A method, system, computer-readable storage medium, and computer program product for point cloud registration based on point cloud segmentation

Abstract

The application relates to the technical field of three-dimensional vision, point cloud processing and space geometry registration, and discloses a point cloud registration method and system based on point cloud segmentation, a computer readable storage medium and a computer program product. The method comprises the following steps: performing multi-scale downsampling and feature extraction on two point clouds to be registered respectively; extracting global features of the most sparse layer point sets of the two point clouds and calculating conditional features of cross-cloud feature fusion; dividing the two point clouds into k sub-point clouds respectively and pairing the sub-point clouds; evaluating the coincidence degrees of the sub-point cloud pairs, and screening out a plurality of sub-point cloud pairs with the highest coincidence degrees as anchor point cloud pairs; merging the anchor point cloud pairs, and estimating a final matching transformation matrix. According to the application, the overall registration is converted into a plurality of high-overlap sub-tasks of sub-point cloud pairs, the sub-point cloud pairs with high overlap can be efficiently isolated under the condition of low overlap, coarse and fine two-stage matching and multi-scale feature consistency are combined, and the robustness and precision of point cloud registration are effectively improved.

Description

Technical Field

[0001] This invention relates to the fields of three-dimensional vision, point cloud processing, and spatial geometric registration technology, and particularly to a method, system, computer-readable storage medium, and computer program product for point cloud registration based on point cloud segmentation. Background Technology

[0002] With the increasing demand for applications such as large-scale 3D reconstruction, mobile robot navigation, and autonomous driving, point cloud registration has become a fundamental and critical issue.

[0003] Existing methods typically rely on iterative global ICP (Iterative Closest Point) or end-to-end learning to extract point matching. This approach works well when the overlap rate is high, but once the overlapping area is sparse, the number of incorrect correspondences caused by geometric / semantic ambiguity increases significantly, and the registration failure rate rises dramatically. Although some works have attempted to directly predict overlap weights or enhance global feature consistency, none of them can fundamentally solve the matching ambiguity problem under low overlap conditions.

[0004] To alleviate matching ambiguity caused by low overlap, early studies explicitly detected overlapping regions by predicting point-level overlap weights. However, this approach relied excessively on the consistency of low-level features, limiting its stability. Subsequent methods shifted towards implicitly enhancing global feature representations and utilizing attention mechanisms to learn transformation-invariant features. Drawing on this idea, SemReg (Semantics Constrained Point Cloud Registration) introduces general semantic-aware features; ColorPCR (Colored point cloud registration) captures cross-cloud associations through multi-stage geometry-color fusion. Although these strategies improve low-overlap registration performance to some extent, their effectiveness significantly degrades as the overlap rate decreases further.

[0005] In scenarios such as industrial vision, robot grasping, indoor and outdoor 3D reconstruction, motion measurement, and alignment of CAD models with measured point clouds, the point clouds to be registered often suffer from problems such as occlusion, viewpoint differences, scan gaps, uneven density, and local repetitive structures. In such scenarios, if a global correspondence is directly established for the entire point cloud, non-overlapping areas and pseudo-corresponding points can easily dominate the optimization process, leading to significant rotation or translational deviations in the registration results.

[0006] Therefore, for those skilled in the art, how to reduce the interference of non-overlapping regions on matching and effectively identify and isolate highly overlapping sub-point cloud pairs has become a technical problem that urgently needs to be solved. Furthermore, there is a need for a registration scheme that can first identify highly overlapping candidate regions at the local sub-point cloud level and then use reliable sub-point cloud pairs as anchor points for global transformation estimation, so as to maintain stable matching accuracy in low overlap, partial occlusion, cross-sensor and complex industrial scenarios. Summary of the Invention

[0007] The purpose of this invention is to provide a method, system, computer-readable storage medium, and computer program product for point cloud registration based on point cloud segmentation, so as to solve or at least partially solve the technical problems mentioned in the background art.

[0008] To achieve this objective, the present invention adopts the following technical solution: In a first aspect, the present invention provides a point cloud registration method based on point cloud segmentation, comprising: Multi-scale downsampling and feature extraction are performed on the two point clouds to be registered to obtain multi-layer point sets and their corresponding feature sets. Extract global features from the sparsest layer point set of two point clouds, and calculate conditional features for cross-cloud feature fusion based on the global features from the sparsest layer point set of two point clouds. The two point clouds are each divided into k sub-point clouds and paired into k×k sub-point cloud pairs; Evaluate the overlap of each sub-point cloud pair and select the sub-point cloud pairs with the highest overlap as anchor point cloud pairs. The anchor point cloud pairs are merged, and the final matching transformation matrix is ​​estimated.

[0009] Optionally, the step of performing multi-scale downsampling and feature extraction on the two point clouds to be registered to obtain multi-layer point sets and their corresponding feature sets for the two point clouds is as follows: A kernel-point convolutional network is used to process the reference point cloud X∈R. m×3 With source point cloud Y∈R n×3 Hierarchical downsampling is performed to obtain four layers of point sets and corresponding feature sets: , ; in, Let be the layer number of the point set, where =1 corresponds to the densest layer of points. =4 corresponds to the sparsest layer point set.

[0010] Optionally, the step of extracting the global features of the sparsest layer point set of the two point clouds, and calculating the conditional features for cross-cloud feature fusion based on the global features of the sparsest layer point set of the two point clouds, specifically includes: In the sparsest layer, first apply the self-attention layer Λ self Extracting global contextual features: , ; Subsequently, through the cross-attention layer Λ cross Achieve cross-cloud feature fusion to obtain conditional features: , .

[0011] Optionally, the step of dividing the two point clouds into k sub-point clouds and pairing them into k×k sub-point cloud pairs specifically includes: In the sparsest layer point set Above, k centers are selected using the farthest point sampling method: , ; in, For the farthest sampling operator, C X The center selected by the farthest point sampling method The set of C Y The center selected by the farthest point sampling method A set; For each center point, a radius threshold is applied using KNN. Constructing a sub-point cloud: , for Sub-dot clouds; , for Sub-dot clouds; one by one The sub-point clouds are respectively with Each sub-point cloud is paired one-to-one to obtain k×k sub-point cloud pairs.

[0012] Optionally, the evaluation of the overlap of each sub-point cloud pair and the selection of several sub-point cloud pairs with the highest overlap as anchor point cloud pairs specifically includes: For each sub-point cloud pair, calculate the Gaussian correlation matrix on the conditional feature: The Gaussian correlation matrix is ​​subjected to double soft normalization to obtain the matching confidence matrix; By weighting the matching set of corresponding point pairs in each sub-point cloud pair with the matching confidence matrix, the preliminary rigid transformation between the two sub-point clouds of each sub-point cloud pair is estimated, and the preliminary transformation set is obtained. First, the overlap prior weight of each sub-point cloud pair is calculated based on the matching confidence matrix, and several sub-point cloud pairs with the largest weights are selected as candidate sub-point cloud pairs; then, the feature consensus weight of each sub-point cloud pair is calculated based on the preliminary transformation set, and several candidate sub-point cloud pairs with weights greater than a predetermined threshold are selected as anchor cloud pairs.

[0013] Optionally, merging the anchor point cloud pairs and estimating the final matching transformation matrix specifically includes: For each anchor point cloud pair, extract the corresponding interior points in the sparsest layer: ; Among them, C t For the interior point correspondence of the anchor point cloud pair The set, For the initial transformation set elements, It is a set threshold. For the initial transformation set Element; Merge all C t The coarse matching set is obtained, and the coarse matching set is finally optimized by the RANSAC or LGR method to estimate the final matching transformation matrix.

[0014] Optionally, for each sub-point cloud pair, calculating the Gaussian correlation matrix on the conditional features specifically involves: Let the sub-point cloud be ,in, and They are respectively: , ; Pairs of Clouds The Gaussian correlation matrix B is calculated on the aforementioned conditional features using the following method: ; in, For matrix matrix elements, for The matrix elements, x u for The point in the middle, y v for The point in the middle; set up The Gaussian correlation matrix B is the matching confidence matrix after double soft normalization. The calculation method for the preliminary rigid transformation is as follows: Computing sub-point cloud pairs Two children dotted the clouds and The weighted centroid is obtained by the following method: , , , Where Z is the normalization factor; Paired Clouds and Decentralize the points in the data: ; Construct the weighted covariance matrix: ; Perform SVD decomposition on the weighted covariance matrix to calculate the rotation matrix R. ij Translation vector t ij ; The initial rigid transformation is then: ,in, .

[0015] Optionally, the method for calculating the overlapping prior weights is as follows: Assuming that overlapping sub-point cloud pairs have strong cross-correlation in features, then the prior mutual information set M is defined as follows: ,in and To match the elements of the confidence matrix, Indicates Top- operate; The formula for calculating the overlapping prior weights is: .

[0016] Optionally, the method for calculating the feature consensus weight is as follows: Using a multi-scale posterior weighted network, overlapping point pairs are retrieved at the mesoscale layer for candidate sub-point cloud pairs, resulting in a set of medium-granularity overlapping points. ; For the retrieved overlapping point pairs, For the initial rigid transformation of candidate sub-point cloud pairs, The threshold value for determining overlap is set. Based on the medium-granularity set of overlapping points, a fine-granularity set of overlapping points is constructed using a cross-scale nearest neighbor strategy: ; Based on the sets of medium-grained and fine-grained overlapping points, the multi-scale features of the two scales are concatenated and multiplied element-wise to generate affinity features: ; We obtain the following through lightweight attention and residual normalization: , ; Where LayerNorm represents layer normalization, max pool represents max pooling, and GELU is an activation function designed based on the Gaussian error function; Then, the feature consensus weights are output through a multilayer perceptron combined with a normalized exponential function: ; Where Softmax is the normalized exponential function and MLP represents a multilayer perceptron.

[0017] Optionally, the alignment-aware weighted loss of the multi-scale posterior weighted network is set as follows: ; in, For weighted alignment loss, For point matching loss, For overlapping sensing circle loss; Weighted alignment loss The calculation method is as follows: ; in, and The constructed positive and negative samples are as follows: , ; ; and These are the feature consensus weights corresponding to positive and negative samples, respectively.

[0018] Optionally, the two-point cloud is obtained through at least one of the following methods: Data is collected via a 3D sensor; Discretize the CAD model or 3D mesh; Through multi-view reconstruction or three-dimensional reconstruction; Before performing multi-scale downsampling and feature extraction on the two point clouds to be registered to obtain the multi-layer point set and its corresponding feature set of the two point clouds, the process also includes: Perform at least one preprocessing operation on the two point clouds; the preprocessing operation includes at least one of denoising processing, voxel downsampling processing, normal estimation processing, coordinate normalization processing, or intensity / color feature stitching processing.

[0019] Optionally, the multi-scale downsampling and feature extraction are implemented using any one or more combinations of kernel-point convolutional networks, point convolutional networks, graph convolutional networks, sparse convolutional networks, and Transformer coding networks.

[0020] Optionally, in the process of dividing the two point clouds into k sub-point clouds respectively, the value of k and / or the radius threshold are adaptively determined based on the point cloud density, scene scale, estimated overlap rate and computing resource constraints. Point cloud segmentation is achieved using at least one of the following methods: farthest point sampling, random sampling, voxel grid clustering, region growing clustering, superpoint segmentation, and semantic segmentation.

[0021] Optionally, the similarity measure of the Gaussian correlation matrix includes at least one of Gaussian kernel similarity, negative Euclidean distance, cosine similarity, dot product similarity, Mahalanobis distance similarity, and learned correlation score; The double soft normalization operation is implemented using at least one of row normalization, column normalization, Sinkhorn normalization, or Softmax normalization with temperature parameters.

[0022] Optionally, in the process of selecting candidate sub-point cloud pairs or anchor point cloud pairs, candidate sub-point cloud pairs with low confidence are also eliminated based on at least one of the following: peak matching confidence, bidirectional consistency, Top-k sparsity, overlap area estimation, inlier ratio, or transformation consistency.

[0023] Optionally, when estimating the final matching transformation matrix, the method further includes performing at least one optimization operation on the coarse matching set, namely, weighted singular value decomposition, iterative reweighted least squares, TEASER++ robust estimation, ICP, GICP, NDT, or local geometric consistency optimization, and using the overlapping prior weights and / or feature consensus weights of the anchor cloud pairs as the weights corresponding to the inliers.

[0024] Secondly, the present invention provides a point cloud registration system based on point cloud segmentation, comprising: The multi-scale sampling unit is used to perform multi-scale downsampling and feature extraction on the two point clouds to be registered, so as to obtain the multi-layer point set of the two point clouds and its corresponding feature set. The feature extraction unit, electrically connected to the multi-scale sampling unit, is used to extract the global features of the sparsest layer point set of the two point clouds, and to calculate the conditional features for cross-cloud feature fusion based on the global features of the sparsest layer point set of the two point clouds. The point cloud segmentation unit is electrically connected to the feature extraction unit and is used to segment two point clouds into k sub-point clouds respectively, and pair them into k×k sub-point cloud pairs; The overlap evaluation unit is electrically connected to the point cloud segmentation unit and is used to evaluate the overlap of each sub-point cloud pair and select the sub-point cloud pairs with the highest overlap as anchor point cloud pairs. The matching unit, electrically connected to the overlap evaluation unit, is used to merge the anchor point cloud pairs and estimate the final matching transformation matrix.

[0025] Optionally, the multi-scale sampling unit is specifically used for: A kernel-point convolutional network is used to process the reference point cloud X∈R. m×3 With source point cloud Y∈R n×3 Hierarchical downsampling is performed to obtain four layers of point sets and corresponding feature sets: , ; in, Let be the layer number of the point set, where =1 corresponds to the densest layer of points. =4 corresponds to the sparsest layer point set.

[0026] Optionally, the feature extraction unit is specifically used for: In the sparsest layer, first apply the self-attention layer Λ self Extracting global contextual features: , ; Subsequently, through the cross-attention layer Λ cross Achieve cross-cloud feature fusion to obtain conditional features: , .

[0027] Optionally, the point cloud segmentation unit is specifically used for: In the sparsest layer point set Above, k centers are selected using the farthest point sampling method: , ; in, For the farthest sampling operator, C X The center selected by the farthest point sampling method The set of C Y The center selected by the farthest point sampling method A set; For each center point, a radius threshold is applied using KNN. Constructing a sub-point cloud: , for Sub-dot clouds; , for Sub-dot clouds; one by one The sub-point clouds are respectively with Each sub-point cloud is paired one-to-one to obtain k×k sub-point cloud pairs.

[0028] Optionally, the overlap assessment unit is specifically used for: For each sub-point cloud pair, calculate the Gaussian correlation matrix on the conditional feature: The Gaussian correlation matrix is ​​subjected to double soft normalization to obtain the matching confidence matrix; By weighting the matching set of corresponding point pairs in each sub-point cloud pair with the matching confidence matrix, the preliminary rigid transformation between the two sub-point clouds of each sub-point cloud pair is estimated, and the preliminary transformation set is obtained. First, the overlap prior weight of each sub-point cloud pair is calculated based on the matching confidence matrix, and several sub-point cloud pairs with the largest weights are selected as candidate sub-point cloud pairs; then, the feature consensus weight of each sub-point cloud pair is calculated based on the preliminary transformation set, and several candidate sub-point cloud pairs with weights greater than a predetermined threshold are selected as anchor cloud pairs.

[0029] Optionally, the matching unit is specifically used for: For each anchor point cloud pair, extract the corresponding interior points in the sparsest layer: ; Among them, C t For the interior point correspondence of the anchor point cloud pair The set, For the initial transformation set elements, It is a set threshold. For the initial transformation set Element; Merge all C t The coarse matching set is obtained, and the coarse matching set is finally optimized by the RANSAC or LGR method to estimate the final matching transformation matrix.

[0030] Optionally, for each sub-point cloud pair, calculating the Gaussian correlation matrix on the conditional features specifically involves: Let the sub-point cloud be ,in, and They are respectively: , ; Pairs of Clouds The Gaussian correlation matrix B is calculated on the aforementioned conditional features using the following method: ; in, For matrix matrix elements, for The matrix elements, x u for The point in the middle, y v for The point in the middle; set up The Gaussian correlation matrix B is the matching confidence matrix after double soft normalization. The calculation method for the preliminary rigid transformation is as follows: Computing sub-point cloud pairs Two children dotted the clouds and The weighted centroid is obtained by the following method: , , , Where Z is the normalization factor; Paired Clouds and Decentralize the points in the data: ; Construct the weighted covariance matrix: ; Perform SVD decomposition on the weighted covariance matrix to calculate the rotation matrix R. ij Translation vector t ij ; The initial rigid transformation is then: ,in, .

[0031] Optionally, the method for calculating the overlapping prior weights is as follows: Assuming that overlapping sub-point cloud pairs have strong cross-correlation in features, then the prior mutual information set M is defined as follows: ,in and To match the elements of the confidence matrix, Indicates Top- operate; The formula for calculating the overlapping prior weights is: .

[0032] Optionally, the method for calculating the feature consensus weight is as follows: Using a multi-scale posterior weighted network, overlapping point pairs are retrieved at the mesoscale layer for candidate sub-point cloud pairs, resulting in a set of medium-granularity overlapping points. ; For the retrieved overlapping point pairs, For the initial rigid transformation of candidate sub-point cloud pairs, The threshold value for determining overlap is set. Based on the medium-granularity set of overlapping points, a fine-granularity set of overlapping points is constructed using a cross-scale nearest neighbor strategy: ; Based on the sets of medium-grained and fine-grained overlapping points, the multi-scale features of the two scales are concatenated and multiplied element-wise to generate affinity features: ; We obtain the following through lightweight attention and residual normalization: , ; Where LayerNorm represents layer normalization, max pool represents max pooling, and GELU is an activation function designed based on the Gaussian error function; Then, the feature consensus weights are output through a multilayer perceptron combined with a normalized exponential function: ; Where Softmax is the normalized exponential function and MLP represents a multilayer perceptron.

[0033] Optionally, the alignment-aware weighted loss of the multi-scale posterior weighted network is set as follows: ; in, For weighted alignment loss, For point matching loss, For overlapping sensing circle loss; Weighted alignment loss The calculation method is as follows: ; in, and The constructed positive and negative samples are as follows: , ; ; and These are the feature consensus weights corresponding to positive and negative samples, respectively.

[0034] Thirdly, the present invention also provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement a point cloud registration method based on point cloud segmentation as described above.

[0035] Fourthly, the present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the point cloud registration method based on point cloud segmentation described above.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows: The point cloud registration method based on point cloud segmentation provided in this application transforms the overall registration into a highly overlapping subtask of multiple sub-point cloud pairs, which can efficiently isolate overlapping sub-point cloud pairs under low overlap conditions. Combining coarse and fine two-stage matching and multi-scale feature consistency, it effectively improves the robustness and accuracy of point cloud registration. Furthermore, compared to schemes that rely solely on global features or point-level overlap prediction, this application transforms the low-overlap registration problem into several locally high-overlap sub-tasks through a process of "segmentation-pairing-overlap evaluation-anchor merging-robust optimization," thereby giving reliable local regions higher weight in the final pose estimation and improving stability under low-overlap, occlusion, and cross-sensor conditions. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart illustrating a point cloud registration method based on point cloud segmentation, provided as an embodiment of the present invention.

[0039] Figure 2-1 This is a schematic diagram of the second-layer point cloud obtained through multi-scale downsampling and feature extraction, provided in an embodiment of the present invention.

[0040] Figure 2-2 This is a schematic diagram of the third-layer point cloud obtained through multi-scale downsampling and feature extraction, provided in an embodiment of the present invention.

[0041] Figure 2-3 This is a schematic diagram of the fourth-layer point cloud obtained through multi-scale downsampling and feature extraction, provided in an embodiment of the present invention.

[0042] Figure 3 This is a schematic diagram of a sub-point cloud obtained by segmenting a point cloud, as provided in an embodiment of the present invention.

[0043] Figure 4 This is a schematic diagram of anchor point cloud pairs filtered through a two-stage weighted process, provided in an embodiment of the present invention.

[0044] Figure 5 This is a schematic diagram of the registration result after merging anchor point cloud pairs according to an embodiment of the present invention.

[0045] Figure 6This image shows a comparison of the registration results of a point cloud registration method based on point cloud segmentation provided in this embodiment of the invention with those of other methods.

[0046] Figure 7 This is a schematic diagram of the architecture of a point cloud registration system based on point cloud segmentation, provided for an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0048] Example 1: For details, please refer to Figure 1 , Figure 1 A flowchart of a point cloud registration method based on point cloud segmentation provided in this embodiment of the invention is shown. The method specifically includes: Step 100: Perform multi-scale downsampling and feature extraction on the two point clouds to be registered to obtain multi-layer point sets and their corresponding feature sets.

[0049] Please refer to Figure 2, where, Figure 2-1 This is a schematic diagram of the second-layer point cloud obtained through multi-scale downsampling and feature extraction, provided in an embodiment of the present invention. Figure 2-2 This is a schematic diagram of the third-layer point cloud obtained through multi-scale downsampling and feature extraction, provided in an embodiment of the present invention. Figure 2-3 This is a schematic diagram of the fourth-layer point cloud obtained through multi-scale downsampling and feature extraction, provided in an embodiment of the present invention.

[0050] In this embodiment, step 100 is implemented as follows: A kernel-point convolutional network is used to process the reference point cloud X∈R. m×3 With source point cloud Y∈R n×3 Hierarchical downsampling is performed to obtain four layers of point sets and corresponding feature sets: , ; in, Let be the layer number of the point set, where =1 corresponds to the densest layer of points. =4 corresponds to the sparsest layer point set.

[0051] It should be noted that since this embodiment does not use the densest layer (first layer) point set, a schematic diagram of the first layer point cloud is not provided here.

[0052] Step 120: Extract the global features of the sparsest layer point set of the two point clouds, and calculate the conditional features for cross-cloud feature fusion based on the global features of the sparsest layer point set of the two point clouds.

[0053] Specifically, step 120 includes: Step 121: In the sparsest layer, apply the self-attention layer Λ self Extracting global contextual features: , ; in, That is, the global context features of point cloud X extracted in the sparsest layer. That is, the global context features of point cloud Y extracted from the sparsest layer.

[0054] Step 122, through the cross-attention layer Λ cross Achieve cross-cloud feature fusion to obtain conditional features: , .

[0055] Step 130: Divide the two point clouds into k sub-point clouds respectively, and pair them into k×k sub-point cloud pairs.

[0056] Step 130 specifically includes: In the sparsest layer point set Above, k centers are selected using the farthest point sampling method: , ; in, For the farthest sampling operator, C X The center selected by the farthest point sampling method The set of C Y The center selected by the farthest point sampling method A set; For each center point, a radius threshold is applied using KNN (K-Nearest Neighbors). Constructing a sub-point cloud: , for Sub-dot clouds; , for Sub-dot clouds; one by one The sub-point clouds are respectively with Each sub-point cloud is paired one-to-one to obtain k×k sub-point cloud pairs.

[0057] Please refer to Figure 3 , Figure 3 This is a schematic diagram of a sub-point cloud obtained by segmenting a point cloud, as provided in an embodiment of the present invention.

[0058] Step 140: Evaluate the overlap of each sub-point cloud pair and select the sub-point cloud pairs with the highest overlap as anchor point cloud pairs.

[0059] Specifically, step 140 includes: Step 141: For each sub-point cloud pair, calculate the Gaussian correlation matrix on the conditional features.

[0060] Specifically, let the sub-point cloud pair be... ,in, and They are respectively: , ; Pairs of dots The Gaussian correlation matrix B is calculated based on the conditional features using the following method: ; in, For matrix matrix elements, for The matrix elements, x u for The point in the middle, y v for The point in the middle.

[0061] Step 142: Perform double soft normalization on the Gaussian correlation matrix to obtain the matching confidence matrix.

[0062] Dual-softmax is a method for calculating the matching probability between two feature sets. It obtains a matching matrix through bidirectional softmax operations (on rows and columns respectively). Specifically, it can be implemented using the bidirectional normalization matching probability calculation method. This embodiment focuses on using the normalization result for sub-point cloud pair overlap prior, candidate selection, and subsequent rigid transformation estimation.

[0063] Step 143: By weighting the matching set of corresponding point pairs in each sub-point cloud pair using the matching confidence matrix, estimate the preliminary rigid transformation between the two sub-point clouds of each sub-point cloud pair, and obtain the preliminary transformation set.

[0064] Specifically, let's set The Gaussian correlation matrix B is the matching confidence matrix after double soft normalization. The calculation method for the initial rigid transformation is as follows: Computing sub-point cloud pairs Two children dotted the clouds and The weighted centroid is obtained by the following method: , , , Where Z is the normalization factor; Paired Clouds and Decentralize the points in the data: ; Construct the weighted covariance matrix: ; Perform Singular Value Decomposition (SVD) on the weighted covariance matrix to calculate the rotation matrix R. ij Translation vector t ij ; The initial rigid transformation is then: ,in, .

[0065] Step 144: First, calculate the overlap prior weight of each sub-point cloud pair based on the matching confidence matrix, and select several sub-point cloud pairs with the largest weights as candidate sub-point cloud pairs; then, calculate the feature consensus weight of each sub-point cloud pair based on the preliminary transformation set, and select several candidate sub-point cloud pairs with weights greater than a predetermined threshold as anchor cloud pairs.

[0066] Specifically, the method for calculating overlapping prior weights is as follows: Assuming that overlapping sub-point cloud pairs have strong cross-correlation in features, then the prior mutual information set M is defined as follows: ,in and To match the elements of the confidence matrix, Indicates Top- operate; The formula for calculating the overlapping prior weights is: .

[0067] The method for calculating feature consensus weights is as follows: Using a multi-scale posterior weighted network, overlapping point pairs are retrieved at the mesoscale layer for candidate sub-point cloud pairs, resulting in a set of medium-granularity overlapping points. ; For the retrieved overlapping point pairs, For the initial rigid transformation of candidate sub-point cloud pairs, The threshold value for determining overlap is set. Based on the medium-granularity set of overlapping points, a fine-granularity set of overlapping points is constructed using a cross-scale nearest neighbor strategy: ; Based on the sets of medium-grained and fine-grained overlapping points, the multi-scale features of the two scales are concatenated and multiplied element-wise to generate affinity features: ; We obtain the following through lightweight attention and residual normalization: , ; Where LayerNorm represents layer normalization, max pool represents max pooling, and GELU is an activation function designed based on the Gaussian error function; Then, the feature consensus weights are output through a multilayer perceptron combined with a normalized exponential function: ; Where Softmax is the normalized exponential function and MLP represents a multilayer perceptron.

[0068] Please refer to Figure 4 , Figure 4 This is a schematic diagram of anchor point cloud pairs filtered through a two-stage weighted process, provided in an embodiment of the present invention.

[0069] Step 150: Merge the anchor point cloud pairs and estimate the final matching transformation matrix.

[0070] Step 150 specifically includes: Step 151: For each anchor point cloud pair, extract the corresponding interior points in the sparsest layer: ; Among them, C t For the interior point correspondence of the anchor point cloud pair The set, For the initial transformation set elements, It is a set threshold. For the initial transformation set Element.

[0071] Step 152: Merge all C t The coarse matching set is obtained, and the coarse matching set is finally optimized by the RANSAC or LGR method to estimate the final matching transformation matrix.

[0072] Among them, RANSAC (Random Sample Consensus) and LGR (Local Geometric Regression) are two methods commonly used in point cloud processing for model fitting and outlier removal, so their optimization process will not be described in detail in this embodiment.

[0073] Please refer to Figure 5 , Figure 5 This is a schematic diagram of the registration result after merging anchor point cloud pairs according to an embodiment of the present invention.

[0074] Furthermore, the alignment-aware weighted loss of the multi-scale posterior weighted network in step 140 is set as follows: ; a and β take values ​​from (0 to 1); in, For weighted alignment loss, For point matching loss, For overlapping sensing circle loss; Weighted alignment loss The calculation method is as follows: ; in, and The constructed positive and negative samples are as follows: , ; ; and These are the feature consensus weights corresponding to positive and negative samples, respectively. yes The corresponding error value, δ and δ are custom thresholds, both taking values ​​between 0 and 1.

[0075] Please continue to refer to this. Figure 6 , Figure 6 This image shows a comparison of the registration results of a point cloud registration method based on point cloud segmentation provided in this embodiment of the invention with those of other methods.

[0076] In summary, this embodiment provides a point cloud registration method based on point cloud segmentation, which transforms the overall registration into a highly overlapping sub-task of multiple sub-point cloud pairs, reducing the interference of non-overlapping regions on the matching; through a two-stage weighted strategy of overlapping prior and multi-scale posterior, it effectively identifies and isolates highly overlapping sub-point cloud pairs; after establishing a fine correspondence, it combines RANSAC or local geometric optimization to achieve high-precision, low-latency rigid transformation estimation; it supports point cloud data acquired by different resolutions and different sensors, and has good generalization and real-time performance.

[0077] Furthermore, in one implementation for low-overlap industrial point clouds, the reference point cloud X and the source point cloud Y can be derived from CAD model sampled point clouds, structured light scanned point clouds, binocular reconstructed point clouds, ToF point clouds, line laser profilometer stitched point clouds, or lidar point clouds, respectively. The system first performs outlier removal, voxel downsampling, normal estimation, and scale normalization on the two point clouds before entering a multi-scale feature extraction network. For point clouds containing color or reflection intensity, RGB, grayscale, intensity, curvature, or normal features can be combined with geometric coordinates as input features.

[0078] Furthermore, the value of k can be adaptively set based on the number of points and the estimated overlap rate. When the point cloud density is high or the scene scale is large, k is increased to obtain more local candidate regions; when the overlap rate is high or real-time requirements are strong, k is decreased to reduce the computational cost of enumerating k×k sub-point cloud pairs. The radius threshold τ can also be adaptively adjusted based on the average distance between points at each scale, voxel size, or local density statistics to avoid sub-point clouds being too sparse or too large.

[0079] Furthermore, in the candidate sub-point cloud pair screening stage, Top-k sparsification and bidirectional consistency checks can be performed on the matching confidence matrix first, retaining only corresponding points that are high-confidence candidates; then, candidate ranking is generated based on the proportion of interior points, consistency of local transformations, and overlapping prior weights. For scenarios with many repetitive structures, the feature consensus weight threshold can be increased, or the candidate sub-point cloud pairs can be required to satisfy consistency constraints at multiple scales.

[0080] Furthermore, in the final transformation estimation stage, in addition to RANSAC or LGR, weighted SVD, iterative reweighted least squares, TEASER++, ICP, GICP, or NDT can be used for refinement. The weights corresponding to interior points can be jointly determined by the overlapping prior weights of the anchor cloud pair, the feature consensus weights, and the point-level matching confidence, so as to reduce the impact of isolated mismatched points on the final rigid transformation matrix.

[0081] Through the above implementation methods, this application is not only applicable to standard point cloud registration datasets, but also to low-overlap or partially occluded scenarios such as industrial parts inspection, robot grasping and positioning, CAD to scanned point cloud alignment, indoor and outdoor 3D reconstruction, and mobile robot positioning and mapping.

[0082] Example 2: Please refer to Figure 7 , Figure 7 This is a schematic diagram illustrating the architecture of a point cloud registration system based on point cloud segmentation, provided as an embodiment of the present invention. The system specifically includes: The multi-scale sampling unit 10 is used to perform multi-scale downsampling and feature extraction on the two point clouds to be registered, so as to obtain the multi-layer point set of the two point clouds and its corresponding feature set. The feature extraction unit 20 and the electrically connected multi-scale sampling unit 10 are used to extract the global features of the sparsest layer point set of the two point clouds, and to calculate the conditional features for cross-cloud feature fusion based on the global features of the sparsest layer point set of the two point clouds. Point cloud segmentation unit 30 and electrical connection feature extraction unit 20 are used to segment two point clouds into k sub-point clouds respectively and pair them into k×k sub-point cloud pairs; The overlap evaluation unit 40 and the electrical connection point cloud segmentation unit 30 are used to evaluate the overlap of each sub-point cloud pair and select the sub-point cloud pairs with the highest overlap as anchor point cloud pairs. Matching unit 50 and electrical connection overlap evaluation unit 40 are used to merge anchor point cloud pairs and estimate the final matching transformation matrix.

[0083] The multi-scale sampling unit 10 is specifically used for: A kernel-point convolutional network is used to process the reference point cloud X∈R. m×3 With source point cloud Y∈R n×3 Hierarchical downsampling is performed to obtain four layers of point sets and corresponding feature sets: , ; in, Let be the layer number of the point set, where =1 corresponds to the densest layer of points. =4 corresponds to the sparsest layer point set.

[0084] Feature extraction unit 20 is specifically used for: In the sparsest layer, first apply the self-attention layer Λ self Extracting global contextual features: , ; Subsequently, through the cross-attention layer Λ cross Cross-cloud feature fusion is achieved to obtain conditional features: , .

[0085] The point cloud segmentation unit 30 is specifically used for: In the sparsest layer point set Above, k centers are selected using the farthest point sampling method: , ; in, For the farthest sampling operator, C X The center selected by the farthest point sampling method The set of C Y The center selected by the farthest point sampling method A set; For each center point, a radius threshold is applied using KNN. Constructing a sub-point cloud: , for Sub-dot clouds; , for Sub-dot clouds; one by one The sub-point clouds are respectively with Each sub-point cloud is paired one-to-one to obtain k×k sub-point cloud pairs.

[0086] Overlap assessment unit 40 is specifically used for: For each sub-point cloud pair, calculate the Gaussian correlation matrix on the conditional features: Double soft normalization is applied to the Gaussian correlation matrix to obtain the matching confidence matrix; By matching the confidence matrix and weighting the matching set of corresponding point pairs in each sub-point cloud pair, the preliminary rigid transformation between the two sub-point clouds of each sub-point cloud pair is estimated, and the preliminary transformation set is obtained. First, the overlap prior weight of each sub-point cloud pair is calculated based on the matching confidence matrix, and several sub-point cloud pairs with the largest weights are selected as candidate sub-point cloud pairs; then, the feature consensus weight of each sub-point cloud pair is calculated based on the preliminary transformation set, and several candidate sub-point cloud pairs with weights greater than a predetermined threshold are selected as anchor cloud pairs.

[0087] Matching unit 50 is specifically used for: For each anchor point cloud pair, extract the corresponding interior points in the sparsest layer: ; Among them, C t For the interior point correspondence of the anchor point cloud pair The set, For the initial transformation set elements, It is a set threshold. For the initial transformation set Element; Merge all Ct The coarse matching set is obtained, and the coarse matching set is finally optimized by the RANSAC or LGR method to estimate the final matching transformation matrix.

[0088] It is understood that the calculation methods of Gaussian correlation matrix, preliminary rigid transformation, overlapping prior weights, feature consensus weights, and alignment-aware weighted loss of multi-scale posterior weighted network have all been described in Example 1, so they will not be repeated in this example.

[0089] Compared with existing technologies, the point cloud registration system based on point cloud segmentation provided by the embodiments of the present invention has at least the following beneficial effects: (1) High robustness: The overall registration is decomposed into sub-tasks to effectively avoid interference in non-overlapping areas; (2) Improved accuracy: Dual-stage weighted summaries combined with multi-scale features enable fine matching between sub-cloud pairs; (3) Real-time performance: The coarse matching stage adopts efficient SVD, and the a posteriori network is lightweight, making it suitable for online deployment; (4) Good generalization: Supports point clouds with different sensors and resolutions, and parameters can be adaptively adjusted.

[0090] Furthermore, the system may also include a preprocessing unit, an adaptive parameter determination unit, a similarity measurement unit, and a robust optimization unit. The preprocessing unit performs denoising, voxel downsampling, normal estimation, coordinate normalization, or color / intensity feature concatenation on the input point cloud. The adaptive parameter determination unit determines the k-value and radius threshold based on point cloud density, scene scale, and computational resource constraints. The similarity measurement unit selects between Gaussian kernel similarity, cosine similarity, dot product similarity, Mahalanobis distance similarity, or learning relevance scoring. The robust optimization unit performs at least one of weighted SVD, RANSAC, LGR, TEASER++, ICP, GICP, or NDT.

[0091] Example 3: This embodiment also provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement a point cloud registration method based on point cloud segmentation as described in Embodiment 1.

[0092] Since the method of point cloud registration based on point cloud segmentation has been described in detail in Example 1, it will not be repeated in this example.

[0093] Example 4: The present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements a point cloud registration method based on point cloud segmentation as described in Embodiment 1.

[0094] Since the method of point cloud registration based on point cloud segmentation has been described in detail in Example 1, it will not be repeated in this example.

[0095] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware, or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0096] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A point cloud registration method based on point cloud segmentation, characterized in that, include: Multi-scale downsampling and feature extraction are performed on the two point clouds to be registered to obtain multi-layer point sets and their corresponding feature sets. Extract global features from the sparsest layer point set of two point clouds, and calculate conditional features for cross-cloud feature fusion based on the global features from the sparsest layer point set of two point clouds. The two point clouds are each divided into k sub-point clouds and paired into k×k sub-point cloud pairs; Evaluate the overlap of each sub-point cloud pair and select the sub-point cloud pairs with the highest overlap as anchor point cloud pairs. The anchor point cloud pairs are merged, and the final matching transformation matrix is ​​estimated.

2. The point cloud registration method based on point cloud segmentation according to claim 1, characterized in that, The process of performing multi-scale downsampling and feature extraction on the two point clouds to be registered, resulting in multi-layer point sets and their corresponding feature sets for the two point clouds, is as follows: A kernel-point convolutional network is used to process the reference point cloud X∈R. m×3 With source point cloud Y∈R n×3 Hierarchical downsampling is performed to obtain four layers of point sets and corresponding feature sets: ， ； in, Let be the layer number of the point set, where =1 corresponds to the densest layer of points. =4 corresponds to the sparsest layer point set.

3. The point cloud registration method based on point cloud segmentation according to claim 2, characterized in that, The extraction of global features from the sparsest layer point set of the two point clouds, and the calculation of conditional features for cross-cloud feature fusion based on the global features from the sparsest layer point set of the two point clouds, specifically include: In the sparsest layer, first apply the self-attention layer Λ self Extracting global contextual features: ， ； Subsequently, through the cross-attention layer Λ cross Cross-cloud feature fusion is achieved to obtain conditional features: ， 。 4. The point cloud registration method based on point cloud segmentation according to claim 3, characterized in that, The step of dividing the two point clouds into k sub-point clouds and pairing them into k×k sub-point cloud pairs specifically includes: In the sparsest layer point set Above, k centers are selected using the farthest point sampling method: ， ； in, For the farthest sampling operator, C X The center selected by the farthest point sampling method The set of C Y The center selected by the farthest point sampling method A set; For each center point, a radius threshold is applied using KNN. Constructing a sub-point cloud: , for Sub-dot clouds; , for Sub-dot clouds; one by one The sub-point clouds are respectively with Each sub-point cloud is paired one-to-one to obtain k×k sub-point cloud pairs.

5. The point cloud registration method based on point cloud segmentation according to claim 4, characterized in that, The evaluation of the overlap of each sub-point cloud pair, and the selection of the sub-point cloud pairs with the highest overlap as anchor point cloud pairs, specifically includes: For each sub-point cloud pair, calculate the Gaussian correlation matrix on the conditional feature: The Gaussian correlation matrix is ​​subjected to double soft normalization to obtain the matching confidence matrix; By weighting the matching set of corresponding point pairs in each sub-point cloud pair with the matching confidence matrix, the preliminary rigid transformation between the two sub-point clouds of each sub-point cloud pair is estimated, and the preliminary transformation set is obtained. First, the overlap prior weight of each sub-point cloud pair is calculated based on the matching confidence matrix, and several sub-point cloud pairs with the largest weights are selected as candidate sub-point cloud pairs; then, the feature consensus weight of each sub-point cloud pair is calculated based on the preliminary transformation set, and several candidate sub-point cloud pairs with weights greater than a predetermined threshold are selected as anchor cloud pairs.

6. The point cloud registration method based on point cloud segmentation according to claim 5, characterized in that, The process of merging the anchor point cloud pairs and estimating the final matching transformation matrix specifically includes: For each anchor point cloud pair, extract the corresponding interior points in the sparsest layer: ； Among them, C t For the interior point correspondence of the anchor point cloud pair The set, For the initial transformation set elements, It is a set threshold. For the initial transformation set Element; Merge all C t The coarse matching set is obtained, and the coarse matching set is finally optimized by the RANSAC or LGR method to estimate the final matching transformation matrix.

7. The point cloud registration method based on point cloud segmentation according to claim 5, characterized in that, For each sub-point cloud pair, the Gaussian correlation matrix is ​​calculated on the conditional features, specifically as follows: Let the sub-point cloud be ,in, and They are respectively: ， ； Pairs of dots The Gaussian correlation matrix B is calculated on the aforementioned conditional features using the following method: ； in, For matrix matrix elements, for The matrix elements, x u for The point in the middle, y v for The point in the middle; set up The Gaussian correlation matrix B is the matching confidence matrix after double soft normalization. The calculation method for the preliminary rigid transformation is as follows: Computing sub-point cloud pairs Two children dotted the clouds and The weighted centroid is obtained by the following method: ， ， ， Where Z is the normalization factor; Paired Clouds and Decentralize the points in the data: ； Construct the weighted covariance matrix: ； Perform SVD decomposition on the weighted covariance matrix to calculate the rotation matrix R. ij Translation vector t ij ; The initial rigid transformation is then: ,in, .

8. The point cloud registration method based on point cloud segmentation according to claim 7, characterized in that, The method for calculating the overlapping prior weights is as follows: Assuming that overlapping sub-point cloud pairs have strong cross-correlation in features, then the prior mutual information set M is defined as follows: ,in and To match the elements of the confidence matrix, Indicates Top- operate; The formula for calculating the overlapping prior weights is: 。 9. The point cloud registration method based on point cloud segmentation according to claim 8, characterized in that, The method for calculating the feature consensus weight is as follows: Using a multi-scale posterior weighted network, overlapping point pairs are retrieved at the mesoscale layer for candidate sub-point cloud pairs, resulting in a set of medium-granularity overlapping points. ; For the retrieved overlapping point pairs, For the initial rigid transformation of candidate sub-point cloud pairs, The threshold value for determining overlap is set. Based on the medium-granularity set of overlapping points, a fine-granularity set of overlapping points is constructed using a cross-scale nearest neighbor strategy: ； Based on the sets of medium-grained and fine-grained overlapping points, the multi-scale features of the two scales are concatenated and multiplied element-wise to generate affinity features: ； We obtain the following through lightweight attention and residual normalization: ， ； Where LayerNorm represents layer normalization, max pool represents max pooling, and GELU is an activation function designed based on the Gaussian error function; Then, the feature consensus weights are output through a multilayer perceptron combined with a normalized exponential function: ； Where Softmax is the normalized exponential function and MLP represents a multilayer perceptron.

10. A point cloud registration method based on point cloud segmentation according to claim 9, characterized in that, The alignment-aware weighted loss of the multi-scale posterior weighted network is set as follows: ； in, For weighted alignment loss, For point matching loss, For overlapping sensing circle loss; Weighted alignment loss The calculation method is as follows: ； in, and The constructed positive and negative samples are as follows: ， ； ； and These are the feature consensus weights corresponding to positive and negative samples, respectively.

11. The point cloud registration method based on point cloud segmentation and anchor point cloud pair filtering according to claim 1, characterized in that, The two-point cloud is obtained through at least one of the following methods: Data is collected via a 3D sensor; Discrete sampling of CAD models or 3D meshes; Through multi-view reconstruction or three-dimensional reconstruction; Before performing multi-scale downsampling and feature extraction on the two point clouds to be registered to obtain the multi-layer point set and its corresponding feature set of the two point clouds, the process also includes: Perform at least one preprocessing operation on the two point clouds; the preprocessing operation includes at least one of denoising processing, voxel downsampling processing, normal estimation processing, coordinate normalization processing, or intensity / color feature stitching processing.

12. The point cloud registration method based on point cloud segmentation and anchor point cloud pair filtering according to claim 1, characterized in that, The multi-scale downsampling and feature extraction are implemented using any one or more combinations of kernel-point convolutional networks, point convolutional networks, graph convolutional networks, sparse convolutional networks, and Transformer coding networks.

13. The point cloud registration method based on point cloud segmentation and anchor point cloud pair filtering according to claim 1, characterized in that, In the process of dividing two point clouds into k sub-point clouds, the value of k and / or the radius threshold are adaptively determined based on the point cloud density, scene scale, estimated overlap rate and computing resource constraints. Point cloud segmentation is achieved using at least one of the following methods: farthest point sampling, random sampling, voxel grid clustering, region growing clustering, superpoint segmentation, and semantic segmentation.

14. The point cloud registration method based on point cloud segmentation and anchor point cloud pair filtering according to claim 5, characterized in that, The similarity measure of the Gaussian correlation matrix includes at least one of Gaussian kernel similarity, negative Euclidean distance, cosine similarity, dot product similarity, Mahalanobis distance similarity, and the learned correlation score; The double soft normalization operation is implemented using at least one of row normalization, column normalization, Sinkhorn normalization, or Softmax normalization with temperature parameters.

15. The point cloud registration method based on point cloud segmentation and anchor point cloud pair filtering according to claim 5, characterized in that, In the process of selecting candidate sub-point cloud pairs or anchor point cloud pairs, candidate sub-point cloud pairs with low confidence are also eliminated based on at least one of the following: peak matching confidence, bidirectional consistency, Top-k sparsity, overlapping area estimation, inlier ratio, or transformation consistency.

16. The point cloud registration method based on point cloud segmentation and anchor point cloud pair filtering according to claim 6, characterized in that, When estimating the final matching transformation matrix, the method further includes performing at least one optimization operation on the coarse matching set, such as weighted singular value decomposition, iterative reweighted least squares, TEASER++ robust estimation, ICP, GICP, NDT, or local geometric consistency optimization, and using the overlapping prior weights and / or feature consensus weights of the anchor cloud pairs as the weights corresponding to the inliers.

17. A point cloud registration system based on point cloud segmentation, characterized in that, include: The multi-scale sampling unit is used to perform multi-scale downsampling and feature extraction on the two point clouds to be registered, so as to obtain the multi-layer point set of the two point clouds and its corresponding feature set. The feature extraction unit, electrically connected to the multi-scale sampling unit, is used to extract the global features of the sparsest layer point set of the two point clouds, and to calculate the conditional features for cross-cloud feature fusion based on the global features of the sparsest layer point set of the two point clouds. The point cloud segmentation unit is electrically connected to the feature extraction unit and is used to segment two point clouds into k sub-point clouds respectively, and pair them into k×k sub-point cloud pairs; The overlap evaluation unit is electrically connected to the point cloud segmentation unit and is used to evaluate the overlap of each sub-point cloud pair and select the sub-point cloud pairs with the highest overlap as anchor point cloud pairs. The matching unit, electrically connected to the overlap evaluation unit, is used to merge the anchor point cloud pairs and estimate the final matching transformation matrix.

18. A point cloud registration system based on point cloud segmentation according to claim 17, characterized in that, The multi-scale sampling unit is specifically used for: A kernel-point convolutional network is used to process the reference point cloud X∈R. m×3 With source point cloud Y∈R n×3 Hierarchical downsampling is performed to obtain four layers of point sets and corresponding feature sets: ， ； in, Let be the layer number of the point set, where =1 corresponds to the densest layer of points. =4 corresponds to the sparsest layer point set.

19. A point cloud registration system based on point cloud segmentation according to claim 18, characterized in that, The feature extraction unit is specifically used for: In the sparsest layer, first apply the self-attention layer Λ self Extracting global contextual features: ， ； Subsequently, through the cross-attention layer Λ cross Cross-cloud feature fusion is achieved to obtain conditional features: ， 。 20. A point cloud registration system based on point cloud segmentation according to claim 19, characterized in that, The point cloud segmentation unit is specifically used for: In the sparsest layer point set Above, k centers are selected using the farthest point sampling method: ， ； in, For the farthest sampling operator, C X The center selected by the farthest point sampling method The set of C Y The center selected by the farthest point sampling method A set; For each center point, a radius threshold is applied using KNN. Constructing a sub-point cloud: , for Sub-dot clouds; , for Sub-dot clouds; one by one The sub-point clouds are respectively with Each sub-point cloud is paired one-to-one to obtain k×k sub-point cloud pairs.

21. A point cloud registration system based on point cloud segmentation according to claim 20, characterized in that, The overlap assessment unit is specifically used for: For each sub-point cloud pair, calculate the Gaussian correlation matrix on the conditional feature: The Gaussian correlation matrix is ​​subjected to double soft normalization to obtain the matching confidence matrix; By weighting the matching set of corresponding point pairs in each sub-point cloud pair with the matching confidence matrix, the preliminary rigid transformation between the two sub-point clouds of each sub-point cloud pair is estimated, and the preliminary transformation set is obtained. First, the overlap prior weight of each sub-point cloud pair is calculated based on the matching confidence matrix, and several sub-point cloud pairs with the largest weights are selected as candidate sub-point cloud pairs; then, the feature consensus weight of each sub-point cloud pair is calculated based on the preliminary transformation set, and several candidate sub-point cloud pairs with weights greater than a predetermined threshold are selected as anchor cloud pairs.

22. The point cloud registration system based on point cloud segmentation according to claim 21, characterized in that, The matching unit is specifically used for: For each anchor point cloud pair, extract the corresponding interior points in the sparsest layer: ； Among them, C t For the interior point correspondence of the anchor point cloud pair The set, For the initial transformation set elements, It is a set threshold. For the initial transformation set Element; Merge all C t The coarse matching set is obtained, and the coarse matching set is finally optimized by the RANSAC or LGR method to estimate the final matching transformation matrix.

23. The point cloud registration system based on point cloud segmentation according to claim 21, characterized in that, For each sub-point cloud pair, the Gaussian correlation matrix is ​​calculated on the conditional features, specifically as follows: Let the sub-point cloud be ,in, and They are respectively: ， ； Pairs of Clouds The Gaussian correlation matrix B is calculated on the aforementioned conditional features using the following method: ； in, For matrix matrix elements, for The matrix elements, x u for The point in the middle, y v for The point in the middle; set up The Gaussian correlation matrix B is the matching confidence matrix after double soft normalization. The calculation method for the preliminary rigid transformation is as follows: Computing sub-point cloud pairs Two children dotted the clouds and The weighted centroid is obtained by the following method: ， ， ， Where Z is the normalization factor; Paired Clouds and Decentralize the points in the data: ； Construct the weighted covariance matrix: ； Perform SVD decomposition on the weighted covariance matrix to calculate the rotation matrix R. ij Translation vector t ij ; The initial rigid transformation is then: ,in, .

24. A point cloud registration system based on point cloud segmentation according to claim 23, characterized in that, The method for calculating the overlapping prior weights is as follows: Assuming that overlapping sub-point cloud pairs have strong cross-correlation in features, then the prior mutual information set M is defined as follows: ,in and To match the elements of the confidence matrix, Indicates Top- operate; The formula for calculating the overlapping prior weights is: 。 25. A point cloud registration system based on point cloud segmentation according to claim 24, characterized in that, The method for calculating the feature consensus weight is as follows: Using a multi-scale posterior weighted network, overlapping point pairs are retrieved at the mesoscale layer for candidate sub-point cloud pairs, resulting in a set of medium-granularity overlapping points. ; For the retrieved overlapping point pairs, For the initial rigid transformation of candidate sub-point cloud pairs, The threshold value for determining overlap is set. Based on the medium-granularity set of overlapping points, a fine-granularity set of overlapping points is constructed using a cross-scale nearest neighbor strategy: ； Based on the sets of medium-grained and fine-grained overlapping points, the multi-scale features of the two scales are concatenated and multiplied element-wise to generate affinity features: ； We obtain the following through lightweight attention and residual normalization: ， ； Where LayerNorm represents layer normalization, max pool represents max pooling, and GELU is an activation function designed based on the Gaussian error function; Then, the feature consensus weights are output through a multilayer perceptron combined with a normalized exponential function: ； Where Softmax is the normalized exponential function and MLP represents a multilayer perceptron.

26. A point cloud registration system based on point cloud segmentation according to claim 25, characterized in that, The alignment-aware weighted loss of the multi-scale posterior weighted network is set as follows: ； in, For weighted alignment loss, For point matching loss, For overlapping sensing circle loss; Weighted alignment loss The calculation method is as follows: ； in, and The constructed positive and negative samples are as follows: ， ； ； and These are the feature consensus weights corresponding to positive and negative samples, respectively.

27. A computer-readable storage medium storing at least one instruction, characterized in that, The instructions are loaded and executed by the processor to implement a point cloud registration method based on point cloud segmentation as described in any one of claims 1-16.

28. A computer program product comprising a computer program / instructions, characterized in that, When a computer program / instruction is executed by a processor, it implements a point cloud registration method based on point cloud segmentation as described in any one of claims 1-16.

Images