A point cloud registration method based on hierarchical dynamic graph convolution and accelerated nearest neighbor search

Abstract

A point cloud registration method based on hierarchical dynamic graph convolution and accelerated nearest neighbor search: input source point cloud and target point cloud, first, the hierarchical dynamic graph convolution network HDGNet is used to extract deep geometric features, through multi-scale edge convolution, dense connection and residual connection, the multi-level feature representation from local to global is obtained. In feature extraction, KD-Tree is used instead of traditional KNN for nearest neighbor search, the complexity is reduced from O(N square) to O(N log N), and batch parallel search is supported. In feature matching, the attention mechanism and the transformer are used to calculate the soft pointer matrix, and the soft correspondence relationship is generated, and the rigid transformation is solved by the differentiable SVD, avoiding the local optimal problem of ICP. Finally, the cyclic consistency loss and multi-scale feature fusion are used to enhance the robustness under low overlap rate and noise. The present application considers the registration accuracy and computational efficiency.

Description

Technical Field

[0001] This invention belongs to the field of 3D point cloud registration, and specifically relates to a method based on deep learning and geometric feature matching. It proposes a point cloud registration method based on hierarchical dynamic graph convolution and accelerated nearest neighbor search. Background Technology

[0002] As a fundamental task in the fields of computer vision and robotics, 3D point cloud registration aims to align point cloud data acquired from different perspectives or coordinate systems into a unified global coordinate system. Its registration accuracy directly affects the performance of applications such as 3D reconstruction, autonomous driving, robot navigation, and augmented reality.

[0003] Early point cloud registration methods primarily relied on manually designed geometric features, such as point feature histograms and fast point feature histograms, combined with iterative optimization strategies like random sampling consistency to find initial corresponding points, followed by fine registration using the iterative nearest-point algorithm. However, these traditional methods heavily depend on initial pose estimation. When the point cloud overlap rate is low, noise is high, or the initial pose deviation is large, the iterative nearest-point algorithm is prone to getting trapped in local optima, leading to registration failure.

[0004] With the development of deep learning technology, point cloud registration methods based on deep neural networks have made significant progress. Point cloud registration networks extract deep geometric features of point clouds through deep convolutional networks, calculate soft correspondences using attention mechanisms, and solve rigid transformation parameters in one step using a differentiable singular value decomposition module, avoiding the iterative optimization process of traditional iterative nearest-point algorithms. However, existing methods still have the following technical bottlenecks: First, feature extraction capabilities are limited. Existing methods mostly use dynamic graph convolutional networks as feature extractors, which capture the geometric features of point clouds by constructing local neighborhood graphs and stacking multiple layers of edge convolutions. However, traditional dynamic graph convolutional networks lack effective information reuse mechanisms between layers, and low-level detailed features are easily lost during layer-by-layer transmission. Furthermore, single-scale feature representations cannot simultaneously capture local geometric details and global shape structure, limiting further improvements in registration accuracy. Second, computational efficiency is low. Existing methods generally use the K-nearest neighbor search algorithm when constructing local neighborhood graphs, which has a time complexity of O(N²), where N is the number of points in the point cloud. When processing large-scale point cloud data, the computational cost of K-nearest neighbor search becomes the performance bottleneck of the entire registration process, making it difficult to meet real-time requirements. Third, the feature fusion method is simplistic. Existing methods only use the output of the last layer of the dynamic graph convolutional network as the point cloud feature representation, failing to fully utilize multi-scale feature information at different levels, resulting in insufficient fusion of local detail features and global structural features.

[0005] In summary, how to improve the computational efficiency of feature extraction and neighborhood search while maintaining registration accuracy, and how to fully utilize multi-scale feature information to enhance the geometric representation capability of point clouds, are technical problems that urgently need to be solved in this field. Summary of the Invention

[0006] To address the limitations of existing technologies in feature extraction, computational efficiency, and multi-scale feature fusion, this invention proposes a point cloud registration method based on hierarchical dynamic graph convolution and accelerated nearest neighbor search. This method effectively captures the local geometric details and global shape structure of point clouds, significantly improves neighborhood search efficiency, and achieves end-to-end high-precision and high-efficiency rigid transformation estimation.

[0007] This invention is achieved through the following technical solution: a point cloud registration method based on hierarchical dynamic graph convolution and accelerated nearest neighbor search, comprising the following steps: Step 1 Point cloud feature extraction: The hierarchical dynamic graph convolutional network HDGNet is used as the point cloud feature encoder to convert the source point cloud and target point cloud to be registered into a multi-level geometric feature representation. The source point cloud and target point cloud to be registered are represented as follows: Where N is the number of points in the point cloud; firstly, a spatial index tree is constructed for the source point cloud P and the target point cloud Q using a KD-Tree data structure, and for each target point... By using a KD-Tree to search for the k nearest neighbors in the source point cloud, a neighborhood index set is obtained. Computational complexity A replacement for traditional KNN Complexity; secondly, for each point And its neighboring points, construct edge features To obtain the edge feature set edge features The input is fed into HDGNet, which contains multiple edge convolutional modules. The calculation process of the l-th edge convolutional module is as follows: , in , To further refine the neighborhood difference features, the outputs of edge convolutions at different levels are concatenated through channels: in Finally, an output convolutional layer maps the fused features to the target dimension:

[0008] in These are the geometric feature representations of the source point cloud and the target point cloud, respectively.

[0009] Step 2 Feature Interaction and Soft Pointer Matching: In the feature interaction channel, the source point cloud features and target point cloud features extracted in Step 1 are input into the Transformer encoder. The multi-head self-attention mechanism is used to capture the global context information inside the point cloud. At the same time, the soft pointer matrix between the source point cloud and the target point cloud is calculated by using the cross-attention mechanism to generate a feature-based soft correspondence. Feature interaction and enhancement between point clouds are achieved through feature weighted aggregation. The source point cloud features and target point cloud features extracted in step 1 are used as inputs to the Transformer encoder, which contains multiple multi-head self-attention layers and feedforward neural network layers. For the source point cloud features, the global context information inside the point cloud is captured through a self-attention mechanism, and the calculation formula is as follows:

[0010] The query matrix, key matrix, and value matrix are respectively...

[0011] The weight matrix is ​​learnable; the multi-head self-attention mechanism computes multiple independent attention heads in parallel and concatenates the results, as shown in the following formula:

[0012] in , The output projection matrix is ​​given by h, which represents the number of attention heads. The subsequent layer is a normalization and residual connection, followed by a nonlinear transformation through a feedforward neural network layer. The formulas are as follows:

[0013] in

[0014] Similarly, for target point cloud features Perform the same self-attention processing to obtain the enhanced source point cloud features. and target point cloud features Next, a cross-attention mechanism is used to calculate the soft pointer matrix between the source and target point clouds. The source point cloud features are used as queries, and the target point cloud features are used as keys and values. The matching probability of each point in the source point cloud with all points in the target point cloud is calculated.

[0015] Furthermore, soft pointer matrix Calculated using the following formula:

[0016] in Let represent the matching probability between the i-th point in the source point cloud and the j-th point in the target point cloud, satisfying ... Finally, the target point cloud features are weighted and aggregated using a soft pointer matrix to generate a feature representation aligned with the source point cloud.

[0017] By concatenating or adding the source point cloud features with the aligned target point cloud features, feature interaction and enhancement between point clouds can be achieved:

[0018] This completes the feature interaction and soft pointer matching between the source point cloud and the target point cloud, and outputs the enhanced feature representation and soft pointer matrix for subsequent rigid transformation parameter estimation steps.

[0019] Step 3: Stiff Transformation Parameter Estimation: In the transformation estimation channel, based on the soft pointer matrix generated in Step 2, the rotation matrix and translation vector from the source point cloud to the target point cloud are solved in one go from the correspondence through the differentiable singular value decomposition module; at the same time, the rotation matrix and translation vector of the reverse registration are calculated by using the orthogonal matrix property, avoiding the local optimum problem caused by the alternating iterative optimization in the traditional iterative nearest point algorithm; Based on the soft pointer matrix generated in step 2 and enhanced feature representation The rigid transformation parameters from the source point cloud to the target point cloud are solved in one go from the correspondence through the differentiable singular value decomposition module, and the transformation parameters for reverse registration are calculated at the same time. First, establish a soft correspondence between the source point cloud and the target point cloud, and calculate the weighted correspondence points:

[0020] Calculate the centroids of the source and target point clouds:

[0021] After centralization, construct the covariance matrix. ; Singular value decomposition of H Solve for the rotation matrix and translation vector:

[0022] Simultaneously calculate the reverse registration parameters:

[0023] The entire solution process is a closed-form solution, requiring no iterative optimization. This avoids the local optimum problem caused by alternating iterations in the traditional ICP algorithm, and end-to-end gradient backpropagation is achieved through differentiable SVD.

[0024] Step 4 Cyclic Consistency Constraint: In the cyclic consistency constraint channel, the bidirectional registration parameters estimated in Step 3 are constrained by the cyclic consistency loss function. The transformation consistency loss from the source point cloud to the target point cloud and back to the source point cloud is calculated, and the transformation consistency loss from the target point cloud to the source point cloud and back to the target point cloud is also calculated to ensure the geometric consistency between forward and reverse registration and enhance the robustness of the model. Based on the forward registration parameters estimated in step 3 and reverse registration parameters Construct a cycle consistency loss function to ensure that the source point cloud can return to its original position after a forward transformation and a reverse transformation. First, the source point cloud P is transformed into the transformed point cloud through a forward transformation. :

[0025] Then The reconstructed source point cloud is obtained through inverse transformation. :

[0026] Cyclic consistency loss Defined as the mean square error between the original source point cloud and the reconstructed source point cloud:

[0027] Similarly, the cycle consistency of the target point cloud can be calculated to obtain...

[0028] The eventual cycle consistency loss is the sum of the losses from both directions:

[0029] The loss function is jointly optimized with the registration loss, and the total loss is:

[0030] Where λ is the weight coefficient, with a value of 0.1. By minimizing the total loss, the geometric consistency between forward and backward registration is ensured, thereby enhancing the robustness of the model.

[0031] Step 5 Multi-scale feature fusion and registration output: In the multi-scale feature fusion channel, the multi-scale feature maps output by edge convolution at different levels in Step 1 are spliced ​​together by channel dimension. After fusing local detail features and global structural features, they are input into the transformation estimation module, and finally the optimized rigid transformation parameters are output to complete the accurate registration from the source point cloud to the target point cloud. The multi-scale feature maps output from edge convolution at different levels in step 1 The data is stitched together along the channel dimension, fusing local detail features and global structural features:

[0032] The fused multi-scale features The input is fed into the transformation estimation module, which consists of steps 2, 3, and 4, and finally outputs the optimized rigid transformation parameters. ; The source point cloud P is registered using the finally estimated transformation parameters to obtain the registered point cloud. :

[0033] Output the registered point cloud and transformation parameters This completes the precise registration from the source point cloud to the target point cloud.

[0034] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention introduces hierarchical dynamic graph convolutional networks into the field of point cloud registration. Through multi-scale edge convolutional modules and dense connection strategies, it solves the problems of insufficient feature reuse and limited single-scale representation in traditional dynamic graph convolutional networks, effectively capturing the local geometric details and global shape structure of point clouds. Simultaneously, the introduction of the KD-Tree data structure in the feature extraction stage significantly improves the efficiency of neighborhood construction, solving the performance bottleneck of real-time processing of large-scale point clouds. Furthermore, for multi-scale features output from edge convolutions at different levels, fusion is achieved through channel-dimensional concatenation, fully combining low-level detail features with high-level semantic features, enhancing the richness of input information for the transform estimation module. Through joint optimization of soft pointer matching, differentiable singular value decomposition, and cycle consistency constraints, end-to-end rigid transform estimation is achieved, avoiding dependence on initial pose and local optima problems, significantly improving the robustness and accuracy of registration in low-overlap and high-noise scenes. Attached Figure Description

[0035] Figure 1 : A system architecture diagram of a point cloud registration method based on hierarchical dynamic graph convolution and accelerated nearest neighbor search; Figure 2 A model structure diagram of a point cloud registration method based on hierarchical dynamic graph convolution and accelerated nearest neighbor search. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0037] A point cloud registration method based on hierarchical dynamic graph convolution and accelerated nearest neighbor search is described below. Figure 1 As shown, it includes the detailed process of the specific algorithm: 1. Point cloud feature extraction.

[0038] The task of this invention is defined as: for a given source point cloud and target point cloud , where N is the number of points in the point cloud, each point contains three-dimensional spatial coordinate information of x, y, z, and predicts the rigid transformation parameters (rotation matrix and translation vector) from the source point cloud to the target point cloud, so that the transformed source point cloud and the target point cloud are accurately aligned. First, a hierarchical dynamic graph convolutional network (HDGNet) is used as a point cloud feature encoder to convert the input source point cloud and target point cloud into a multi-level geometric feature representation. The source point cloud and target point cloud to be registered are represented as follows: Where N is the number of points in the point cloud; firstly, a spatial index tree is constructed for the source point cloud P and the target point cloud Q using a KD-Tree data structure, and for each target point... By using a KD-Tree to search for the k nearest neighbors in the source point cloud, a neighborhood index set is obtained. Computational complexity A replacement for traditional KNN Complexity; secondly, for each point And its neighboring points, construct edge features To obtain the edge feature set edge features The input is fed into HDGNet, which contains multiple edge convolutional modules. The calculation process of the l-th edge convolutional module is as follows: , in , To further refine the neighborhood difference features, the outputs of edge convolutions at different levels are concatenated through channels: in Finally, an output convolutional layer maps the fused features to the target dimension:

[0039] in These are the geometric feature representations of the source point cloud and the target point cloud, respectively.

[0040] 2. Feature interaction and soft pointer matching.

[0041] (1) Self-attention feature enhancement: The source point cloud features extracted in step 1 are enhanced. and target point cloud features As input, it is fed into a Transformer encoder, which contains multiple multi-head self-attention layers and feedforward neural network layers; for the source point cloud features, the global context information inside the point cloud is captured through a self-attention mechanism, and the calculation formula is as follows:

[0042] The query matrix, key matrix, and value matrix are as follows:

[0043] in The weight matrix is ​​a learnable matrix. For the dimension of attention head;

[0044] The multi-head self-attention mechanism computes multiple independent attention heads in parallel and then concatenates the results, as shown in the following formula:

[0045] in , The output projection matrix is ​​given by h, where h is the number of attention heads. The self-attention layer is followed by a normalization layer and a residual connection, then a nonlinear transformation is performed through a feedforward neural network layer, as shown in the following formulas:

[0046] in ; Similarly, the same self-attention processing is performed on the target point cloud features to obtain the enhanced source point cloud features. and target point cloud features ; (2) Constructing a soft pointer matrix: Calculate the soft pointer matrix between the source point cloud and the target point cloud using a cross-attention mechanism. Use the enhanced source point cloud features as the query and the enhanced target point cloud features as the key and value. Calculate the matching probability between each point in the source point cloud and all points in the target point cloud:

[0047] Furthermore, soft pointer matrix Calculated using the following formula:

[0048] in Let represent the matching probability between the i-th point in the source point cloud and the j-th point in the target point cloud, satisfying ...

[0049] (3) Feature interaction and enhancement: The target point cloud features are weighted and aggregated using a soft pointer matrix to generate a feature representation aligned with the source point cloud.

[0050] The source point cloud features are concatenated with the aligned target point cloud features to achieve feature interaction and enhancement between point clouds:

[0051] This completes the feature interaction and soft pointer matching between the source and target point clouds, and outputs the enhanced feature representation. The soft pointer matrix M is used in the subsequent rigid transformation parameter estimation step.

[0052] 3. Stiff transformation parameter estimation.

[0053] (1) Establishment of soft correspondence: based on the soft pointer matrix generated in step 2 and enhanced feature representation The rigid transformation parameters from the source point cloud to the target point cloud are solved in one go from the correspondence through the differentiable singular value decomposition module, and the transformation parameters for reverse registration are calculated at the same time. First, establish the source cloud. With target point cloud The soft correspondence between them is calculated by using a soft pointer matrix to determine the weighted correspondence points, as shown in the following formula:

[0054] in Let represent the matching probability between the i-th point in the source point cloud and the j-th point in the target point cloud, satisfying ... ;

[0055] (2) Centroid Calculation and Centering: Calculate the centroid of the corresponding point sets of the source and target point clouds using the following formula:

[0056] Centralizing the point cloud yields decentralized point cloud coordinates:

[0057] (3) Construction of covariance matrix and singular value decomposition: Construction of cross-covariance matrix

[0058] Perform singular value decomposition on the cross-covariance matrix H:

[0059] in and It is an orthogonal matrix. It is a diagonal matrix containing singular values; (4) Solving for rigid transformation parameters: Solving for the rotation matrix and translation vector from the source point cloud to the target point cloud:

[0060] To ensure the orthogonality of the rotation matrix and the right-handed coordinate system constraint, the rotation matrix is ​​modified as follows:

[0061] in, For determinant operations, ensure ; (5) Calculation of reverse registration parameters: At the same time, the rotation matrix and translation vector of the reverse registration are calculated using the properties of orthogonal matrices:

[0062] The entire solution process is a closed-form solution, requiring no iterative optimization. This avoids the local optima problem caused by alternating iterations in traditional iterative nearest-point algorithms. Furthermore, end-to-end gradient backpropagation is achieved through a differentiable singular value decomposition module, outputting positive registration parameters. and reverse registration parameters This is used for subsequent cycle consistency constraint steps.

[0063] 4. Cyclic consistency constraint.

[0064] Considering that the geometric consistency of forward and backward registration has a significant impact on the robustness of the model in point cloud registration tasks, relying solely on registration loss is insufficient to guarantee the reversibility and consistency of bidirectional transformation. Therefore, a cyclic consistency loss function is used to constrain the bidirectional registration parameters to ensure that the source point cloud can return to its original position after a forward transformation and then a backward transformation.

[0065] First, for the forward registration parameters estimated in step 3... and reverse registration parameters Construct the cycle consistency loss function, as shown in the following formula: The source point cloud P is transformed into the transformed point cloud through a forward transformation. :

[0066] Then The reconstructed source point cloud is obtained through inverse transformation. :

[0067] Cyclic consistency loss Defined as the mean square error between the original source point cloud and the reconstructed source point cloud:

[0068] Similarly, calculate the cycle consistency loss of the target point cloud. The eventual cycle consistency loss is the sum of the losses from both directions:

[0069] The cycle consistency loss and registration loss are jointly optimized, and the total loss function is:

[0070] The weighting coefficient for the λ-cycle consistency loss is 0.1. The registration loss is calculated using the following formula:

[0071] Where I is the identity matrix, It is the Frobenius norm. This is the actual translation vector.

[0072] 5. Multi-scale fusion and registration output.

[0073] After completing feature interaction, transform estimation, and cycle consistency constraints through steps 2, 3, and 4, the contributions of feature information at different levels to the final registration accuracy vary. Relying solely on a single-scale feature representation makes it difficult to simultaneously capture local geometric details and global shape structure. Therefore, a multi-scale feature fusion mechanism is used to concatenate feature maps at different levels along their channel dimensions, enhancing the richness of the input information for the transform estimation module.

[0074] First, the multi-scale feature maps output from edge convolution at different levels in step 1 are... The data is stitched together along the channel dimension, fusing local detail features with global structural features, as shown in the following formula:

[0075] Where N is the number of points in the point cloud, and the feature dimension after stitching is 64+64+128+256=512.

[0076] The fused multi-scale features The input is fed into the transformation estimation module, which consists of steps 2 (feature interaction and soft pointer matching), 3 (rigid transformation parameter estimation), and 4 (cycle consistency constraint). The final output is the optimized rigid transformation parameters.

[0077] Finally, the source point cloud P is registered using the final estimated transformation parameters to obtain the registered point cloud.

[0078] Output the registered point cloud and transformation parameters This completes the precise registration from the source point cloud to the target point cloud.

Claims

1. A point cloud registration method based on hierarchical dynamic graph convolution and accelerated nearest neighbor search, characterized in that, Includes the following steps: Step 1 Point Cloud Feature Extraction: The source and target point clouds to be registered are taken as input. The hierarchical dynamic graph convolutional network HDGNet is used as the point cloud feature encoder. The KD-Tree data structure is used to replace the traditional KNN for nearest neighbor search, reducing the computational complexity from O(N²) to O(N log N). Local neighborhood features are aggregated layer by layer by constructing a multi-scale edge convolution module. The output of edge convolution at different levels is concatenated by a dense connection strategy. Residual connections are introduced to enhance gradient flow. Finally, the multi-level geometric feature representation of the source and target point clouds is output. Step 2 Feature Interaction and Soft Pointer Matching: The extracted source point cloud features and target point cloud features are input into the Transformer encoder. The global context information inside the point cloud is captured through the self-attention mechanism. Then, the soft pointer matrix between the source point cloud and the target point cloud is calculated using the cross-attention mechanism to generate a feature-based soft correspondence, thereby realizing feature interaction and enhancement between point clouds. Step 3: Stiff Transformation Parameter Estimation: Based on the soft pointer matrix generated in Step 2, the rotation matrix and translation vector from the source point cloud to the target point cloud are solved in one go from the correspondence using the differentiable singular value decomposition module. At the same time, the rotation matrix and translation vector for the reverse registration are calculated. Step 4 Cyclic Consistency Constraint: Constrain the rotation matrix and translation vector of the bidirectional registration estimated in Step 3 by using the cyclic consistency loss function, and calculate the transformation consistency loss from the source point cloud to the target point cloud and back to the source point cloud; Step 5 Multi-scale feature fusion and registration output: The local neighborhood features output from edge convolution at different levels in Step 1 are concatenated along the channel dimension. After fusing local detail features and global structural features, they are input into the transformation estimation module. Finally, the optimized rigid transformation parameters are output to complete the accurate registration from the source point cloud to the target point cloud.

2. The point cloud registration method based on hierarchical dynamic graph convolution and accelerated nearest neighbor search according to claim 1, characterized in that, In step 1, the specific method is as follows: The source point cloud and target point cloud to be registered are represented as follows: Where N is the number of points in the point cloud; firstly, a spatial index tree is constructed for the source point cloud P and the target point cloud Q using a KD-Tree data structure, and for each target point... By using a KD-Tree to search for the k nearest neighbors in the source point cloud, a neighborhood index set is obtained. Computational complexity A replacement for traditional KNN Complexity; Secondly, for each point And its neighboring points, construct edge features To obtain the edge feature set edge features The input is fed into HDGNet, which contains multiple edge convolutional modules. The calculation process of the l-th edge convolutional module is as follows: ， in , For neighborhood difference features, the outputs of edge convolutions at different levels are concatenated through channels: in Finally, an output convolutional layer maps the fused features to the target dimension: in These are the geometric feature representations of the source point cloud and the target point cloud, respectively.

3. The point cloud registration method based on hierarchical dynamic graph convolution and accelerated nearest neighbor search according to claim 1, characterized in that, In step 2, the specific method is as follows: The multi-level geometric features of the source point cloud and the target point cloud extracted in step 1 are used as inputs to the Transformer encoder, which contains multiple multi-head self-attention layers and feedforward neural network layers. For the source point cloud features, the global context information inside the point cloud is captured through a self-attention mechanism, and the calculation formula is as follows: The query matrix, key matrix, and value matrix are respectively... The weight matrix is ​​learnable; the multi-head self-attention mechanism computes multiple independent attention heads in parallel and concatenates the results, as shown in the following formula: in , The output projection matrix is ​​given by h, which represents the number of attention heads. The subsequent layer is a normalization and residual connection, followed by a nonlinear transformation through a feedforward neural network layer. The formulas are as follows: in Target point cloud features Perform the same self-attention processing to obtain the enhanced source point cloud features. and target point cloud features Next, a cross-attention mechanism is used to calculate the soft pointer matrix between the source and target point clouds. The source point cloud features are used as queries, and the target point cloud features are used as keys and values. The matching probability of each point in the source point cloud with all points in the target point cloud is calculated. Furthermore, soft pointer matrix Calculated using the following formula: in Let represent the matching probability between the i-th point in the source point cloud and the j-th point in the target point cloud, satisfying ... Finally, the target point cloud features are weighted and aggregated using a soft pointer matrix to generate a feature representation aligned with the source point cloud. By concatenating or adding the source point cloud features with the aligned target point cloud features, feature interaction and enhancement between point clouds can be achieved: This completes the feature interaction and soft pointer matching between the source point cloud and the target point cloud, and outputs the enhanced feature representation and soft pointer matrix for subsequent rigid transformation parameter estimation steps.

4. The point cloud registration method based on hierarchical dynamic graph convolution and accelerated nearest neighbor search according to claim 1, characterized in that, In step 3, the specific method is as follows: Based on the soft pointer matrix generated in step 2 and enhanced feature representation The rigid transformation parameters from the source point cloud to the target point cloud are solved in one go from the correspondence through the differentiable singular value decomposition module, and the transformation parameters for reverse registration are calculated at the same time. First, establish a soft correspondence between the source point cloud and the target point cloud, and calculate the weighted correspondence points: Calculate the centroids of the source and target point clouds: After centralization, construct the covariance matrix. ; Singular value decomposition of H Solve for the rotation matrix and translation vector: Simultaneously calculate the reverse registration parameters: The entire solution process is a closed-form solution, requiring no iterative optimization. This avoids the local optimum problem caused by alternating iterations in the traditional ICP algorithm, and end-to-end gradient backpropagation is achieved through differentiable SVD.

5. The point cloud registration method based on hierarchical dynamic graph convolution and accelerated nearest neighbor search according to claim 1, characterized in that, In step 4, the specific method is as follows: Based on the forward registration parameters estimated in step 3 and reverse registration parameters Construct a cycle consistency loss function to ensure that the source point cloud can return to its original position after a forward transformation and a reverse transformation. First, the source point cloud P is transformed into the transformed point cloud through a forward transformation. : Then The reconstructed source point cloud is obtained through inverse transformation. : Cyclic consistency loss Defined as the mean square error between the original source point cloud and the reconstructed source point cloud: Similarly, the cycle consistency of the target point cloud can be calculated to obtain... The eventual cycle consistency loss is the sum of the losses from both directions: The loss function is jointly optimized with the registration loss, and the total loss is: Where λ is the weighting coefficient, with a value of 0.1, the geometric consistency of forward and backward registration is ensured by minimizing the total loss.

6. The point cloud registration method based on hierarchical dynamic graph convolution and accelerated nearest neighbor search according to claim 1, characterized in that, In step 5, the specific method is as follows: The multi-scale feature maps output from edge convolution at different levels in step 1 The data is stitched together along the channel dimension, fusing local detail features and global structural features: The fused multi-scale features The input is fed into the transformation estimation module, which consists of steps 2, 3, and 4, and finally outputs the optimized rigid transformation parameters. ; The source point cloud P is registered using the finally estimated transformation parameters to obtain the registered point cloud. : Output the registered point cloud and transformation parameters This completes the precise registration from the source point cloud to the target point cloud.

Images