A point cloud registration method based on fusion of global features and local features

Abstract

The application discloses a point cloud registration method based on global feature and local feature fusion, comprising the following steps: obtaining a reference point cloud and a point cloud to be registered; extracting point cloud features of the reference point cloud and the point cloud to be registered respectively to form an initial reference point cloud feature set and an initial point cloud to be registered feature set; introducing self-attention mechanism and cross-attention mechanism, and introducing correlation information between the reference point cloud and the point cloud to be registered to form a final reference point cloud feature set and a final point cloud to be registered feature set; obtaining similarity information between each point pair; constructing a matching point pair relationship matrix according to the similarity information between the point pairs; obtaining a transformation model parameter of the reference point cloud and the point cloud to be registered; and registering the reference point cloud and the point cloud to be registered. The application fully excavates potential spatial topological structure information of point cloud data by combining local features and global features, obtains point cloud fusion features with rich semantic information, and can effectively improve the accuracy of point cloud registration.

Description

Technical Field

[0001] This invention belongs to the field of point cloud registration technology, specifically relating to a point cloud registration method based on the fusion of global and local features. Background Technology

[0002] Point cloud registration, a crucial step in 3D data processing, has been widely applied in fields such as 3D reconstruction, 3D classification, 3D segmentation, and autonomous driving. Point cloud registration technology refers to the process of aligning two or more point clouds of the same scene acquired at different times, from different viewpoints, or by different sensors by transforming their spatial coordinates.

[0003] Traditional point cloud registration methods can be divided into probability density-based registration methods and ICP (Iterative Closest Point)-based registration methods. Probability density-based registration methods represent the matching relationship between two point clouds using a probabilistic model, thus transforming the solution of the transformation relationship between the two point clouds into solving for the parameters of the probabilistic model. Among these, CPD (Coherent Point Drift) is the most representative algorithm. CPD uses a Gaussian mixture model to fit the matching relationship between point clouds and uses the expectation-maximization algorithm to find the maximum likelihood estimate of the parameters, thereby obtaining the transformation parameters between the point clouds. The ICP-based registration method is based on iterative nearest neighbor. It searches for the nearest neighbor in two point cloud datasets under certain constraints, estimates the rotation matrix and translation vector based on the obtained nearest neighbor, and merges the overlapping parts of the two point clouds by minimizing a designed loss function. This iterative process of finding the nearest neighbor and solving for transformation parameters aims to determine the optimal model parameters for the relationship between the two point clouds. It does not require feature extraction from the point clouds, thus offering the advantage of high processing speed. However, it lacks utilization of point cloud structural information and is sensitive to the initial position of the point clouds, making it prone to getting trapped in local optima. Subsequent research has addressed this shortcoming of ICP by proposing algorithms such as Go-ICP (Globally Optimal ICP) and NICP (Normal ICP). These algorithms effectively utilize point cloud structural information by introducing parameters such as normal vectors and local curvature, but they still suffer from the problem of easily getting trapped in local optima.

[0004] Deep learning-based point cloud registration methods leverage the powerful self-learning capabilities and high-dimensional information extraction abilities of deep neural networks to obtain feature descriptors containing high-dimensional semantic information. This avoids the shortcomings of traditional manually designed feature descriptors and effectively improves the performance of point cloud registration methods. Depending on the format of the raw data processed by the deep neural network, deep learning-based point cloud registration methods can be divided into two main categories: those based on voxel grid data and those based on raw point cloud data.

[0005] Point cloud registration methods based on voxel grid data require voxelizing the point cloud data to be processed before inputting the voxelized data into a neural network for feature extraction and representation. This method has two obvious drawbacks: (1) Due to the uneven distribution of point cloud data, a large number of empty voxels are generated after voxelization, and these invalid network units occupy a lot of storage space; (2) It is sensitive to rotational changes.

[0006] The registration method based on raw point cloud data uses neural networks to extract features from the point cloud data, thereby obtaining global or local features that can effectively represent the point cloud. Based on this, the subsequent registration process is carried out. The problems with this method are: (1) It fails to make full use of the structural information contained in the point cloud data. (2) It isolates the reference point cloud and the point cloud to be registered, ignoring the connection information between the two.

[0007] Traditional methods lack the utilization of point cloud structural information. Other deep learning-based methods only learn local or global information of point clouds through the network to represent the topological structure of point clouds, which does not make high use of the spatial information of point clouds themselves. Summary of the Invention

[0008] To address the aforementioned problems in existing technologies, this invention provides a point cloud registration method based on the fusion of global and local features. It utilizes a combination of self-attention and cross-attention mechanisms to introduce inter-point cloud correlation information into a single point cloud feature set. Simultaneously, it incorporates graph convolution concepts to provide a method for extracting local point cloud features, effectively enriching the semantic information of the obtained point cloud features and improving registration accuracy. The technical problem to be solved by this invention is achieved through the following technical solution:

[0009] This invention provides a point cloud registration method based on the fusion of global and local features, comprising:

[0010] S1: Obtain the baseline point cloud and the point cloud to be matched;

[0011] S2: Extract the point cloud features of the reference point cloud and the point cloud to be registered respectively to form an initial reference point cloud feature set and an initial point cloud feature set to be registered;

[0012] S3: Based on the initial reference point cloud feature set and the initial point cloud feature set to be registered, by introducing a self-attention mechanism and a cross-attention mechanism, and by introducing the correlation information between the reference point cloud and the point cloud to be registered, the final reference point cloud feature set and the final point cloud feature set to be registered are formed.

[0013] S4: Calculate the similarity between the final reference point cloud feature set and the final point cloud feature set to be registered, and obtain the similarity information between each pair of points;

[0014] S5: Construct a matching point pair relationship matrix based on the similarity information between each point pair;

[0015] S6: Obtain the transformation model parameters of the reference point cloud and the point cloud to be registered based on the matching point pair relationship matrix;

[0016] S7: Register the reference point cloud and the point cloud to be registered according to the obtained transformation model parameters.

[0017] In one embodiment of the present invention, S2 includes:

[0018] S2.1: The reference point cloud and the point cloud to be registered are encoded using a multilayer perceptron, and multi-dimensional feature information is introduced into the final feature through skip connections between different levels, so as to obtain the global features of the reference point cloud and the point cloud to be registered for perceiving data information of different dimensions respectively.

[0019] S2.2: The local structure inside the reference point cloud and the point cloud to be registered is constructed by using the k nearest neighbor algorithm to obtain the local features of the reference point cloud and the point cloud to be registered, respectively.

[0020] S2.3: An adaptive fusion strategy is adopted to fuse the obtained global features and local features of the point cloud to obtain the initial reference point cloud feature set and the initial point cloud feature set to be registered.

[0021] In one embodiment of the present invention, S2.2 includes:

[0022] S2.21: The local structure inside the point cloud is constructed by using the k-nearest neighbor algorithm on the reference point cloud and the point cloud to be registered, and the edge information is obtained. The edge information is represented as follows:

[0023]

[0024] in, The center point x of the reference point cloud or the point cloud to be registered represents... i The edge formed with the k-th point, The center point x of the reference point cloud or the point cloud to be registered represents...i The k-th point found by the K-nearest neighbor algorithm;

[0025] S2.22: By extracting the edge information of the reference point cloud or the point cloud to be registered using a convolutional network, the local features of the reference point cloud and the point cloud to be registered are obtained.

[0026] In one embodiment of the present invention, in step S2.3, the fused initial reference point cloud feature set or the initial point cloud feature set to be registered is represented as follows:

[0027] F mix =α1F global +α2F local

[0028]

[0029] Among them, F mix For the fusion features of the reference point cloud or the point cloud to be registered, F global For the global features of the reference point cloud or the point cloud to be registered, F local For the local features of the reference point cloud or the point cloud to be registered, α i For normalized weights, ∑α i =1, w i and w j This indicates the initialization of the weight parameters.

[0030] In one embodiment of the present invention, S3 includes:

[0031] S3.1: Using a self-attention mechanism, weights are assigned to the initial reference point cloud feature set and the initial point cloud feature set to be registered according to their saliency, to obtain the feature set enhanced by the self-attention mechanism:

[0032] F new =softmax(mlp(softmax(FF) T softmax(FF) T )F

[0033] Wherein, F represents the fusion feature of the reference point cloud feature set or the point cloud feature set to be registered, F new For the reference point cloud feature set or the point cloud feature set to be registered after being enhanced by the self-attention mechanism, mlp() represents multilayer perceptron processing;

[0034] S3.2: A cross-attention mechanism is used to introduce correlation information between point clouds into the reference point cloud feature set or the point cloud feature set to be registered, which are enhanced by the self-attention mechanism respectively. This is represented as:

[0035] F final，1 =f1(Fsource ,F target )+F source

[0036]

[0037] F final,2 =f2(F source ,F target )+F target

[0038]

[0039] Among them, w q =V θ F target , w v =V γ F source V θ , and V γ F is a set of learnable weight matrices. target F represents the baseline point cloud feature set enhanced by the self-attention mechanism. source F represents the feature set of the point cloud to be registered after being enhanced by the self-attention mechanism. final,1 For the final point cloud feature set to be registered after enhancement by the cross-attention mechanism, F final,2 This is the final baseline point cloud feature set enhanced by the cross-attention mechanism.

[0040] In one embodiment of the present invention, S5 includes:

[0041] The similarity information between the reference point cloud and the point cloud to be registered is transformed into a matching probability matrix between [0,1] using the normalization function softmax, thereby obtaining the matching point pair relationship matrix between the reference point cloud and the point cloud to be registered.

[0042] In one embodiment of the present invention, S6 includes:

[0043] S6.1: Perform singular value decomposition on the obtained point cloud matching point pair relationship matrix to obtain the rotation matrix parameters of the point cloud transformation model;

[0044] S6.2: Obtain the translation parameters of the transformation model based on the obtained rotation matrix parameters and the matching point pair relationship matrix between point clouds.

[0045] Another aspect of the present invention provides a storage medium storing a computer program for executing the steps of the point cloud registration method based on the fusion of global and local features as described in any of the above embodiments.

[0046] Another aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor, when calling the computer program in the memory, implements the steps of the point cloud registration method based on the fusion of global and local features as described in any of the above embodiments.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] 1. This invention combines local and global features to fully explore the potential spatial topological structure information of point cloud data, and obtains point cloud fusion features with rich semantic information, thereby improving the accuracy of point cloud registration.

[0049] 2. This invention autonomously learns global features of point clouds through neural networks, and simultaneously constructs local topological structures between point clouds using the K-nearest neighbor algorithm. Information is extracted from the constructed local topological structures through deep networks to obtain local features of the point clouds. An adaptive fusion strategy is used to fuse global and local features to obtain fused features with more complete structural information. Furthermore, by introducing self-attention and cross-attention mechanisms, the solution of correct matching relationships in the registration algorithm process is improved, which can effectively improve the accuracy of point cloud registration.

[0050] 3. The point cloud registration method of the present invention balances the proportion of local and global features in the final mixed features through an adaptive strategy, so that the obtained mixed feature set can adaptively adjust the contribution of local and global features in the final features according to the distribution pattern of the point cloud to be registered. This can effectively avoid the decrease in accuracy caused by manually setting parameters when the point cloud distribution is irregular.

[0051] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of a point cloud registration method based on the fusion of local and global features provided in an embodiment of the present invention. Detailed Implementation

[0053] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following describes in detail a point cloud registration method based on the fusion of global and local features proposed in accordance with the present invention, in conjunction with the accompanying drawings and specific embodiments.

[0054] The foregoing and other technical contents, features, and effects of the present invention will be clearly presented in the following detailed description of specific embodiments in conjunction with the accompanying drawings. Through the description of the specific embodiments, a more in-depth and concrete understanding can be gained of the technical means and effects adopted by the present invention to achieve its intended purpose. However, the accompanying drawings are for reference and illustration only and are not intended to limit the technical solutions of the present invention.

[0055] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or apparatus comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or apparatus that includes said element.

[0056] Please see Figure 1 , Figure 1 This is a schematic diagram of a point cloud registration method based on the fusion of local and global features provided in an embodiment of the present invention. The point cloud registration method includes:

[0057] S1: Obtain the baseline point cloud and the point cloud to be registered.

[0058] In this embodiment, two point clouds are first obtained. One point cloud is used as the reference point cloud, and the other is used as the point cloud to be registered. The reference point cloud refers to the target template point cloud, and the point cloud to be registered refers to the point cloud that needs to be processed to coincide with the target template point cloud after coordinate transformation.

[0059] S2: Extract the point cloud features of the reference point cloud and the point cloud to be registered respectively to form an initial reference point cloud feature set and an initial point cloud feature set to be registered.

[0060] Specifically, step S2 in this embodiment includes:

[0061] S2.1: The reference point cloud and the point cloud to be registered are encoded using a multilayer perceptron, and multi-dimensional feature information is introduced into the final feature through skip connections between different layers, so as to obtain the global features of the reference point cloud and the point cloud to be registered for perceiving data information of different dimensions.

[0062] Specifically, a multilayer perceptron is used to encode the reference point cloud and the point cloud to be registered, respectively. The original coordinates of the point cloud are used as the input data of the perceptron. In the feature extraction, the present invention adopts a six-layer perceptron design, and the outputs of the first, third and fifth layers are spliced ​​with the input features through skip connections, so as to introduce multi-dimensional feature information into the final features, so as to obtain the global features of the reference point cloud and the point cloud to be registered for perceiving low, medium and high dimension data information respectively.

[0063] It's important to note that a multilayer perceptron introduces one or more hidden layers into a single-layer neural network. These hidden layers are located between the input and output layers. Neurons in the hidden layers are fully connected to all inputs in the input layer, and neurons in the output layer are also fully connected to all neurons in the hidden layers. Therefore, both the hidden and output layers in a multilayer perceptron are fully connected layers.

[0064] S2.2: The local structure inside the point cloud is constructed by using the k-nearest neighbor algorithm to obtain the local features of the reference point cloud and the point cloud to be registered.

[0065] Specifically, firstly, the local structure inside the point cloud is constructed using the k-nearest neighbor algorithm for the reference point cloud and the point cloud to be registered, and edge information is obtained. The edge information is represented as follows:

[0066]

[0067] in, The center point x of the reference point cloud or the point cloud to be registered represents... i The edge formed with the k-th point, The center point x of the reference point cloud or the point cloud to be registered represents... i The k-th point found by the K-nearest neighbor algorithm is preferably k=20.

[0068] S2.22: By extracting the edge information of the reference point cloud or the point cloud to be registered using a convolutional network, the local features of the reference point cloud and the point cloud to be registered are obtained, wherein point x i The edge features are specifically represented as follows:

[0069] edge i =max{h θ (x i ,e i 1 ),h θ (x i ,e i 2 ),…,hθ (x i ,e i k )}

[0070] Among them, h θ Let θ represent the aggregation function, where θ is a learnable parameter of the convolutional neural network. To ensure that the extracted features are permutation-invariant, max pooling is performed on the extracted features to obtain the edge. i That is, point x i Edge features.

[0071] It should be noted that in this step, the edge features of each point in the reference point cloud and the point cloud to be registered are calculated using the expression of the table features mentioned above, thereby forming the local features of the reference point cloud and the point cloud to be registered, respectively. That is, the local features are the set of edge features of all points in each point cloud.

[0072] S2.3: An adaptive fusion strategy is adopted to fuse the obtained global features and local features of the point cloud to obtain the initial reference point cloud feature set and the initial point cloud feature set to be registered.

[0073] Specifically, the fused initial reference point cloud feature set or the initial point cloud feature set to be registered is represented as follows:

[0074] F mix =α1F global +α2F local

[0075]

[0076] Among them, F mix For the fusion features of the reference point cloud or the point cloud to be registered, F global For the global features of the reference point cloud or the point cloud to be registered, F local For the local features of the reference point cloud or the point cloud to be registered, α i For normalized weights, ∑α i =1, w i and w j This indicates the initialization of the weight parameters.

[0077] S3: Based on the initial reference point cloud feature set and the initial point cloud feature set to be registered, the network is suppressed by introducing a self-attention mechanism and a cross-attention mechanism to suppress the bias of the network towards features with low saliency within the point cloud. At the same time, the correlation information between the reference point cloud and the point cloud to be registered is introduced to form the final reference point cloud feature set and the final point cloud feature set to be registered.

[0078] In this embodiment, S3 includes:

[0079] S3.1: Using a self-attention mechanism, weights are assigned to the initial reference point cloud feature set and the initial point cloud feature set to be registered according to their saliency, to obtain the feature set enhanced by the self-attention mechanism, as follows:

[0080] F new =softmax(mlp(softmax(FF) T softmax(FF) T )F

[0081] Wherein, F represents the fusion feature of the reference point cloud feature set or the point cloud feature set to be registered, that is, the initial reference point cloud feature set and the initial point cloud feature set to be registered, F new For the reference point cloud feature set or the point cloud feature set to be registered after being enhanced by the self-attention mechanism, mlp() represents multilayer perceptron processing;

[0082] S3.2: A cross-attention mechanism is used to introduce correlation information between point clouds into the reference point cloud feature set or the point cloud feature set to be registered, which are enhanced by the self-attention mechanism respectively. This is represented as:

[0083] F final，1 =f1(F source ,F target )+F source

[0084]

[0085] F final,2 =f2(F source ,F target )+F target

[0086]

[0087] Among them, w q =V θ F target , w v =V γ F source V θ , and V γ F is a set of learnable weight matrices. target F represents the baseline point cloud feature set enhanced by the self-attention mechanism. source F represents the feature set of the point cloud to be registered after being enhanced by the self-attention mechanism. final,1 For the final point cloud feature set to be registered after enhancement by the cross-attention mechanism, F final,2 This is the final baseline point cloud feature set enhanced by the cross-attention mechanism.

[0088] S4: Calculate the similarity between the final reference point cloud feature set and the final point cloud feature set to be registered, and obtain the similarity information between each pair of points.

[0089] Specifically, by calculating the similarity between each point feature in the reference point cloud (i.e., the final reference point cloud feature set) and each point feature in the point cloud to be registered (i.e., the final point cloud feature set), the similarity information between the reference point cloud and the point cloud to be registered is obtained. The similarity calculation rule is determined by those skilled in the art according to business needs, and this invention does not impose any restrictions on it. In this embodiment, the similarity calculation rule is the dot product method, which can obtain the similarity information between the reference point cloud and the point cloud to be registered.

[0090] S5: Construct a matching point pair relationship matrix based on the similarity information between point pairs.

[0091] Specifically, the similarity information between the reference point cloud and the point cloud to be registered is transformed into a matching probability matrix between [0,1] by using the normalization function softmax, which yields the matching point pair relationship matrix between the reference point cloud and the point cloud to be registered.

[0092] Specifically, assuming that the current reference point cloud and the point cloud to be registered each have m points, the final matching point pair relationship matrix is ​​an m*m matrix. The data in the first column represents the similarity between point 1 in the reference point cloud and all (m points) in the point cloud to be registered, the data in the second column represents the similarity between point 2 in the reference point cloud and all (m points) in the point cloud to be registered, and so on.

[0093] S6: Obtain the transformation model parameters of the reference point cloud and the point cloud to be registered based on the matching point pair relationship matrix.

[0094] In this embodiment, S6 includes:

[0095] S6.1: Perform singular value decomposition on the obtained point cloud matching point pair relationship matrix to obtain the rotation matrix parameters of the point cloud transformation model;

[0096] S6.2: Obtain the translation parameters of the transformation model based on the obtained rotation matrix parameters and the matching point pair relationship matrix between point clouds.

[0097] Specifically, further processing of the aforementioned matching point pair relation matrix yields the transformation model parameters. In this embodiment, singular value decomposition (SVD) is used to convert the matching point pair relation matrix into transformation model parameters. The affine transformation parameters, obtained through SVD, are the transformation model parameters. In other embodiments, other methods can be used for the transformation, and there are no limitations on this.

[0098] S7: Register the reference point cloud and the point cloud to be registered according to the obtained transformation model parameters.

[0099] This invention can adaptively fuse local and global features of a point cloud to obtain a fused feature that includes both local and global information. Based on this fused feature, self-attention and cross-attention mechanisms are introduced to enhance the registration accuracy and efficiency of the point cloud. In summary, this invention can effectively combine global and local information of a point cloud, thereby improving the accuracy and efficiency of point cloud registration.

[0100] The point cloud registration method of this invention uses an adaptive strategy to balance the proportion of local and global features in the final mixed features. This allows the obtained mixed feature set to adaptively adjust the contribution of local and global features in the final features according to the distribution pattern of the point cloud to be registered. This effectively avoids the decrease in accuracy caused by manually setting parameters when the point cloud distribution is irregular.

[0101] Another embodiment of the present invention provides a storage medium storing a computer program for executing the steps of the point cloud registration method based on the fusion of global and local features described in the above embodiments. A further aspect of the present invention provides an electronic device including a memory and a processor. The memory stores a computer program, and the processor, when calling the computer program in the memory, implements the steps of the point cloud registration method based on the fusion of global and local features as described in the above embodiments. Specifically, the integrated modules implemented as software functional modules can be stored in a computer-readable storage medium. The software functional modules stored in a storage medium include several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0102] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A point cloud registration method based on the fusion of global and local features, characterized in that, include: S1: Obtain the baseline point cloud and the point cloud to be matched; S2: Extract the point cloud features of the reference point cloud and the point cloud to be registered respectively to form an initial reference point cloud feature set and an initial point cloud feature set to be registered; S3: Based on the initial reference point cloud feature set and the initial point cloud feature set to be registered, by introducing a self-attention mechanism and a cross-attention mechanism, and by introducing the correlation information between the reference point cloud and the point cloud to be registered, the final reference point cloud feature set and the final point cloud feature set to be registered are formed. S4: Calculate the similarity between the final reference point cloud feature set and the final point cloud feature set to be registered, and obtain the similarity information between each pair of points; S5: Construct a matching point pair relationship matrix based on the similarity information between each point pair; S6: Obtain the transformation model parameters of the reference point cloud and the point cloud to be registered based on the matching point pair relationship matrix; S7: Register the reference point cloud and the point cloud to be registered according to the obtained transformation model parameters; S2 includes: S2.1: A multilayer perceptron is used to encode the reference point cloud and the point cloud to be registered, respectively. Skip connections are used between different layers to introduce multi-dimensional feature information into the final feature, so as to obtain the global features of the reference point cloud and the point cloud to be registered for perceiving data information of different dimensions. The multilayer perceptron includes a six-layer perceptron, and the outputs of the first, third and fifth layers of the perceptron are concatenated with the input features through skip connections to introduce multi-dimensional feature information into the final feature, so as to obtain the global features of the reference point cloud and the point cloud to be registered for perceiving data information of different dimensions. The multilayer perceptron introduces one or more hidden layers on the basis of a single-layer neural network. The hidden layer is located between the input layer and the output layer. The neurons in the hidden layer are fully connected to each input in the input layer, and the neurons in the output layer are fully connected to each neuron in the hidden layer. The hidden layer and the output layer in the multilayer perceptron are both fully connected layers. S2.2: The local structure inside the reference point cloud and the point cloud to be registered is constructed by using the k nearest neighbor algorithm to obtain the local features of the reference point cloud and the point cloud to be registered, respectively. S2.3: An adaptive fusion strategy is adopted to fuse the obtained global features and local features of the point cloud to obtain the initial reference point cloud feature set and the initial point cloud feature set to be registered.

2. The point cloud registration method based on the fusion of global and local features according to claim 1, characterized in that, S2.2 includes: S2.21: The local structure inside the point cloud is constructed by using the k-nearest neighbor algorithm on the reference point cloud and the point cloud to be registered, and the edge information is obtained. The edge information is represented as follows: in, This refers to the center point of the reference point cloud or the point cloud to be registered. With the k The edge formed by the points, This refers to the center point of the reference point cloud or the point cloud to be registered. The first one found by the K nearest neighbor algorithm k One point; S2.22: By extracting the edge information of the reference point cloud or the point cloud to be registered using a convolutional network, the local features of the reference point cloud and the point cloud to be registered are obtained.

3. The point cloud registration method based on the fusion of global and local features according to claim 2, characterized in that, In step S2.3, the fused initial reference point cloud feature set or the initial point cloud feature set to be registered is represented as follows: in, The fusion features of the reference point cloud or the point cloud to be registered. The global features of the reference point cloud or the point cloud to be registered. For the local features of the reference point cloud or the point cloud to be registered, To normalize the weights, , w i and w j This indicates the initialization of the weight parameters.

4. The point cloud registration method based on the fusion of global and local features according to claim 1, characterized in that, S3 includes: S3.1: Using a self-attention mechanism, weights are assigned to the initial reference point cloud feature set and the initial point cloud feature set to be registered according to their saliency, to obtain the feature set enhanced by the self-attention mechanism: in, F This represents the fusion feature of the reference point cloud feature set or the point cloud feature set to be registered. F new The reference point cloud feature set or the point cloud feature set to be registered is enhanced by the self-attention mechanism. mlp () indicates multilayer perceptron processing; S3.2: A cross-attention mechanism is used to introduce correlation information between point clouds into the reference point cloud feature set or the point cloud feature set to be registered, which are enhanced by the self-attention mechanism respectively. This is represented as: in, = V , = V , = V , , and A set of learnable weight matrices, This represents the baseline point cloud feature set enhanced by the self-attention mechanism. This represents the feature set of the point cloud to be registered after being enhanced by the self-attention mechanism. This is the final feature set of the point cloud to be registered after enhancement by the cross-attention mechanism. This is the final baseline point cloud feature set enhanced by the cross-attention mechanism.

5. The point cloud registration method based on the fusion of global and local features according to claim 1, characterized in that, S5 includes: The similarity information between the reference point cloud and the point cloud to be registered is transformed into a matching probability matrix between [0,1] using the normalization function softmax, thereby obtaining the matching point pair relationship matrix between the reference point cloud and the point cloud to be registered.

6. The point cloud registration method based on the fusion of global and local features according to claim 1, characterized in that, S6 includes: S6.1: Perform singular value decomposition on the obtained point cloud matching point pair relationship matrix to obtain the rotation matrix parameters of the point cloud transformation model; S6.2: Obtain the translation parameters of the transformation model based on the obtained rotation matrix parameters and the matching point pair relationship matrix between point clouds.

7. A storage medium, characterized in that, The storage medium stores a computer program for executing the steps of the point cloud registration method based on the fusion of global and local features as described in any one of claims 1 to 6.

8. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor, when calling the computer program in the memory, implements the steps of the point cloud registration method based on the fusion of global and local features as described in any one of claims 1 to 6.

Images