Point cloud registration method and system based on local geometric features and clustering strategy

Abstract

The invention relates to a point cloud registration method and system based on local geometric features and a clustering strategy, and the method comprises the steps: obtaining to-be-registered three-dimensional point cloud information which comprises a source point cloud image and a target point cloud image; performing feature extraction on the to-be-registered three-dimensional point cloud information to obtain high-dimensional point cloud feature information; according to the high-dimensional point cloud feature information, feature filtering is carried out on the three-dimensional point cloud information to be registered, a corresponding point pair is obtained, and the corresponding point pair is a nearest corresponding point in the source point cloud and the target point cloud; the corresponding point pairs are subjected to two-space clustering for screening, high-confidence seeds are obtained, and the high-confidence seeds are point pairs obtained after the corresponding point pairs are filtered; and screening the high-confidence seeds, obtaining a consensus set, decomposing the consensus set, and completing point cloud registration, the consensus set being a final corresponding point pair set. According to the method, the spatial geometric features can be better utilized, the high-confidence seed is obtained, and the point cloud registration precision and efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud registration, and in particular to a point cloud registration method and system based on local geometric features and clustering strategy. Background Art

[0002] Point cloud registration is the process of aligning multiple point cloud datasets into the same coordinate system, which has important applications in 3D reconstruction, robot navigation, virtual reality, and medical imaging. The main goal of point cloud registration is to determine the optimal transformation (including translation and rotation) between point clouds so that point clouds from different sources overlap as much as possible in space, thereby forming a complete 3D scene. Although traditional registration algorithms have high registration accuracy in a small range, they still face challenges such as noise, occlusion, and point cloud sparsity when dealing with large-scale or complex scenes. In recent years, point cloud registration methods based on deep learning have gradually emerged, but achieving both high efficiency and high accuracy in point cloud registration remains a challenging task.

[0003] Mainstream deep learning-based methods often use deep learning strategies to obtain better spatial information, but this impairs the network's ability to capture geometric feature details. The loss of detail information can lead to incorrect matching of small objects, thin structures, and edge regions, resulting in a significant reduction in accuracy. Summary of the invention

[0004] The purpose of the present invention is to provide a point cloud registration method and system based on local geometric features and clustering strategy, which can better utilize spatial geometric features, obtain high-confidence seeds, and improve the accuracy and efficiency of point cloud registration.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] Point cloud registration methods based on local geometric features and clustering strategies include:

[0007] Acquire three-dimensional point cloud information to be registered, wherein the three-dimensional point cloud information to be registered includes a source point cloud image and a target point cloud image;

[0008] Extracting features of the three-dimensional point cloud information to be registered to obtain high-dimensional point cloud feature information;

[0009] Performing feature filtering on the three-dimensional point cloud information to be registered according to the high-dimensional point cloud feature information to obtain corresponding point pairs, wherein the corresponding point pairs are the corresponding points with the closest distance between the source point cloud and the target point cloud;

[0010] The corresponding point pairs are clustered in two spaces to obtain a consensus set, and the consensus set is decomposed to complete the point cloud registration, wherein the consensus set is the final corresponding point pair set.

[0011] Optionally, obtaining high-dimensional point cloud feature information includes: using a feature extractor to extract local features of the three-dimensional point cloud information to be registered to obtain the high-dimensional point cloud feature information, and the feature extractor is constructed based on a U-Net network.

[0012] Optionally, feature filtering is performed on the three-dimensional point cloud information to be registered to obtain corresponding point pairs, including: geometric feature aggregation is performed on the high-dimensional point cloud feature information of the source point cloud and the target point cloud according to the attention mechanism to obtain a feature matrix, Euclidean distance calculation is performed on the feature matrix to obtain a similarity matrix, and the topk corresponding point pairs in each row of the similarity matrix are selected to obtain the corresponding point pairs.

[0013] Optionally, performing two-space clustering on the corresponding point pairs to obtain a consensus set includes:

[0014] According to the high-dimensional point cloud feature information and the source point cloud in the corresponding point pairs, a spatial consistency matrix and an angular compatibility matrix are obtained, and according to the product of the spatial consistency matrix and the angular compatibility matrix, a local geometric feature spatial compatibility matrix is ​​obtained;

[0015] Screening the corresponding point pairs according to the local geometric feature space compatibility matrix to obtain initial seeds;

[0016] The initial seeds are screened according to the feature similarity threshold and the distance threshold to obtain a corresponding point pair set in the feature space and a corresponding point pair set in the geometric space;

[0017] The corresponding point pairs common to the corresponding point pair set in the feature space and the corresponding point pair set in the geometric space are selected to obtain the consensus set.

[0018] Optionally, obtaining the spatial consistency matrix of the corresponding point pairs includes:

[0019] Using the KNN nearest neighbor method to process the high-dimensional point cloud feature information corresponding to the corresponding point pair, to obtain a first local block composed of a plurality of adjacent points;

[0020] Distance differences are calculated for the corresponding point pairs in the first local block to obtain the spatial consistency matrix.

[0021] Optionally, obtaining the angle compatibility matrix includes:

[0022] According to the source point cloud corresponding to the corresponding point pair, a KNN nearest neighbor method is used to perform a k-nearest neighbor search in the original point cloud space to obtain a second local block;

[0023] Calculating a feature vector in the second local block according to the high-dimensional point cloud feature information to obtain the feature vector;

[0024] The angular difference of the eigenvector is calculated to obtain the angular compatibility matrix.

[0025] Optionally, screening the corresponding point pairs for two-space clustering according to the local geometric feature space compatibility matrix to obtain the initial seeds includes:

[0026] The topk corresponding point pairs in each row of the local geometric feature space compatibility matrix are selected to obtain the initial seeds.

[0027] Optionally, the consensus set is:

[0028] S final =S feat ∩S geo

[0029] S feat ={(x i ,y j )|Sim feat (x i ,y j )≥θ feat}

[0030] S geo ={(x i ,y j )|dist feat (x i ,y j )≤θ geo}

[0031] Among them, S final is the final corresponding point pair set, S feat is the set of corresponding point pairs in the feature space, S geo is the set of corresponding point pairs in geometric space, x i ,y j are the corresponding point pairs in the source point cloud and the target point cloud, Sim feat To calculate the feature similarity of corresponding points in the feature space, θ feat is the feature similarity threshold, dist feat To calculate the distance between corresponding points in geometric space, θ geo is the distance threshold.

[0032] Optionally, decomposing the consensus set to complete the point cloud registration includes: performing singular value decomposition on the consensus set, obtaining a transformation matrix, and completing the point cloud registration.

[0033] The present invention also provides a point cloud registration system based on local geometric features and clustering strategy, including: a collection module, an extraction module, a screening module and a registration module;

[0034] The acquisition module is used to obtain the three-dimensional point cloud information to be registered, wherein the three-dimensional point cloud information to be registered includes a source point cloud image and a target point cloud image;

[0035] The extraction module is used to extract features from the three-dimensional point cloud information to be registered to obtain high-dimensional point cloud feature information;

[0036] The screening module is used to perform feature filtering on the three-dimensional point cloud information to be registered according to the feature information of the high-dimensional point cloud, and obtain corresponding point pairs between the three-dimensional point cloud information to be registered, wherein the corresponding point pairs are the corresponding points with the closest distance between the source point cloud and the target point cloud;

[0037] The registration module is used to perform two-space clustering on the corresponding point pairs, obtain a consensus set, decompose the consensus set, and complete point cloud registration, wherein the consensus set is a final corresponding point pair set.

[0038] The beneficial effects of the present invention are as follows: (1) The present invention further deepens the feature information by constructing a local geometric feature spatial compatibility matrix through spatial feature aggregation guided by geometric features, so as to obtain initial seeds. Seed screening is performed using the local geometric feature spatial compatibility matrix and the two-space clustering strategy. This makes it possible to fully extract and utilize geometric features and spatial features, making the seeds more reliable and more informative. By introducing the local geometric feature spatial compatibility matrix and the two-space clustering strategy, spatial geometric features can be better utilized to obtain high-confidence seeds, thereby improving the accuracy and efficiency of point cloud registration.

[0039] (2) The present invention proposes a two-space clustering strategy that can effectively capture the feature dependencies between corresponding point pairs, greatly reducing mismatches and computing requirements, ensuring operational efficiency while also improving registration accuracy. This innovative mechanism not only has significant advantages in network design, but also performs well in practical applications, providing a more efficient and reliable solution for solving complex problems. By introducing the spatial clustering mechanism, the model can better understand and utilize the correlation between geometric features and spatial features, thereby achieving more accurate registration and reasoning, bringing new ideas and possibilities to research and application in related fields.

[0040] (3) The present invention fully extracts and utilizes the spatial information and local geometric details in the three-dimensional point cloud to achieve highly accurate alignment while maintaining excellent efficiency, and exhibits competitive performance on multiple datasets, especially in challenging areas, including areas with low overlap rates. It is able to achieve accurate point cloud alignment while maintaining high efficiency, thereby further demonstrating its robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0042] Figure 1 A flow chart of a point cloud registration method based on local geometric features and clustering strategy according to an embodiment of the present invention;

[0043] Figure 2 It is a schematic diagram of the main framework of a point cloud registration method based on local geometric features and clustering strategy according to an embodiment of the present invention;

[0044] Figure 3 Schematic diagram of the structure of a feature extractor according to an embodiment of the present invention;

[0045] Figure 4 Schematic diagram of a local geometric feature space compatibility matrix according to an embodiment of the present invention, wherein (a) is an example of a three-dimensional scene point cloud, (b) is a spatial consistency matrix, (c) is a local angular compatibility matrix, and (d) is a local geometric feature space compatibility matrix;

[0046] Figure 5 Schematic diagram of the two-space clustering strategy of an embodiment of the present invention, wherein (a) is a feature space point pair compatibility diagram; (b) is a feature space compatibility matrix derived from (a); (c) is a geometric space point pair compatibility diagram; (d) is a geometric space compatibility matrix derived from (c); (e) is a two-space point pair compatibility diagram; and (f) is a two-space compatibility matrix derived from (e). DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] Embodiment 1:

[0050] like Figure 2 As shown, this embodiment provides a point cloud registration method based on local geometric features and clustering strategy, including:

[0051] Acquire three-dimensional point cloud information to be registered, wherein the three-dimensional point cloud information to be registered includes a source point cloud image and a target point cloud image;

[0052] Perform feature extraction on the three-dimensional point cloud information to be registered to obtain high-dimensional point cloud feature information;

[0053] Perform feature filtering on the three-dimensional point cloud information to be registered according to the feature information of the high-dimensional point cloud to obtain corresponding point pairs, where the corresponding point pairs are the corresponding points with the closest distance between the source point cloud and the target point cloud;

[0054] The corresponding point pairs are clustered in two spaces to obtain a consensus set, which is then decomposed to complete the point cloud registration. The consensus set is the final set of corresponding point pairs.

[0055] Specifically, this embodiment further deepens the feature information by guiding the local geometric feature spatial compatibility matrix of the aggregated spatial features with geometric features, so as to obtain the corresponding seeds. The local geometric feature spatial compatibility matrix and the two-space clustering strategy are used to screen high-confidence seeds. This makes it possible to fully extract and utilize geometric features and spatial features, making the seeds more reliable and more informative. By introducing the local geometric feature spatial compatibility matrix and the two-space clustering strategy, spatial geometric features can be better utilized to obtain high-confidence seeds and improve the accuracy and efficiency of point cloud registration.

[0056] Furthermore, obtaining high-dimensional point cloud feature information includes: using a feature extractor to perform feature aggregation on local features of the three-dimensional point cloud information to be registered, and extracting high-dimensional point cloud feature information, and the feature extractor is constructed based on a U-Net network.

[0057] Specifically, in this embodiment, a feature extractor is constructed based on a U-Net structure of a convolutional neural network, and a residual connection of the U-Net structure and aggregation of high-dimensional information are used to perform delicate feature extraction to obtain high-dimensional feature information.

[0058] A three-dimensional convolution layer with a convolution kernel size of 7×7×7 and a stride of 1 is used to obtain the initial feature information of the point cloud image. The three-layer residual block is used to downsample the initial feature information of the point cloud image to obtain high-dimensional feature information. The residual block consists of two three-dimensional convolution layers with a convolution kernel size of 3×3×3 and a stride of 1, and one three-dimensional convolution layer with a convolution kernel size of 1×1×1. The output channels of each residual block are 64, 128, and 256, respectively.

[0059] Furthermore, feature filtering is performed on the three-dimensional point cloud information to be registered to obtain corresponding point pairs, including: geometric feature aggregation of high-dimensional point cloud feature information of the source point cloud and the target point cloud according to the attention mechanism, obtaining a feature matrix, performing Euclidean distance calculation on the feature matrix, obtaining a similarity matrix, and selecting corresponding point pairs for the topk corresponding point pairs in each row of the similarity matrix.

[0060] Furthermore, the corresponding point pairs are clustered in two spaces to obtain the consensus set including:

[0061] According to the high-dimensional point cloud feature information and the source point cloud in the corresponding point pairs, a spatial consistency matrix and an angular compatibility matrix are obtained, and according to the product of the spatial consistency matrix and the angular compatibility matrix, a local geometric feature spatial compatibility matrix is ​​obtained;

[0062] The corresponding point pairs are screened according to the local geometric feature space compatibility matrix to obtain the initial seeds;

[0063] The initial seeds are screened according to the feature similarity threshold and the distance threshold to obtain a corresponding point pair set in the feature space and a corresponding point pair set in the geometric space;

[0064] Select the corresponding point pairs shared by the corresponding point pair set in the feature space and the corresponding point pair set in the geometric space to obtain a consensus set.

[0065] Furthermore, obtaining the spatial consistency matrix of corresponding point pairs includes:

[0066] The KNN nearest neighbor method is used to process the high-dimensional point cloud feature information corresponding to the corresponding point pairs to obtain the first local block composed of several adjacent points;

[0067] Calculate the corresponding point pairs in the first local block to obtain a spatial consistency matrix.

[0068] Specifically, after obtaining the high-dimensional point cloud information features, K adjacent points are selected to form a local patch through the KNN nearest neighbor algorithm, and the spatial consistency of the initial corresponding point pairs in the patch is calculated to describe the correctness of the point pairs.

[0069] Furthermore, obtaining the angle compatibility matrix includes:

[0070] According to the corresponding source point cloud in the corresponding point pair, a KNN nearest neighbor method is used to perform a k-nearest neighbor search in the original point cloud space to obtain a second local block;

[0071] Calculate a feature vector in the second local block according to the feature information of the high-dimensional point cloud to obtain a feature vector;

[0072] The angular difference of the eigenvectors is calculated to obtain the angular compatibility matrix.

[0073] Specifically, this embodiment performs a k-nearest neighbor search in the original point cloud space to form a local patch. We calculate feature vectors in the local patch, which capture the local geometry of the point. For each pair of matching point pair feature vectors, their angular difference is calculated to obtain an angular compatibility matrix.

[0074] Furthermore, the corresponding point pairs are screened for two-space clustering according to the local geometric feature space compatibility matrix, and obtaining the initial seeds includes: selecting topk corresponding point pairs in each row of the local geometric feature space compatibility matrix, and obtaining the filtered corresponding point pairs.

[0075] Furthermore, the consensus set is:

[0076] S final =S feat ∩S geo

[0077] S feat ={(x i ,y j )|Sim feat (x i ,y j )≥θ feat}

[0078] S geo ={(x i ,y j )|dist feat (x i ,y j )≤θ geo}

[0079] Among them, S final is the final corresponding point pair set, S feat is the set of corresponding point pairs in the feature space, S geo is the set of corresponding point pairs in geometric space, x i ,y j For the corresponding point pairs in the source point cloud and the target point cloud, Sim feat To calculate the feature similarity of corresponding points in the feature space, θ feat is the feature similarity threshold, dist feat To calculate the distance between corresponding points in geometric space, θ geo is the distance threshold.

[0080] In this embodiment, a two-space clustering strategy is used to further construct a consensus set using high-confidence seeds. The consensus set is constructed by using the common high-confidence seeds between the spatial features of deep learning and the traditional spatial features of the three-dimensional point cloud itself.

[0081] Furthermore, the consensus set is decomposed to complete the point cloud registration, including: performing singular value decomposition on the consensus set, obtaining a transformation matrix, and completing the point cloud registration.

[0082] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0083] like Figure 1 As shown, it is a schematic diagram of the method flow of this embodiment, and the steps include: obtaining the point cloud image information required for point cloud registration; performing feature extraction on the point cloud image information to obtain high-dimensional point cloud feature information; obtaining corresponding point pairs based on the high-dimensional point cloud feature information; performing local geometric feature space compatibility matrix calculation on the corresponding point pairs to obtain initial seeds; based on the two-space clustering strategy, obtaining high-confidence seeds and constructing a consensus set; based on singular value decomposition, obtaining a transformation matrix to complete point cloud registration.

[0084] This embodiment obtains source point cloud images and target point cloud images required for point cloud registration, using mainstream standard data sets in the field of point cloud registration.

[0085] Specifically, you can download and obtain them by visiting the official websites of commonly used public datasets such as 3DMatch, KITTI, 3DLoMatch, etc. These datasets are widely used in this field and can provide important support for the development and evaluation of point cloud registration methods.

[0086] The three-dimensional point cloud information includes: source point cloud image and target point cloud image; the high-dimensional point cloud feature information includes: features with local geometric information between batches and features with global spatial information. Obtaining the high-dimensional point cloud feature information includes: constructing a feature extractor based on a U-net network of a convolutional neural network and aggregating features guided by local features to obtain high-dimensional point cloud feature information; obtaining corresponding point pairs includes: aggregating geometric features of high-dimensional feature information to obtain a feature matrix, calculating the Euclidean distance algorithm to obtain a similarity matrix, filtering the three-dimensional point source information to be registered according to the similarity matrix, and obtaining corresponding point pairs, wherein the corresponding point pairs are the corresponding points with the closest distance between the source point cloud and the target point cloud; the method for constructing the initial seed includes: filtering and updating the corresponding point pairs based on the local geometric feature space compatibility matrix to obtain the initial seed, wherein the initial seed is the corresponding point pair after filtering.

[0087] The difference between the local geometric feature space compatibility matrix and the geometric feature is whether to pay attention to learning high-dimensional features. The local geometric feature space compatibility matrix and the geometric feature information have the same scale. The feature extractor is constructed with the U-Net structure based on the convolutional neural network; specifically, the point cloud features of the geometric information are used to construct local features with the local receptive field; the local geometric feature space compatibility matrix and the geometric feature information obtained by feature extraction have the same scale, only whether to pay attention to learning high-dimensional features. The feature extractor includes: a three-dimensional convolution layer with a convolution kernel size of 7×7×7 and a step size of 1, which is used to obtain the initial feature information of the point cloud image; a three-layer residual block is used to downsample the initial feature information of the point cloud image to obtain high-dimensional feature information; among them, the residual block is composed of two layers of three-dimensional convolution layers with a convolution kernel size of 3×3×3 and a step size of 1, and one layer of three-dimensional convolution layer with a convolution kernel size of 1×1×1. The output channels of each residual block are 64, 128, and 256 respectively. Specifically, Figure 3 shown.

[0088] The method for obtaining corresponding point pairs comprises the following steps:

[0089] The feature extractor is constructed using the U-Net structure based on the convolutional neural network. The local receptive field is used to extract point cloud features containing geometric information, and local features are constructed. This is used as a guide for feature aggregation to obtain high-dimensional point cloud feature information. Subsequently, the high-dimensional feature information is filtered according to the Euclidean distance algorithm to obtain corresponding point pairs.

[0090] Specifically, it includes: using a feature extractor to extract features from the source point cloud image and the target point cloud image to obtain high-dimensional point cloud feature information, and using spatial consistency and angle consistency to guide the network to aggregate spatial information based on the high-dimensional point cloud feature information of the source point cloud image and the target point cloud image to complete the learning of high-dimensional information. Under the guidance of spatial consistency and angle consistency, the features are further aggregated through non-local geometric feature extraction. Given the input feature matrix, that is, the high-dimensional point cloud feature information F∈R of the source point cloud and the target point cloud N×d , output two feature matrices Z∈R N×d It is the weighted sum of all feature matrices. R represents an N*d matrix, where N is the number of points. Both the source point cloud and the target point cloud have a feature information matrix. i，j is obtained by calculating the attention score e i，j The softmax is calculated row by row. This process can be summarized as follows:

[0091]

[0092] e i，j =softamax(θβ(F i W Q )(Fj W K ))

[0093] Where β represents the spatial consistency term defined by the length constraint derived from the three-dimensional rigid transformation, and θ represents the angular consistency term defined by the local angle constraint under the three-dimensional rigid transformation proposed in this embodiment. Here, W Q , W k , W V are the projection matrices of query, key, and value respectively, N is the number of points in the point cloud, a i，j is the weight coefficient, F j is the feature of the jth point in the target point cloud, Fi is the feature of the i-th point in the source point cloud, e i，j It is the feature correlation calculation between the source point cloud and the target point cloud, where i and j are the indexes of the points in the source point cloud and the target point cloud respectively.

[0094] The Euclidean distance between the two output feature matrices is calculated to obtain the similarity matrix.

[0095] Select the topk corresponding point pairs in each row of the similarity matrix to obtain the corresponding point pairs.

[0096] Due to the presence of noise, the obtained corresponding point pairs may also contain erroneous point pairs, resulting in a decrease in registration accuracy. Therefore, this embodiment uses the local geometric feature space compatibility matrix to further obtain the initial seed and complete the filtering of the corresponding point pairs.

[0097] The method for filtering corresponding point pairs based on the local geometric feature space compatibility matrix includes: extracting effective spatial information based on high-dimensional point cloud feature information obtained through deep learning and traditional spatial feature information of the three-dimensional point cloud itself.

[0098] After obtaining the corresponding point pairs, we select the high-dimensional point cloud information features corresponding to the corresponding point pairs, select K adjacent points to form a local patch through the KNN nearest neighbor algorithm, and calculate the spatial consistency matrix of the corresponding point pairs in the patch to describe the correctness of the point pairs. This process can be summarized as the following formula:

[0099]

[0100] Among them, d ij is the distance difference between corresponding point pairs, SC ij is the spatial consistency matrix (used to measure point pairs), represents a monotonically decreasing kernel function, x represents a point in the source point cloud, and y represents a point in the target point cloud.

[0101] Due to the influence of noise and randomness, the point pairs obtained based on the spatial consistency matrix may not be robust and may contain erroneous point pairs. The present invention uses the local angle consistency and the spatial consistency matrix to work together to obtain the local geometric feature space compatibility matrix. Figure 4 As shown in (a)-(d).

[0102] The feature information matrix of the source point cloud and the target point cloud has been obtained through the feature extractor, which means that each point has a fixed-dimensional feature descriptor to represent it. The local patch is formed by k-nearest neighbor search, and the distance is used as the weight to weight the aggregation into a local feature vector. Specifically, after obtaining the corresponding point pair, the original point cloud contained in the corresponding point pair is selected, and the k-nearest neighbor search is performed in its original point cloud space to form a local patch. The feature vectors are calculated in the local patch through the feature descriptor, and these feature vectors capture the local geometry of the point.

[0103] For each pair of matching point pair feature vectors, calculate their angular difference to obtain the angular compatibility matrix. Then dot product with the spatial consistency matrix to obtain the local geometric feature space compatibility matrix. This process can be summarized as:

[0104]

[0105] GSC ij =SC ij *AC ij

[0106] in, is the local feature vector of the i-th point in the source point cloud, is the local feature vector of the jth point in the target point cloud, AC ij is the angle-consistent matrix.

[0107] The local geometric feature space compatibility matrix stores the point pair scores, and each row takes the topk corresponding point pairs to perform filtering and obtain the initial seeds.

[0108] After obtaining the initial seeds, in order to achieve the best effect. This embodiment adopts a two-space clustering strategy to further construct a consensus set with high-confidence seeds. By combining the spatial features extracted by deep learning with the traditional geometric features of the three-dimensional point cloud itself, high-confidence seed points are jointly identified in the feature space and the geometric space to construct a consensus set. Figure 5 As shown in (a)-(e), compatibility refers to the score of corresponding point pairs, and different color depths are used to indicate the score.

[0109] Specifically, the corresponding point pair set in the feature space is selected from the initial seed by the feature similarity threshold, the corresponding point pair set in the geometric space is selected from the initial seed by the distance threshold, and the high-confidence seed is obtained. The common high-confidence seeds are selected from the corresponding point pair set in the feature space and the corresponding point pair set in the geometric space to obtain the consensus set. This process can be summarized as follows:

[0110] S feat ={(x i ,y j )|Sim feat (x i ,y j )≥θ feat}

[0111] S geo ={(x i ,y j )|dist feat (x i ,y j )≤θ geo}

[0112] S final =S feat ∩S geo

[0113] Among them, S final is the final corresponding point pair set, S feat is the set of corresponding point pairs in the feature space, S geo is the set of corresponding point pairs in geometric space, x i ,y j are the corresponding points in the source point cloud and the target point cloud, Sim feat To calculate the feature similarity of corresponding points in the feature space, θ feat is the feature similarity threshold, dist feat To calculate the distance between corresponding points in geometric space, θ geo The distance threshold is used. The corresponding point pairs are screened based on the local geometric feature space compatibility matrix and the two-space clustering strategy to obtain high-confidence seeds and build a consensus set.

[0114] Finally, the singular value decomposition algorithm is used to obtain the transformation matrix to complete the point cloud registration.

[0115] Embodiment 2:

[0116] This embodiment provides a point cloud registration system based on local geometric features and clustering strategy, including: an acquisition module, an extraction module, a screening module and a registration module;

[0117] An acquisition module is used to obtain the three-dimensional point cloud information to be registered, wherein the three-dimensional point cloud information to be registered includes a source point cloud image and a target point cloud image;

[0118] An extraction module is used to extract features from the three-dimensional point cloud information to be registered and obtain high-dimensional point cloud feature information;

[0119] A screening module is used to perform feature filtering on the three-dimensional point cloud information to be registered according to the feature information of the high-dimensional point cloud, and obtain corresponding point pairs between the corresponding points and the three-dimensional point cloud information to be registered, wherein the corresponding point pairs are the corresponding points with the closest distance between the source point cloud and the target point cloud;

[0120] The registration module is used to perform two-space clustering on corresponding point pairs, obtain a consensus set, decompose the consensus set, and complete point cloud registration, where the consensus set is the final set of corresponding point pairs.

[0121] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A point cloud registration method based on local geometric features and clustering strategy, characterized in that: include: Acquire three-dimensional point cloud information to be registered, wherein the three-dimensional point cloud information to be registered includes a source point cloud image and a target point cloud image; Extracting features of the three-dimensional point cloud information to be registered to obtain high-dimensional point cloud feature information; Performing feature filtering on the three-dimensional point cloud information to be registered according to the high-dimensional point cloud feature information to obtain corresponding point pairs, wherein the corresponding point pairs are the corresponding points with the closest distance between the source point cloud and the target point cloud; The corresponding point pairs are clustered in two spaces to obtain a consensus set, and the consensus set is decomposed to complete the point cloud registration, wherein the consensus set is the final corresponding point pair set.

2. The point cloud registration method based on local geometric features and clustering strategy according to claim 1, characterized in that: Acquiring high-dimensional point cloud feature information includes: using a feature extractor to extract local features of the three-dimensional point cloud information to be registered to obtain the high-dimensional point cloud feature information, and the feature extractor is constructed based on a U-Net network.

3. The point cloud registration method based on local geometric features and clustering strategy according to claim 1, characterized in that: Performing feature filtering on the three-dimensional point cloud information to be registered to obtain corresponding point pairs includes: performing geometric feature aggregation on the high-dimensional point cloud feature information of the source point cloud and the target point cloud according to the attention mechanism to obtain a feature matrix, performing Euclidean distance calculation on the feature matrix to obtain a similarity matrix, and selecting the topk corresponding point pairs in each row of the similarity matrix to obtain the corresponding point pairs.

4. The point cloud registration method based on local geometric features and clustering strategy according to claim 1, characterized in that: Performing two-space clustering on the corresponding point pairs to obtain a consensus set includes: According to the high-dimensional point cloud feature information and the source point cloud in the corresponding point pairs, a spatial consistency matrix and an angular compatibility matrix are obtained, and according to the product of the spatial consistency matrix and the angular compatibility matrix, a local geometric feature spatial compatibility matrix is ​​obtained; Screening the corresponding point pairs according to the local geometric feature space compatibility matrix to obtain initial seeds; The initial seeds are screened according to a feature similarity threshold and a distance threshold to obtain a corresponding point pair set in a feature space and a corresponding point pair set in a geometric space; The corresponding point pairs common to the corresponding point pair set in the feature space and the corresponding point pair set in the geometric space are selected to obtain the consensus set.

5. The point cloud registration method based on local geometric features and clustering strategy according to claim 4, characterized in that: Obtaining the spatial consistency matrix of the corresponding point pairs includes: Using the KNN nearest neighbor method to process the high-dimensional point cloud feature information corresponding to the corresponding point pair, to obtain a first local block composed of a plurality of adjacent points; Distance differences are calculated for the corresponding point pairs in the first local block to obtain the spatial consistency matrix.

6. The point cloud registration method based on local geometric features and clustering strategy according to claim 4, characterized in that: Obtaining the angle compatibility matrix includes: According to the source point cloud corresponding to the corresponding point pair, a KNN nearest neighbor method is used to perform a k-nearest neighbor search in the original point cloud space to obtain a second local block; Calculating a feature vector in the second local block according to the high-dimensional point cloud feature information to obtain the feature vector; The angular difference of the eigenvector is calculated to obtain the angular compatibility matrix.

7. The point cloud registration method based on local geometric features and clustering strategy according to claim 4, characterized in that: The corresponding point pairs are screened and clustered in two spaces according to the local geometric feature space compatibility matrix to obtain initial seeds, including: The topk corresponding point pairs in each row of the local geometric feature space compatibility matrix are selected to obtain the initial seeds.

8. The point cloud registration method based on local geometric features and clustering strategy according to claim 4, characterized in that: The consensus set is: S final =S feat ∩S geo S feat ={(x i ,y j )|Sim feat (x i ,y j )≥θ feat } S geo ={(x i ,y j )|dist feat (x i ,y j )≤θ geo } Among them, S final is the final corresponding point pair set, S feat is the set of corresponding point pairs in the feature space, S geo is the set of corresponding point pairs in geometric space, x i ,y j For the corresponding point pairs in the source point cloud and the target point cloud, Sim feat To calculate the feature similarity of corresponding points in the feature space, θ feat is the feature similarity threshold, dist feat To calculate the distance between corresponding points in geometric space, θ geo is the distance threshold.

9. The point cloud registration method based on local geometric features and clustering strategy according to claim 1, characterized in that: Decomposing the consensus set to complete the point cloud registration includes: performing singular value decomposition on the consensus set, obtaining a transformation matrix, and completing the point cloud registration.

10. A point cloud registration system based on local geometric features and clustering strategy, used to implement the method according to any one of claims 1 to 9, characterized in that: include: Acquisition module, extraction module, screening module and registration module; The acquisition module is used to obtain the three-dimensional point cloud information to be registered, wherein the three-dimensional point cloud information to be registered includes a source point cloud image and a target point cloud image; The extraction module is used to extract features from the three-dimensional point cloud information to be registered to obtain high-dimensional point cloud feature information; The screening module is used to perform feature filtering on the three-dimensional point cloud information to be registered according to the feature information of the high-dimensional point cloud, and obtain corresponding point pairs between the three-dimensional point cloud information to be registered, wherein the corresponding point pairs are the corresponding points with the closest distance between the source point cloud and the target point cloud; The registration module is used to perform two-space clustering on the corresponding point pairs, obtain a consensus set, decompose the consensus set, and complete point cloud registration, wherein the consensus set is a final corresponding point pair set.

Images