A Multi-Scale Dynamically Connected 3D Point Cloud Registration Method

Abstract

This invention relates to a multi-scale dynamically connected 3D point cloud registration method, belonging to the field of computer vision and 3D data processing technology. To extract local and global features, this invention designs a multi-scale dynamically connected module to extract local and global features at different scales from the point cloud. To further enhance the feature fusion capability between point clouds, a recurrent attention module is used to fuse the features of two point clouds, generating hybrid features. This invention can effectively capture the local and global features of point clouds, ensuring rich representation of the point cloud data. Simultaneously, the introduction of recurrent attention enhances the feature fusion capability between point clouds, further improving the accuracy and robustness of registration. The multi-scale dynamically connected 3D point cloud registration method proposed in this invention uses deep learning technology to achieve registration between point clouds, exhibiting high accuracy and robustness, and has significant application value.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and 3D data processing technology, specifically relating to a multi-scale dynamically connected 3D point cloud registration method. Background Technology

[0002] Point cloud registration is a fundamental and crucial technology in the field of 3D computer vision. Its core objective is to achieve precise alignment of point cloud data from different coordinate systems through spatial transformation. With the rapid development of 3D sensing technology, point cloud registration is playing an increasingly important role in areas such as autonomous driving environmental perception, industrial 3D inspection, and digital twin modeling.

[0003] Traditional point cloud registration methods are mainly divided into two categories: iterative optimization-based methods and feature matching-based methods. The Iterative Closest Point (ICP) algorithm, as the most representative iterative method, achieves registration by alternating between nearest-point search and least-squares optimization. However, its performance heavily relies on initial pose estimation and is prone to getting trapped in local optima. To address this issue, researchers have proposed various improved algorithms, such as feature-weighted ICP variants and branch-and-bound Go-ICP, but these still suffer from low computational efficiency when processing large-scale point clouds. Feature matching-based methods (such as FPFH and SHOT) establish point correspondences by extracting hand-designed local feature descriptors. These methods perform well in feature-rich regions, but their performance drops sharply in flat regions or under noise interference. In recent years, deep learning methods have brought revolutionary progress to point cloud registration. PointNet was the first to achieve end-to-end point cloud feature learning, but its global feature extraction method ignores local geometric structures. Subsequent versions of PointNet++ improved local feature representation capabilities through scaled feature extraction. Works such as DCP and RPMNet have further introduced attention mechanisms into point cloud registration, improving the accuracy of feature matching. However, existing deep learning methods still have three main limitations: first, the feature extraction network does not adequately model the local geometric relationships of point clouds; second, the design of feature interaction mechanisms between point clouds is relatively simple; and third, robustness to noise and partially overlapping scenes needs to be improved. These technical bottlenecks limit the application effectiveness of existing methods in complex real-world scenarios.

[0004] Currently, with the increasing demands for 3D perception accuracy from applications such as autonomous driving and Industry 4.0, developing point cloud registration algorithms with stronger feature representation capabilities and higher robustness has become a key research direction of common concern in academia and industry. Especially when dealing with large-scale, high-noise, and low-overlap-rate real-world point cloud data, existing methods still need to find a better balance between algorithm accuracy and computational efficiency. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-scale dynamically connected 3D point cloud registration method. It employs a multi-scale feature joint extraction strategy, significantly improving the accuracy of point cloud registration in complex scenes through the synergistic fusion of local, regional, and global features. Simultaneously, a cyclic attention optimization mechanism is used to enhance the feature matching ability between point clouds through multiple rounds of feature interaction, avoiding the performance degradation of traditional methods under partial overlap or noise interference. Furthermore, a dynamic weight adjustment mechanism is used to accurately solve the registration parameters, improving the accuracy and robustness of point cloud registration.

[0006] This invention is achieved through the following method: a multi-scale dynamically connected 3D point cloud registration method, the steps of which are as follows:

[0007] 1) Import point cloud dataset: A point cloud is a collection of a large number of three-dimensional points, each containing three-dimensional coordinates (x, y, z), which together form a point cloud dataset.

[0008] The ModelNet40 dataset is imported into the model. The ModelNet40 dataset consists of multiple txt files, each of which represents a complete point cloud. A point cloud is a collection of three-dimensional points.

[0009] 2) Divide the point cloud dataset: Use 70% of the point cloud dataset as the training dataset, 10% as the test dataset, and 20% as the validation dataset.

[0010] 3) Preprocess the point cloud data: Denote the point clouds in the point cloud dataset from 1) as... , For source point cloud; A rigid transformation is denoted as , Using a reference point cloud, perform rotation and translation operations on the point cloud, randomly drawing a rigid transformation; for rotations along each axis, sample uniformly within [0, 45°], and for translations, sample within [-0.5, 0.5]. The rigid transformation uses a 4×4 homogeneous transformation matrix. This means that the matrix can simultaneously describe rotation and translation operations. The true transformation matrix; source point cloud Reference point cloud As input to the model.

[0011] 4) Construct a point cloud registration model based on deep learning and multi-scale dynamic connectivity: The model input is the point cloud obtained in step 2). , The output is the transformation matrix obtained after model training. Multi-scale dynamic connectivity connects features at different scales, fusing local and global features to extract effective features. After extraction, these features are input into a recurrent attention module, interacting with features from two point clouds to ensure the features possess both unique characteristics within and common features between the point clouds. This iterative interaction yields the final features, which are then used to calculate the transformation matrix Rt using the SVD algorithm. The transformation matrix predicted by the model is then used to... With the true transformation matrix By comparing the two matrices, the Adam optimization algorithm is used to backpropagate and update the weight matrix, thereby updating the model, reducing the error between the two matrices, and obtaining the final transformation matrix.

[0012] 4.1 Using multi-scale dynamic connectivity to extract features at different scales

[0013] For point clouds any point in Selected by K-nearest neighbor algorithm The nearest points constitute neighborhood , This means that subtracting the coordinates of the center point from the coordinates of the nearest points yields the coordinates of the nearest points within a local area. Represented as a normal feature, it describes the angle information between the center point and its neighboring points. and Represented as and The normal information;

[0014] ;

[0015] ; ;

[0016] for With the A 10-dimensional feature vector formed by connecting its nearest neighbors;

[0017] ;

[0018] feature This is represented as the feature information of each point;

[0019] ;

[0020] Obtain initial features Then, the local features are obtained by multilayer perceptron (MLP) for dimensionality upscaling. These local features are then concatenated with the initial features and used as input features in the next MLP layer. For point clouds... Each time MLP changes The size is used to adjust the range of the initial features. Multiple MLPs are performed to obtain local features at different scales. Finally, the features at each scale are concatenated and passed through the last MLP layer to obtain the final global features. Reference point cloud Perform the same operation;

[0021] 4.2 Extracting Mixed Features Using Recurrent Attention

[0022] After feature extraction via a convolutional network, the feature information of a single point cloud is fully learned, but the cross-fusion features between two point clouds are not learned. The input is fed into a recurrent attention module to extract cross-fusion features, enhancing the cross-fusion features of the point cloud. The recurrent attention module consists of a self-attention module for learning features within the point cloud and a cross-fusion attention module for extracting mixed features between point clouds. The two modules interleave N times, finally extracting the mixed features. and ;

[0023] 4.2.1 Self-Attention Module

[0024] The self-attention module is used to learn the correlation features between points within each point cloud. Given the input feature matrix Output feature matrix It is a weighted sum of all projected input features, which enhances the features of points by learning feature relationships within the point cloud. For the number of points, For feature dimensions, refer to point clouds The calculation method is the same;

[0025] ;

[0026] Weighting coefficient By scoring attention Obtained by row-by-row softmax. The calculation results are as follows:

[0027] ;

[0028] These represent query, key, and value, respectively. This is the matrix transpose.

[0029] 4.2.2 Cross-integration Attention Module

[0030] A feature-based cross-fusion attention module is used to fuse two point cloud features, extracting common information from the two point cloud features, based on... Calculate the self-attention feature matrix ,calculate Cross-fusion attention feature matrix Point cloud The method of cross-fusion feature calculation and same;

[0031] ;

[0032] ;

[0033] The self-attention feature module further extracts relevant features for each point cloud, while the cross-fusion attention feature module extracts mixed features between two point clouds. This process is repeated multiple times in an alternating loop. and Adding them together yields the final feature. Point cloud Similarly;

[0034] ;

[0035] 4.3 SVD Solution

[0036] According to the source cloud and reference point cloud The correspondence between the two point clouds is as follows:

[0037] ;

[0038] Where R is the rotation matrix, T is the translation vector, and Ni is the noise vector added to the experiment to match reality; the solution steps are as follows: find the centroids of the two point clouds respectively, then find the displacement vector of each point relative to the centroid, use the matrix obtained by the centroid displacement vector to perform SVD decomposition, and obtain the transformation matrix based on the decomposition result. ;

[0039] Will and The centroid is defined as:

[0040] ;

[0041] ;

[0042] cross-covariance matrix for:

[0043] ;

[0044] After continuously optimizing the features of the initial point cloud to obtain the final transformation matrix, it is input into the Singular Value Decomposition (SVD) module for decomposition. The transformation matrix Rt is obtained;

[0045] ;

[0046] because The transformation matrix contains information about the rotation and translation directions. Decompose it into a rotation matrix R and a translation vector t. The specific method is as follows: take the transformation matrix... The first three rows and first three columns are the rotation matrix R, and the first three elements of the fourth column are the translation vector t; R and t are the final outputs.

[0047] 5) Input the point cloud dataset into the point cloud registration model in 4) for training and testing, and save the model with the best performance.

[0048] The Adam optimization algorithm is used to train the model weight matrix, and the loss function is:

[0049] ;

[0050] For actual rotation and translation, and The rotation and translation parameters output by the model are used to evaluate the model's performance by assessing the errors between the actual rotation and the predicted rotation, and between the actual translation and the predicted translation, thereby reducing the complexity of the evaluation function.

[0051] 6) Input the point cloud data to be detected into the trained registration model, perform point cloud registration, and save the registration results.

[0052] During training, the loss of the test results is calculated, using mean squared error (MSE), root mean square error (RMSE), and mean absolute error (MAE) to measure the error between the true and predicted values; MSE(R) is the mean squared error of the rotation matrix, while MSE(t) is the average error of the translation matrix; if the rigid alignment meets the criteria, these error metrics are zero, as shown in the following formula:

[0053] ;

[0054] ;

[0055] ;

[0056] in, For the true value, For fitted values, The number of points.

[0057] The beneficial effects of this invention are as follows:

[0058] 1. By fusing local and global features at different scales through a multi-scale dynamic connection mechanism, the richness and effectiveness of feature representation are effectively improved, enabling the model to capture multi-level spatial structure information in point cloud data more accurately, thereby significantly improving point cloud registration accuracy and significantly reducing registration error compared to traditional methods.

[0059] 2. The introduction of the recurrent attention module enables deep interaction between dual point cloud features, which not only preserves the unique features within the point cloud but also strengthens the common features between point clouds. This significantly enhances the model's adaptability to complex scenes (such as noise interference and partial occlusion) and improves the robustness of the registration results.

[0060] 3. By combining the efficient backpropagation mechanism of the Adam optimization algorithm, model parameter updates become more targeted, significantly shortening the training cycle while ensuring registration accuracy and improving the overall efficiency of point cloud registration. This is particularly suitable for real-time processing scenarios involving large-scale point cloud data. This invention provides a more accurate, efficient, and robust point cloud registration solution for fields such as autonomous driving, 3D reconstruction, and robot navigation, effectively promoting the implementation and development of related technologies in practical applications. Attached Figure Description

[0061] Figure 1 This is a diagram of the overall architecture of the model.

[0062] Figure 2 This is the feature extraction map.

[0063] Figure 3 This is a cyclic attention graph.

[0064] Figure 4 This is a diagram of the attention structure. Detailed Implementation

[0065] A multi-scale dynamically connected 3D point cloud registration method, the steps of which are as follows:

[0066] 1) Import point cloud dataset: A point cloud is a collection of a large number of three-dimensional points, each containing three-dimensional coordinates (x, y, z), which together form a point cloud dataset.

[0067] The ModelNet40 dataset is imported into the model. The ModelNet40 dataset consists of multiple txt files, each of which represents a complete point cloud. A point cloud is a collection of three-dimensional points.

[0068] 2) Divide the point cloud dataset: Use the first 70% of the point cloud dataset as the training dataset, 10% as the test dataset, and the last 20% as the validation dataset.

[0069] 3) Preprocess the point cloud data: Denote the point clouds in the point cloud dataset from 1) as... , For source point cloud; A rigid transformation is denoted as , Using a reference point cloud, perform rotation and translation operations on the point cloud, randomly drawing a rigid transformation; for rotations along each axis, sample uniformly within [0, 45°], and for translations, sample within [-0.5, 0.5]. The rigid transformation uses a 4×4 homogeneous transformation matrix. This means that the matrix can simultaneously describe rotation and translation operations. The true transformation matrix; source point cloud Reference point cloud As input to the model.

[0070] 4) Construct a point cloud registration model based on deep learning with multi-scale dynamic connectivity. The model architecture is as follows: Figure 1 As shown: The point cloud obtained by model input 2). , The output is the transformation matrix obtained after model training. Multi-scale dynamic connectivity connects features at different scales, fusing local and global features to extract effective features. After feature extraction, these features are input into a recurrent attention module, such as... Figure 3 As shown, the features of two point clouds are interacted to include both unique features within the point clouds and common features between them. This interaction is repeated to obtain the final features, which are then used to calculate the transformation matrix Rt using the SVD algorithm. The transformation matrix predicted by the model is then used to calculate the final features. With the true transformation matrix By comparing the two matrices, the Adam optimization algorithm is used to backpropagate and update the weight matrix, thereby updating the model, reducing the error between the two matrices, and obtaining the final transformation matrix.

[0071] 4.1 Using multi-scale dynamic connectivity to extract features at different scales

[0072] For point clouds any point in Selected by K-nearest neighbor algorithm The nearest points constitute neighborhood , This means that subtracting the coordinates of the center point from the coordinates of the nearest points yields the coordinates of the nearest points within a local area. Represented as a normal feature, it describes the angle information between the center point and its neighboring points. and Represented as and The normal information;

[0073] ;

[0074] ;

[0075] ;

[0076] for With the A 10-dimensional feature vector formed by connecting its nearest neighbors;

[0077] ;

[0078] The features are represented as the feature information of each point;

[0079] ;

[0080] After obtaining the initial features, they are upscaled using a multilayer perceptron (MLP) to obtain local features. These local features are then concatenated with the initial features and used as input features in the next MLP layer. For each point cloud iteration, the MLP modifies... The size is used to adjust the range of the initial features. Multiple MLPs are performed to obtain local features at different scales. Finally, the features at each scale are concatenated and passed through the last MLP layer to obtain the final global features. Reference point cloud Perform the same operation;

[0081] 4.2 Extracting Mixed Features Using Recurrent Attention

[0082] Feature extraction is performed using a convolutional network, and the feature extraction process is as follows: Figure 2 As shown, the feature information of a single point cloud is fully learned, but the cross-fusion features between two point clouds are not learned, and the extracted features... The input is fed into the recurrent attention module to extract cross-fusion features, enhancing the cross-fusion features of the point cloud. The recurrent attention consists of a self-attention module for learning features within the point cloud and a cross-fusion attention module for extracting mixed features between point clouds. The attention structure is as follows: Figure 4 As shown, the two modules iterate over each other N times, and finally extract the hybrid features. and ;

[0083] 4.2.1 Self-Attention Module

[0084] The self-attention module is used to learn the correlation features between points within each point cloud. Given the input feature matrix Output feature matrix It is a weighted sum of all projected input features, which enhances the features of points by learning feature relationships within the point cloud. For the number of points, For feature dimensions, refer to point clouds The calculation method is the same;

[0085] ;

[0086] Weighting coefficient By scoring attention Obtained by row-by-row softmax. The calculation results are as follows:

[0087] ;

[0088] These represent query, key, and value, respectively. This is the matrix transpose.

[0089] 4.2.2 Cross-integration Attention Module

[0090] A feature-based cross-fusion attention module is used to fuse two point cloud features, extracting common information from the two point cloud features, based on... Calculate the self-attention feature matrix ,calculate Cross-fusion attention feature matrix Point cloud The method of cross-fusion feature calculation and same;

[0091] ;

[0092] ;

[0093] The self-attention feature module further extracts relevant features for each point cloud, while the cross-fusion attention feature module extracts mixed features between two point clouds. This process is repeated multiple times in an alternating loop. and Adding them together yields the final features; point cloud Similarly;

[0094] ;

[0095] 4.3 SVD Solution

[0096] According to the source cloud and reference point cloud The correspondence between the two point clouds is as follows:

[0097] ;

[0098] Where R is the rotation matrix, T is the translation vector, and Ni is the noise vector added to the experiment to match reality; the solution steps are as follows: find the centroids of the two point clouds respectively, then find the displacement vector of each point relative to the centroid, use the matrix obtained by the centroid displacement vector to perform SVD decomposition, and obtain the transformation matrix based on the decomposition result. ;

[0099] Will and The centroid is defined as:

[0100] ;

[0101] ;

[0102] cross-covariance matrix for:

[0103] ;

[0104] After continuously optimizing the features of the initial point cloud to obtain the final transformation matrix, it is input into the Singular Value Decomposition (SVD) module for decomposition. The transformation matrix Rt is obtained;

[0105] ;

[0106] because The transformation matrix contains information about the rotation and translation directions. Decompose it into a rotation matrix R and a translation vector t. The specific method is as follows: take the transformation matrix... The first three rows and first three columns are the rotation matrix R, and the first three elements of the fourth column are the translation vector t; R and t are the final outputs.

[0107] 5) Input the point cloud dataset into the point cloud registration model in 4) for training and testing, and save the model with the best performance.

[0108] The Adam optimization algorithm is used to train the model weight matrix, and the loss function is:

[0109] ;

[0110] For actual rotation and translation, and The rotation and translation parameters output by the model are used to evaluate the model's performance by assessing the errors between the actual rotation and the predicted rotation, and between the actual translation and the predicted translation, thereby reducing the complexity of the evaluation function.

[0111] 6) Input the point cloud data to be detected into the trained registration model, perform point cloud registration, and save the registration results.

[0112] During training, the loss of the test results is calculated, using mean squared error (MSE), root mean square error (RMSE), and mean absolute error (MAE) to measure the error between the true and predicted values; MSE(R) is the mean squared error of the rotation matrix, while MSE(t) is the average error of the translation matrix; if the rigid alignment meets the criteria, these error metrics are zero, as shown in the following formula:

[0113] ;

[0114] ;

[0115] ;

[0116] in, For the true value, For fitted values, The number of points.

[0117] Example: Training and testing using the ModelNet40 and ShapeNet datasets:

[0118] Step 1: Import ModelNet40 and ShapeNet data, process the data, and divide it into training, validation and test sets in a ratio of 7:1:2.

[0119] Step 2: Set up the model running environment and parameters

[0120] The model evaluation experiment was implemented in PyTorch and Python 3.8, using an Intel® Core™ i9-10900K CPU @3.70GHz processor and an NVIDIA GeForce RTX 3080 Ti (12GB) graphics card. The model parameters were set as follows: 150 iterations, 16 batches, and an optimizer learning rate of 0.001, which decayed by a factor of 0.1 every 75 epochs.

[0121] Step 3: Construct a deep learning-based multi-scale dynamic connectivity point cloud registration model. Model structure: A multi-scale dynamic connectivity module is used to extract local and global features from point clouds at different scales. To further enhance the feature fusion capability between point clouds, a recurrent attention mechanism module is used to fuse the features of two point clouds to generate hybrid features. The final features are input into the SVD module to obtain the final rotation matrix and translation vector. The final result is compared with the true result, and the Adam optimization algorithm is used for backpropagation to update the weight matrix.

[0122] The results show that the model proposed in this invention has smaller errors in the rotation direction and translation direction on the point cloud dataset used, and has better performance.

[0123] The point cloud registration method based on multi-scale dynamic connectivity proposed in this invention uses deep learning technology to achieve registration between point clouds, which has high accuracy and robustness and has good application value.

Claims

1. A multi-scale dynamically connected 3D point cloud registration method, characterized in that, The steps are as follows: Step 1) Import point cloud dataset: A point cloud is a collection of a large number of three-dimensional points, each containing three-dimensional coordinates (x, y, z), which together form a point cloud dataset; Step 2) Divide the point cloud dataset: Divide the point cloud dataset into a training dataset, a test dataset, and a validation dataset; Step 3) Preprocess the point cloud data: Denote the point cloud data in the point cloud dataset from Step 1) as... , For source point cloud; A rigid transformation is denoted as , For reference point cloud; rigid transformation uses a 4×4 homogeneous transformation matrix. This means that the matrix can simultaneously describe rotation and translation operations. The true transformation matrix; source point cloud Reference point cloud As input to the model; Step 4) Construct a point cloud registration model based on deep learning and multi-scale dynamic connectivity: The model input is the result obtained in step 2). , The output is the transformation matrix obtained after model training. Multi-scale dynamic connectivity connects features at different scales, fusing local and global features to extract effective features. After extraction, these features are input into a recurrent attention module, interacting with features from two point clouds to ensure the features possess both unique characteristics within each cloud and common features between them. This iterative interaction yields the final features, which are then used to calculate the transformation matrix using the SVD algorithm. ; Transformation matrix predicted by the model With the true transformation matrix By comparing the two matrices, the Adam optimization algorithm is used to backpropagate and update the weight matrix, thereby updating the model, reducing the error between the two matrices, and obtaining the final transformation matrix. Step 5) Input the point cloud dataset into the point cloud registration model in Step 4) for training and testing, and save the model with the best performance; Step 6) Input the point cloud data to be detected into the trained registration model, perform point cloud registration, and save the registration results.

2. The multi-scale dynamic connection three-dimensional point cloud registration method according to claim 1, characterized in that, In step 1), the ModelNet40 dataset is imported into the model. The ModelNet40 dataset consists of multiple txt files, each of which represents a complete point cloud. A point cloud is a collection of three-dimensional points.

3. The multi-scale dynamic connection three-dimensional point cloud registration method according to claim 1, characterized in that, In step 1), the point cloud is rotated and translated, and a rigid transformation is randomly drawn; the rotation along each axis is uniformly sampled within [0, 45°], and the translation is sampled within [-0.5, 0.5].

4. The multi-scale dynamic connection three-dimensional point cloud registration method according to claim 1, characterized in that, In step 3), the first 70% of the point cloud dataset is used as the training dataset, 10% as the test dataset, and the last 20% as the validation dataset.

5. The multi-scale dynamic connection three-dimensional point cloud registration method according to claim 1, characterized in that, In step 4), the specific method is as follows: Step 4.1 Extract features at different scales using multi-scale dynamic connectivity. For point clouds any point in Selected by K-nearest neighbor algorithm The nearest points constitute The neighborhood, This means that subtracting the coordinates of the center point from the coordinates of the nearest points yields the coordinates of the nearest points within a local area. Represented as normal features, describing the angular information between the center point and its neighboring points, and Represented as and Normal information; ； ； ； for With the A 10-dimensional feature vector formed by connecting its nearest neighbors; ； feature This is represented as the feature information of each point; ； Obtain initial features Then, the local features are obtained by multilayer perceptron (MLP) for dimensionality upscaling. These local features are then concatenated with the initial features and used as input features in the next MLP layer. For point clouds... Each time MLP changes The size is used to adjust the range of the initial features. Multiple MLPs are performed to obtain local features at different scales. Finally, the features at each scale are concatenated and passed through the last MLP layer to obtain the final global features. Reference point cloud Perform the same operation; Step 4.2 Extracting mixed features using recurrent attention After feature extraction via a convolutional network, the feature information of a single point cloud is fully learned, but the cross-fusion features between two point clouds are not learned. The input is fed into a recurrent attention module to extract cross-fusion features, enhancing the cross-fusion features of the point cloud. The recurrent attention module consists of a self-attention module for learning features within the point cloud and a cross-fusion attention module for extracting mixed features between point clouds. The two modules interleave N times, finally extracting the mixed features. and ; Step 4.2.1 Self-attention module The self-attention module is used to learn the correlation features between points within each point cloud. Given the input feature matrix Output feature matrix It is a weighted sum of all projected input features, which enhances the features of points by learning feature relationships within the point cloud. For the number of points, For feature dimensions, refer to point clouds The calculation method is the same; ； Weighting coefficient By scoring attention Obtained by row-by-row softmax. The calculation results are as follows: ； These represent query, key, and value, respectively. This is the matrix transpose. Step 4.2.2 Cross-fusion attention module A feature-based cross-fusion attention module is used to fuse two point cloud features, extracting common information from the two point cloud features, based on... Calculate the self-attention feature matrix ,calculate Cross-fusion attention feature matrix Point cloud The method of cross-fusion feature calculation and same; ； The self-attention feature module further extracts relevant features for each point cloud, while the cross-fusion attention feature module extracts mixed features between two point clouds. This process is repeated multiple times in an alternating loop. and Adding them together yields the final feature. Point cloud Similarly; ； Step 4.3 SVD Solution According to the source cloud and reference point cloud The correspondence between the two point clouds is as follows: ； Where R is the rotation matrix, T is the translation vector, and Ni is the noise vector added to the experiment to match reality; the solution steps are as follows: find the centroids of the two point clouds respectively, then find the displacement vector of each point relative to the centroid, use the matrix obtained by the centroid displacement vector to perform SVD decomposition, and obtain the transformation matrix based on the decomposition result. ; Will and The centroid is defined as: ； ； cross-covariance matrix for: ； After continuously optimizing the features of the initial point cloud to obtain the final transformation matrix, it is input into the Singular Value Decomposition (SVD) module for decomposition. The transformation matrix Rt is obtained; ； because The transformation matrix contains information about the rotation and translation directions. Decompose it into a rotation matrix R and a translation vector t. The specific method is as follows: take the transformation matrix... The first three rows and first three columns are the rotation matrix R, and the first three elements of the fourth column are the translation vector t; R and t are the final outputs.

6. The multi-scale dynamic connection three-dimensional point cloud registration method according to claim 1, characterized in that, In step 5), the specific method is as follows: The Adam optimization algorithm is used to train the model weight matrix, and the loss function is: ； For actual rotation and translation, and The rotation and translation parameters output by the model are used to evaluate the model's performance by assessing the errors between the actual rotation and the predicted rotation, and between the actual translation and the predicted translation, thereby reducing the complexity of the evaluation function.

7. The multi-scale dynamic connection three-dimensional point cloud registration method according to claim 1, characterized in that, In step 6), the specific method is as follows: During training, the loss of the test results is calculated, using mean squared error (MSE), root mean square error (RMSE), and mean absolute error (MAE) to measure the error between the true and predicted values; MSE(R) is the mean squared error of the rotation matrix, while MSE(t) is the average error of the translation matrix; if the rigid alignment meets the criteria, these error metrics are zero, as shown in the following formula: ； ； ； in, For the true value, These are the fitted values. The number of points.

Images