Point cloud registration method based on outlier suppression
By building feature descriptors and graph feature extractors between point clouds, combining length consistency and LRF compatibility, using rigid transformation of coarse to fine sampling and weighted SVD to generate consensus sets, the problems of low point cloud registration efficiency and low accuracy in the prior art are solved, and more efficient and accurate point cloud registration effects are achieved.
Patent Information
- Application Number
- CN202411568315.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-05
- Publication Date
- 2025-02-25
AI Technical Summary
Existing point cloud registration methods are inefficient and inaccurate when dealing with outliers, especially in scenarios with high outliers.
By obtaining the feature descriptor between the source point cloud and the target point cloud, a correspondence relationship is constructed and feature embedding is obtained using a graph feature extractor. Combining length consistency and local coordinate system (LRF) compatibility, a rigid transformation of the consensus set is generated using coarse to fine sampling and weighted SVD, and finally the optimal transformation that satisfies the intrapoint correspondence relationship is selected.
It significantly improves the accuracy and reliability of point cloud registration, especially in complex scenarios, with higher recall and lower rotation and translation errors compared to traditional methods such as ICP algorithms.
Smart Images
Figure CN119515934A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of point cloud registration, and in particular relates to a point cloud registration method based on outlier suppression. Background Art
[0002] With the rapid development of modern information technology and scanning equipment, point cloud data (PCD) has become the main data format for representing the 3D world. Point clouds have a wide range of applications in many fields, including robotics, biomedical imaging, road and architecture mapping, urban modeling, autonomous driving, and augmented reality. 3D point cloud registration is a critical and complex problem in point cloud data processing. Point cloud registration usually takes a pair of unregistered point clouds (source point cloud and target point cloud) as input, and the goal is to determine a rigid transformation that allows the two point clouds to be aligned, thereby generating a complete 3D point cloud.
[0003] At present, point cloud registration methods are mainly divided into deep learning-based point cloud registration methods and traditional point cloud registration methods. Traditional point cloud registration methods such as RANSAC can achieve good performance in some point cloud registration scenarios, but due to the existence of outliers, it often takes a lot of time to converge and the accuracy of the results cannot be guaranteed. Ratio test and reciprocal test methods only identify internal correspondences based on descriptor similarity without considering their spatial coherence. In the 3D domain, spatial consistency (SC) is one of the most important geometric properties that inliers should follow. This is a strong constraint obtained through rigid transformation because for inliers, they are always compatible with each other, while outliers are almost incompatible with each other due to randomness. Spectral matching (SM) is a traditional algorithm that relies on spatial consistency to achieve outlier rejection. SM has a fast execution speed, but the accuracy is low in scenarios with high outlier rates.
[0004] In recent years, many learning-based 3D outlier suppression algorithms have emerged, such as DGR and 3DRegNet, which transform the outlier rejection problem into an internal / outlier binary classification problem, where the network embeds deep features from the corresponding input and predicts the internal probability of each correspondence for outlier removal. However, these methods only rely on general operators such as sparse convolution and multi-layer perceptron to capture contextual information, while ignoring the basic 3D spatial relationships, so their performance in complex scenes is not optimistic.
[0005] In view of the above problems, the present invention proposes a point cloud registration method based on outlier suppression. Summary of the invention
[0006] The purpose of the present invention is to provide a point cloud registration method based on outlier suppression, aiming to solve the problems raised in the above background technology.
[0007] The purpose of the present invention is achieved through the following technical solutions:
[0008] A point cloud registration method based on outlier suppression includes the following steps:
[0009] Step 1: Obtain feature descriptors between the source point cloud and the target point cloud, construct the correspondence between the source point cloud and the target point cloud by using the feature descriptors, take the correspondence as input and obtain feature embedding between each correspondence through a graph feature extractor; the graph feature extractor includes a graph edge generator and a feature aggregation layer;
[0010] Step 2: After obtaining the features between the corresponding relationships, use coarse-to-fine sampling to obtain the consensus set: first, obtain the core points by sampling the grid points, and then gather the corresponding relationships with consistent features around the core points to form a consensus set;
[0011] Step 3: After obtaining the consensus set, perform weighted SVD for each consensus set to generate the corresponding rigid transformation. The weighted SVD considers the weight of each corresponding relationship, and the weight is obtained by local spectral matching.
[0012] Step 4: Select the best rigid transformation from all rigid transformations generated in the consensus set. The selection criterion is based on the interior point correspondence satisfied by each transformation.
[0013] Furthermore, the specific process of step 1 is as follows:
[0014] First, a low-dimensional local feature descriptor P is constructed for each correspondence using the coordinates of the correspondence. i The local feature descriptor P i It is expressed as:
[0015]
[0016] Among them C n Represents C i Neighbor, K i Represents C i k nearest neighbors in Euclidean space;
[0017] Through the nonlinear function f θ :R K×12 →R D The low-dimensional local feature descriptor P i Mapped to a high-dimensional feature space, where D is the feature dimension, R is the dimension representation, K is the number of k nearest neighbors; θ represents the parameters of the nonlinear function, including multi-layer perceptron, maximum pooling and concatenation operations; the obtained high-dimensional features are used as the node features F of the initial input graph, for the corresponding relationship C i The node feature F i It is expressed as:
[0018] F i =f θ (P i )
[0019] Secondly, the edge adjacency matrix is generated using the graph edge generator. Assume that the graph constructed by the correspondence is represented by G = {C, E}, where G is the graph, C is the correspondence, and E is the edge. The graph edge generator consists of a standard transformer architecture and a sotfmax operation. The edge adjacency matrix generated is represented as:
[0020] τ1=f transformer (F)
[0021] E = softmax(τ1 T ×τ1)
[0022] Where τ1 is the transformed node feature and T is the transpose;
[0023] Finally, the feature aggregation layer is used to aggregate the corresponding graph node features, with the node feature F and the graph edge feature E as input, and the autocorrelation feature F of the output node Corr , the autocorrelation feature is aggregated from the nodes along the edge of graph G; the node autocorrelation feature of graph G It is calculated by the following formula:
[0024]
[0025] in is the node autocorrelation feature after aggregation, N is the number of nodes, is the L1 column normalized matrix computed from E, E ij represents the edge feature between the i-th node and the j-th node; f μ and are two different message passing functions, consisting of a fully connected layer and ReLu; F j and F i These are two different node features.
[0026] Furthermore, the attention component in the transformer architecture is replaced by the consistency score attention, which explicitly uses the length consistency and LRF compatibility between the corresponding relations to learn the feature consistency between the corresponding relations; the specific process is as follows:
[0027] First, the feature matrix F j Projected to query Q j , key K j ,value V j :
[0028] Q j =F j W Q ,K j =F j W K ,V j =F j W V
[0029] Where W Q ,W K ,W V ∈R D×D , which are the linear projection functions of query, key and value respectively;
[0030] Second, the attention scores in the original attention calculation are re-weighted using the consistency score matrix SSM:
[0031] Z j =softmax(ωSSM)V j
[0032] Where Z j is the reweighted attention score, ω is Q j and K j The dot product of , SSM is composed of length consistency and LRF compatibility.
[0033] Furthermore, the length consistency is calculated by measuring the difference between the length of the point pair in the source point cloud X and the length of the corresponding point pair in the target point cloud Y:
[0034] α ij =|||x i -x j ||-||y i -y j |||
[0035] where α ij is the length consistency, x i and x j are two different points in the first point cloud, y i Yes and y j are two different points in the second point cloud;
[0036] LRF compatibility is calculated by computing the LRF difference between the length of a point pair in the source point cloud X and the corresponding point pair in the target point cloud Y:
[0037] β ij =|∈ LRF (x i ,x j )-∈ LRF (y i ,yj )|
[0038] where β ij is LRF compatible, ∈ LRF (x i ,x j ) is the LRF compatibility of two points in the first point cloud, ∈ LRF (y i ,y j ) is the LRF compatibility of two points in the second point cloud;
[0039] ∈ LRF (x i ,x j ) is expressed as follows:
[0040]
[0041] Where acos is the arccos function, trace is the trace of the matrix, is x j The inverse of the LRF, Represents x i LRF;
[0042] By polymerizing α ij and β ij , the final consistency score is expressed as:
[0043]
[0044] where α ij is the length consistency constraint between the i-th point and the j-th point, β ij LRF compatibility constraints between the i-th and j-th points, SSM ij is the consistency score, δ α and δ β are the parameters for length consistency and LRF compatibility, respectively.
[0045] Furthermore, the specific process of step 2 is as follows:
[0046] First, divide the point cloud into P grids, select a point with the highest confidence in each grid, and then cyclically sample in each grid, selecting the point with the highest confidence except the previously selected one, until the number of points reaches K, forming K core points;
[0047] Secondly, the K core points are expanded into K consensus sets. Specifically, first, K1 nearest neighbors are searched in the feature space to form a coarse set. On the basis of the coarse set, the consistency score of the coarse set is further obtained, and then the K2 corresponding points with the highest consistency scores are taken to form a fine set, where K2 <K1。
[0048] Furthermore, the specific process of step 3 is as follows:
[0049] First, the compatibility matrix is obtained using the consistency score and feature consistency score:
[0050] M ij =SSM ij *γ ij
[0051] Among them, M ij is the compatibility matrix, SSM ij is the consistency score, γ ij is the feature consistency score, expressed as follows:
[0052]
[0053] Where MAX[·] represents the maximum value operation. express The L2 normalized features, and are the autocorrelation features of two different aggregated nodes. express The L2 normalized features, δ γ is a parameter that controls the sensitivity of feature consistency;
[0054] Secondly, the compatibility matrix M is solved by the power method ij The main eigenvector of is regarded as the probability of the internal point, and the rigid transformation is solved by weighted SVD.
[0055] Furthermore, the specific process of step 4 is as follows:
[0056] First, for the kth consensus set R K and t K The estimate of R is calculated by a predefined error threshold (τ) K and t K The number of corresponding relations of the constraints is as follows:
[0057]
[0058] Where R K and t K is the rotation R and translation t corresponding to the k-th consensus set, Num K is the inlier count of the kth consensus set, N is the number of points in the point cloud, i represents the i-th point, and x i is the i-th point in the first point cloud, y i is the i-th point in the second point cloud, τ is the threshold, and [·] is the Iverson bracket;
[0059] Second, select the R with the highest inlier count K and t K As the final result.
[0060] Compared with the prior art, the present invention has the following beneficial effects:
[0061] The present invention enhances the performance of the learning-based outlier suppression algorithm in point cloud registration by explicitly utilizing length consistency and local coordinate system (LRF) compatibility, making the result more accurate and reliable. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION
[0063] In order to have a clearer understanding of the technical features, purposes and beneficial effects of the present invention, the technical solution of the present invention is now described in detail below, but it should not be understood as limiting the scope of the present invention. The experimental methods described in the following examples are all conventional methods unless otherwise specified.
[0064] like Figure 1 As shown, a point cloud registration method based on outlier suppression provided by an embodiment of the present invention includes the following steps:
[0065] Step 1: Obtain feature descriptors between the source point cloud and the target point cloud, use the feature descriptors to construct the correspondence between the source point cloud and the target point cloud, take the correspondence as input and obtain feature embedding between each correspondence through a graph feature extractor; the graph feature extractor includes a graph edge generator and a feature aggregation layer.
[0066] The specific process of step 1 is as follows:
[0067] Use the coordinates of the corresponding relationship to construct a low-dimensional local feature descriptor P for each corresponding relationship; for the corresponding relationship C i The local feature descriptor P i It is expressed as:
[0068]
[0069] Among them C n Represents C i Neighbor, K i Represents C i k nearest neighbors in Euclidean space;
[0070] Through the nonlinear function f θ :R K×12 →R D The low-dimensional local feature descriptor P iMapped to a high-dimensional feature space, where D is the feature dimension, R is the dimensional representation, K is the number of k nearest neighbors; θ represents the parameters of the nonlinear function, including multi-layer perceptron (MLP), maximum pooling and concatenation operations; the obtained high-dimensional features are used as the node features F of the initial input graph, and for the corresponding relationship C i The node feature F i It is expressed as:
[0071] F i =f θ (P i )
[0072] If only local features are used to represent the relationship between correspondences, it is very easy to obtain incorrect relationships, especially when local outliers are high, because it does not learn the global relationship. Therefore, we use correspondences to construct the graph and establish better correlation between correspondences.
[0073] After obtaining the feature embedding of the correspondence, the edge adjacency matrix is generated using the graph edge generator. Assume that the graph constructed by the correspondence is represented as G = {C, E}, where G is the graph, C is the correspondence, and E is the edge. The graph edge generator consists of a standard transformer architecture and a sotfmax operation. The edge adjacency matrix generated is represented as:
[0074] τ1=f transformer (F)
[0075] E = softmax(τ1 T ×τ1)
[0076] Where τ1 is the transformed node feature and T is the transpose;
[0077] After obtaining the features of the edge adjacency matrix and the corresponding relationship, the feature aggregation layer is used to aggregate the corresponding graph node features. It accepts the node feature F and the graph edge feature E as input and outputs the node autocorrelation feature F Corr , these features are aggregated from the nodes along the edges of the graph. It is calculated by the following formula:
[0078]
[0079] in is the node autocorrelation feature after aggregation, N is the number of nodes, is the L1 column normalized matrix computed from E, E ij represents the edge feature between the i-th node and the j-th node; f μ and are two different message passing functions, consisting of a fully connected layer and ReLu; Fj and F i These are two different node features.
[0080] Among them, the attention component in the transformer architecture is replaced by the consistency score attention, which explicitly uses the length consistency and LRF compatibility between corresponding relations to learn the feature consistency between corresponding relations, because the inliers are always compatible with each other and the outliers are almost not compatible with each other, so the obtained consistency score matrix indicates whether the corresponding relations are compatible with each other. Using the consistency score, the graph edge generator can capture the relationship between different corresponding relations to obtain more representative features.
[0081] The specific process is as follows:
[0082] First, the feature matrix F j Projected to query Q j , key K j ,value V j :
[0083] Q j =F j W Q ,K j =F j W K ,V j =F j W V
[0084] Where W Q ,W K ,W V ∈R D×D , which are the linear projection functions of query, key and value respectively;
[0085] Second, the attention scores in the original attention calculation are re-weighted using the consistency score matrix (SSM):
[0086] Z j =softmax(ωSSM)V j
[0087] Where Z j is the reweighted attention score, ω is Q j and K j The SSM is composed of length consistency (α) and LRF compatibility (β).
[0088] Specifically, length consistency is calculated by measuring the difference between the length of a point pair in the source point cloud X and the length of the corresponding point pair in the target point cloud Y:
[0089] α ij =|||x i -x j ||-||y i -y j ||
[0090] where α ij is the length consistency, x i and x j are two different points in the first point cloud, y i Yes and y j are two different points in the second point cloud;
[0091] LRF compatibility is calculated by computing the LRF difference between the length of a point pair in the source point cloud X and the corresponding point pair in the target point cloud Y:
[0092] β ij =|∈ LRF (x i ,x j )-∈ LRF (y i ,y j )|
[0093] where β ij is LRF compatible, ∈ LRF (x i ,x j ) is the LRF compatibility of two points in the first point cloud, ∈ LRF (y i ,y j ) is the LRF compatibility of two points in the second point cloud;
[0094] ∈ LRF (x i ,x j ) is expressed as follows:
[0095]
[0096] Where acos is the arccos function, trace is the trace of the matrix, is x i The inverse of the LRF, Represents x i LRF;
[0097] By polymerizing α ij and β ij , the final consistency score is expressed as:
[0098]
[0099] where α ijis the length consistency constraint between the i-th point and the j-th point, β ij the LRF compatibility constraint between the i-th point and the j-th point, SSM ij is the consistency score, δ α and δ β are the parameters of length consistency and LRF compatibility respectively, used to control their sensitivities. If the length consistency is greater than δ α or the LRF compatibility is greater than δ β both are considered incompatible. In contrast, only when both the length consistency and the LRF compatibility are small, SSM will give a large value. When both the length consistency and the LRF compatibility are 0, SSM will be 1.
[0100] Step 2: After obtaining the features between the correspondences, use coarse-to-fine sampling to obtain the consensus set: First, obtain the kernel points through grid point sampling, and then gather the correspondences with consistent features around the kernel points to form the consensus set.
[0101] The specific process of the said Step 2 is as follows:
[0102] First, divide the point cloud into O grids, select a point with the highest confidence in each grid, and then sample cyclically in each grid, choosing points other than the previously selected point with the highest confidence until the number of points is K, forming K kernel points;
[0103] Secondly, expand the K kernel points into K consensus sets. Specifically: First, search for K1 nearest neighbors in the feature space to form a coarse set; based on the coarse set, further obtain the consistency score of the coarse set, and then select the highest K2 correspondences (K2 < K1) of the consistency score to form a fine set.
[0104] Step 3: After obtaining the consensus set, perform weighted SVD for each consensus set to generate the corresponding rigid transformation. The weighted SVD considers the weights of each correspondence. Here, the weights of the correspondences are obtained by local spectral matching.
[0105] The specific process of the said Step 3 is as follows:
[0106] First, use the consistency score and the feature consistency score to obtain the compatibility matrix:
[0107] M ij = SSM ij *γ ij
[0108] where M ij is the compatibility matrix, SSM ij is the consistency score, γ ij is the feature consistency score, which is expressed as follows:
[0109]
[0110] Where MAX[·] represents the maximum value operation. express The L2 normalized features, and are the autocorrelation features of two different aggregated nodes. express The L2 normalized features, δ γ is a parameter that controls the sensitivity of feature consistency;
[0111] Secondly, we transform the compatibility matrix M ij The main eigenvector of is associated with the weight of the corresponding relationship. The higher the main eigenvector value, the higher the probability that it is an inlier. The main eigenvector can be solved by the power method, and finally the main eigenvector is regarded as the probability of the inlier, and the rigid transformation is solved by weighted SVD.
[0112] Step 4: Select the best rigid transformation from all rigid transformations generated in the consensus set. The selection criterion is based on the internal point correspondence satisfied by each transformation.
[0113] The specific process of step 4 is as follows:
[0114] First, for the kth consensus set R K and t K The estimate of R is calculated by a predefined error threshold (τ) K and t K The number of corresponding relations of the constraints is as follows:
[0115]
[0116] Among them, R K and t K is the rotation R and translation t corresponding to the k-th consensus set, Num K is the inlier count of the kth consensus set, N is the number of points in the point cloud, i represents the i-th point, and x i is the i-th point in the first point cloud, y i is the i-th point in the second point cloud, τ is the threshold, and [·] is the Iverson bracket;
[0117] Second, select the R with the highest inlier count K and t K As the final result.
[0118] Embodiment 1. The present invention uses point cloud data collected in a real environment. These point cloud data contain real coordinates in the 3D world. The feature descriptor will extract the correspondence relationship of these point clouds in advance, because our method uses the correspondence relationship as input. The calculated correspondence relationship will be sent to the network model we proposed for further outlier removal, so the pre-extracted correspondence relationship contains a large number of outliers. After the outliers are removed, the probability of whether the correspondence relationship is an inner point is obtained by spectral matching, and then the rigid transformation is obtained through weighted SVD.
[0119] The input correspondence can be represented as a 6-tuple:
[0120] C=(Sx,Sy,Sz,Tx,Ty,Tz)
[0121] Where Sx, Sy, Sz are the xyz coordinates of the source point cloud, and Tx, Ty, Tz are the xyz coordinates of the target point cloud. Because the point cloud contains a large number of points, C will be a set. The entire dataset will be divided into training set, validation set and test set.
[0122] Taking a set of corresponding relationships under the training set as an example, the technical solution of the present invention is clearly and completely described. The methods of the present invention include but are not limited to the embodiments given.
[0123] The input set of correspondences is first passed through the graph feature extractor to obtain a unique feature embedding, and then the feature embedding is used to obtain a reliable consensus set using a coarse-to-fine sampling method. Secondly, spectral matching and weighted SVD are used to calculate a rigid transformation for each consensus set, and finally the rigid transformation with the best result is selected.
[0124] Specifically, the corresponding relationship C of the input i Through the feature extractor to obtain its feature F i , the feature represents the connection between each corresponding relationship and other corresponding relationships, and then we regard each corresponding relationship as a node, and its node feature is F i , assuming that the graph composed of correspondences can be represented as G = {C, E}, where G represents the graph, C represents the correspondence, and E represents the edge. Then the graph edge generator is used to obtain the edge features. In this process, the consistency score attention is used to obtain more unique edge features. The consistency score attention module explicitly uses length consistency and local coordinate system (LRF) compatibility to form a consistency score matrix as a constraint of the attention mechanism. After obtaining the edge features and node features, the feature aggregation layer is used to obtain the aggregated node autocorrelation features:
[0125]
[0126] where Fj and F i are two different node features. is the L1 column normalized matrix computed from E, E ij represents the edge feature between the i-th node and the j-th node, is the node autocorrelation feature after aggregation, f μ and Both are message passing functions, consisting of a fully connected layer and ReLu.
[0127] After obtaining the autocorrelation features of the nodes, the unique features representing the corresponding relationships are obtained. Then, grid point sampling is used to obtain reliable and well-distributed core points. Then, a coarse-to-fine sampling method is used around the core points to generate a consensus set. The coarse-to-fine sampling method is as follows: first, the nearest neighbors are searched to form a coarse set. Then, on this basis, neighbor points with length consistency and local coordinate system compatibility are selected. Finally, the core points and neighbor points form a consensus set.
[0128] Each consensus set can be regarded as a part of the original point cloud. Spectral matching can be used to obtain the probability of whether the corresponding relationship in the consensus set is an internal point. Then, weighted SVD will use the probability as a weight to calculate the rigid transformation of the consensus set. Each consensus set will produce a rigid transformation, and finally the rigid transformation that will result in the most aligned points in the original point cloud will be selected as the result output.
[0129] In the 3DMatch dataset of FPFH descriptor, the registration results of this method are compared with the classic point cloud registration algorithm - ICP algorithm. The comparison results are shown in Table 1:
[0130] Table 1 Registration results of this method and ICP algorithm
[0131] FPF RR (%↑) RE(°↓) TE(cm↓) ICP 5.79 7.93 17.59 Ours 83.94 2.06 6.47
[0132] From the data in Table 1, we can see that the ICP algorithm performs poorly on the 3DMatch dataset under the FPFH descriptor, and its RR (recall rate) is only 5.79%, which means that only a small number of point pairs are correctly matched. At the same time, RE (rotation error) is 7.93° and TE (translation error) is 17.59cm, both of which are relatively high. In contrast, our method (Ours) performs better on the 3DMatch dataset under the FPFH descriptor, with a RR of up to 83.94%, indicating that most point pairs are correctly matched. In addition, RE is only 2.06° and TE is 6.47cm, both of which are relatively low. The above results show that our method is significantly better than the ICP algorithm in terms of registration results. Our method can basically achieve alignment, while the ICP algorithm cannot achieve the same alignment effect.
[0133] The above are only preferred embodiments of the present invention. It should be pointed out that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention. These should also be regarded as the protection scope of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent.
Claims
1. A point cloud registration method based on outlier suppression, characterized in that: The following steps are involved: Step 1: Obtain feature descriptors between the source point cloud and the target point cloud, construct the correspondence between the source point cloud and the target point cloud by using the feature descriptors, take the correspondence as input and obtain feature embedding between each correspondence through a graph feature extractor; the graph feature extractor includes a graph edge generator and a feature aggregation layer; Step 2: After obtaining the features between the corresponding relationships, use coarse-to-fine sampling to obtain the consensus set: first, obtain the core points by sampling the grid points, and then gather the corresponding relationships with consistent features around the core points to form a consensus set; Step 3: After obtaining the consensus set, perform weighted SVD for each consensus set to generate the corresponding rigid transformation. The weighted SVD considers the weight of each corresponding relationship, and the weight is obtained by local spectral matching. Step 4: Select the best rigid transformation from all rigid transformations generated in the consensus set. The selection criterion is based on the interior point correspondence satisfied by each transformation.
2. The point cloud registration method based on outlier suppression according to claim 1, characterized in that: The specific process of step 1 is as follows: First, a low-dimensional local feature descriptor P is constructed for each correspondence using the coordinates of the correspondence. i The local feature descriptor P i It is expressed as: Among them C n Represents C i Neighbor, K i Represents C i k nearest neighbors in Euclidean space; Through the nonlinear function f θ :R K×12 →R D The low-dimensional local feature descriptor P i Mapped to a high-dimensional feature space, where D is the feature dimension, R is the dimension representation, K is the number of k nearest neighbors; θ represents the parameters of the nonlinear function, including multi-layer perceptron, maximum pooling and concatenation operations; the obtained high-dimensional features are used as the node features F of the initial input graph, for the corresponding relationship C i The node feature F i It is expressed as: F i =f θ (P i ) Secondly, the edge adjacency matrix is generated using the graph edge generator. Assume that the graph constructed by the correspondence is represented by G = {C, E}, where G is the graph, C is the correspondence, and E is the edge. The graph edge generator consists of a standard transformer architecture and a sotfmax operation. The edge adjacency matrix generated is represented as: τ1=f transformer (F) E=softmax(τ1 T ×τ1) Where τ1 is the transformed node feature and T is the transpose; Finally, the feature aggregation layer is used to aggregate the corresponding graph node features, with the node feature F and the graph edge feature E as input, and the autocorrelation feature F of the output node Corr , the autocorrelation feature is aggregated from the nodes along the edge of graph G; the node autocorrelation feature of graph G It is calculated by the following formula: in is the node autocorrelation feature after aggregation, N is the number of nodes, is the L1 column normalized matrix computed from E, E ij represents the edge feature between the i-th node and the j-th node; f μ and are two different message passing functions, consisting of a fully connected layer and ReLu; F j and F i These are two different node features.
3. The point cloud registration method based on outlier suppression according to claim 2, characterized in that: The attention component in the transformer architecture is replaced by consistency score attention, which explicitly uses the length consistency and LRF compatibility between correspondences to learn feature consistency between correspondences; The specific process is as follows: First, the feature matrix F j Projected to query Q j , key K j ,value V j : Q j =F j W Q ,K j =F j W K ,V j =F j W V Where W Q ,Q K ,W V ∈R D×D , which are the linear projection functions of query, key and value respectively; Second, the attention scores in the original attention calculation are re-weighted using the consistency score matrix SSM: Z j =softmax(ωSSM)V j Where Z j is the reweighted attention score, ω is Q j and K j The dot product of , SSM is composed of length consistency and LRF compatibility.
4. The point cloud registration method based on outlier suppression according to claim 3, characterized in that: The length consistency is calculated by measuring the difference between the length of a point pair in the source point cloud X and the length of the corresponding point pair in the target point cloud Y: a ij =||x i -x j ||-||and i -and j ||| where α ij is the length consistency, x i and x j are two different points in the first point cloud, y i Yes and y j are two different points in the second point cloud; LRF compatibility is calculated by computing the LRF difference between the length of a point pair in the source point cloud X and the corresponding point pair in the target point cloud Y: β ij =|∈ LRF (x i ,x j )-∈ LRF (and i ,and j )| where β ij is LRF compatible, ∈ LRF (x i ,x j ) is the LRF compatibility of two points in the first point cloud, ∈ LRF (y i ,y j ) is the LRF compatibility of two points in the second point cloud; ∈ LRF (x i ,x j ) is expressed as follows: Where acos is the arccos function, trace is the trace of the matrix, is x j The inverse of the LRF, Represents x i LRF; By polymerizing α ij and β ij , the final consistency score is expressed as: where α ij is the length consistency constraint between the i-th point and the j-th point, β ij LRF compatibility constraints between the i-th and j-th points, SSM ij is the consistency score, δ α and δ β are the parameters for length consistency and LRF compatibility, respectively.
5. The point cloud registration method based on outlier suppression according to claim 1, characterized in that: The specific process of step 2 is as follows: First, divide the point cloud into O grids, select a point with the highest confidence in each grid, and then cyclically sample in each grid, selecting the point with the highest confidence except the one selected before, until the number of points reaches K, forming K core points; Secondly, the K core points are expanded into K consensus sets. Specifically, first, K1 nearest neighbors are searched in the feature space to form a coarse set. On the basis of the coarse set, the consistency score of the coarse set is further obtained, and then the K2 corresponding points with the highest consistency scores are taken to form a fine set, where K2 <K1。 6. The point cloud registration method based on outlier suppression according to claim 1, characterized in that: The specific process of step 3 is as follows: First, the compatibility matrix is obtained using the consistency score and feature consistency score: M ij =SSM ij *γ ij Among them, M ij is the compatibility matrix, SSM ij is the consistency score, γ ij is the feature consistency score, expressed as follows: Where MAX[·] represents the maximum value operation. express The L2 normalized features, and are the autocorrelation features of two different aggregated nodes. express The L2 normalized features, δ γ is a parameter that controls the sensitivity of feature consistency; Secondly, the compatibility matrix M is solved by the power method ij The main eigenvector of is regarded as the probability of the internal point, and the rigid transformation is solved by weighted SVD.
7. The point cloud registration method based on outlier suppression according to claim 1, characterized in that: The specific process of step 4 is as follows: First, for the kth consensus set R K and t K The estimate of R is calculated by a predefined error threshold (τ) K and t K The number of corresponding relations of the constraints is as follows: Where R K and t K is the rotation R and translation t corresponding to the k-th consensus set, Num K is the inlier count of the kth consensus set, N is the number of points in the point cloud, i represents the i-th point, and x i is the i-th point in the first point cloud, y i is the i-th point in the second point cloud, τ is the threshold, and [·] is the Iverson bracket; Second, select the R with the highest inlier count K and t K As the final result.