Point pair convolution-based low-overlap three-dimensional point cloud registration method and system

Abstract

The application relates to a low-overlap point cloud registration method and device based on point pair feature convolution. First, an encoder capable of extracting rotation-invariant features is proposed, and the performance of the network under large rotation is enhanced. Secondly, a method for aggregating global features is proposed, which improves the limitation that the performance is affected by the receptive field size of the convolution layer, and makes the network have stronger ability to identify the overlapping part of the point cloud. Through the combination of the two, a low-overlap point cloud registration method based on point pair feature convolution is researched and explored.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a low-overlap point cloud registration method and system based on point-pair feature convolution. Background Art

[0002] Point cloud registration involves aligning two or more 3D point clouds to determine the optimal correspondence between them. By identifying and matching common features between different point cloud collections, point cloud registration enables the integration and fusion of spatial data. Therefore, point cloud registration technology is of great significance for a variety of applications, including 3D reconstruction, autonomous driving, and cultural heritage preservation.

[0003] With the development of deep learning technology, deep learning-based methods have been widely used in point cloud registration research. Leveraging the powerful fitting capabilities of neural networks, these methods have achieved promising results. Existing deep learning-based point cloud registration techniques mostly employ feature learning-based approaches. These methods employ deep neural networks to learn robust feature correspondence search. They then estimate the transformation matrix between the two point clouds based on corresponding point pairs. This method can learn deep point cloud features and is highly robust and interpretable. However, current feature learning-based point cloud registration networks also suffer from the following issues.

[0004] First, most point cloud descriptors are designed for specific poses or viewpoints, which means they are sensitive to rotation changes. When there are large rotations between point clouds, these descriptors may not be able to effectively match corresponding points because their feature representations will change significantly due to the rotation. This sensitivity to rotation will lead to degraded network performance in the presence of large rotations.

[0005] Furthermore, when using convolutional neural networks to process point cloud data, the receptive field size of the convolutional layer significantly impacts its performance. A limited receptive field means that the network can only consider information from a limited area surrounding each point when processing it. This locality can limit the network's accuracy in identifying overlapping areas within the point cloud, particularly when the point cloud has small overlapping areas or unevenly distributed features. Summary of the Invention

[0006] To improve the accuracy of point cloud registration, particularly the network's ability to extract rotationally invariant features and predict the overlap of low-overlap point cloud pairs, this paper proposes a low-overlap point cloud registration method and system based on point-pair feature convolution. The invented method takes as input a source point cloud X and a target point cloud Y, outputs point-by-point feature descriptors for X and Y, and finally matches feature points and calculates transformation parameters.

[0007] The low-overlap point cloud registration method based on point-pair feature convolution of the present invention comprises the following steps:

[0008] Step S1: Construct an encoder layer based on point kernel convolution to downsample the input source point cloud and the target point cloud to obtain a lower density point cloud and convolved point cloud features;

[0009] Step S2: Construct a geometry-guided coding layer, which includes two branches: geometric coding and semantic coding. The geometric coding branch takes the original point cloud and the downsampled point cloud as input and outputs the geometric features of the downsampled point cloud. The semantic coding branch takes the downsampled point cloud and the convolutional point cloud features as input and outputs the semantic features of the downsampled point cloud. The obtained geometric features and semantic features are input into the fully connected layer to obtain the source point cloud and target point cloud features as the final output of this layer.

[0010] Step S3: Construct a global aggregation layer, which is used to aggregate the global features of the source point cloud and the target point cloud. The input of the global aggregation layer is the source point cloud and target point cloud features output by S2, and the output is the features of the downsampled point cloud after global aggregation.

[0011] Step S4: Construct a decoder layer based on nearest neighbor upsampling to restore the downsampled point cloud features to the dense features of the original point cloud, and pass the encoder features to the decoder through skip connection; the input of this layer is the features of the downsampled point cloud output by S1, and the output is the point-by-point feature descriptor of the original point cloud;

[0012] Step S5: Perform feature matching through the point-by-point feature descriptors of the original point cloud to obtain the matching point set between the source point cloud and the target point cloud, and use post-processing methods such as RANSAC to calculate the transformation parameters through the matching point set.

[0013] Step S6: Use the data set with the real matching point set labels to train the deep learning network constructed by S1-S5;

[0014] Step S7: Use the network trained in step S6 to perform point cloud registration;

[0015] Furthermore, the step S1 adopts the improved point kernel convolution module to extract the rotation invariant features of the point cloud; the specific process of S1 is:

[0016] S1-1, the convolution kernel is a spherical area, and each kernel point has a weight W k :

[0017]

[0018] in, represents the kernel point, r represents the radius of the convolution kernel;

[0019] S1-2, for each area that needs convolution, let the point obtained after convolution be x0, N xis the set of points in the convolution area, f is the feature of the points in the convolution area, PPF(x i ,x0) is x i And the point pair features of x0, the convolution process is:

[0020]

[0021]

[0022] The input original point cloud X is output through step S1 as the downsampled point cloud X' and feature F X’ , the points in X' are called super points, y i =x i -x0, mlp is the multi-layer perceptron layer, PPF represents point pair features, g and h are intermediate variables.

[0023] Furthermore, the specific process of step S2 is as follows:

[0024] S2-1, input the original point cloud X and the downsampled point cloud X' into the geometry encoding branch in the geometry guidance layer, and use the K-nearest neighbor method to estimate the normal of the downsampled point cloud. That is, the normal direction of a point in the downsampled point cloud is the average of the normal directions of its k neighboring points in the original point cloud;

[0025] S2-2, select k neighbor points for each super point in X' and construct PPF features, N x’ is the normal of the point:

[0026]

[0027] Combine the PPF features with x i ' and x i '-x ij 'For splicing, each point can obtain features with a dimension of k×10;

[0028] S2-3, input the features obtained in S2-2 into the one-dimensional convolution layer and the maximum pooling layer to obtain the geometric features G output by the geometric coding branch X’ ;

[0029] S2-4, downsample the point cloud X' and feature F X’ Input the semantic encoding branch in the geometric guidance layer, pass through the KNN graph neural network and input it into the fully connected layer to obtain the semantic feature H output by the semantic encoding branch. X’ :

[0030]

[0031] S2-5, concatenates the features obtained from S2-3 and S2-4, inputs them into the fully connected layer, and obtains the geometric and semantically enhanced point cloud features output by the geometry-guided encoding layer

[0032]

[0033] Furthermore, the specific process of step 3 is as follows:

[0034] S3-1, input the enhanced features of the source point cloud and target point cloud output in step 2 into the global aggregation layer respectively, and after point-by-point maximum pooling, obtain two 256-dimensional feature vectors respectively;

[0035] S3-2, repeat and concatenate the two feature vectors obtained in the previous step to obtain features with dimensions of N'×1024 and M'×1024 respectively. These two feature matrices are input into the fully connected layer to obtain super-point features after aggregating global features, with dimensions of N'×256 and M'×256.

[0036] Furthermore, the decoder in step S4 is used to generate a dense feature descriptor of the original point cloud; the super-point features obtained in step S1 are fed into the decoder, and the decoder part uses nearest neighbor upsampling to obtain the final point-by-point feature descriptor; jumps are used to pass the encoder features to the decoder; these features are connected to the upsampled features and processed by MLP; the output of the decoder is N×256 and M×256 dense original point features.

[0037] Furthermore, the specific process of step S5 is as follows:

[0038] S5-1, calculate the Gaussian correlation between the super-point features in point clouds X' and Y':

[0039]

[0040] Then, through the bidirectional normalization operation, the ambiguous matching is further suppressed. X' is the downsampled point cloud of the original point cloud, Y' is the number of super points of the target point cloud, and H i is the super-point feature of the original point cloud, H j is the super-point feature of the target point cloud;

[0041] S5-2, according to S ij Select the most reliable P pairs of superpoints; for each pair of superpoints, perform fine-grained matching between the two point sets around the superpoint, and finally obtain the matching point set of the original point cloud;

[0042] S5-3, the transformation parameters R,t,are calculated using the random sampling consensus algorithm RANSAC from the P matching point sets obtained in the previous step, and P groups R,t,can be obtained as candidates;

[0043] In S5-4, we use the probability sampling method to select the key points of the source point cloud and the target point cloud according to the matching factor, and use RANSAC to estimate a set of global R,t; a total of P+1 sets of R,t are obtained;

[0044] S5-5, select the optimal R,t as the final result.

[0045] Furthermore, the specific process of step S6 is as follows:

[0046] S6-1, uses H samples for training, each sample contains two point clouds, the source point cloud and the target point cloud, and their corresponding real matching point set and real transformation parameters;

[0047] S6-2, calculate the feature loss for the point cloud features output by S4; let the overlapping part of point cloud X and point cloud Y be X overlap and Y overlap ; For point x in point cloud X, ε x Y overlap The corresponding point set in is a positive sample; y Y overlap With ε x The difference between them is a negative sample; is the Euclidean distance between point features, Δ p , Δ n is the distance between positive samples and negative samples, Different weights are defined for each positive and negative sample; the feature loss is calculated using the ring loss:

[0048]

[0049]

[0050] S6-3, calculate the point matching loss for the super point matching point set obtained in S5. For each super point correspondence C i * , a soft assignment matrix Z is calculated through the optimal transport layer i * , represents the confidence of the point correspondence; random sampling N g Set the real super-point correspondence, set M i is the set of real point correspondences under a given superpoint correspondence, I i and J i Calculate the negative log-likelihood loss for the set of points that do not match in point cloud X' and point cloud Y':

[0051]

[0052] S6-5, add the above two losses to calculate the total network loss:

[0053] L=L c +L p

[0054] Based on the same inventive concept, the present invention also provides an electronic device, which is special in that it includes:

[0055] one or more processors;

[0056] a storage device for storing one or more programs;

[0057] When one or more programs are executed by the one or more processors, the one or more processors implement the low-overlap point cloud registration method based on point-pair feature convolution as described above.

[0058] Based on the same inventive concept, the present invention also designs a computer-readable medium on which a computer program is stored. The special feature of the computer-readable medium is that when the program is executed by a processor, the low-overlap point cloud registration method based on point-pair feature convolution as described above is implemented.

[0059] Advantages of the present invention:

[0060] 1. This invention innovatively changes the convolution method of point kernel convolution, modifying the point coordinate features that extract rotational variability to point pair features that extract rotational invariance, thereby improving the performance of the network in scenarios with large angle rotation.

[0061] 2. This paper innovatively proposes a geometry-guided coding layer, which fuses geometric and semantic features to increase the proportion of internal corresponding points. In the geometric coding branch, point-pair features of K nearest neighbors are innovatively constructed to improve the network's robustness to rotation. In the semantic coding branch, a KNN graph neural network is innovatively used. Furthermore, a normal smoothing method is proposed, which makes the calculation of superpoint normal directions more accurate.

[0062] 3. This invention innovatively introduces a global aggregation layer, which improves the prediction accuracy of overlapping parts of point clouds. Due to the limited receptive field of convolution, the convolutional features in the encoder alone may not be sufficient to accurately predict overlapping parts. The global aggregation module effectively improves prediction accuracy by introducing global features into the prediction process. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 2 is a diagram of the encoder-decoder layer neural network architecture of the present invention.

[0064] Figure 2 This is a diagram of the neural network architecture of the geometrically guided coding layer of the present invention.

[0065] Figure 3 This is a diagram of the global aggregation layer neural network architecture of the present invention. DETAILED DESCRIPTION

[0066] In order to facilitate ordinary technicians in this field to understand and implement the present invention, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0067] Example 1

[0068] The method provided by the present invention designs a low-overlap point cloud registration method based on point-pair feature convolution, which includes the following steps:

[0069] In step S1, the original point cloud is first downsampled using voxel downsampling. After downsampling three times, four point clouds with different densities are obtained, and the density of the point clouds decreases in sequence. The improved point kernel convolution module is used to convolve the input original point cloud three times. The point cloud obtained after each convolution is the same as the point cloud of the corresponding stage of voxel downsampling. The convolution radius is 2.5 times the voxel downsampling grid unit size, and the convolution kernel radius is 1.5 times the voxel downsampling grid unit size. The specific implementation process is described as follows:

[0070] S1-1, the convolution kernel is a spherical area, where K = 15 kernel points are preset, and each kernel point has a weight W k :

[0071]

[0072] S1-2, for each area that needs convolution, its center point is x0, PPF(x i ,x0) is x i And the point pair features of x0, the convolution process is:

[0073]

[0074] The input original point cloud X is output through step S1 as the downsampled point cloud X' and feature F X’ , and the points in X' are called super points.

[0075] Step S2: Construct a geometry-guided coding layer, which contains two branches: geometric coding and semantic coding. The geometric coding branch takes the original point cloud and the downsampled point cloud as input, and outputs the geometric features of the downsampled point cloud. The semantic coding branch takes the downsampled point cloud and the convolutional point cloud features as input, and outputs the semantic features of the downsampled point cloud. The obtained geometric features and semantic features are input into the fully connected layer to obtain the source point cloud and target point cloud features as the final output of this layer. The specific implementation process is described as follows:

[0076] S2-1, input the original point cloud X and the downsampled point cloud X' into the geometry encoding branch in the geometry guidance layer, and use the K-nearest neighbor method to estimate the normal of the downsampled point cloud. That is, the normal direction of a point in the downsampled point cloud is the average of the normal directions of its k neighboring points in the original point cloud;

[0077] S2-2, select k neighbor points for each super point in X' and construct PPF features, N x’ is the normal of the point:

[0078]

[0079] Combine the PPF features with x i ' and x i '-x ij 'For each point, two 3D coordinates are concatenated with a 4D point pair feature to obtain a feature with a dimension of k×10;

[0080] S2-3, input the features obtained in S2-2 into the one-dimensional convolution layer and the maximum pooling layer to obtain the geometric features G output by the geometric coding branch X’ ;

[0081] S2-4, downsample the point cloud X' and feature F X’ Input the semantic encoding branch in the geometric guidance layer, pass through the KNN graph neural network and input it into the fully connected layer to obtain the semantic feature H output by the semantic encoding branch. X’ :

[0082]

[0083] S2-5, concatenates the features obtained from S2-3 and S2-4, inputs them into the fully connected layer, and obtains the geometric and semantically enhanced point cloud features F output by the geometry-guided encoding layer X’ gge :

[0084]

[0085] Step S3: construct a global aggregation layer, which is used to aggregate the global features of the source point cloud and the target point cloud. The global aggregation layer inputs the source point cloud and the target point cloud features output by S2, and outputs the point-by-point feature descriptors of the downsampled point cloud. The specific implementation process is described as follows:

[0086] S3-1, input the enhanced features of the source point cloud and target point cloud output in step 2 into the global aggregation layer respectively, and after point-by-point maximum pooling, obtain two 256-dimensional feature vectors respectively;

[0087] S3-2, repeat and concatenate the two feature vectors obtained in the previous step to obtain features with dimensions of N'×1024 and M'×1024 respectively. These two feature matrices are input into the fully connected layer to obtain super-point features after aggregating global features, with dimensions of N'×256 and M'×256.

[0088] Step S4 constructs the decoder layer. The decoder generates dense feature descriptors for the original point cloud. The super-point features obtained in step S1 are fed into the decoder, which uses nearest neighbor upsampling to obtain the final point-by-point feature descriptors. Skip connections are used to pass encoder features to the decoder. These features are concatenated with the upsampled features and processed by the MLP. The decoder outputs are N×256 and M×256 dense original point features. The specific implementation process is described as follows:

[0089] In S4, the downsampled point cloud is restored to its original density using nearest neighbor upsampling, and the features of the corresponding density point cloud are obtained. After three upsampling steps, the dense features of the original point cloud are obtained.

[0090] In step S5, feature matching is performed using the point-by-point feature descriptors of the original point cloud to obtain the matching point set between the source point cloud and the target point cloud, and the transformation parameters are calculated using post-processing methods such as RANSAC. The input of this step is the matching point set of the original source point cloud and the target point cloud, and the output is the transformation parameters R,t from the source point cloud to the target point cloud. The specific implementation process is described as follows:

[0091] S5-1, calculate the Gaussian correlation between the super-point features in point clouds X' and Y':

[0092]

[0093] Then, through the bidirectional normalization operation, ambiguous matches are further suppressed;

[0094] S5-2, according to S ij Select the most reliable P pairs of superpoints; for each pair of superpoints, perform fine-grained matching between the two point sets around the superpoint, and finally obtain the matching point set of the original point cloud;

[0095] S5-3, the transformation parameters R,t,are calculated using the random sampling consensus algorithm RANSAC from the P matching point sets obtained in the previous step, and P groups R,t,can be obtained as candidates;

[0096] In S5-4, we use the probability sampling method to select the key points of the source point cloud and the target point cloud according to the matching factor, and use RANSAC to estimate a set of global R,t; a total of P+1 sets of R,t are obtained;

[0097] S5-5, select the optimal R,t as the final result.

[0098] Step S6: Use the point cloud registration datasets 3DMatch and KITTI to train the network constructed in steps S1-S5. The 3DMatch dataset is trained for 40 cycles, and the KITTI dataset is trained for 40 cycles. The specific implementation process is described as follows:

[0099] S6-1, uses H samples for training, each sample contains two point clouds, the source point cloud and the target point cloud, and their corresponding real matching point set and real transformation parameters;

[0100] S6-2, calculate the feature loss for the point cloud features output by S4; let the overlapping part of point cloud X and point cloud Y be X overlap and Y overlap ; For point x in point cloud X, ε x Y overlap The corresponding point set in is a positive sample; y Y overlap With ε x The difference between them is a negative sample; is the Euclidean distance between point features, Δ p , Δ n is the distance between positive samples and negative samples, Different weights are defined for each positive sample and negative sample; circle loss is used to calculate feature loss:

[0101]

[0102] S6-3, calculate the point matching loss for the super point matching point set obtained in S5. For each super point correspondence C i * , a soft assignment matrix Z is calculated through the optimal transport layer i * , represents the confidence of the point correspondence; random sampling N g Set the real super-point correspondence, set M i is the set of real point correspondences under a given superpoint correspondence, I i and J i Calculate the negative log-likelihood loss for the set of points that do not match in point cloud X' and point cloud Y':

[0103]

[0104] S6-5, add the above two losses to calculate the total network loss:

[0105] L=L c +L p

[0106] In step S7, the model trained in step S6 is used to perform point cloud registration. First, the source point cloud and the target point cloud are input, and the model outputs the matching point set of the source point cloud and the target point cloud. RANSAC is then used for post-processing to obtain the point cloud registration result.

[0107] This method constructs and trains a point cloud registration neural network, which improves the network's robustness to rotation; and improves the accuracy of network feature matching by extracting rotation-invariant features and aggregating global features.

[0108] Example 2

[0109] Based on the same inventive concept, the present invention also provides an electronic device, comprising one or more processors; a storage device for storing one or more programs; when one or more programs are executed by the one or more processors, the one or more processors implement the method described in Example 1.

[0110] Since the device described in the second embodiment of the present invention is an electronic device used to implement the low-overlap point cloud registration method based on point-pair feature convolution in the first embodiment of the present invention, those skilled in the art will be able to understand the specific structure and deformation of the electronic device based on the method described in the first embodiment of the present invention, and therefore will not be described in detail here. All electronic devices used in a method of the embodiment of the present invention fall within the scope of protection of the present invention.

[0111] Example 3

[0112] Based on the same inventive concept, the present invention further provides a computer-readable medium having a computer program stored thereon, which implements the method described in the first embodiment when the program is executed by a processor.

[0113] Since the device described in Example 3 of the present invention is a computer-readable medium used to implement the low-overlap point cloud registration method based on point-pair feature convolution in Example 1 of the present invention, those skilled in the art will be able to understand the specific structure and deformation of the electronic device based on the method described in Example 1 of the present invention, and therefore will not be described in detail here. All electronic devices used in a method of Example 1 of the present invention fall within the scope of protection of the present invention.

[0114] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Persons skilled in the art may make various modifications, additions, or substitutions to the described specific embodiments without departing from the spirit of the present invention or exceeding the scope of the appended claims.

Claims

1. A low-overlap 3D point cloud registration method based on point-to-point convolution, characterized in that: The following steps are involved: Step S1: Construct an encoder layer based on point kernel convolution to downsample the input source point cloud and the target point cloud to obtain a lower density point cloud and convolved point cloud features; Step S2 includes: constructing a geometry-guided coding layer, which includes two branches: geometric coding and semantic coding; the geometric coding branch inputs the original point cloud and the downsampled point cloud, and outputs the geometric features of the downsampled point cloud; the semantic coding branch inputs the downsampled point cloud and the convolution point cloud features, and outputs the semantic features of the point cloud; the obtained geometric features and semantic features are input into the fully connected layer to obtain the final output of the layer, the source point cloud and the target point cloud r fusion features, which specifically include: S2-1, input the original point cloud X and the downsampled point cloud X' into the geometry encoding branch in the geometry guidance layer, and use the K nearest neighbor method to estimate the normal of the downsampled point cloud; S2-2, select k neighbor points for each super point in X' and construct PPF point pair features, N x′ is the normal of the point: The PPF characteristics are compared with and - Splicing to obtain multi-dimensional features, for Neighbor points; S2-3, input the features obtained in S2-2 into the one-dimensional convolution layer and the maximum pooling layer to obtain the geometric features G output by the geometric coding branch X’ ; S2-4, the original point cloud is output through step S1 to obtain the downsampled point cloud X′ and feature F X′ Input the semantic encoding branch in the geometric guidance layer, pass through the KNN graph neural network and input it into the fully connected layer to obtain the semantic feature H output by the semantic encoding branch. X′ : S2-5, concatenates the features obtained from S2-3 and S2-4, inputs them into the fully connected layer, and obtains the geometric and semantically enhanced point cloud features F output by the geometry-guided encoding layer X′ gge : Among them, MLP is a multi-layer perceptron layer, and cat is feature concatenation; Step S3: Construct a global aggregation layer, which is used to aggregate the global features of the source point cloud and the target point cloud; the input of the global aggregation layer is the fusion feature of step S2, and the output is the feature of the point cloud after global aggregation; Step S4: Construct a decoder layer based on nearest neighbor upsampling to restore the features aggregated in step S3 to the dense features of the original point cloud, and pass the encoder features to the decoder through skip connection; the input of this layer is the features aggregated in S3, and the output is the point-by-point feature descriptor of the original point cloud; Step S5: performing feature matching using the point-by-point feature descriptors of the original point cloud to obtain a matching point set between the source point cloud and the target point cloud, and using a post-processing method to calculate the transformation parameters using the matching point set; Step S6: Use the data set with the real matching point set labels to train the deep learning network constructed by S1-S5; Step S7: Use the network trained in step S6 to perform point cloud registration.

2. The low-overlap 3D point cloud registration method based on point-pair convolution according to claim 1, characterized in that: The specific process of constructing the encoder layer based on point kernel convolution in step S1 is as follows: S1-1, the convolution kernel is a spherical area, and each kernel point has a weight W k : in, represents the core point, r Represents the radius of the convolution kernel; S1-2, for each area that needs convolution, let the point obtained after convolution be x 0 ，N x is the set of points in the convolution area, f is the feature of the point in the convolution area, PPF( x i , x 0) x i and x 0 point pair features, y i =x i -x 0, x 0 is the center point of the convolution area, that is, the point obtained after convolution, x i For other points in the area, the convolution process is: The input original point cloud X is output through step S1 as the downsampled point cloud X' and feature F X′ , the points in X′ are called super points, g and h is an intermediate variable, mlp It is a multi-layer perceptron layer.

3. The low-overlap 3D point cloud registration method based on point-pair convolution according to claim 1, characterized in that: The specific process of step 3 is as follows: S3-1, input the enhanced features of the source point cloud and target point cloud output in step 2 into the global aggregation layer respectively, and obtain two feature vectors after point-by-point maximum pooling; S3-2, repeat and concatenate the two feature vectors obtained in the previous step, input the two concatenated feature matrices into the fully connected layer to obtain the super-point features after aggregating the global features.

4. The low-overlap 3D point cloud registration method based on point-pair convolution according to claim 3, characterized in that: The specific process of step S4 is as follows: the decoder is used to generate a dense feature descriptor of the original point cloud; the super-point features obtained in step S1 are fed into the decoder, and the decoder part uses nearest neighbor upsampling to obtain the final point-by-point feature descriptor; the jump connection is used to pass the encoder features to the decoder; These features are concatenated to the upsampled features and processed by MLP; the output of the decoder is dense raw point features.

5. The low-overlap 3D point cloud registration method based on point-pair convolution according to claim 1, characterized in that: The specific process of step S5 is as follows: S5-1, calculate the Gaussian correlation between the super-point features in point clouds X′ and Y′: Then, through the bidirectional normalization operation, the ambiguous matching is further suppressed. X′ is the downsampled point cloud of the original point cloud, and Y′ is the number of super points of the target point cloud. H i is the super-point feature of the original point cloud, H j is the super-point feature of the target point cloud; S5-2, according to S ij Select the most reliable P pair superpoint pairing; for For each pair of superpoints, fine-grained matching is performed between the two point sets around the superpoint, and finally the matching point set of the original point cloud is obtained; S5-3, the transformation parameters R,t,are calculated using the random sampling consensus algorithm from the P matching point sets obtained in the previous step, and P groups R,t,are obtained as candidates; S5-4, use the probabilistic sampling method to select the key points of the source point cloud and the target point cloud according to the matching factor, and use the random sampling consensus algorithm to estimate a set of global R,t; a total of P+1 sets of R,t are obtained; S5-5, select the optimal R,t as the final result.

6. The low-overlap 3D point cloud registration method based on point-pair convolution according to claim 1, characterized in that: The specific process of step S6 is as follows: S6-1, uses H samples for training, each sample contains two point clouds, the source point cloud and the target point cloud, and their corresponding real matching point set and real transformation parameters; S6-2, calculate the feature loss for the point cloud features output by S4; let the overlapping part of point cloud X and point cloud Y be X overlap and Y overlap ; For point x in point cloud X, Y overlap The corresponding point set in is a positive sample; Y overlap and The difference between them is a negative sample; is the Euclidean distance between point features, is the distance between positive samples and negative samples, Different weights are defined for each positive and negative sample; the feature loss is calculated using the ring loss: S6-3, calculate the point matching loss for the super point matching point set obtained in S5, for each super point correspondence C i * , a soft assignment matrix Z is calculated through the optimal transmission layer i * , represents the confidence of the point correspondence; random sampling N g The correspondence relationship between the real super points is M i is the set of real point correspondences under a given super-point correspondence, I i and J i For the set of points that do not match in point cloud X′ and point cloud Y′, calculate the negative log-likelihood loss: S6-5, add the above two losses to calculate the total network loss: 。 7. An electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 6.

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

Images