Multi-view point cloud registration method, system and device and storage medium

Abstract

The invention discloses a multi-view point cloud registration method, system and device and a storage medium, and belongs to the field of point cloud registration. The method comprises the following steps: acquiring a plurality of single-view-angle point cloud models of a workpiece under a plurality of view angles, and calculating curvature information of each sampling point in the single-view-angle point cloud models; converting the single-view point cloud model into a high-curvature point cloud model; segmenting the high-curvature point cloud model according to the distribution density of sampling points in the high-curvature point cloud model to obtain a plurality of high-curvature point cloud sub-models; one part of the high-curvature point cloud sub-models are used for training the PAR model to obtain a target PAR model; inputting the other part of high-curvature point cloud sub-models into the target PAR model to obtain a plurality of transformation matrixes; determining a rotation matrix according to a distance threshold and the plurality of transformation matrixes; according to the rotation matrix, registration of a plurality of single-view-angle point cloud models of a workpiece is achieved, point cloud data are divided into a plurality of local features, and accurate and rapid point cloud registration is achieved through the local features.

Description

Technical Field

[0001] The present application belongs to the technical field of point cloud registration, and specifically relates to a multi-view point cloud registration method, system, device and storage medium. Background Art

[0002] Multi-view point cloud registration refers to the process of adjusting the pose of point cloud data from multiple viewpoints using transformation matrices to unify them in the same standard world coordinate system. This is an important task because point cloud data obtained from a single viewpoint may miss some information, while data obtained from multiple viewpoints can provide more complete geometric and texture information. Multi-view point cloud registration has high requirements for accuracy when applied to industry. Point cloud registration faces challenges due to multiple factors such as noise and inaccurate initial position.

[0003] At present, the existing multi-view point cloud registration methods will have technical problems such as low point cloud registration accuracy and slow registration speed when the point cloud data volume is huge and the initial position is unclear.

[0004] The present application provides a multi-view point cloud registration deep learning network. When the amount of point cloud data is huge and the initial position is unclear, the point cloud data is segmented into multiple and local features, and accurate and fast point cloud registration is achieved through local features. Summary of the invention

[0005] The purpose of the embodiments of the present application is to provide a multi-view point cloud registration method, system, device and storage medium, which can solve the technical problems of low point cloud registration accuracy and slow registration speed in existing multi-view point cloud registration methods when the amount of point cloud data is huge and the initial position is unclear.

[0006] In order to solve the above technical problems, this application is implemented as follows:

[0007] In a first aspect, an embodiment of the present application provides a multi-view point cloud registration method, the method comprising:

[0008] Acquire multiple single-view point cloud models of the workpiece under multiple viewing angles, and calculate the curvature information of each sampling point in the single-view point cloud model;

[0009] According to the curvature information of each sampling point in each of the single-view point cloud models, the single-view point cloud model is converted into a high-curvature point cloud model;

[0010] Segmenting the high curvature point cloud model according to the distribution density of sampling points in the high curvature point cloud model to obtain a plurality of high curvature point cloud sub-models;

[0011] Using a portion of the point cloud data in the plurality of high curvature point cloud sub-models as input of a PAR model, using a standard pose transformation matrix as output of the PAR model, and training the PAR model to obtain a target PAR model;

[0012] Inputting another part of the point cloud data in the plurality of high curvature point cloud sub-models into the target PAR model to obtain a plurality of transformation matrices;

[0013] Determine a rotation matrix corresponding to the single-view point cloud model according to a distance threshold and the multiple transformation matrices;

[0014] Multiple single-view point cloud models of the workpiece are registered according to the rotation matrices corresponding to multiple view angles.

[0015] As an optional implementation of the first aspect of the present application, the step of calculating the curvature information of each sampling point in the single-view point cloud model is specifically:

[0016] For any sampling point p in the single-view point cloud model i Set the neighborhood radius and determine the sampling point p i The Euclidean distance to other sampling points within the neighborhood radius;

[0017] According to the Euclidean distance, the sampling point p is determined i The weight of

[0018] Each sampling point p in the single-view point cloud model i The weight is compared with the preset target weight, and the sampling point p whose weight is greater than the preset target weight is selected. i as a point cloud cluster, and solving the centroid of the point cloud cluster;

[0019] According to the centroid of the first point cloud cluster, a sampling point p in the single-view point cloud model is established. i The Gaussian distribution covariance matrix formed;

[0020] Solve the eigenvalue of the Gaussian distribution covariance matrix, and calculate each sampling point p according to the eigenvalue i The curvature σ i ;

[0021] The step of converting the single-view point cloud model into a high-curvature point cloud model according to the curvature information of each sampling point in each single-view point cloud model is specifically:

[0022] According to the preset curvature threshold σ0 and each of the sampling points p i The curvature σ i , multiple sampling points p in the single-view point cloud model iDetermine the high curvature point, where the high curvature point is the curvature σ i Sampling points above the curvature threshold σ0;

[0023] According to the high curvature points, obtaining the high curvature point cloud model;

[0024] Among them, p i represents the i-th sampling point in the single-view point cloud model, σ i Represents the curvature of the i-th sampling point in the single-view point cloud model.

[0025] As an optional implementation manner of the first aspect of the present application, the step of segmenting the high curvature point cloud model according to the distribution density of sampling points in the high curvature point cloud model to obtain a plurality of high curvature point cloud sub-models is specifically:

[0026] For the sampling point q in the high curvature point cloud model i Perform distance measurement to obtain the sampling point q i The set of sampling points N in the neighborhood with a neighborhood radius of ε ε (q i ), if |N ε (q i )|≥MinPts, then the sampling point q i As the core point, traverse all sampling points in the high curvature point cloud model to obtain a core point set Ω;

[0027] If the core point set Ω is not an empty set, then select a core point q from the core point set Ω j ; The core point queue is Ω cur =q j , the current high curvature point cloud sub-model is C k =q j , update the sampling point set in the high curvature point cloud model to Γ = Γ-q j ;

[0028] If the core point queue Ω cur is not an empty set, then in the core point queue Ω cur Select the core point q j ′, search within the neighborhood radius ε to obtain the core point q j ′ Neighborhood sampling point set N ε (q j ′);

[0029] Let Δ = N ε (q j ′)∩Γ, update the high curvature point cloud sub-model to C k =C k∪Δ, update the sampling point set in the high curvature point cloud model to Γ = Γ-Δ, and update the core point queue to Ω cur =Ω cur ∪(Δ∩Ω)-q j ′, update the core point set to Ω=Ω-Ω cur ;

[0030] If the core point queue Ω cur is an empty set, then the high curvature point cloud sub-model C k Generation completed;

[0031] Among them, the core point queue Ω cur Represents the core points in the high curvature point cloud model for which the domain sampling point set search is required;

[0032] qi represents the i-th sampling point in the high curvature point cloud model, N ε (q i ) represents the neighborhood sampling point set of qi with a neighborhood radius of ε, MinPts represents the threshold of the number of sampling point sets, q j represents the jth core point in the core point set, k represents the sequence number of the high curvature point cloud sub-model, C k represents the kth high curvature point cloud sub-model, Γ represents the set of sampling points in the high curvature point cloud model, q j ′ represents the jth core point in the core point queue, N ε (q j ′) represents point q j ′ is a set of neighborhood sampling points with a neighborhood radius of ε, and Δ represents the number of sampling points in the high curvature point cloud model that are related to the core point q j The set of sampling points whose distance between them is within ε.

[0033] As an optional implementation of the first aspect of the present application, the PAR model is obtained by improving the existing Pointnet model, and the PAR model includes a group sampling layer, a convolutional layer, an attention layer, a pooling layer and a fully connected layer in sequence.

[0034] As an optional implementation of the first aspect of the present application, a portion of the point cloud data in the plurality of high curvature point cloud sub-models is used as the input of the PAR model, and a standard pose transformation matrix is ​​used as the output of the PAR model, and the PAR model is trained to obtain a target PAR model; specifically:

[0035] The multiple high curvature point cloud sub-models are respectively used as input data of the PAR model, and the multiple high curvature point cloud sub-models are processed according to the group sampling layer, the convolution layer and the attention layer to obtain multiple attention weighted feature matrices corresponding to the multiple high curvature point cloud sub-models;

[0036] The pooling layer performs pooling processing on the multiple attention weighted feature matrices through the MaxPooling function to obtain multiple global feature matrices corresponding to the multiple attention weighted feature matrices;

[0037] The fully connected layer learns the multiple global feature matrices through a loss function with the standard pose transformation matrix as the target to obtain the target PAR model.

[0038] As an optional implementation of the first aspect of the present application, the grouped sampling layer includes a sampling sublayer and a search sublayer; the processing of the multiple high curvature point cloud sub-models according to the grouped sampling layer, the convolution layer and the attention layer is specifically:

[0039] Performing farthest point sampling on a single high curvature point cloud sub-model through the sampling sub-layer to obtain a plurality of sampling points in the high curvature point cloud sub-model;

[0040] The search sublayer obtains a set of neighboring points within a neighborhood radius of each of the plurality of sampling points by a ball search method;

[0041] The convolution layer processes the plurality of neighboring point sets corresponding to the plurality of sampling points through a convolution unit to obtain a basic feature matrix corresponding to the single high curvature point cloud sub-model;

[0042] The attention layer operates the basic feature matrix through a self-attention mechanism to obtain an attention weighted feature matrix corresponding to the single high curvature point cloud sub-model;

[0043] By sequentially passing the multiple high curvature point cloud sub-models through the group sampling layer, the convolution layer and the attention layer, a plurality of attention weighted feature matrices corresponding to the high curvature point cloud sub-models are obtained.

[0044] As an optional implementation manner of the first aspect of the present application, the loss function is specifically:

[0045]

[0046] Among them, m represents the number of rows of the global feature matrix, n represents the number of columns of the global feature matrix, and Y ij Represents the element of the i-th row and j-th column of the global feature matrix, T ij Represents the element in the i-th row and j-th column of the standard transformation matrix.

[0047] In a second aspect, an embodiment of the present application provides a multi-view point cloud registration system, the system comprising:

[0048] An acquisition module is used to acquire a single-view point cloud model of the workpiece at different viewpoints;

[0049] A first processing module, configured to process the point cloud model acquired by the first acquisition module according to preset curvature information, delete low-curvature point clouds in the point cloud model, and obtain a high-curvature point cloud model;

[0050] A second processing module, configured to perform component segmentation on the high curvature point cloud model obtained by the first processing module to obtain a plurality of high curvature point cloud sub-models;

[0051] An establishment module, used to establish a PAR model and a training target of the PAR model;

[0052] A training module, used for inputting the multiple high curvature point cloud sub-models obtained by the second processing module into the PAR model established by the first establishment module, training the PAR model according to the training target, and obtaining a target PAR model;

[0053] The error analysis module inputs the multiple high curvature point cloud sub-models into the PAR model trained by the first training module to obtain multiple transformation matrices, analyzes and deletes the transformation matrices whose errors are greater than a preset distance threshold among the multiple transformation matrices, and obtains a rotation matrix;

[0054] The registration module registers the rotation matrices of the multiple viewing angles obtained by the error analysis module to achieve multi-view point cloud registration of the workpiece.

[0055] In a third aspect, an embodiment of the present application provides an electronic device, which includes a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the method described in the first aspect.

[0056] In a fourth aspect, an embodiment of the present application provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the method described in the first aspect are implemented.

[0057] In the embodiments of the present application, by calculating curvature information and screening high curvature points, it is possible to focus on the feature area of ​​the object, thereby improving the accuracy of registration; the use of point cloud segmentation and PAR model enables the method to adapt to workpieces of different complexities and shapes. When the amount of point cloud data is huge and the initial position is unclear, the point cloud data is segmented into multiple and local features, and accurate and fast point cloud registration is achieved through local features; the process of automatically training the PAR model and predicting the transformation matrix reduces manual intervention and improves the efficiency of registration; this method is not only suitable for multi-view point cloud registration of workpieces, but can also be extended to other application fields that require three-dimensional reconstruction and registration, such as cultural relics protection, medical image processing, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is a flow chart of a multi-view point cloud registration method provided by some embodiments of the present application;

[0059] Figure 2 This is a structural diagram of a PAR model in a multi-view point cloud registration method provided in some embodiments of the present application. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0061] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here. In addition, the "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally represents that the objects associated with each other are in an "or" relationship.

[0062] In the following, in combination with the accompanying drawings, a multi-view point cloud registration method, system, device and storage medium provided by the embodiments of the present application are described in detail through specific embodiments and their application scenarios.

[0063] Example

[0064] A multi-view point cloud registration method comprises the following steps:

[0065] S100: Acquire multiple single-view point cloud models of the workpiece under multiple view angles, and calculate the curvature information of each sampling point in the single-view point cloud model;

[0066] It should be understood that the point cloud data of the workpiece is obtained from multiple perspectives through the scanning device, and a single-view point cloud model is generated for each perspective; the curvature information of the sampling points in each single-view point cloud model is calculated. The curvature reflects the degree of curvature of the point cloud surface and is a key feature of the subsequent processing steps.

[0067] Furthermore, the step of calculating the curvature information of each sampling point in the single-view point cloud model in S100 is specifically as follows:

[0068] S110: For any sampling point p in the single-view point cloud model i Set the neighborhood radius and determine the sampling point p i The Euclidean distance to other sampling points within the neighborhood radius;

[0069] S120: Determine the sampling point p based on the Euclidean distance i The weight of

[0070] S130: Each sampling point p in the single-view point cloud model i The weight of the sample point p is compared with the preset target weight, and the sample point p with a weight greater than the preset target weight is selected. i As a point cloud cluster, and solve the centroid of the point cloud cluster;

[0071] S140: According to the centroid of the first point cloud cluster, establish the sampling point p in the single-view point cloud model i The Gaussian distribution covariance matrix is ​​constructed;

[0072] S150: Solve the eigenvalues ​​of the Gaussian distribution covariance matrix and calculate each sampling point p according to the eigenvalues i The curvature σ i ;

[0073] Specifically, the sampling point p in S120 i The weight is calculated by the following formula:

[0074]

[0075] Among them, p i represents the i-th sampling point, p ij represents the sampling point p i The jth sampling point in the neighborhood radius of ij represents the sampling point p i The jth sampling point in the neighborhood radius is relative to the sampling point p i The weight of ||p i -p ij || represents the sampling point p i The jth sampling point in the neighborhood radius is relative to the sampling point p i The Euclidean distance of , ε represents the neighborhood radius;

[0076] The sampling point p whose S130 weight is greater than the preset target weight i Denoted as p ik , the sampling point p in the single-view point cloud model is established in S140 i The Gaussian distribution covariance matrix Σ i as follows:

[0077]

[0078] Among them, ν i represents the point cloud cluster, μ i Represents the centroid of the point cloud cluster;

[0079] The covariance matrix Σ for the Gaussian distribution in S150 i Performing feature decomposition, we get:

[0080]

[0081] Curvature σ i The calculation formula is as follows:

[0082]

[0083] Among them, λ i1 , i2 , i3 is the covariance matrix Σ i Eigenvalues ​​in ascending order, R i is the approximate point cloud cluster ν i The matrix composed of the eigenvectors of the three axes of the ellipsoid.

[0084] It should be understood that, based on the above feature definition, a neighborhood radius is set for each sampling point in the single-view point cloud model. This radius defines a local area around the sampling point for subsequent calculations; then, the Euclidean distance between the sampling point and all other sampling points within its neighborhood radius is calculated; the Euclidean distance is the straight-line distance between two points in space, which is used to measure the relative position relationship between sampling points. The neighborhood radius is set to capture the local geometric features around each sampling point; a weight is assigned to each sampling point based on the Euclidean distance between the sampling point and other sampling points in its neighborhood; the weight is usually related to a certain function of the Euclidean distance (such as an inverse relationship) to reflect the closeness between the sampling point and its neighborhood points; the weight is used to emphasize or weaken the influence of different neighborhood points on the curvature calculation of the sampling point. Neighborhood points with closer distances usually contribute more to the curvature; the weight is compared with the preset target weight, and the sampling points with weights greater than the target weight are screened out to form a point cloud cluster. Then, the centroid of these point cloud clusters is calculated, which is the average position of all points in the point cloud cluster; the point cloud cluster represents a group of sampling points with similar local features. The centroid, as the center of the point cloud cluster, is helpful for the subsequent establishment of the Gaussian distribution model; with the centroid of the point cloud cluster as the center, the covariance matrix of the Gaussian distribution composed of the sampling points is calculated. The covariance matrix describes the degree of discreteness of the sampling points in all directions; Gaussian distribution is a commonly used statistical model used to describe the distribution characteristics of data; the covariance matrix is ​​an important parameter of the Gaussian distribution, which reflects the correlation of the data in all dimensions; solve the eigenvalues ​​of the covariance matrix of the Gaussian distribution, which represent the degree of discreteness (i.e. variance) of the data in all directions. Then, the curvature of each sampling point is calculated based on the eigenvalue; the size of the eigenvalue reflects the speed of change of the data in all directions; in three-dimensional space, the curvature can be calculated by a certain combination of eigenvalues, which describes the degree of curvature of the surface at the sampling point.

[0085] It should be noted that through careful neighborhood analysis and weight calculation, the local geometric features of each sampling point can be captured more accurately, thereby calculating more precise curvature information; the Gaussian distribution model can handle the noise and uncertainty in the data, making the curvature calculation more stable and reliable; this method is suitable for point cloud models of different shapes and complexities, and can flexibly process various three-dimensional data.

[0086] S200: converting the single-view point cloud model into a high-curvature point cloud model according to the curvature information of each sampling point in each single-view point cloud model;

[0087] It should be understood that based on the calculated curvature information, points with higher curvature are screened out, because these points are usually located at the edges or feature parts of the object and are more important to the alignment process; these high curvature points are reassembled into a new point cloud model, namely the high curvature point cloud model.

[0088] Furthermore, in S200, the step of converting the single-view point cloud model into a high-curvature point cloud model according to the curvature information of each sampling point in each single-view point cloud model is specifically as follows:

[0089] S210: According to the preset curvature threshold σ0 and each sampling point p i The curvature σ i , multiple sampling points p in the single-view point cloud model i Determine the high curvature point, where the high curvature point is the curvature σ i Sampling points above the curvature threshold σ0;

[0090] S220: obtaining a high curvature point cloud model according to the high curvature points;

[0091] Among them, p i represents the i-th sampling point in the single-view point cloud model, σ i represents the curvature of the i-th sampling point in the single-view point cloud model, and σ0 represents the preset curvature threshold.

[0092] It should be understood that, based on the above feature definition, for each sampling point p in the single-view point cloud model i , according to its curvature σ i Compare with the preset curvature threshold σ0; if σ i If it is greater than σ0, then the sampling point p i The curvature threshold σ0 is a standard for screening high curvature points. In three-dimensional point cloud data, high curvature points are usually located at the edge, corner or feature-obvious area of ​​the object. These areas are particularly important for subsequent tasks such as registration and reconstruction because they provide more geometric information. By traversing all sampling points in the single-view point cloud model and comparing their curvatures with the curvature threshold one by one, high curvature points can be efficiently screened out. All high curvature points determined in S210 are combined to form a new point cloud model, namely, a high curvature point cloud model. The high curvature point cloud model only contains high curvature points in the original point cloud, so it is more concise and focuses on the feature area of ​​the object. This model usually has higher efficiency and accuracy in subsequent tasks such as registration and feature extraction. Data structures (such as arrays, lists or data structures in point cloud libraries) can be used to store high curvature points and construct corresponding high curvature point cloud models.

[0093] It should be noted that by screening high curvature points, the high curvature point cloud model is simpler than the original point cloud model, reducing the amount of data for subsequent processing; high curvature points are usually located in the characteristic areas of the object, so the high curvature point cloud model can more accurately reflect the geometric characteristics of the object; in subsequent alignment, reconstruction and other tasks, the use of high curvature point cloud models can usually reduce the amount of calculation and improve processing efficiency.

[0094] S300: segmenting the high curvature point cloud model according to the distribution density of sampling points in the high curvature point cloud model to obtain multiple high curvature point cloud sub-models;

[0095] It should be understood that the high curvature point cloud model is divided into multiple smaller sub-models according to the distribution density of the sampling points in the model; the purpose of segmentation is to further simplify the problem so that each sub-model contains fewer points but retains important feature information, which is convenient for subsequent processing steps.

[0096] Furthermore, the segmentation process in S300 is specifically as follows:

[0097] S310: Sampling points q in high curvature point cloud model i Perform distance measurement to obtain the sampling point q i The set of sampling points N in the neighborhood with a neighborhood radius of ε ε (q i ), if N ε (q i )|≥MinPts, then the sampling point q i As the core point, traverse all sampling points in the high curvature point cloud model to obtain the core point set Ω;

[0098] S320: If the core point set Ω is not an empty set, select a core point q from the core point set Ω j ; The core point queue is Ω cur =q j , the current high curvature point cloud sub-model is C k =q j , update the sampling point set in the high curvature point cloud model to Γ = Γ-q j ;

[0099] S321: If the core point queue Ω cur is not an empty set, then in the core point queue Ω cur Select the core point q j ′, search within the neighborhood radius ε to obtain the core point q j ′ Neighborhood sampling point set N ε (q j ′);

[0100] Let Δ = N ε (q j ′)∩Γ, update the high curvature point cloud sub-model to C k =C k ∪Δ, update the sampling point set in the high curvature point cloud model to Γ = Γ-Δ, and update the core point queue to Ω cur =Ω cur ∪(Δ∩Ω)-q j ′, update the core point set to Ω=Ω-Ωcur ;

[0101] S330 If the core point queue Ω cur is an empty set, then the high curvature point cloud sub-model C k Generation completed;

[0102] Among them, the core point queue Ω cur Indicates the core points that need to be searched for domain sampling point set in the high curvature point cloud model;

[0103] qi represents the i-th sampling point in the high curvature point cloud model, N ε (q i ) represents the neighborhood sampling point set of qi with a neighborhood radius of ε, MinPts represents the threshold of the number of sampling point sets, q j represents the jth core point in the core point set, k represents the sequence number of the high curvature point cloud sub-model, C k represents the kth high curvature point cloud sub-model, Γ represents the set of sampling points in the high curvature point cloud model, q j ′ represents the jth core point in the core point queue, N ε (q j ′) represents point q j ′ is a set of neighborhood sampling points with a neighborhood radius of ε, and Δ represents the number of sampling points in the high curvature point cloud model that are related to the core point q j The set of sampling points whose distance between them is within ε.

[0104] It should be understood that, based on the above feature definition, for each sampling point q in the high curvature point cloud model i Perform distance measurement to determine the set of neighborhood sampling points N within a given neighborhood radius ε ε (q i ); if N ε (q i ) is greater than or equal to the preset point threshold MinPts, then the sampling point q i q is regarded as the core point; the entire high curvature point cloud model is traversed, and all core points are collected to form a core point set Ω; the definition of core points is based on their local density in space. By setting the neighborhood radius and point number threshold, those relatively dense points in space can be screened out as core points. These points are usually located in the key parts or feature areas of the object; a core point q is selected from the core point set Ω j As the starting point, add it to the core point queue Ω cur and the current high curvature point cloud sub-model C k And remove the point from the sampling point set Γ of the high curvature point cloud model; when the core point queue Ω cur When it is not empty, select a core point q from the queue j′, search for the neighborhood sampling point set N within its neighborhood radius ε ε (q j ′). Then, calculate N ε (q j The intersection Δ of the current unassigned sampling point set Γ is calculated by adding the points in Δ to the current high curvature point cloud sub-model C. k and remove these points from Γ and Ω; at the same time, update the core point queue Ω cur , add the points in Δ that belong to the core point set Ω to the queue, but remove the processed core points q j ′; Through neighborhood search and intersection calculation, the high curvature point cloud sub-model is gradually expanded. This method is based on the concept of density connectivity, which can ensure that the points in the sub-model are closely connected to each other and reflect the continuity of the local structure of the object; by iteratively processing the core point queue until the queue is empty, the generation of all high curvature point cloud sub-models is completed; when the core point queue Ω cur When empty, it indicates the current high curvature point cloud sub-model C k The generation is completed. At this point, you can start the generation process of the next sub-model or end the entire segmentation process; the core point queue is empty, which means that all core points that are densely connected to the current sub-model have been processed, so it can be considered that a complete, locally continuous high curvature point cloud sub-model has been generated.

[0105] It should be noted that through the segmentation method based on density connectivity, a high-curvature point cloud sub-model that is locally continuous and reflects the characteristic structure of the object can be generated; by adjusting the neighborhood radius ε and the point count threshold MinPts, the granularity and accuracy of the segmentation can be flexibly controlled to adapt to objects of different shapes and complexities; using efficient data structures and algorithms (such as neighborhood search, set operations, etc.), the segmentation process of the high-curvature point cloud model can be completed quickly.

[0106] S400: using a portion of point cloud data in the plurality of high curvature point cloud sub-models as input of the PAR model, using a standard pose transformation matrix as output of the PAR model, and training the PAR model to obtain a target PAR model;

[0107] It is important to understand that the PAR (Point Cloud Area Regression) model is trained using a portion of the high curvature point cloud sub-model data as input and the standard pose transformation matrix as output; the PAR model can predict the pose transformation of the new input point cloud by learning the relationship between the input point cloud and the corresponding pose transformation.

[0108] Furthermore, the PAR model in S400 is improved by the existing Pointnet model. The PAR model includes a group sampling layer, a convolutional layer, an attention layer, a pooling layer and a fully connected layer in sequence.

[0109] It should be understood that, based on the above-mentioned feature limitations, the group sampling layer may be a key step in the preliminary processing of point cloud data, similar to the sampling layer in Pointnet; it may be responsible for grouping and sampling point cloud data to reduce data redundancy while retaining important geometric and shape information; through group sampling, the model can process large-scale point cloud data more efficiently, laying the foundation for subsequent feature extraction and classification tasks; the convolution layer is the core component of the convolutional neural network (CNN), and is also suitable for processing point cloud data; it extracts features from the input data through the convolution kernel matrix, and can capture local features in point cloud data; in the PAR model, the convolution layer is used to extract fine features of point cloud data, providing strong support for subsequent attention mechanisms and classification tasks; the attention layer is an important innovation in the PAR model, which draws on the attention mechanism in natural language processing (NLP); the attention mechanism can simulate long-distance dependencies in sequences , enabling the model to better capture contextual information; in the PAR model, the attention layer may be responsible for capturing global and local dependencies in the point cloud data, thereby improving the classification and recognition performance of the model; the pooling layer is a commonly used operation in CNN, which is used to reduce the spatial size (i.e., height and width) of the feature map while maintaining important feature information; in the PAR model, the pooling layer may be used to reduce the dimension and select features of the feature map output by the convolution layer to reduce the amount of calculation and the number of parameters of the model; at the same time, the pooling layer can also improve the robustness and generalization ability of the model, enabling it to better adapt to different point cloud data; the fully connected layer is a linear feature mapping process in the neural network, which maps multi-dimensional feature inputs to two-dimensional feature outputs; in the PAR model, the fully connected layer may be responsible for combining and classifying the features extracted by the convolution layer and the pooling layer; through the processing of the fully connected layer, the model can output the final classification result or feature vector for subsequent tasks and applications.

[0110] Furthermore, in S400, the PAR model is trained to obtain a target PAR model, specifically:

[0111] S410: using the multiple high curvature point cloud sub-models as input data of the PAR model respectively, processing the multiple high curvature point cloud sub-models according to the group sampling layer, the convolution layer and the attention layer, and obtaining multiple attention weighted feature matrices corresponding to the multiple high curvature point cloud sub-models;

[0112] S420: The pooling layer performs pooling processing on multiple attention weighted feature matrices through the MaxPooling function to obtain multiple global feature matrices corresponding to the multiple attention weighted feature matrices;

[0113] S430: The fully connected layer uses the standard pose transformation matrix as the target to learn multiple global feature matrices through the loss function to obtain the target PAR model;

[0114] Specifically, the loss function is:

[0115]

[0116] Among them, m represents the number of rows of the global feature matrix, n represents the number of columns of the global feature matrix, and Y ij Represents the element of the i-th row and j-th column of the global feature matrix, T ij Represents the element in the i-th row and j-th column of the standard transformation matrix.

[0117] It should be understood that, based on the above-mentioned feature limitations, multiple high-curvature point cloud sub-models are used as inputs of the PAR model respectively; the input point cloud sub-models are grouped and sampled by the group sampling layer to reduce the data size and retain key information; the convolution layer performs convolution operations on the sampled point cloud data to extract local features; the attention layer uses the attention mechanism to weight the features output by the convolution layer, emphasizes important features and suppresses unimportant features, thereby obtaining an attention-weighted feature matrix; the attention-weighted feature matrix is ​​pooled using the MaxPooling function; MaxPooling reduces the spatial dimension of the feature matrix by selecting the maximum value in each feature channel while retaining the most important feature information; after pooling, multiple global feature matrices corresponding to multiple attention-weighted feature matrices are obtained; these global feature matrices contain the overall shape and structural information of the input point cloud sub-models; the standard pose transformation matrix is ​​used as the target, which is usually a known, predefined transformation matrix used to represent the correct posture or position of the point cloud data; a loss function is designed to measure the difference between the global feature matrix output by the model and the standard pose transformation matrix. This loss function can be based on Euclidean distance, cosine similarity or other metrics; the fully connected layer takes the global feature matrix as input, and performs feature combination and classification through the fully connected layer; according to the gradient information of the loss function, the weights and parameters of the PAR model are updated through the back propagation algorithm; the above process is repeated until the loss function converges to a smaller value, and the PAR model obtained at this time is the target PAR model.

[0118] It should be noted that through the combination of group sampling, convolution and attention mechanism, the PAR model can efficiently extract local and global features of point cloud data; the use of MaxPooling for pooling processing improves the robustness of the model to changes in input data; training with the standard pose transformation matrix as the target ensures the accuracy and reliability of the model output.

[0119] Furthermore, in S410, the grouped sampling layer includes a sampling sublayer and a search sublayer; and the plurality of high curvature point cloud submodels are processed according to the grouped sampling layer, the convolution layer, and the attention layer, specifically:

[0120] S411: performing farthest point sampling on a single high curvature point cloud sub-model through a sampling sub-layer to obtain multiple sampling points in the high curvature point cloud sub-model;

[0121] S412: The search sublayer obtains a set of neighboring points within a neighborhood radius by using a ball search method for each of the multiple sampling points;

[0122] S413: The convolution layer processes multiple neighboring point sets corresponding to multiple sampling points through convolution units to obtain a basic feature matrix corresponding to a single high curvature point cloud sub-model;

[0123] S414: The attention layer operates the basic feature matrix through the self-attention mechanism to obtain the attention weighted feature matrix corresponding to the single high curvature point cloud sub-model;

[0124] S415: By repeating S411-S414 in sequence for multiple high curvature point cloud sub-models, multiple attention weighted feature matrices corresponding to the high curvature point cloud sub-models are obtained.

[0125] It should be understood that, based on the above definition, the sampling sublayer is responsible for selecting representative sampling points from a single high curvature point cloud submodel; the farthest point sampling strategy is adopted to gradually select the points farthest from the selected points from the point cloud submodel as new sampling points. This method can ensure that the sampling points are distributed throughout the point cloud space, thereby capturing more geometric information; multiple sampling points in the high curvature point cloud submodel are obtained; the search sublayer is responsible for finding points in its neighborhood for each sampling point to form a set of neighboring points; for each sampling point, a neighborhood radius is set with the point as the center, and then all points are searched within the radius to form a set of neighboring points for the sampling point; for each sampling point, a set of neighboring points in its neighborhood is obtained; the convolution layer is responsible for processing the sampling points and their corresponding neighboring point sets to extract local features; for each sampling point and its neighboring point set, a convolution unit is used for feature extraction. The convolution unit sequentially includes a convolution function (such as Conv2d), a normalization function (such as BatchNorm2d), and a nonlinear activation function (such as ReLU) for feature transformation to obtain a basic feature matrix corresponding to a single high curvature point cloud submodel. This matrix contains the local feature information of the sampling point and its neighboring points; the attention layer is responsible for weighting the basic feature matrix using the self-attention mechanism, emphasizing important features and suppressing unimportant features; the self-attention mechanism assigns a weight to each feature vector by calculating the similarity or correlation between feature vectors. These weights are used to adjust the importance of feature vectors to obtain a weighted feature matrix; the attention-weighted feature matrix corresponding to a single high-curvature point cloud sub-model is obtained. The feature vectors in this matrix have been weighted according to their importance; the above steps are repeated for multiple high-curvature point cloud sub-models; the attention-weighted feature matrix corresponding to each high-curvature point cloud sub-model is obtained. These matrices will be used as inputs for subsequent processing steps to further extract global features and perform classification tasks.

[0126] It should be noted that the farthest point sampling and ball search methods can efficiently extract local features of point cloud data. The self-attention mechanism is used to weight features, which improves the model's ability to capture important features. This processing flow can be applied to multiple high-curvature point cloud sub-models, improving the model's generalization ability and ability to process large-scale point cloud data.

[0127] S500: inputting another part of the point cloud data in the multiple high curvature point cloud sub-models into the target PAR model to obtain multiple transformation matrices;

[0128] It should be understood that the remaining high curvature point cloud sub-model data is input into the trained target PAR model, and the model outputs multiple transformation matrices; these transformation matrices represent the transformation of each sub-model relative to the standard pose.

[0129] S600: Determine a rotation matrix corresponding to the single-view point cloud model according to a distance threshold and a plurality of transformation matrices;

[0130] It should be understood that the rotation matrix corresponding to each single-view point cloud model is calculated based on the distance threshold and multiple transformation matrices; the distance threshold is used to filter and merge similar transformations to obtain a more accurate rotation matrix.

[0131] S700: registering multiple single-view point cloud models of the workpiece according to the rotation matrices corresponding to the multiple viewpoints;

[0132] It should be understood that multiple single-view point cloud models are registered using rotation matrices corresponding to multiple viewpoints, that is, they are transformed into a unified coordinate system to form a complete three-dimensional model.

[0133] According to a multi-view point cloud registration method of this embodiment, first, multiple single-view point cloud models of the workpiece are obtained from different perspectives. These point cloud models are a collection of points in three-dimensional space, representing the shape of the workpiece surface; the curvature information of the sampling points in each single-view point cloud model is calculated; curvature is a quantity that describes the degree of curvature of a curve or surface. For point cloud data, the curvature of each point can be estimated by the geometric relationship of the neighborhood points; based on the curvature information of each sampling point, points with higher curvature are screened out to construct a high curvature point cloud model; high curvature points are usually located at the edge of the object or in areas with obvious features, which are particularly important for the subsequent registration process; according to the distribution density of sampling points in the high curvature point cloud model, it is divided into multiple high curvature point cloud sub-models; this step aims to decompose complex point cloud data into smaller pieces that are easier to handle, so as to perform more accurate registration later; a portion of the high curvature point cloud sub-model data is used as input to train a PAR model, the goal of which is to output a standard pose transformation matrix. The pose transformation matrix is ​​used to describe the position and posture of an object in space. Through training, the model can learn the mapping relationship from point cloud data to the pose transformation matrix. The remaining high-curvature point cloud sub-model data is input into the trained target PAR model, and the model outputs multiple transformation matrices. These transformation matrices reflect the relative position and posture relationship between point cloud models under different perspectives. According to the distance threshold and multiple transformation matrices, the rotation matrix corresponding to each single-view point cloud model is determined. The distance threshold is used to screen the transformation matrix that meets the requirements to ensure the accuracy of the registration. Finally, the rotation matrices corresponding to multiple perspectives are used to register multiple single-view point cloud models of the workpiece. The registration process aligns the point cloud data under different perspectives into a unified coordinate system to generate a complete three-dimensional model.

[0134] Beneficial effects: By calculating curvature information and screening high curvature points, it is possible to focus on the feature area of ​​the object, thereby improving the accuracy of registration; the use of point cloud segmentation and PAR model enables the method to adapt to workpieces of different complexity and shapes. When the amount of point cloud data is huge and the initial position is unclear, the point cloud data is segmented into multiple and local features, and accurate and fast point cloud registration is achieved through local features; the process of automatically training the PAR model and predicting the transformation matrix reduces manual intervention and improves the efficiency of registration; this method is not only suitable for multi-view point cloud registration of workpieces, but can also be extended to other application fields that require three-dimensional reconstruction and registration, such as cultural relics protection, medical image processing, etc.

[0135] It should be noted that the multi-view point cloud registration method provided in the embodiment of the present application can be executed by a multi-view point cloud registration system, or a control module in a multi-view point cloud registration system for executing and loading a multi-view point cloud registration method. In the embodiment of the present application, a multi-view point cloud registration method provided in the embodiment of the present application is described by taking a multi-view point cloud registration system executing and loading a multi-view point cloud registration method as an example.

[0136] A multi-view point cloud registration system, comprising:

[0137] An acquisition module is used to acquire a single-view point cloud model of the workpiece at different viewpoints;

[0138] A first processing module, used to process the point cloud model acquired by the first acquisition module according to preset curvature information, delete low-curvature point clouds in the point cloud model, and obtain a high-curvature point cloud model;

[0139] A second processing module is used to perform component segmentation on the high curvature point cloud model obtained by the first processing module to obtain a plurality of high curvature point cloud sub-models;

[0140] Establishing a module for establishing a PAR model and a training target of the PAR model;

[0141] A training module, used to input the multiple high curvature point cloud sub-models obtained by the second processing module into the PAR model established by the first establishment module, train the PAR model according to the training target, and obtain a target PAR model;

[0142] The error analysis module inputs multiple high curvature point cloud sub-models into the PAR model trained by the first training module to obtain multiple transformation matrices, analyzes and deletes the transformation matrices whose errors are greater than a preset distance threshold among the multiple transformation matrices, and obtains a rotation matrix;

[0143] The registration module registers the rotation matrices of multiple viewpoints obtained by the error analysis module to achieve multi-viewpoint point cloud registration of the workpiece.

[0144] According to a multi-view point cloud registration system of this embodiment, the acquisition module is responsible for acquiring a single-view point cloud model of the workpiece at different viewpoints; this is achieved through three-dimensional scanning technology, which can capture the three-dimensional information of the workpiece surface and form point cloud data; the first processing module processes the acquired point cloud model, deletes the low curvature point cloud according to the preset curvature information, and obtains a high curvature point cloud model. This is because the high curvature point cloud usually contains more feature information and is more critical for the subsequent registration process; the second processing module performs component segmentation on the high curvature point cloud model to obtain multiple high curvature point cloud sub-models. The purpose of component segmentation is to distinguish different parts of the workpiece so that each part can be processed more accurately later; the establishment module is responsible for establishing a PAR model (which may refer to a specific point cloud registration algorithm or model) and its training target. The specific form and training target of the PAR model depend on the design and application requirements of the system; the training module inputs multiple high curvature point cloud sub-models into the PAR model respectively, and trains the PAR model according to the training target to obtain a target PAR model. The training process is to continuously adjust the parameters of the PAR model so that it can better process point cloud data and achieve more accurate registration; the error analysis module inputs multiple high-curvature point cloud sub-models into the trained PAR model to obtain multiple transformation matrices. Then, it analyzes these transformation matrices, deletes the transformation matrices with errors greater than the preset distance threshold, and finally obtains the rotation matrix. The purpose of this step is to ensure that only accurate transformation matrices are used in the subsequent registration process; the registration module registers the rotation matrices of multiple perspectives obtained by the error analysis module to achieve multi-perspective point cloud registration of the workpiece. The registration process is to convert point cloud data from different perspectives into the same coordinate system so that they can align with each other to form a complete three-dimensional model.

[0145] A multi-view point cloud registration system in an embodiment of the present application may be a device, or a component, integrated circuit, or chip in a terminal. The device may be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device may be a mobile phone, a tablet computer, a laptop computer, a PDA, an in-vehicle electronic device, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA), etc., and the non-mobile electronic device may be a server, a network attached storage (NAS), a personal computer (PC), etc., which is not specifically limited in the embodiment of the present application.

[0146] A multi-view point cloud registration system in the embodiment of the present application may be a device having an operating system. The operating system may be an Android operating system, an iOS operating system, or other possible operating systems, which are not specifically limited in the embodiment of the present application.

[0147] The multi-view point cloud registration system provided in the embodiment of the present application can achieve Figure 1 to Figure 2 In order to avoid repetition, the various processes of implementing a multi-view point cloud registration method in the method embodiment will not be repeated here.

[0148] According to a multi-view point cloud registration system of this embodiment, by deleting low-curvature point clouds and performing component segmentation, the system can process point cloud data more accurately, reduce the impact of noise and redundant information, and thus improve the registration accuracy; the PAR model is used for training and optimization, so that the system can adapt to workpieces of different shapes and sizes and enhance the robustness of the registration process; the entire registration process is automated, reducing the need for manual intervention and improving work efficiency; the system can process multi-view point cloud data and is suitable for three-dimensional reconstruction and positioning tasks in complex scenes.

[0149] Optionally, an embodiment of the present application also provides an electronic device, including a processor, a memory, and a program or instruction stored in the memory and executable on the processor. When the program or instruction is executed by the processor, each process of the above-mentioned multi-view point cloud registration method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0150] An embodiment of the present application also provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, each process of the above-mentioned multi-view point cloud registration method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0151] The processor is the processor in the electronic device in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0152] It should be noted that, in this article, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0153] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0154] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A multi-view point cloud registration method, characterized in that: The method comprises: Acquire multiple single-view point cloud models of the workpiece under multiple viewing angles, and calculate the curvature information of each sampling point in the single-view point cloud model; According to the curvature information of each sampling point in each of the single-view point cloud models, the single-view point cloud model is converted into a high-curvature point cloud model; Segmenting the high curvature point cloud model according to the distribution density of sampling points in the high curvature point cloud model to obtain a plurality of high curvature point cloud sub-models; Using a portion of the point cloud data in the plurality of high curvature point cloud sub-models as input of a PAR model, using a standard pose transformation matrix as output of the PAR model, and training the PAR model to obtain a target PAR model; Inputting another part of the point cloud data in the plurality of high curvature point cloud sub-models into the target PAR model to obtain a plurality of transformation matrices; Determine a rotation matrix corresponding to the single-view point cloud model according to a distance threshold and the multiple transformation matrices; Multiple single-view point cloud models of the workpiece are registered according to the rotation matrices corresponding to multiple view angles.

2. A multi-view point cloud registration method according to claim 1, characterized in that: The step of calculating the curvature information of each sampling point in the single-view point cloud model is specifically as follows: For any sampling point p in the single-view point cloud model i Set the neighborhood radius and determine the sampling point p i The Euclidean distance to other sampling points within the neighborhood radius; According to the Euclidean distance, the sampling point p is determined i The weight of Each sampling point p in the single-view point cloud model i The weight is compared with the preset target weight, and the sampling point p whose weight is greater than the preset target weight is selected. i as a point cloud cluster, and solving the centroid of the point cloud cluster; According to the centroid of the first point cloud cluster, a sampling point p in the single-view point cloud model is established. i The Gaussian distribution covariance matrix formed; Solve the eigenvalue of the Gaussian distribution covariance matrix, and calculate each sampling point p according to the eigenvalue i The curvature σ i ; The step of converting the single-view point cloud model into a high-curvature point cloud model according to the curvature information of each sampling point in each single-view point cloud model is specifically: According to the preset curvature threshold σ0 and each of the sampling points p i The curvature σ i , multiple sampling points p in the single-view point cloud model i Determine the high curvature point, where the high curvature point is the curvature σ i Sampling points above the curvature threshold σ0; According to the high curvature points, obtaining the high curvature point cloud model; Among them, p i represents the i-th sampling point in the single-view point cloud model, σ i Represents the curvature of the i-th sampling point in the single-view point cloud model.

3. A multi-view point cloud registration method according to claim 2, characterized in that: The step of segmenting the high curvature point cloud model according to the distribution density of sampling points in the high curvature point cloud model to obtain a plurality of high curvature point cloud sub-models is specifically as follows: For the sampling point q in the high curvature point cloud model i Perform distance measurement to obtain the sampling point q i In the neighborhood sampling point set N with a neighborhood radius of ε ε (q i ), if |N ε (q i )|≥MinPts, then the sampling point q i As the core point, traverse all sampling points in the high curvature point cloud model to obtain a core point set Ω; If the core point set Ω is not an empty set, then select a core point q from the core point set Ω j ; The core point queue is Ω cur =q j , the current high curvature point cloud sub-model is C k =q j , update the sampling point set in the high curvature point cloud model to Γ = Γ-q j ; If the core point queue Ω cur is not an empty set, then in the core point queue Ω cur Select the core point q j ′, search within the neighborhood radius ε to obtain the core point q j ′ Neighborhood sampling point set N ε (q j ′); Let Δ = N ε (q j ′)∩Γ, update the high curvature point cloud sub-model to C k =C k ∪Δ, update the sampling point set in the high curvature point cloud model to Γ = Γ-Δ, and update the core point queue to Ω cur =Ω cur ∪(Δ∩Ω)-q j ′, update the core point set to Ω=Ω-Ω cur ; If the core point queue Ω cur is an empty set, then the high curvature point cloud sub-model C k Generation completed; Among them, the core point queue Ω cur Represents the core points in the high curvature point cloud model for which the domain sampling point set search is required; qi represents the i-th sampling point in the high curvature point cloud model, N ε (q i ) represents the neighborhood sampling point set of qi with a neighborhood radius of ε, MinPts represents the threshold of the number of sampling point sets, q j represents the jth core point in the core point set, k represents the sequence number of the high curvature point cloud sub-model, C k represents the kth high curvature point cloud sub-model, Γ represents the set of sampling points in the high curvature point cloud model, q j ′ represents the jth core point in the core point queue, N ε (q j ′) represents point q j ′ is a set of neighborhood sampling points with a neighborhood radius of ε, and Δ represents the number of sampling points in the high curvature point cloud model that are related to the core point q j The set of sampling points whose distance between them is within ε.

4. A multi-view point cloud registration method according to claim 3, characterized in that: The PAR model is obtained by improving the existing Pointnet model. The PAR model includes a group sampling layer, a convolutional layer, an attention layer, a pooling layer and a fully connected layer in sequence.

5. A multi-view point cloud registration method according to claim 4, characterized in that: The method uses a portion of the point cloud data in the plurality of high curvature point cloud sub-models as the input of the PAR model, uses the standard pose transformation matrix as the output of the PAR model, and trains the PAR model to obtain a target PAR model; specifically: The multiple high curvature point cloud sub-models are respectively used as input data of the PAR model, and the multiple high curvature point cloud sub-models are processed according to the group sampling layer, the convolution layer and the attention layer to obtain multiple attention weighted feature matrices corresponding to the multiple high curvature point cloud sub-models; The pooling layer performs pooling processing on the multiple attention weighted feature matrices through the MaxPooling function to obtain multiple global feature matrices corresponding to the multiple attention weighted feature matrices; The fully connected layer learns the multiple global feature matrices through a loss function with the standard pose transformation matrix as the target to obtain the target PAR model.

6. A multi-view point cloud registration method according to claim 5, characterized in that: The grouped sampling layer includes a sampling sublayer and a search sublayer; the processing of the multiple high curvature point cloud submodels according to the grouped sampling layer, the convolution layer and the attention layer is specifically as follows: Performing farthest point sampling on a single high curvature point cloud sub-model through the sampling sub-layer to obtain a plurality of sampling points in the high curvature point cloud sub-model; The search sublayer obtains a set of neighboring points within the neighborhood radius for each of the plurality of sampling points by a ball search method; The convolution layer processes the plurality of neighboring point sets corresponding to the plurality of sampling points through a convolution unit to obtain a basic feature matrix corresponding to the single high curvature point cloud sub-model; The attention layer operates the basic feature matrix through a self-attention mechanism to obtain an attention weighted feature matrix corresponding to the single high curvature point cloud sub-model; By sequentially passing the multiple high curvature point cloud sub-models through the group sampling layer, the convolution layer and the attention layer, a plurality of attention weighted feature matrices corresponding to the high curvature point cloud sub-models are obtained.

7. A multi-view point cloud registration method according to claim 6, characterized in that: The loss function is specifically: Among them, m represents the number of rows of the global feature matrix, n represents the number of columns of the global feature matrix, and Y ij Represents the element of the i-th row and j-th column of the global feature matrix, T ij Represents the element in the i-th row and j-th column of the standard transformation matrix.

8. A multi-view point cloud registration system, capable of implementing a multi-view point cloud registration method according to any one of claims 1 to 7, characterized in that: The system comprises: An acquisition module is used to acquire a single-view point cloud model of the workpiece at different viewpoints; A first processing module, configured to process the point cloud model acquired by the first acquisition module according to preset curvature information, delete low-curvature point clouds in the point cloud model, and obtain a high-curvature point cloud model; A second processing module, configured to perform component segmentation on the high curvature point cloud model obtained by the first processing module to obtain a plurality of high curvature point cloud sub-models; An establishment module, used to establish a PAR model and a training target of the PAR model; A training module, used for inputting the multiple high curvature point cloud sub-models obtained by the second processing module into the PAR model established by the first establishment module, training the PAR model according to the training target, and obtaining a target PAR model; The error analysis module inputs the multiple high curvature point cloud sub-models into the PAR model trained by the first training module to obtain multiple transformation matrices, analyzes and deletes the transformation matrices whose errors are greater than a preset distance threshold among the multiple transformation matrices, and obtains a rotation matrix; The registration module registers the rotation matrices of the multiple viewing angles obtained by the error analysis module to achieve multi-view point cloud registration of the workpiece.

9. An electronic device, characterized in that: It includes a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein when the program or instruction is executed by the processor, the steps of a multi-view point cloud registration method as described in claims 1-7 are implemented.

10. A readable storage medium, characterized in that: The readable storage medium stores a program or instruction, and when the program or instruction is executed by the processor, the steps of a multi-view point cloud registration method as described in claims 1-7 are implemented.

Images