, the present invention adds a CPS module, defined as Var2. The results show that Var2 improves by 0.9pp on 3DMatch and 3.0pp on 3DLoMatch compared to Var1, indicating that the module can effectively extract the overlapping area between two point clouds, especially on 3DLoMatch. This also indirectly proves that explicitly utilizing the overlapping area can improve the registration performance. Finally, the present invention adds a PMF module to complete QER. The results show that QER outperforms Var2 by 1.3pp on 3DMatch and by 1.0pp on 3DLoMatch, both of which are higher than Geo Transformer.

[0177] In summary, the model performs best when all three modules exist simultaneously, and deleting any one of them leads to a decrease in performance, indicating that the synergistic effect of these three modules is crucial.

[0178] Example 2

[0179] This embodiment provides a device for implementing a robust point cloud registration method based on common view query embedding, the device comprising: a data acquisition module, a feature extraction module, an inter-block matching module, a point matching module, and a pose screening and refinement module;

[0180] The data acquisition module is used to acquire three-dimensional point clouds generated by scanning the same scene at different angles;

[0181] The feature extraction module is used to extract point-by-point local features of the three-dimensional point cloud;

[0182] The inter-block matching module performs inter-block matching on the sampling points and the corresponding features to obtain corresponding point sets of the source point cloud and the target point cloud;

[0183] The inter-block matching module includes a geometric attention transformation module, a common view attention transformation module, a common viewpoint segmentation module and a block matching module with filtering;

[0184] The geometric attention transform is used to encode the geometric structure information of the point cloud and realize the interaction of geometric features within and across point clouds. It specifically includes: a geometric structure encoding layer, a self-attention layer, and a cross-attention layer;

[0185] Geometric Attention Transformation Module: The geometric structure encoding layer is responsible for encoding the relative position information between superpoints in the point cloud. The self-attention layer is used to learn the global correlation of superpoints in the feature space and geometric space within a single point cloud. The cross-attention layer, as a typical module for point cloud registration tasks, focuses on feature exchange and geometric consistency encoding between two input point clouds.

[0186] The common view attention transformation module performs common view transformation based on the extracted super point geometric features to achieve interaction and alignment of dual point cloud features, and outputs the common view features of the source point cloud and the target point cloud;

[0187] The common viewpoint segmentation module further calculates the overlap score of the superpoints of the two point clouds based on the common view feature, and extracts the common view area of ​​the two point cloud pairs and represents it as a common view probability map;

[0188] The block matching module with filtering constructs an overlap matrix through the common view probability map and fuses it with the feature similarity matrix to filter out false matches in non-overlapping areas and obtain a super point correspondence set;

[0189] The point matching module assigns the dense points and their dense features of the source point cloud and the target point cloud to the superpoint closest to them in the geometric space, forming a patch corresponding to each superpoint; then, the similarity matrix within each patch in the source point cloud and the target point cloud is calculated to obtain a local dense point correspondence set;

[0190] The pose screening and refinement module is used to obtain a series of local candidate rigid transformation matrices from the local dense point correspondence set using weighted SVD, calculate the inlier rate of each candidate transformation, and select the local candidate rigid transformation matrix with the largest inlier rate as the final rigid transformation matrix.

[0191] Example 3

[0192] This embodiment also provides a computer-readable storage medium storing executable instructions, which, when executed, enable the machine to perform the robust point cloud registration method based on common view query embedding as described above.

[0193] Specifically, a system or device equipped with a readable storage medium can be provided, on which software program codes that implement the functions of any of the above-mentioned embodiments are stored, and a computer or processor of the system or device can read and execute instructions stored in the readable storage medium.

[0194] In this case, the program code itself read from the computer-readable medium can realize the function of any one of the above embodiments, and thus the computer-readable code and the computer-readable storage medium storing the computer-readable code constitute part of this specification.

[0195] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (e.g., CD-ROMs, CD-Rs, CD-RWs, DVD-ROMs, DVD-RAMs, DVD-RWs, DVD-RWs), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, the program code may be downloaded from a server computer or a cloud via a communication network.

[0196] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0197] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0198] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0199] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0200] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation methods of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A robust point cloud registration method based on common view query embedding, characterized by: The method comprises: S1. Obtain a 3D point cloud pair, i.e., a source point cloud and a target point cloud, generated by scanning the same scene at different angles. The sampled points, i.e., super points and dense points, are obtained through encoder-point sampling. At the same time, point-by-point local features are extracted to obtain the sampled points and the corresponding local features, i.e., super point features and dense point features. S2, perform inter-block matching between the sampling points and the corresponding local features to obtain the super-point correspondence set of the source point cloud and the target point cloud; S3. Assign the dense points and dense features of the source point cloud and the target point cloud to the superpoint closest to them in the geometric space, forming a patch corresponding to each superpoint; then, calculate the similarity matrix within each patch in the source point cloud and the target point cloud to obtain the local dense point correspondence set: use the local to global registration method to filter the local dense point correspondence set to obtain the final rigid transformation matrix, so that the source point cloud and the target point cloud are successfully registered; The S2 specifically includes: S21. First stage: Perform geometric attention transformation on super-points and their corresponding super-point features, encoding the relative position information of each super-point in the point cloud into the super-point features of each point cloud; at the same time, encode the geometric and contextual information within and across point clouds, capture local and global feature relationships, and output super-point geometric features that are transformation invariant; S22, the second stage: performing common view feature transformation and common view point segmentation on the super-point geometric features, and outputting the common view features and common view probability maps of the source point cloud and the target point cloud; S23, the third stage: based on the obtained common view features, a secondary geometric attention transformation is performed again to achieve feature enhancement to obtain global geometric features; S24. Finally, the global geometric features and their common view probability maps are matched between blocks, and the super-point correspondence between the source point cloud and the target point cloud is finally output; The common view feature transformation and common viewpoint segmentation include two parts: common view attention transformation and common viewpoint segmentation: The common view attention transformation performs common view transformation based on the extracted super-point geometric features to achieve the interaction and alignment of dual point cloud features, and outputs the common view features of the source point cloud and the target point cloud; Common viewpoint segmentation further calculates the overlap fraction of the superpoints of the two point clouds based on the common view features, and extracts the common view area of ​​the two point cloud pairs. The overlap fraction effectively reflects the distribution of the common view area between the two point clouds, that is, a common view probability map with clear geometric meaning is then generated based on the transformed common view features.

2. A robust point cloud registration method based on common view query embedding according to claim 1, characterized in that: Said S1 specifically includes: S11: Use KPConv-FPN as the backbone network to perform downsampling and local feature extraction on the input source point cloud P and target point cloud Q at different scales. The output downsampled points are called super points, and { , }, the corresponding local feature, that is, the super-point feature is expressed as , where d represents the dimension of local features, Represents a feature set; The super-point features are calculated by aggregating the features of the neighboring points of each super-point and using a distance-weighted approach; S12: After the backbone network completes downsampling of the source point cloud and the target point cloud, it gradually upsamples the downsampled points. During the upsampling process, local point features are extracted for each upsampled point. After upsampling, the output upsampled points are dense points with the resolution of the original 3D point cloud. and 1 / 2 of the corresponding local point features, that is, dense point features, are recorded as and , d represents the dimension of the feature, Represents a feature set.

3. The robust point cloud registration method based on common view query embedding according to claim 2, characterized in that: The geometric attention transformation specifically includes: a geometric structure encoding layer, a self-attention layer, and a cross-attention layer; First, the superpoints of the source point cloud and the target point cloud and their corresponding superpoint features are taken as input, and the relative position information between the superpoints is processed by the geometric structure encoding layer. Then, the global superpoint features of each point cloud are encoded using the self-attention layer. Finally, these global features are input into the cross-attention layer to complete the feature interaction of the two point clouds, and finally the hybrid features with transformation invariance, namely the superpoint geometric features, are output, which are expressed as { },in, and represents the super points after downsampling of the source point cloud and the target point cloud, Indicates the first stage.

4. The robust point cloud registration method based on common view query embedding according to claim 3, characterized in that: The co-viewing attention transformation specifically includes: First, the common view information of the super-point geometric features is decoded to obtain the common view information of the two point clouds: by initializing the query vector Q and using the Transformer decoder to decode the common view information of the two point clouds , expressed as { }: （1）； In formula (1), the output { } represents the geometric features from the super point { The common view information decoded in} represents the overlapping information of two point cloud pairs, Q represents the initialized query vector, { } represents the output super-point geometric features, Indicates the first stage, and Represents the super points after downsampling of the source point cloud and the target point cloud, q, k, v represent The receiving variable of the input; Then, the decoded common view information { }As a priori information and the superpoint geometric features{ } is input into the cross attention layer to obtain the common view feature of the two point clouds represented as { },in Indicates the second stage; Finally, in order to make the common view feature { } is more discriminative, and the obtained common view features { } is used as input and a secondary geometric attention transformation is performed to achieve deep encoding and optimization of features, and the final output is the enhanced global geometric features and ,in Represents the global geometric features of the third stage after being enhanced by the second group of structural attention transformation modules; The common viewpoint segmentation is specifically as follows: First, the decoded common view information is transformed into Projection, constructing a weight map, which is used to enhance the common view feature and Next, a 1×1 convolution operation is applied to the obtained weight map and processed by the Sigmoid activation function to output the overlap score of the source point cloud and the target point cloud superpoints, as follows: （2）； （3）； In formulas (2)~(3), represents the superpoint after downsampling of the i-th source point cloud and target point cloud, and Represents the common view feature of two point clouds, { } represents the common view information of two point clouds, represents a weight map, which is used to represent the probability that each superpoint is located in the overlapping area, represents element-wise multiplication, represents the Sigmoid activation function, express and The common view probability map is specifically expressed as { , }, Represents a 1×1 convolution operation.

5. The robust point cloud registration method based on common view query embedding according to claim 4, characterized in that: The inter-block matching specifically includes: Patch matching with filtering using common view probability map { , Construct an overlap matrix and fuse it with the feature similarity matrix to filter out false matches in non-overlapping regions; Given global geometric features and , calculate their similarity matrix and the overlap matrix : （4）； In formula (4), and They are super points { , } midpoint index, and represents the inner product operation; Next, these two matrices are fused by element-wise multiplication to obtain the feature matching matrix : （5）； In formula (5), The operation is to convert the similarity matrix and overlap matrix Multiply the elements of the corresponding positions to filter out invalid matches outside the common view area. Finally, from the feature matching matrix Select the first k matching pairs as the super-point correspondence set: （6）； In formula (6), is the feature matching matrix between the source point cloud and the target point cloud, n and m represent the super point indexes of the two respectively, represents the superpoint matching set of the source point cloud filtered from the matching matrix, is a collection The elements in is the corresponding superpoint matching set in the target point cloud, is a collection Elements in Finally, these matching pairs will be point matched for further refined matching. Represents the set of superpoint correspondences.

6. The robust point cloud registration method based on common view query embedding according to claim 5, characterized in that: The S3 is specifically: S31. First, the dense points and dense features of the source point cloud and the target point cloud are assigned to the nearest superpoints in the geometric space to form a patch corresponding to each superpoint; Given a superpoint , the patch assigned to it is denoted as ,and The associated feature group is defined as ,in ; Similarly, calculate and express Patches in and feature groups , for each superpoint correspondence , first calculate the similarity matrix : （7）； In formula (7), d is the feature dimension, and Represents the source point cloud and target point cloud Corresponding superpoint index S32, calculate the similarity matrix in each patch of the source point cloud and the target point cloud, and obtain the local dense point correspondence set: use the Sinkhorn algorithm to calculate the similarity matrix Optimize and get the allocation matrix ; Then, from Extract the first k correspondences and construct a local dense point correspondence set S33, then aggregate the local dense point correspondences of all patches, all local dense point correspondences are set Merge into the final global dense point correspondence set ,in is the number of superpoint correspondences; S34. In order to avoid the slow convergence and instability of RANSAC iteration, a local to global registration scheme is used to estimate the transformation of the final output, using weighted SVD from the local dense point correspondence set. Compute candidate transformations Set, then calculate the number of inliers for each transformation candidate, and select the transformation with the largest number of inliers as the final rigid transformation matrix The calculation can be expressed as follows: （8）； In formula (8), τ represents the acceptance radius, and [⋅] is the Iverson bracket; The source point cloud and target point cloud are successfully registered through the rigid transformation matrix T.

7. A device for implementing a robust point cloud registration method based on common view query embedding, characterized in that: The device comprises: Data acquisition module, feature extraction module, inter-block matching module, point matching module and pose screening and refinement module; The data acquisition module is used to acquire three-dimensional point clouds generated by scanning the same scene at different angles; The feature extraction module is used to extract point-by-point local features of the three-dimensional point cloud; The inter-block matching module performs inter-block matching on the sampling points and the corresponding features to obtain corresponding point sets of the source point cloud and the target point cloud; The inter-block matching module includes a geometric attention transformation module, a common view attention transformation module, a common viewpoint segmentation module and a block matching module with filtering; The geometric attention transform is used to encode the geometric structure information of the point cloud and realize the interaction of geometric features within and across point clouds. It specifically includes: a geometric structure encoding layer, a self-attention layer, and a cross-attention layer; Geometric Attention Transformation Module: The geometric structure encoding layer is responsible for encoding the relative position information between superpoints in the point cloud. The self-attention layer is used to learn the global correlation of superpoints in the feature space and geometric space within a single point cloud. The cross-attention layer, as a typical module for point cloud registration tasks, focuses on feature exchange and geometric consistency encoding between two input point clouds. The common view attention transformation module performs common view transformation based on the extracted super point geometric features to achieve interaction and alignment of dual point cloud features, and outputs the common view features of the source point cloud and the target point cloud; The common viewpoint segmentation module further calculates the overlap score of the superpoints of the two point clouds based on the common view feature, and extracts the common view area of ​​the two point cloud pairs and represents it as a common view probability map; The block matching module with filtering constructs an overlap matrix through the common view probability map and fuses it with the feature similarity matrix to filter out false matches in non-overlapping areas and obtain a super point correspondence set; The point matching module assigns the dense points and their dense features of the source point cloud and the target point cloud to the superpoint closest to them in the geometric space, forming a patch corresponding to each superpoint; then, the similarity matrix within each patch in the source point cloud and the target point cloud is calculated to obtain the local dense point correspondence set: The pose screening and refinement module is used to obtain a series of local candidate rigid transformation matrices from the local dense point correspondence set using weighted SVD, calculate the inlier rate of each candidate transformation, and select the local candidate rigid transformation matrix with the largest inlier rate as the final rigid transformation matrix.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images