Point cloud registration method and system

Abstract

The present invention provides a point cloud registration method and system. This method processes source and target point cloud features through self-attention and cross-attention layers to obtain interacted source and target point cloud features. When searching for corresponding points, the method no longer relies on the similarity of local features. Instead, it generates source and target point cloud soft points based on the source and target point clouds, respectively. This allows for the formation of corresponding soft points, with the number of generated soft points being fewer than the point cloud to be registered. Subsequently, rotation and translation parameters can be directly estimated from the corresponding soft points. This method does not rely on explicit or implicit feature matching to establish correspondences, but rather automatically generates soft points to establish correspondences. This eliminates the need for a feature matching process, thus avoiding the problem of numerous mismatches caused by feature matching in low-overlap areas. Furthermore, the method eliminates the need for outlier rejection, effectively increasing operational speed and shortening point cloud registration time.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional vision, and in particular to a point cloud registration method and system. Background Art

[0002] Point cloud registration is an important 3D vision task, commonly used in 3D reconstruction, robotic pose estimation, and remote sensing observation. Point cloud registration aims to find the optimal pose and align two point clouds to obtain a complete object or scene. To estimate the pose transformation parameters between two point clouds, a common approach is to select key points and then describe the features of the local region where the key points reside. Based on the feature similarity between the key points in the scene point cloud and the target point cloud, feature matching is performed to obtain a set of corresponding points. The rotation matrix and translation vector are then estimated based on the correspondence between the corresponding points.

[0003] In the existing technology, several methods have been proposed for describing features at key points, such as PFH, FPFH, SHOT, RoPS, and Spin Image. The features extracted by these methods have a certain degree of rigid transformation invariance and perform well on small datasets. However, because these methods extract low-level features that are manually designed, they suffer from problems such as poor robustness and insufficient description capabilities. In recent years, an increasing number of large-scale labeled 3D datasets have become available, and methods based on deep neural networks (3DMatch, CGF, PerfectMatch, FCGF, DIP, D3Feat, SpinNet, etc.) have achieved certain results in learning local feature descriptions. However, in practice, the source and target point clouds usually have only a small amount of overlap, and the geometric structures of the corresponding points obtained through feature matching are not necessarily exactly the same. This results in explicit local feature matching generating a large number of erroneous corresponding points.

[0004] A large number of erroneous correspondences results in a significant waste of time in estimating rotation and translation parameters. Therefore, other methods are needed to eliminate these erroneous correspondences. RANSAC (RAndom SAmple Consensus) is the most commonly used outlier removal method, which searches for the optimal transformation parameters through multiple iterations. In each iteration, RANSAC randomly selects a subset of corresponding points to estimate the parameters, and finally selects the parameters with the highest inlier rate from the multiple sets of parameters as the final estimate. Currently, some more advanced outlier removal methods have emerged, such as first mapping the corresponding points and then finding the largest cluster in the map to eliminate erroneous correspondences; and using boundary rules to reduce the input correspondences to fewer correspondences with a higher inlier rate. This method has a time complexity that is the square of the number of point correspondences. Although these methods can effectively eliminate erroneous correspondences, they are all time-consuming.

[0005] To avoid the excessive time consumption caused by outlier removal, several RANSAC-free registration methods have emerged, often based on implicit feature matching. Typical implicit feature matching methods include DCP, RPMNet, CoFiNet, and GeoTransformer. These methods typically construct a matching matrix based on feature similarity and optimal transfer theory. This matrix describes not a one-to-one correspondence but a many-to-many relationship. Rotation and translation parameters are calculated using weighted SVD, with weights provided by the matching matrix. Compared to explicit feature matching methods, these methods improve the robustness of feature matching and avoid the RANSAC process, achieving end-to-end registration algorithms. However, the construction of the matching matrix still relies on the similarity of local features, which can still lead to mismatches in some registration tasks due to inconsistencies in the local geometric structure of corresponding points. Both explicit and implicit feature matching methods have limitations in registration tasks involving small amounts of overlapping regions.

[0006] Therefore, how to avoid the large number of iterations required by RANSAC, which consumes a lot of time, and how to solve the problem of a large number of false matches in the feature matching process due to insufficient local feature description, are technical problems that need to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides a point cloud registration method and system to avoid the large number of iterations required by RANSAC, which consumes a lot of time, and solves the problem of a large number of false matches in the feature matching process due to insufficient local feature description.

[0008] One aspect of the present invention provides a point cloud registration method, which is applied to a point cloud registration system. The point cloud registration system includes a feature extraction and interaction module, a soft point generation module, and a rotation and translation parameter estimation module. The method includes:

[0009] Step S1: The feature extraction and interaction module takes the source point cloud and the target point cloud as input, linearly projects the source point cloud and the target point cloud into source point cloud features and target point cloud features respectively, and then transmits the source point cloud features and the target point cloud features to the Transformer model. The Transformer model includes 12 Transformer layers. The first 6 Transformer layers use self-attention layers, and the last 6 Transformer layers use cross-attention layers. The source point cloud features and the target point cloud features are processed by the self-attention layers and the cross-attention layers to obtain the source point cloud features and the target point cloud features after interaction;

[0010] Step S2, in the soft point generation module, all points in the source point cloud feature after interaction are generated according to kThe weight of the group source point cloud is weighted k The source point cloud soft point is composed of all points in the target point cloud feature after interaction according to k The weight of the group target point cloud is weighted to obtain k Target point cloud soft points, k The group source point cloud weight is predicted by the source point cloud features after interaction. k The weight of the group target point cloud is predicted by the target point cloud features after interaction;

[0011] Step S3, in the rotation and translation parameter estimation module, the Kabsch-Umeyama algorithm is used to obtain k Source point cloud software and k The rotation and translation parameters are directly estimated from the soft points of the target point cloud by singular value decomposition, and then the point cloud registration is performed based on the rotation and translation parameters.

[0012] Another aspect of the present invention provides a point cloud registration system, comprising a feature extraction and interaction module, a soft point generation module, and a rotation and translation parameter estimation module;

[0013] The feature extraction and interaction module takes the source point cloud and the target point cloud as input, linearly projects the source point cloud and the target point cloud into source point cloud features and target point cloud features respectively, and then transmits the source point cloud features and the target point cloud features to the Transformer model. The Transformer model includes 12 Transformer layers. The first 6 Transformer layers use self-attention layers, and the last 6 Transformer layers use cross-attention layers. The source point cloud features and the target point cloud features are processed by the self-attention layers and cross-attention layers to obtain the source point cloud features and the target point cloud features after interaction.

[0014] In the soft point generation module, all points in the source point cloud feature after interaction are calculated based on k The weight of the group source point cloud is weighted k The source point cloud soft point is composed of all points in the target point cloud feature after interaction according to k The weight of the group target point cloud is weighted to obtain k Target point cloud soft points, k The group source point cloud weight is predicted by the source point cloud features after interaction. k The weight of the group target point cloud is predicted by the target point cloud features after interaction;

[0015] In the rotation and translation parameter estimation module, the Kabsch-Umeyama algorithm is used to obtain k Source point cloud software and k The rotation and translation parameters are directly estimated from the soft points of the target point cloud by singular value decomposition, and then the point cloud registration is performed based on the rotation and translation parameters.

[0016] According to the point cloud registration method and system provided by the present invention, the source point cloud features and the target point cloud features are processed by the self-attention layer and the cross-attention layer to obtain the source point cloud features after interaction and the target point cloud features after interaction. When finding corresponding points, the present invention is no longer based on the similarity of local features, but is based on the source point cloud and the target point cloud generated separately. k Source point cloud software and k target point cloud soft points, which can form k The soft points corresponding to the group are generated, and the number of soft points generated is less than the point cloud to be registered. Then, the rotation parameters and translation parameters can be directly obtained from k The soft points corresponding to the group are estimated. Different from the existing methods, the present invention does not establish correspondence based on explicit or implicit feature matching, but automatically generates k By grouping soft points into corresponding pairs, the present invention eliminates the need for feature matching, thus avoiding the problem of numerous mismatches caused by feature matching in low-overlap areas. Furthermore, the present invention eliminates the need for outlier removal, significantly increasing the speed and shortening the point cloud registration time. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flowchart of a point cloud registration method provided by an embodiment of the present invention;

[0018] Figure 2 is an exemplary registration effect diagram;

[0019] Figure 3 This is a structural block diagram of the point cloud registration system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the embodiments of the present invention, and should not be construed as limiting the present invention.

[0021] See also Figure 1 The embodiment of the present invention provides a point cloud registration method, which is applied to a point cloud registration system. The point cloud registration system includes a feature extraction and interaction module, a soft point generation module, and a rotation and translation parameter estimation module. The core of this method is soft point registration. The point cloud registration system focuses on finding kThe set of reliable corresponding points is used to quickly estimate the rotation and translation parameters. Since explicit or implicit feature matching is based on the similarity of local features, but the local geometric structures of corresponding points are not completely consistent in low-overlap scenes, the present invention no longer uses the similarity of local features when finding corresponding points, but uses the attention mechanism to generate the source point cloud and target point cloud separately. k A soft spot, forming k Then, the rotation and translation parameters can be directly obtained from k The corresponding points in the group are estimated.

[0022] Specifically, the method includes steps S1 to S3:

[0023] In step S1, the feature extraction and interaction module takes the source point cloud and the target point cloud as input, linearly projects the source point cloud and the target point cloud into source point cloud features and target point cloud features respectively, and then transmits the source point cloud features and the target point cloud features to the Transformer model. The Transformer model includes 12 Transformer layers. The first 6 Transformer layers use self-attention layers, and the last 6 Transformer layers use cross-attention layers. The source point cloud features and the target point cloud features are processed by the self-attention layers and the cross-attention layers to obtain the source point cloud features after interaction and the target point cloud features after interaction.

[0024] Among them, the feature extraction and interaction modules are based on the source point cloud and target point cloud For input, , , represents the set of real numbers; is the total number of points in the source point cloud and the total number of points in the target point cloud. The self-attention layer and cross-attention layer in Transformer are used to extract and interact the features of the source point cloud and the target point cloud.

[0025] Among them, the self-attention layer is used to perceive the geometric structure of the source point cloud and the target point cloud and obtain point-wise features, and the cross-attention layer is used to allow the features of the source point cloud to interact with the features of the target point cloud so that they can perceive each other's geometric structure. Specifically, first and is linearly projected to d dimensional embedding to obtain the source point cloud features and target point cloud features , , , the present invention sets d =256. Then, and After being added to the position code, it is fed into the Transformer model. The Transformer model consists of 12 Transformer layers, each of which contains two sublayers: a multi-head attention layer and a point-wise feedforward network. The self-attention layer is used in the first 6 Transformer layers, which allows points to interact with other points in the same point cloud, thereby perceiving the geometric structures of the source point cloud and the target point cloud respectively. The cross-attention layer is used in the last 6 Transformer layers, which is used to interact with the features of the two point clouds, so that they exchange information and perceive each other's geometric structure. The multi-head attention operation in each Transformer layer is defined as follows:

[0026] ;

[0027] ;

[0028] in, represents the multi-head attention mechanism, 、 and They are query matrix, keyword matrix and content matrix respectively. 、 、 Respectively represent the first and h , H Size, represents splicing along the channel dimension, represents the weight matrix that can be learned through the network, 、 、 Respectively represent h The weight matrix corresponding to the query matrix, keyword matrix, and content matrix corresponding to each head.

[0029] , , The present invention sets H =8, d head = d / H =32.

[0030] In traditional single-head attention, and The dot product between them introduces a computational cost that grows quadratically with the length of the input sequence. When there are many points in the point cloud, it is impractical to directly apply the vanilla version of single-head attention. This paper uses the attention layer proposed in the linear Transformer to reduce the computational complexity:

[0031] ;

[0032] represents the attention mechanism, is the improved activation function, , , is the activation function, Indicates transpose.

[0033] In the first 6 Transformer layers, the self-attention layer uses the following formula to calculate the source point cloud features: and target point cloud features Processing:

[0034] ;

[0035] ;

[0036] in, is the source point cloud feature after processing by the self-attention layer, It is the target point cloud feature after processing by the self-attention layer;

[0037] In the last 6 Transformer layers, the cross attention layer is calculated as follows: and Processing:

[0038] ;

[0039] ;

[0040] in, is the source point cloud feature after interaction, is the target point cloud feature after interaction.

[0041] Step S2, in the soft point generation module, all points in the source point cloud feature after interaction are generated according to k The weight of the group source point cloud is weighted k The source point cloud soft point is composed of all points in the target point cloud feature after interaction according to k The weight of the group target point cloud is weighted to obtain k Target point cloud soft points, k The group source point cloud weight is predicted by the source point cloud features after interaction. k The group target point cloud weight is predicted by the target point cloud features after interaction.

[0042] The present invention uses and Generate for source point cloud and target point cloud respectively k A soft spot, Source point cloud soft points and The soft points of the target point cloud are recorded as and , , Each pair and They should have the ability to automatically find a corresponding part of the source point cloud and the target point cloud. As an example, the calculation process of The calculation principle is the same). It is obtained by weighting all points in the source point cloud, and its weight is recorded as ,Depend on Predicted, specifically Depend on It is mapped to 1D by MLP and obtained after softmax calculation. Has been from The information of is updated, so when assigning weights to each point in the source point cloud, it can automatically give and The local areas of common interest are given higher weights, while points with inconsistent local geometric structures are assigned low weights. The calculation formula is:

[0043] ;

[0044] in, Indicates the Source point cloud soft points, Indicates the first The weight of the point, Indicates the first point in the source point cloud Points, is the total number of points in the source point cloud.

[0045] In this embodiment, it is possible to calculate k Group weight and k Soft spots:

[0046] ;

[0047] ;

[0048] in, express k Group source point cloud weights, , is the normalization function, Indicates that the feature dimension is transformed from d Dimensionality reduction k Dimensional operations, express k Source point cloud soft points, , Represents the source point cloud.

[0049] Understandably, k Target point cloud soft points It can be calculated in the same way.

[0050] Step S3, in the rotation and translation parameter estimation module, the Kabsch-Umeyama algorithm is used to obtain k Source point cloud software and k The rotation and translation parameters are directly estimated from the soft points of the target point cloud by singular value decomposition, and then the point cloud registration is performed based on the rotation and translation parameters.

[0051] The rotation and translation parameter estimation module satisfies the following formula:

[0052] ;

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] in, express k Target point cloud soft points, Indicates the Target point cloud soft points, express The center of mass, express The center of mass, 、 、 Represent left singular vectors, singular values, and right singular vectors, respectively. represents the rotation parameter, Represents the translation parameter.

[0058] In this embodiment, the rotation and translation error is used as the loss function to minimize and The L1 error between Is the source point cloud The result obtained by transforming the estimated parameters into rotation and translation parameters is: Is the source point cloud The result obtained after transformation based on the true value of the rotation parameter and the true value of the translation parameter.

[0059] Specifically, the loss function of the point cloud registration system is for:

[0060] ;

[0061] in, Represents the source point cloud The The points are rotated according to the estimated rotation parameters and translation parameters The result after transformation is Represents the source point cloud The The result obtained by transforming the points according to the true value of the rotation parameter and the true value of the translation parameter.

[0062] The method presented in this paper was tested below. Commonly used point cloud registration evaluation metrics include Registration Recall (RR), Relative Rotation Errors (RRE), and Relative Translation Errors (RTE). The method presented in this paper was compared with other methods using the relatively low-level dataset 3DloMatch.The specific results are shown in Table 1. In Table 1, 3DSN (The Perfect Match: 3D Point Cloud Matching with Smoothed Densities) is a point cloud matching method based on smoothed density; FCGF (Feature Correspondence Graph for Fast Point Cloud Registration) is a point cloud matching method based on feature correspondence graph for fast point cloud registration; D3Feat (Joint Learning of Dense Detection and Description of 3D Local Features) is a point cloud matching method based on joint learning of dense detection and description of 3D local features; Predator (Registration of 3D Point Clouds with Low Overlap) is a 3D point cloud registration method under low overlap, with two different versions Predator-5k and Predator-1k, where 1k and 5k represent the number of key points; OMNet (Learning Overlapping Mask for Partial-to-Partial Point Cloud Registration) is an overlapping mask learning method for partial point cloud to partial point cloud registration; DGR (Deep Global Registration) is a deep global registration method; PCAM (Product of Cross-Attention Matrices for Rigid Registration of Point Clouds) is a point cloud rigid registration method based on the cross-attention matrix product; REGTR (End-to-end Point Cloud Correspondences with Transformers) is a registration method based on the end-to-end point cloud correspondence relationship of Transformer; Geo Transformer (Fast and robust point cloud registration with geometric transformer) is a fast and robust point cloud registration method based on Transformer.

[0063] Table 1 Comparison of results on the 3DloMatch dataset

[0064]

[0065] As can be seen from Table 1, on 3DLoMatch, the method of the present invention has the best performance in the three indicators of RR, RRE and RTE. Since the overlap of point cloud pairs in the 3DLoMatch dataset is very low, less than 30%, many corresponding point pairs in the overlapping area do not have similar geometric structures, which results in the inability of feature matching to produce correct point correspondences. The method of the present invention does not look for point correspondences based on feature matching, but uses soft points generated by weights automatically assigned by the attention mechanism to form correspondences. The weights automatically assigned by the network can automatically focus on similar local geometric structures and filter out inconsistent geometric structures. This greatly improves the robustness of correspondences in low-overlap areas, and makes the method of the present invention more advantageous in low-overlap registration datasets.

[0066] Table 2 compares the time required to complete the registration of two point clouds by the present invention and other methods. In Table 2, SpinNet (Learning a General Surface Descriptor for 3D Point Cloud Registration) is a 3D point cloud registration method based on learning a general surface descriptor.

[0067] Table 2 Time required to register two point clouds using different methods

[0068]

[0069] As can be seen in Table 2, the method of the present invention significantly accelerates the execution speed compared to the RANSAC method. Because the point correspondences formed by soft points are sufficiently reliable, the method of the present invention does not rely on RANSAC to eliminate false correspondences, but directly estimates the transformation parameters based on the soft point correspondences. Therefore, the method of the present invention is a RANSAC-free method and runs faster.

[0070] In addition, the method of the present invention is an end-to-end registration method, which only requires the input of the source point cloud and the target point cloud, and finally obtains the rotation and translation transformation from the source point cloud to the target point cloud. Figure 2 As shown in Figure 1, it is only necessary to input the source point cloud and the target point cloud with overlapping areas, and finally obtain the registered point cloud.

[0071] In summary, according to the point cloud registration method of the above embodiment, the source point cloud features and the target point cloud features are processed by the self-attention layer and the cross-attention layer to obtain the source point cloud features after interaction and the target point cloud features after interaction. When finding corresponding points, the present invention is no longer based on the similarity of local features, but is based on the source point cloud and the target point cloud. k Source point cloud software and k target point cloud soft points, which can form kThe soft points corresponding to the group are generated, and the number of soft points generated is less than the point cloud to be registered. Then, the rotation parameters and translation parameters can be directly obtained from k The soft points corresponding to the group are estimated. Different from the existing methods, the present invention does not establish correspondence based on explicit or implicit feature matching, but automatically generates k By grouping soft points into corresponding pairs, the present invention eliminates the need for feature matching, thus avoiding the problem of numerous mismatches caused by feature matching in low-overlap areas. Furthermore, the present invention eliminates the need for outlier removal, significantly increasing the speed and shortening the point cloud registration time.

[0072] See also Figure 3 ,Another embodiment of the present invention provides a point cloud registration system, comprising a feature extraction and interaction module, a soft point generation module, and a rotation and translation parameter estimation module;

[0073] The feature extraction and interaction module takes the source point cloud and the target point cloud as input, linearly projects the source point cloud and the target point cloud into source point cloud features and target point cloud features respectively, and then transmits the source point cloud features and the target point cloud features to the Transformer model. The Transformer model includes 12 Transformer layers. The first 6 Transformer layers use self-attention layers, and the last 6 Transformer layers use cross-attention layers. The source point cloud features and the target point cloud features are processed by the self-attention layers and cross-attention layers to obtain the source point cloud features and the target point cloud features after interaction.

[0074] In the soft point generation module, all points in the source point cloud feature after interaction are calculated based on k The weight of the group source point cloud is weighted k The source point cloud soft point is composed of all points in the target point cloud feature after interaction according to k The weight of the group target point cloud is weighted to obtain k Target point cloud soft points, k The group source point cloud weight is predicted by the source point cloud features after interaction. k The weight of the group target point cloud is predicted by the target point cloud features after interaction;

[0075] In the rotation and translation parameter estimation module, the Kabsch-Umeyama algorithm is used to obtain k Source point cloud software and k The rotation and translation parameters are directly estimated from the soft points of the target point cloud by singular value decomposition, and then the point cloud registration is performed based on the rotation and translation parameters.

[0076] In this embodiment, the Transformer model satisfies the following formula:

[0077] ;

[0078] ;

[0079] ;

[0080] in, represents the multi-head attention mechanism, 、 and They are query matrix, keyword matrix and content matrix respectively. 、 、 Respectively represent the first and h , H Size, represents splicing along the channel dimension, represents the weight matrix that can be learned through the network, 、 、 Respectively represent h The weight matrix corresponding to the query matrix, keyword matrix, and content matrix, represents the attention mechanism, is the improved activation function, , , is the activation function, represents transpose;

[0081] In the first 6 Transformer layers, the self-attention layer uses the following formula to calculate the source point cloud features: and target point cloud features Processing:

[0082] ;

[0083] ;

[0084] in, is the source point cloud feature after processing by the self-attention layer, It is the target point cloud feature after processing by the self-attention layer;

[0085] In the last 6 Transformer layers, the cross attention layer is calculated as follows: and Processing:

[0086] ;

[0087] ;

[0088] in, is the source point cloud feature after interaction, is the target point cloud feature after interaction.

[0089] In this embodiment, the soft point generation module satisfies the following formula:

[0090] ;

[0091] in, Indicates the Source point cloud soft points, Indicates the first The weight of the point, Indicates the first Points, is the total number of points in the source point cloud.

[0092] In this embodiment, the feature dimensions of the source point cloud features and the target point cloud features are both d dimension;

[0093] The soft point generation module also satisfies the following formula:

[0094] ;

[0095] ;

[0096] in, express k Group source point cloud weights, is the normalization function, Indicates that the feature dimension is transformed from d Dimensionality reduction k Dimensional operations, express k Source point cloud soft points, Represents the source point cloud.

[0097] In this embodiment, the rotation and translation parameter estimation module satisfies the following formula:

[0098] ;

[0099] ;

[0100] ;

[0101] ;

[0102] ;

[0103] in, express k Target point cloud soft points, Indicates the Target point cloud soft points, express The center of mass, express The center of mass, 、 、 Represent left singular vectors, singular values, and right singular vectors, respectively. represents the rotation parameter, Represents the translation parameter.

[0104] In this embodiment, the loss function of the point cloud registration system is for:

[0105] ;

[0106] in, Represents the source point cloud The The points are rotated according to the estimated rotation parameters and translation parameters The result after transformation is Represents the source point cloud The The result obtained by transforming the points according to the true value of the rotation parameter and the true value of the translation parameter.

[0107] In summary, according to the point cloud registration system of the above embodiment, the source point cloud features and the target point cloud features are processed by the self-attention layer and the cross-attention layer to obtain the source point cloud features after interaction and the target point cloud features after interaction. When finding corresponding points, the present invention is no longer based on the similarity of local features, but is based on the source point cloud and the target point cloud. k Source point cloud software and k target point cloud soft points, which can form k The soft points corresponding to the group are generated, and the number of soft points generated is less than the point cloud to be registered. Then, the rotation parameters and translation parameters can be directly obtained from k The soft points corresponding to the group are estimated. Different from the existing methods, the present invention does not establish correspondence based on explicit or implicit feature matching, but automatically generates k By grouping soft points into corresponding pairs, the present invention eliminates the need for feature matching, thus avoiding the problem of numerous mismatches caused by feature matching in low-overlap areas. Furthermore, the present invention eliminates the need for outlier removal, significantly increasing the speed and shortening the point cloud registration time.

[0108] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A point cloud registration method, characterized in that: Applied to a point cloud registration system, the point cloud registration system includes a feature extraction and interaction module, a soft point generation module, and a rotation and translation parameter estimation module, and the method includes: Step S1: The feature extraction and interaction module takes the source point cloud and the target point cloud as input, linearly projects the source point cloud and the target point cloud into source point cloud features and target point cloud features respectively, and then transmits the source point cloud features and the target point cloud features to the Transformer model. The Transformer model includes 12 Transformer layers. The first 6 Transformer layers use self-attention layers, and the last 6 Transformer layers use cross-attention layers. The source point cloud features and the target point cloud features are processed by the self-attention layers and the cross-attention layers to obtain the source point cloud features and the target point cloud features after interaction; Step S2, in the soft point generation module, all points in the source point cloud feature after interaction are generated according to k The weight of the group source point cloud is weighted k The source point cloud soft point is composed of all points in the target point cloud feature after interaction according to k The weight of the group target point cloud is weighted to obtain k Target point cloud soft points, k The group source point cloud weight is predicted by the source point cloud features after interaction. k The weight of the group target point cloud is predicted by the target point cloud features after interaction; Step S3, in the rotation and translation parameter estimation module, the Kabsch-Umeyama algorithm is used to obtain k Source point cloud software and k The rotation and translation parameters are directly estimated from the soft points of the target point cloud by singular value decomposition, and then the point cloud registration is performed based on the rotation and translation parameters.

2. The point cloud registration method according to claim 1, characterized in that: The Transformer model satisfies the following formula: ； ； ； in, represents the multi-head attention mechanism, 、 and They are query matrix, keyword matrix and content matrix respectively. 、 、 Respectively represent the first and h , H Size, represents splicing along the channel dimension, represents the weight matrix that can be learned through the network, 、 、 Respectively represent h The weight matrix corresponding to the query matrix, keyword matrix, and content matrix, represents the attention mechanism, is the improved activation function, , , is the activation function, represents transpose; In the first 6 Transformer layers, the self-attention layer uses the following formula to calculate the source point cloud features: and target point cloud features Processing: ； ； in, is the source point cloud feature after processing by the self-attention layer, It is the target point cloud feature after processing by the self-attention layer; In the last 6 Transformer layers, the cross attention layer is calculated as follows: and Processing: ； ； in, is the source point cloud feature after interaction, is the target point cloud feature after interaction.

3. The point cloud registration method according to claim 2, characterized in that: The soft point generation module satisfies the following formula: ； in, Indicates the Source point cloud soft points, Indicates the first The weight of the point, Indicates the first point in the source point cloud Points, is the total number of points in the source point cloud.

4. The point cloud registration method according to claim 3, characterized in that: The feature dimensions of the source point cloud features and the target point cloud features are d dimension; The soft point generation module also satisfies the following formula: ； ； in, express k Group source point cloud weights, is the normalization function, Indicates that the feature dimension is transformed from d Dimensionality reduction k Dimensional operations, express k Source point cloud soft points, Represents the source point cloud.

5. The point cloud registration method according to claim 4, characterized in that: The rotation and translation parameter estimation module satisfies the following formula: ； ； ； ； ； in, express k Target point cloud soft points, Indicates the Target point cloud soft points, express The center of mass, express The center of mass, 、 、 Represent left singular vectors, singular values, and right singular vectors, respectively. represents the rotation parameter, Represents the translation parameter.

6. The point cloud registration method according to claim 5, characterized in that: The loss function of the point cloud registration system for: ； in, Represents the source point cloud The The points are rotated according to the estimated rotation parameters and translation parameters The result after transformation is Represents the source point cloud The The result obtained by transforming the points according to the true value of the rotation parameter and the true value of the translation parameter.

7. A point cloud registration system, characterized in that: It includes feature extraction and interaction module, soft point generation module, and rotation and translation parameter estimation module; The feature extraction and interaction module takes the source point cloud and the target point cloud as input, linearly projects the source point cloud and the target point cloud into source point cloud features and target point cloud features respectively, and then transmits the source point cloud features and the target point cloud features to the Transformer model. The Transformer model includes 12 Transformer layers. The first 6 Transformer layers use self-attention layers, and the last 6 Transformer layers use cross-attention layers. The source point cloud features and the target point cloud features are processed by the self-attention layers and cross-attention layers to obtain the source point cloud features and the target point cloud features after interaction. In the soft point generation module, all points in the source point cloud feature after interaction are calculated based on k The weight of the group source point cloud is weighted k The source point cloud soft point is composed of all points in the target point cloud feature after interaction according to k The weight of the group target point cloud is weighted to obtain k Target point cloud soft points, k The group source point cloud weight is predicted by the source point cloud features after interaction. k The weight of the group target point cloud is predicted by the target point cloud features after interaction; In the rotation and translation parameter estimation module, the Kabsch-Umeyama algorithm is used to obtain k Source point cloud software and k The rotation and translation parameters are directly estimated from the soft points of the target point cloud by singular value decomposition, and then the point cloud registration is performed based on the rotation and translation parameters.

8. The point cloud registration system according to claim 7, characterized in that: The Transformer model satisfies the following formula: ； ； ； in, represents the multi-head attention mechanism, 、 and They are query matrix, keyword matrix and content matrix respectively. 、 、 Respectively represent the first and h , H Size, represents splicing along the channel dimension, represents the weight matrix that can be learned through the network, 、 、 Respectively represent h The weight matrix corresponding to the query matrix, keyword matrix, and content matrix, represents the attention mechanism, is the improved activation function, , , is the activation function, represents transpose; In the first 6 Transformer layers, the self-attention layer uses the following formula to calculate the source point cloud features: and target point cloud features Processing: ； ； in, is the source point cloud feature after processing by the self-attention layer, It is the target point cloud feature after processing by the self-attention layer; In the last 6 Transformer layers, the cross attention layer is calculated as follows: and Processing: ； ； in, is the source point cloud feature after interaction, is the target point cloud feature after interaction.

9. The point cloud registration system according to claim 8, characterized in that: The soft point generation module satisfies the following formula: ； in, Indicates the Source point cloud soft points, Indicates the first The weight of the point, Indicates the first point in the source point cloud Points, is the total number of points in the source point cloud.

10. The point cloud registration system according to claim 9, characterized in that: The feature dimensions of the source point cloud features and the target point cloud features are d dimension; The soft point generation module also satisfies the following formula: ； ； in, express k Group source point cloud weights, is the normalization function, Indicates that the feature dimension is transformed from d Dimensionality reduction k Dimensional operations, express k Source point cloud soft points, Represents the source point cloud.

Images