A low-overlap 3D dynamic point cloud registration method and system based on attention mechanism
Through the graph neural network based on attention mechanism extraction and sampling key points, combined with local geometric and distance characteristics, the problem of low registration performance of low overlap point clouds is solved, achieving higher registration accuracy and performance improvement.
Patent Information
- Application Number
- CN202111547748.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-16
- Publication Date
- 2024-12-13
- Estimated Expiration
- 2041-12-16
AI Technical Summary
Under low overlap, the registration performance of traditional point cloud registration methods and deep learning-based methods is degraded, making it difficult to effectively improve the registration accuracy of point clouds.
The graph neural network based on attention mechanism extracts local geometric features, combines the similarity scores across the attention network to sample key points, combines the local geometric features and distance features of the key points to build a mixed feature similarity matrix, and decomposes them through SVD singular values. Calculate the registration matrix and registration vector for point cloud registration.
It effectively improves the accuracy and performance of point cloud registration, avoids local optimal traps, and improves registration efficiency in low overlap scenarios.
Smart Images
Figure CN114332175B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of three-dimensional point cloud registration, and more specifically, to a low-overlap 3D dynamic point cloud registration method and system based on an attention mechanism. Background Art
[0002] Point cloud registration is an important task in computer vision. Its goal is to find a rigid transformation to align a 3D source point cloud with another 3D target point cloud. It has a wide range of applications in pose estimation, 3D reconstruction, target positioning and other fields. The most widely used traditional registration method is the iterative closest point algorithm ICP (Iterative Closest Point), but this method is prone to fall into local optimality. In order to improve the shortcomings of the ICP algorithm, deep learning-based methods have been widely used. However, when the overlap rate of point clouds is low, the registration performance of both traditional registration methods and deep learning-based methods will decrease. In the case of low overlap (i.e., the overlap ratio is less than 1 / 3), how to improve the registration performance of point clouds is an urgent problem to be solved.
[0003] Traditional point cloud registration methods and deep learning-based registration methods are mostly performed in high-overlap scenarios. When in low-overlap scenarios, the registration performance of both traditional registration methods and deep learning-based methods needs to be improved. For two partially overlapping point clouds, if feature points can be sampled in the overlapping parts of the two point clouds, the point cloud registration performance can be effectively improved. Therefore, the present invention proposes a key point sampling method for overlapping areas based on an attention mechanism, and combines the local geometric features and distance features of the key points in the iterative matching process to construct an end-to-end point cloud registration network, thereby improving the accuracy of registration.
[0004] The existing registration method for indoor sparse point cloud scenes based on local ICP selects a small amount of overlapping parts in the point cloud scene with low overlap between two frames, and performs local ICP registration on the small amount of overlapping parts to guide the registration of the entire large scene. The amount of data collected is small, but this method is prone to fall into local optimality, reducing the overall registration accuracy. Summary of the invention
[0005] In order to overcome the above technical problems, the present invention provides a low-overlap 3D dynamic point cloud registration method based on an attention mechanism, which has good point cloud registration effect with low overlap rate and is not easy to fall into local optimum.
[0006] The technical solution of the present invention is as follows:
[0007] A low-overlap 3D dynamic point cloud registration method based on an attention mechanism comprises the following steps:
[0008] S1, input the source point cloud and target point cloud to be registered into the pre-trained point cloud registration network, and extract the local geometric features of all the point cloud points in the source point cloud and all the point cloud points in the target point cloud;
[0009] S2, calculating the similarity score between the local geometric features of the source point cloud and the local geometric features of the target point cloud;
[0010] S3, sampling the point cloud points of the source point cloud and the point cloud points of the target point cloud whose similarity scores reach the threshold to obtain key points;
[0011] S4, calculating the distance features of the key points of the source point cloud and the target point cloud, combining the distance features with the local geometric features of the key points to construct a hybrid feature similarity matrix;
[0012] S5. The hybrid matching scores of the key points are calculated according to the hybrid feature similarity matrix, and the hybrid matching scores are subjected to SVD singular value decomposition to obtain the registration matrix and the registration vector. The position of the source point cloud is transformed by using the registration matrix and the registration vector to complete the registration of the source point cloud with the target point cloud.
[0013] Among them, step S1 extracts the local geometric features through the graph neural network based on the attention mechanism in the point cloud registration network, and step S2 calculates the similarity score through the graph neural network based on the attention mechanism and the cross-attention network in the point cloud registration network.
[0014] This technical solution proposes a low-overlap 3D dynamic point cloud registration method based on an attention mechanism. To address the problem of low registration efficiency of traditional ICP algorithms and deep learning-based methods in low-overlap situations, a graph neural network based on an attention mechanism is used to extract local geometric features. The graph neural network based on the attention mechanism and the similarity scores of local geometric features across the attention network are used to sample key points with high similarity scores for point cloud registration, which can effectively improve the accuracy of point cloud registration. At the same time, the present invention combines the local geometric features and distance features of the key points, utilizes the interaction of features of different point pairs, and further improves the performance of point cloud registration.
[0015] Furthermore, the training method of the point cloud registration network is to perform supervised learning through a matching loss function, and the registration matrix and the registration vector described in step S5 are both input into the loss function. If the matching loss function converges, the trained point cloud registration network is obtained. If the matching loss function does not converge, return to step S4, calculate the distance features of the key points of the source point cloud and the target point cloud after the position of the registration matrix and the registration vector is adjusted, and then execute step S5 until the loss function converges and the point cloud registration network training is completed.
[0016] Furthermore, the local geometric feature extraction method in step S1 is: for each point cloud point p of the source point cloud and the target point cloud i Construct a spherical region and use the knn algorithm to select the nearest k points in the region, denoted as p e , where e∈{1, 2, …, k}, through h θ (p i ,p e ) get point p i and p e The edge features of h θ It is a nonlinear function composed of learnable parameters θ. Finally, the k edge features are summed to get point p i Graph features The graph features are processed through several multi-layer perceptrons (MLPs) and a maximum pooling layer to obtain the local geometric features of the point cloud points n×a. n , where a n is the output dimension of the last layer of MLP.
[0017] Furthermore, the similarity score calculation method described in step S2 is to connect each local geometric feature in the source point cloud and all local geometric features in the target point cloud and compare them one by one to obtain the overlapping similarity of the overlapping points and the significant similarity of the points with prominent features, and output the overlapping similarity score and the significant similarity score as the similarity score.
[0018] Furthermore, the calculation formula of the distance feature in step S4 is:
[0019]
[0020] Where [;] represents the connection operation, p i is the key point of the source point cloud, q j is the key point of the target point cloud;
[0021] After calculating the distance feature, the distance feature d(i,j) and the local geometric feature F of the key point are combined. p (i) F q (j) Connect them to get the mixed feature H (n) (i,j)=[F p (i); F q (j); d(i,j)], for the mixed feature H (n) Apply a series of 1×1 two-dimensional convolutions to map it to a high-dimensional space, then input it into the multi-layer perceptron MLP, and apply the Softmax function to each row of the mixed feature to obtain the similarity matrix S (n) (i,j), S (n) (i,j) represents point p in the source point cloud i and point p in the target point cloud jMix and match scores.
[0022] A low-overlap 3D dynamic point cloud registration system based on an attention mechanism, comprising: a point cloud acquisition module, a local geometric feature extraction module, a similarity attention module, a key point sampling module, a hybrid feature generation module, a hybrid matching score calculation module, and a registration module;
[0023] The point cloud acquisition module acquires the source point cloud and target point cloud to be registered; the local geometric feature extraction module extracts the local geometric features of all point cloud points in the source point cloud and all point cloud points in the target point cloud; the similarity attention module calculates the similarity score between the local geometric features of the source point cloud and the local geometric features of the target point cloud; the key point sampling module samples the point cloud points of the source point cloud and the point cloud points of the target point cloud whose similarity scores reach the threshold to obtain key points; the hybrid feature generation module calculates the distance features of the key points of the source point cloud and the target point cloud, and combines the distance features with the local geometric features of the key points to construct a hybrid feature similarity matrix; the hybrid matching score calculation module calculates the hybrid matching scores of the key points according to the hybrid feature similarity matrix, and performs SVD singular value decomposition on the hybrid matching scores to obtain the registration matrix and registration vector; the registration module uses the registration matrix and registration vector to transform the position of the source point cloud so that the source point cloud and the target point cloud are registered.
[0024] Furthermore, the local geometric feature extraction module extracts the local geometric feature by constructing a spherical region for each point pi of the source point cloud and the target point cloud, and selecting the nearest k points in the region by using the knn algorithm and recording them as p e , where e∈{1, 2, …, k}, through h θ (p i ,p e ) get point p i and p e The edge features of h θ It is a nonlinear function composed of learnable parameters θ. Finally, the k edge features are summed to get point p i Graph features The graph features are processed through several multi-layer perceptrons (MLPs) and a maximum pooling layer to obtain the local geometric features of the point cloud points n×a. n , where a n is the output dimension of the last layer of MLP.
[0025] Furthermore, the method in which the similarity attention module calculates the similarity score is: connecting each local geometric feature in the source point cloud and all local geometric features in the target point cloud for comparison one by one, obtaining the overlapping similarity of the overlapping points and the significant similarity of the points with prominent features, and outputting the overlapping similarity score and the significant similarity score as the similarity score.
[0026] Furthermore, the formula for calculating the distance feature by the hybrid feature generation module is:
[0027]
[0028] Where [;] represents the connection operation, p i is the key point of the source point cloud, q j is the key point of the target point cloud.
[0029] Furthermore, the hybrid feature generation module calculates the distance feature and combines the distance feature d(i,j) and the local geometric feature F of the key point p (i) F q (j) Connect them to get the mixed feature H (n) (i,j)=[F p (i); F q (j); d(i,j)], for the mixed feature H (n) Apply a series of 1×1 two-dimensional convolutions to map it to a high-dimensional space, then input it into the multi-layer perceptron MLP, and apply the Softmax function to each row of the mixed feature to obtain the similarity matrix S (n) (i,j), S (n) (i,j) represents point p in the source point cloud i and point p in the target point cloud j Mix and match scores.
[0030] The technical solution proposes a small sample text classification method and system for semi-supervised learning. Compared with the prior art, the beneficial effects of the technical solution of the present invention are as follows: to address the problem of low registration efficiency of traditional ICP algorithms and methods based on deep learning under low overlap conditions, the present invention adopts a graph neural network based on an attention mechanism to extract local geometric features, and utilizes the graph neural network based on the attention mechanism and the similarity scores of local geometric features across the attention network to sample these key points with high similarity scores for point cloud registration, which can effectively improve the accuracy of point cloud registration. At the same time, the present invention combines the local geometric features and distance features of the key points, utilizes the interaction of features of different point pairs, and further improves the performance of point cloud registration. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 Schematic diagram of the steps of the point cloud registration method of the present invention
[0032] Figure 2 This is a schematic diagram of the point cloud registration network structure;
[0033] Figure 3 Flowchart for training point cloud registration network;
[0034] Figure 4It is a structural diagram of the local geometric feature extraction module;
[0035] Figure 5 Schematic diagram of the similarity attention module structure. DETAILED DESCRIPTION
[0036] In order to clearly illustrate the low-overlap 3D dynamic point cloud registration method and system based on the attention mechanism of the present invention, the present invention is further described in combination with the embodiments and drawings, but this should not limit the scope of protection of the present invention.
[0037] Example 1
[0038] A low-overlap 3D dynamic point cloud registration method based on attention mechanism, such as Figure 1 As shown, the following steps are included:
[0039] S1, input the source point cloud and target point cloud to be registered into the pre-trained point cloud registration network, and extract the local geometric features of all the point cloud points in the source point cloud and all the point cloud points in the target point cloud;
[0040] S2, calculating the similarity score between the local geometric features of the source point cloud and the local geometric features of the target point cloud;
[0041] S3, sampling the point cloud points of the source point cloud and the point cloud points of the target point cloud whose similarity scores reach the threshold to obtain key points;
[0042] S4, calculating the distance features of the key points of the source point cloud and the target point cloud, combining the distance features with the local geometric features of the key points to construct a hybrid feature similarity matrix;
[0043] S5. The hybrid matching scores of the key points are calculated according to the hybrid feature similarity matrix, and the hybrid matching scores are subjected to SVD singular value decomposition to obtain the registration matrix and the registration vector. The position of the source point cloud is transformed by using the registration matrix and the registration vector to complete the registration of the source point cloud with the target point cloud.
[0044] Among them, step S1 extracts the local geometric features through the graph neural network based on the attention mechanism in the point cloud registration network, and step S2 calculates the similarity score through the graph neural network based on the attention mechanism and the cross-attention network in the point cloud registration network.
[0045] In order to solve the problem of low registration efficiency of traditional ICP algorithms and deep learning-based methods in low overlap conditions, this embodiment adopts a graph neural network based on an attention mechanism to extract local geometric features, and uses the graph neural network based on the attention mechanism and the similarity scores of local geometric features across the attention network to sample key points with high similarity scores for point cloud registration, which can effectively improve the accuracy of point cloud registration. At the same time, the present invention combines the local geometric features and distance features of the key points, utilizes the interaction of features of different point pairs, and further improves the performance of point cloud registration.
[0046] Example 2
[0047] A low-overlap 3D dynamic point cloud registration method based on an attention mechanism comprises the following steps:
[0048] S1, input the source point cloud and target point cloud to be registered into the pre-trained point cloud registration network, and extract the local geometric features of all the point cloud points in the source point cloud and all the point cloud points in the target point cloud;
[0049] The training method of the point cloud registration network is supervised learning through matching loss function. The training process is as follows: Figure 3 As shown, the registration matrix and the registration vector described in step S5 are input into the loss function. If the matching loss function converges, the trained point cloud registration network is obtained. If the matching loss function does not converge, return to step S4, calculate the distance features of the key points of the source point cloud and the target point cloud after the position adjustment of the registration matrix and the registration vector, and then execute step S5 until the loss function converges and the point cloud registration network training is completed.
[0050] The calculation formula of the matching loss function is:
[0051]
[0052] Where N is the number of sampled key points in the source point cloud, j * is the index of the point in the target point cloud sampling point that is closest to the source point cloud sampling point under the correct transformation, r is a hyperparameter, which represents the distance between the point in the source point cloud and the point in the target point cloud. is the indicator function, when it satisfies The output is 1 if the condition is met, otherwise the output is 0.
[0053] The local geometric feature extraction method is: for each point cloud point p of the source point cloud and the target point cloud i Construct a spherical region and use the knn algorithm to select the nearest k points in the region, denoted as p e , where e∈{1, 2, …, k}, through h θ (p i ,p e ) get point pi and p e The edge features of h θ It is a nonlinear function composed of learnable parameters θ. Finally, the k edge features are summed to get point p i Graph features The graph features are processed through several multi-layer perceptrons (MLPs) and a maximum pooling layer to obtain the local geometric features of the point cloud points n×a. n , where a n is the output dimension of the last layer of MLP.
[0054] S2, calculating the similarity score between the local geometric features of the source point cloud and the local geometric features of the target point cloud;
[0055] The similarity score is calculated by connecting each local geometric feature in the source point cloud and all local geometric features in the target point cloud and comparing them one by one, obtaining the overlapping similarity of the overlapping points and the significant similarity of the points with prominent features, and outputting the overlapping similarity score and the significant similarity score as the similarity score.
[0056] S3, sampling the point cloud points of the source point cloud and the point cloud points of the target point cloud whose similarity scores reach the threshold to obtain key points;
[0057] S4, calculating the distance features of the key points of the source point cloud and the target point cloud, combining the distance features with the local geometric features of the key points to construct a hybrid feature similarity matrix;
[0058] The calculation formula of the distance feature is:
[0059]
[0060] Where [;] represents the connection operation, p i is the key point of the source point cloud, q j is the key point of the target point cloud;
[0061] After calculating the distance feature, the distance feature d(i,j) and the local geometric feature F of the key point are combined. p (i) F q (j) Connect them to get the mixed feature H (n) (i,j)=[F p (i); F q (j); d(i,j)], for the mixed feature H (n) Apply a series of 1×1 two-dimensional convolutions to map it to a high-dimensional space, then input it into the multi-layer perceptron MLP, and apply the Softmax function to each row of the mixed feature to obtain the similarity matrix S (n) (i,j), S (n) (i,j) represents point p in the source point cloudi and point p in the target point cloud j Mix and match scores.
[0062] S5. The hybrid matching scores of the key points are calculated according to the hybrid feature similarity matrix, and the hybrid matching scores are subjected to SVD singular value decomposition to obtain the registration matrix and the registration vector. The position of the source point cloud is transformed by using the registration matrix and the registration vector to complete the registration of the source point cloud with the target point cloud.
[0063] Among them, step S1 extracts the local geometric features through the graph neural network based on the attention mechanism in the point cloud registration network, and step S2 calculates the similarity score through the graph neural network based on the attention mechanism and the cross-attention network in the point cloud registration network.
[0064] In order to solve the problem of low registration efficiency of traditional ICP algorithms and deep learning-based methods in low overlap conditions, this embodiment adopts a graph neural network based on an attention mechanism to extract local geometric features, and uses the graph neural network based on the attention mechanism and the similarity scores of local geometric features across the attention network to sample key points with high similarity scores for point cloud registration, which can effectively improve the accuracy of point cloud registration. At the same time, the present invention combines the local geometric features and distance features of the key points, utilizes the interaction of features of different point pairs, and further improves the performance of point cloud registration.
[0065] Example 3
[0066] A low-overlap 3D dynamic point cloud registration system based on attention mechanism, the system structure is as follows Figure 2 As shown, it includes: point cloud acquisition module, local geometric feature extraction module, similarity attention module, key point sampling module, hybrid feature generation module, hybrid matching score calculation module, and registration module;
[0067] The point cloud acquisition module obtains the source point cloud and target point cloud to be registered; the local geometric feature extraction module structure is as follows Figure 4 As shown, the local geometric feature extraction module extracts local geometric features of all point cloud points in the source point cloud and all point cloud points in the target point cloud;
[0068] The method of extracting local geometric features by the local geometric feature extraction module is as follows: for each point p in each source point cloud i Construct a spherical region and use the knn algorithm to select the nearest k points in the region, denoted as p e , where e∈{1, 2, …, k}, through h θ (p i ,p e ) get point p i and p e The edge features of hθ It is a nonlinear function composed of learnable parameters θ. Finally, the k edge features are summed to get point p i Graph features When the dimension of the input point is n×3, the dimension of the graph feature is n×k×3. The graph feature passes through several multi-layer perceptrons MLP and a maximum pooling layer to obtain the output local geometric features n×a n , where a n is the output dimension of the last layer of MLP.
[0069] The structure of the local geometric feature extraction module is as follows: Figure 4 As shown in Figure 1, it includes several edge convolution layers, normalization layers, and maximum pooling layers. The structure of each layer is the same. Before the last layer, the local geometric features output by each previous layer are connected and then input into the edge convolution layer. After normalization and pooling, the local geometric features F of the source point cloud are obtained. p (i) and the local geometric features F of the target point cloud q (j) Each edge convolution layer in the local geometric feature extraction module executes the entire process of the method for extracting local geometric features. The local geometric feature extraction module enhances the model expression capability through the repeated structure of several edge convolution layers, normalization layers and maximum pooling layers;
[0070] The similarity attention module calculates the similarity score between the local geometric features of the source point cloud and the local geometric features of the target point cloud;
[0071] The structure of the similarity attention module is as follows Figure 5 As shown, the method for calculating the similarity score is: first, the local geometric features are input into the encoder composed of the graph neural network GNN based on the attention mechanism of the similarity attention module to enhance the connection between the features, and then the encoded features are input across the attention network, each feature in the source point cloud and all the features in the target point cloud are connected and compared one by one, and then the comparison results are decoded in the decoder, and the overlapping similarity score and the significant similarity score are output. The overlapping similarity score refers to the similarity score of the overlapping points of the source point cloud and the target point cloud, and the significant similarity score refers to the similarity score of the points with prominent features in the source point cloud and the target point cloud.
[0072] The key point sampling module samples the point cloud points of the source point cloud and the point cloud points of the target point cloud whose similarity scores reach the threshold to obtain key points, and the key points refer to the points with higher similarity in the source point cloud and the target point cloud; the hybrid feature generation module calculates the distance features of the key points of the source point cloud and the target point cloud, and combines the distance features with the local geometric features of the key points to construct a hybrid feature similarity matrix; the hybrid matching score calculation module calculates the hybrid matching scores of the key points according to the hybrid feature similarity matrix, and performs SVD singular value decomposition on the hybrid matching scores to obtain the registration matrix and registration vector;
[0073] The method for calculating the registration matrix and registration vector by the hybrid feature generation module is as follows: first, the distance features of the key points of the source point cloud and the target point cloud are calculated. Where [;] represents the connection operation, which combines the distance feature and the local geometric feature F extracted previously p (i) F q (j) are connected to obtain the mixed feature H (n) (i,j)=[F p (i); F q (j); d(i,j)], for H (n) Apply a series of 1×1 two-dimensional convolutions to map it to a high-dimensional space, then input it into the multi-layer perceptron MLP, and apply the Softmax function to each row of the mixed feature to obtain the similarity matrix S (n) (i,j), S (n) (i,j) represents point p in the source point cloud i and point p in the target point cloud j The matching score is S (n) The point pair (i, j) with the largest matching score in each row is recorded as a set of corresponding points (p i ,p * i ). Substitute the corresponding points into the following formula and use SVD singular value decomposition to solve the rotation matrix R (n) and displacement vector t (n) :
[0074]
[0075] Get the rotation matrix R (n) and displacement vector t (n) That is the registration matrix and registration vector.
[0076] The registration module uses the registration matrix and registration vector to transform the position of the source point cloud so that the source point cloud and the target point cloud are registered.
[0077] Example 4
[0078] A low-overlap 3D dynamic point cloud registration system based on an attention mechanism, comprising: a point cloud acquisition module, a local geometric feature extraction module, a similarity attention module, a key point sampling module, a hybrid feature generation module, a hybrid matching score calculation module, and a registration module;
[0079] The point cloud acquisition module acquires the source point cloud and target point cloud to be registered; the local geometric feature extraction module extracts the local geometric features of all point cloud points in the source point cloud and all point cloud points in the target point cloud; the similarity attention module calculates the similarity score between the local geometric features of the source point cloud and the local geometric features of the target point cloud; the key point sampling module samples the point cloud points of the source point cloud and the point cloud points of the target point cloud whose similarity scores reach the threshold to obtain key points; the hybrid feature generation module calculates the distance features of the key points of the source point cloud and the target point cloud, and combines the distance features with the local geometric features of the key points to construct a hybrid feature similarity matrix; the hybrid matching score calculation module calculates the hybrid matching scores of the key points according to the hybrid feature similarity matrix, and performs SVD singular value decomposition on the hybrid matching scores to obtain the registration matrix and registration vector; the registration module uses the registration matrix and registration vector to transform the position of the source point cloud so that the source point cloud and the target point cloud are registered.
[0080] The local geometric feature extraction module extracts the local geometric feature by: for each point cloud point p of the source point cloud and the target point cloud i Construct a spherical region and use the knn algorithm to select the nearest k points in the region, denoted as p e , where e∈{1, 2, …, k}, through h θ (p i ,p e ) get point p i and p e The edge features of h θ It is a nonlinear function composed of learnable parameters θ. Finally, the k edge features are summed to get point p i Graph features The graph features are processed through several multi-layer perceptrons (MLPs) and a maximum pooling layer to obtain the local geometric features of the point cloud points n×a. n , where a n is the output dimension of the last layer of MLP.
[0081] The method for the similarity attention module to calculate the similarity score is: each local geometric feature in the source point cloud is connected to all local geometric features in the target point cloud for comparison one by one, to obtain the overlapping similarity of the overlapping points and the significant similarity of the points with prominent features, and to output the overlapping similarity score and the significant similarity score as the similarity score.
[0082] The formula for calculating the distance feature by the hybrid feature generation module is:
[0083]
[0084] Where [;] represents the connection operation, p i is the key point of the source point cloud, q j is the key point of the target point cloud.
[0085] Furthermore, the hybrid feature generation module calculates the distance feature and combines the distance feature d(i,j) and the local geometric feature F of the key point p (i) F q (j) Connect them to get the mixed feature H (n) (i,j)=[F p (i); F q (j); d(i,j)], for the mixed feature H (n) Apply a series of 1×1 two-dimensional convolutions to map it to a high-dimensional space, then input it into the multi-layer perceptron MLP, and apply the Softmax function to each row of the mixed feature to obtain the similarity matrix S (n) (i,j), S (n) (i,j) represents point p in the source point cloud i and point p in the target point cloud j Mix and match scores.
[0086] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.
Claims
1. A low-overlap 3D dynamic point cloud registration method based on attention mechanism, characterized in that: The following steps are involved: S1, input the source point cloud and target point cloud to be registered into the pre-trained point cloud registration network, and extract the local geometric features of all the point cloud points in the source point cloud and all the point cloud points in the target point cloud; S2, calculating the similarity score between the local geometric features of the source point cloud and the local geometric features of the target point cloud; S3, sampling the point cloud points of the source point cloud and the point cloud points of the target point cloud whose similarity scores reach the threshold to obtain key points; S4, calculating the distance features of the key points of the source point cloud and the target point cloud, combining the distance features with the local geometric features of the key points to construct a hybrid feature similarity matrix; S5. The hybrid matching scores of the key points are calculated according to the hybrid feature similarity matrix, and the hybrid matching scores are subjected to SVD singular value decomposition to obtain the registration matrix and the registration vector. The position of the source point cloud is transformed by using the registration matrix and the registration vector to complete the registration of the source point cloud with the target point cloud. Among them, step S1 extracts the local geometric features through the graph neural network based on the attention mechanism in the point cloud registration network, and step S2 calculates the similarity score through the graph neural network based on the attention mechanism and the cross-attention network in the point cloud registration network.
2. According to the low-overlap 3D dynamic point cloud registration method based on attention mechanism in claim 1, it is characterized in that: The training method of the point cloud registration network is to perform supervised learning through a matching loss function, and the registration matrix and the registration vector in step S5 are both input into the loss function. If the matching loss function converges, the trained point cloud registration network is obtained. If the matching loss function does not converge, return to step S4, calculate the distance features of the key points of the source point cloud and the target point cloud after the position of the registration matrix and the registration vector are adjusted, and then execute step S5 until the loss function converges and the point cloud registration network training is completed.
3. According to the low-overlap 3D dynamic point cloud registration method based on attention mechanism in claim 1, it is characterized in that: The local geometric feature extraction method in step S1 is: for each point cloud point p of the source point cloud and the target point cloud i Construct a spherical region and use the knn algorithm to select the nearest k points in the region, denoted as p e , where e∈{1, 2, …, k}, through h θ (p i ,p e ) get point p i and p e The edge features of h θ It is a nonlinear function composed of learnable parameters θ. Finally, the k edge features are summed to get point p i Graph features The graph features are processed through several multi-layer perceptrons (MLPs) and a maximum pooling layer to obtain the local geometric features of the point cloud points n×a. n , where a n is the output dimension of the last layer of MLP.
4. The low-overlap 3D dynamic point cloud registration method based on attention mechanism according to claim 1, characterized in that: The similarity score calculation method described in step S2 is to connect each local geometric feature in the source point cloud and all local geometric features in the target point cloud and compare them one by one to obtain the overlapping similarity of the overlapping points and the significant similarity of the points with prominent features, and output the overlapping similarity score and the significant similarity score as the similarity score.
5. The low-overlap 3D dynamic point cloud registration method based on attention mechanism according to claim 1, characterized in that: The calculation formula of the distance feature in step S4 is: Where [;] represents the connection operation, p i is the key point of the source point cloud, q j is the key point of the target point cloud; After calculating the distance feature, the distance feature d(i,j) and the local geometric feature F of the key point are combined. p (i) F q (j) Connect them to get the mixed feature H (n) (i,j)=[F p (i); F q (j); d(i,j)], for the mixed feature H (n) Apply a series of 1×1 two-dimensional convolutions to map it to a high-dimensional space, then input it into the multi-layer perceptron MLP, and apply the Softmax function to each row of the mixed feature to obtain the similarity matrix S (n) (i,j), S (n) (i,j) represents point p in the source point cloud i and point p in the target point cloud j Mix and match scores.
6. A low-overlap 3D dynamic point cloud registration system based on an attention mechanism for executing the low-overlap 3D dynamic point cloud registration method based on an attention mechanism according to any one of claims 1 to 5, characterized in that: include: Point cloud acquisition module, local geometric feature extraction module, similarity attention module, key point sampling module, hybrid feature generation module, hybrid matching score calculation module, and registration module; The point cloud acquisition module acquires the source point cloud and the target point cloud to be registered; the local geometric feature extraction module extracts the local geometric features of all the point cloud points in the source point cloud and all the point cloud points in the target point cloud; the similarity attention module calculates the similarity score between the local geometric features of the source point cloud and the local geometric features of the target point cloud; the key point sampling module samples the point cloud points of the source point cloud and the point cloud points of the target point cloud whose similarity scores reach the threshold to obtain key points; The hybrid feature generation module calculates the distance features of the key points of the source point cloud and the target point cloud, and combines the distance features with the local geometric features of the key points to construct a hybrid feature similarity matrix; The hybrid matching score calculation module calculates the hybrid matching score of the key points based on the hybrid feature similarity matrix, and performs SVD singular value decomposition on the hybrid matching score to obtain the registration matrix and registration vector; the registration module uses the registration matrix and registration vector to transform the position of the source point cloud so that the source point cloud and the target point cloud are registered.
7. The low-overlap 3D dynamic point cloud registration system based on attention mechanism according to claim 6, characterized in that: The local geometric feature extraction module extracts the local geometric feature by: for each point cloud point p of the source point cloud and the target point cloud i Construct a spherical region and use the knn algorithm to select the nearest k points in the region, denoted as p e , where e∈{1, 2, …, k}, through h θ (p i ,p e ) get point p i and p e The edge features of h θ It is a nonlinear function composed of learnable parameters θ. Finally, the k edge features are summed to get point p i Graph features The graph features are processed through several multi-layer perceptrons (MLPs) and a maximum pooling layer to obtain the local geometric features of the point cloud points n×a. n , where a n is the output dimension of the last layer of MLP.
8. The low-overlap 3D dynamic point cloud registration system based on attention mechanism according to claim 6, characterized in that: The method for the similarity attention module to calculate the similarity score is: each local geometric feature in the source point cloud is connected to all local geometric features in the target point cloud for comparison one by one, to obtain the overlapping similarity of the overlapping points and the significant similarity of the points with prominent features, and to output the overlapping similarity score and the significant similarity score as the similarity score.
9. The low-overlap 3D dynamic point cloud registration system based on attention mechanism according to claim 6, characterized in that: The formula for calculating the distance feature by the hybrid feature generation module is: Where [;] represents the connection operation, p i is the key point of the source point cloud, q j is the key point of the target point cloud.
10. The low-overlap 3D dynamic point cloud registration system based on attention mechanism according to claim 6, characterized in that: The hybrid feature generation module calculates the distance feature and combines the distance feature d(i,j) and the local geometric feature F of the key point p (i) F q (j) Connect them to get the mixed feature H (n) (i,j)=[F p (i); F q (j); d(i,j)], for the mixed feature H (n) Apply a series of 1×1 two-dimensional convolutions to map it to a high-dimensional space, then input it into the multi-layer perceptron MLP, and apply the Softmax function to each row of the mixed feature to obtain the similarity matrix S (n) (i,j), S (n) (i,j) represents point p in the source point cloud i and point p in the target point cloud j Mix and match scores.