Method and device for enhancing three-dimensional model registration based on hierarchical dynamic features

Abstract

The invention relates to the technical field of medical image processing and surgical navigation, and particularly discloses a method and a device for enhancing three-dimensional model registration based on hierarchical dynamic features. According to the method, an intraoperative target point cloud and a preoperative source point cloud are acquired, a P-ScConv framework is adopted to perform redundancy removal processing on point cloud features, group normalization and a gating mechanism of an SRU module are utilized to screen and dissect related features, and channel features are optimized through a segmentation-transformation-fusion strategy of a CRU module; a P-DFF module is adopted to carry out channel-space dual dynamic attention fusion on the multi-scale features; establishing an over-point corresponding relation through an enhanced over-point matching module, and optimizing matching distribution by adopting a dual normalization strategy and farthest point sampling; finally, the super points are correspondingly propagated to the dense points, and a rigid transformation matrix is calculated. According to the method, the problems of incomplete point cloud, insufficient real-time performance, poor adaptability, mismatching and the like in the operation are effectively solved, and the accuracy, the real-time performance and the robustness of registration are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image processing and surgical navigation, and in particular to a three-dimensional model registration method and device based on hierarchical dynamic feature enhancement. BACKGROUND

[0002] With the advancement of medical technology, precision medicine has gradually become an important direction of modern medical development. In the field of orthopedic surgery, accurate bone positioning and surgical navigation are of great significance to improve the success rate of surgery and reduce patient trauma. Point cloud registration, as one of the key technologies to achieve accurate surgical navigation, its main purpose is to align two or more point cloud data sets in three-dimensional space, so as to facilitate accurate surgical planning and execution. Point cloud registration technology has a wide range of applications in medical image processing, surgical navigation, robot-assisted surgery and other fields.

[0003] In recent years, methods based on deep learning have made some progress in registration by learning the geometric structure of point clouds. For example, GeoTransformer and other methods extract point cloud features and establish correspondence through geometric transformer architecture, which improves the registration accuracy. These methods use the powerful expression ability of neural networks to learn effective feature representations from a large amount of training data, providing a new technical path for point cloud registration.

[0004] However, the existing point cloud registration methods have the following problems when facing intraoperative dynamic environment, incomplete point cloud data or low overlap, etc. (1) Incomplete point cloud data problem: during surgery, due to factors such as tissue movement and instrument occlusion, the acquired point cloud data is often missing or has low overlap, which makes it difficult for traditional point cloud registration methods to accurately align the point clouds, resulting in registration failure; (2) Insufficient real-time problem: the existing point cloud registration methods have high delay in the calculation process, which cannot meet the needs of intraoperative real-time navigation. Real-time is crucial for surgical navigation, as it directly affects the smoothness and safety of surgery; (3) Poor adaptability problem: different patients have different bone morphologies, and existing registration methods often cannot adapt well to these differences, resulting in unstable registration results among different patients; (4) Mis-matching problem: when performing superpoint matching, existing methods are prone to mis-matching at similar anatomical structures, such as adjacent vertebrae or symmetrical joints, which reduces the accuracy of registration.

[0005] In summary, the existing technology has the technical problems of low accuracy of intraoperative point cloud registration, insufficient real-time, poor adaptability to different patient bone morphologies, and mis-matching at similar anatomical structures. SUMMARY

[0006] This invention provides a hierarchical dynamic feature-enhanced 3D model registration method, which can effectively handle the incompleteness of point cloud data in the dynamic environment of intraoperative surgery, improve the accuracy, real-time performance and robustness of registration, and solve the technical problems of low accuracy, insufficient real-time performance, poor adaptability and mismatch in intraoperative point cloud registration in the prior art.

[0007] In a first aspect, the present invention provides a method for 3D model registration based on hierarchical dynamic feature enhancement, comprising: acquiring a target point cloud Q acquired intraoperatively and a source point cloud P reconstructed preoperatively, wherein the target point cloud... Source Cloud The process involves feature extraction and redundancy removal for both the target and source point clouds. The input features are processed using the P-ScConv framework with SRU and CRU modules to output redundancy-removed point cloud features. These redundancy-removed features are then downsampled to obtain downsampled point cloud features. Hierarchical dynamic fusion is performed on the downsampled point cloud features, using the P-DFF module to apply channel attention weighting and spatial attention weighting to multi-scale features, resulting in fused point cloud features. Superpoint feature extraction and matching are then performed on the fused point cloud features. An enhanced superpoint matching module establishes superpoint correspondences between the target and source point clouds. These superpoint correspondences are then propagated to dense points, and a rigid transformation matrix is ​​calculated based on the dense point correspondences. ,in Let be a rotation matrix. It is a translation vector.

[0008] As described above, this invention creatively combines the P-ScConv redundancy removal framework, the P-DFF dynamic fusion module, and the enhanced superpoint matching module through a hierarchical dynamic feature enhancement technical architecture, forming a complete end-to-end cloud registration solution. The core innovation of this technical solution lies in: firstly, the P-ScConv framework performs redundancy removal at the feature extraction front end, effectively filtering intraoperative environmental noise while preserving key anatomical features; secondly, the P-DFF module achieves dynamic adaptive fusion of multi-scale features, overcoming the limitations of traditional static fusion methods in handling different patient bone morphologies; and finally, the enhanced superpoint matching module establishes high-confidence point correspondences, significantly reducing the mismatch rate at similar anatomical structures. The overall technical solution, while ensuring registration accuracy, significantly reduces computational latency through feature redundancy removal and structural optimization, meeting the stringent requirements of real-time intraoperative navigation.

[0009] Preferably, the input features are processed using the P-ScConv framework through SRU and CRU module processing, including: adding the P-ScConv framework to the front end of the encoder to process the input features X. in The SRU module and CRU module processes the data sequentially, and the output features are then processed. In the SRU module, group normalization is used to group features, and a gating mechanism is used to quantify feature importance and screen anatomical features related to registration. In the CRU module, a segmentation-transformation-fusion strategy is used to differentiate different channel groups, and the processed features are dynamically weighted and fused.

[0010] As described above, this invention organically combines the SRU and CRU modules through a cascaded processing architecture of the P-ScConv framework. The SRU module achieves adaptive evaluation of feature importance through group normalization and gating mechanisms, effectively identifying and retaining key features related to skeletal anatomy while filtering out noisy features introduced by factors such as instrument reflection, tissue occlusion, and electromagnetic interference. The CRU module performs differentiated processing on feature channels using a segmentation-transformation-fusion strategy, effectively compressing redundant information while maintaining feature diversity. This front-end redundancy removal mechanism is one of the key innovations distinguishing this invention from existing technologies, enabling subsequent processing modules to perform calculations based on more compact and high-quality features, thereby simultaneously improving registration accuracy and computational efficiency.

[0011] Preferably, in the SRU module, group normalization is used to group features, and a gating mechanism is used to quantify feature importance, including: grouping input features by channel, and processing each group of features x g Normalization is performed, and the normalized features are calculated as follows: Where g is the group index. Let G be the feature vector of this group, G be the number of groups, and γ be the feature vector of this group. g and β g The parameters are learnable; the gating weights are calculated based on normalized features. , where σ is the Sigmoid function; the feature space is filtered according to the gating weights, retaining features with weights close to 1, and marking features with weights close to 0 as redundant features.

[0012] As described above, this invention achieves precise screening of orthopedic anatomical features through group normalization and gating mechanisms. The group normalization mechanism enables adaptive separation of features from different anatomical structures; for example, subtle structural features such as trabecular orientation and articular surface curvature can be effectively normalized within their respective channel groups, thereby improving the signal-to-noise ratio. The gating mechanism utilizes a learnable parameter γ... g The importance of normalized features is evaluated, and the sigmoid activation function maps feature importance to a continuous interval of 0 to 1, ensuring that important anatomical features are preserved while redundant noise features are suppressed. This gated feature selection mechanism is a significant innovation of this invention in the feature preprocessing stage, effectively solving the problem of low signal-to-noise ratio in intraoperative point cloud data.

[0013] Preferably, in the CRU module, a segmentation-transformation-fusion strategy is used to differentiate different channel groups, including: dividing the channel dimension into high-information channels and low-information channels according to the coefficient α, wherein the number of high-information channels is... Low number of information channels For high-information channels, parallel feature extraction is performed using group convolution and point convolution; for low-information channels, the dimensions are expanded using point convolution and then concatenated with the original features for feature recombination; the features of high-information channels and low-information channels are dynamically weighted and fused to output the fused channel features.

[0014] As described above, this invention achieves differentiated processing and lossless information preservation of feature channels through the segmentation-transformation-fusion strategy of the CRU module. By dynamically dividing high-information channels and low-information channels using a coefficient α, channels containing rich anatomical information can undergo more complex feature transformation processing, while channels with lower information content are effectively utilized through lightweight processing and feature recombination. This differentiated processing strategy reduces computational complexity while ensuring feature integrity, and is one of the key technical means by which this invention achieves a balance between real-time performance and accuracy.

[0015] Preferably, the P-DFF module performs channel attention weighting and spatial attention weighting on the multi-scale features, including: performing one-dimensional convolution on the input multi-scale features X1 and X2 to align the channel dimensions, and obtaining the feature concatenation F. cat ; For F cat Perform global average pooling to obtain F avg Calculate channel attention weights ; For F cat Channel-weighted fusion is performed to obtain channel-weighted features. The spatial weights are calculated by performing one-dimensional convolutions on X1 and X2 respectively. By combining channel attention weights and spatial attention weights, the final fused feature is output. .

[0016] As described above, this invention solves the problems of noise propagation and geometric distortion in multi-scale feature fusion through the dual dynamic attention mechanism of the P-DFF module. Compared with the traditional static feature stitching method F... fused =F L +F HUnlike other methods, the P-DFF module dynamically adjusts the contribution weights of different channels through a channel attention mechanism, which can suppress the contamination of high-level semantic features by intraoperative noise in low-level features; at the same time, it compensates for geometric distortions introduced by the upsampling process through a spatial attention mechanism. This channel-space dual dynamic attention fusion mechanism is an important improvement of this invention compared to existing geometric transformation methods, enabling the registration model to adaptively adjust according to the differences in skeletal morphology in different cases, significantly improving the model's generalization ability.

[0017] Preferably, the superpoint correspondence between the target point cloud and the source point cloud is established through an enhanced superpoint matching module, including: extracting superpoint features from the target point cloud and the source point cloud to obtain reference skeleton superpoint features F respectively. ref Source skeleton superpoint features F src ; Calculate the similarity scoring matrix based on superpoint features ,in The Euclidean distance is represented; the similarity score matrix is ​​double-normalized to obtain the comprehensive analysis score; the Top-K method is used to select the top K candidate corresponding points with the highest matching scores based on the comprehensive analysis score; the spatial distribution of the candidate corresponding points is optimized by sampling the farthest point, and the final super-point correspondence is output.

[0018] As described above, this invention, through an enhanced superpoint matching module, combines a dual normalization strategy with farthest point sampling, effectively solving the problem of mismatches that easily occur at similar anatomical structures in existing superpoint matching methods. The Gaussian kernel function converts feature distances into similarity scores, enabling more accurate quantification of the matching degree between different bony landmarks. The dual normalization strategy normalizes from both the reference point cloud and the source point cloud, effectively balancing bilateral contributions and reducing mismatches caused by similar anatomical structures such as adjacent vertebrae or symmetrical joints. Farthest point sampling ensures that matched superpoints are spatially uniformly distributed, avoiding the problem of corresponding points being over-concentrated in local areas. This enhanced superpoint matching mechanism is one of the core innovations of this invention in improving registration accuracy.

[0019] Preferably, the similarity scoring matrix is ​​subjected to double normalization to obtain a comprehensive anatomical score, including: normalizing the similarity scoring matrix along each row to obtain a reference point cloud normalized score. Normalize the similarity score matrix along its columns to obtain the normalized score of the source point cloud. Multiply the reference point cloud normalized score by the source point cloud normalized score to obtain the comprehensive analysis score. .

[0020] As described above, this invention achieves multi-faceted evaluation of matching scores through a dual row and column normalization strategy. Row normalization evaluates the relative matching degree between each reference superpoint and all source superpoints from the perspective of the reference point cloud, while column normalization performs a symmetrical evaluation from the perspective of the source point cloud. The product of the two normalized scores integrates information from both sides; a high comprehensive analytical score can only be obtained when a matching pair has high scores from both sides. This dual normalization mechanism effectively avoids the bias that may be caused by unilateral normalization and is the technical contribution of this invention in the superpoint matching stage.

[0021] Preferably, spatial distribution optimization is performed on the candidate corresponding points using farthest point sampling, including: adjusting the coordinates P of the candidate superpoints in the target point cloud. ref Candidate superpoint coordinates P in the source point cloud src Input the farthest point sampling function respectively; obtain the sampling index fps ref and FPS src This ensures that the final corresponding point pairs are evenly distributed in space; the final corresponding index j is output based on the sampling index. ref and j src And the corresponding matching confidence level ρ.

[0022] As described above, this invention optimizes the spatial distribution of candidate matching points through farthest point sampling, ensuring that the final selected superpoint pairs represent the complete skeleton structure. In the candidate matching points selected by the Top-K method, there may be multiple high-scoring matching pairs concentrated in similar anatomical regions, which is detrimental to subsequent global registration optimization. The farthest point sampling algorithm effectively ensures uniform coverage of the corresponding points across the entire skeleton surface by iteratively selecting the point furthest from the already selected point set. This spatial distribution optimization strategy, combined with double normalization, constitutes the core technical feature of the enhanced superpoint matching module of this invention.

[0023] Preferably, calculating the rigid transformation matrix based on the correspondence of dense points includes: minimizing the objective function based on the set C of correspondences of dense points. Solve for the optimal rigid transformation matrix T; output the rigid transformation matrix. This is used to align the source point cloud to the target point cloud.

[0024] As described above, this invention solves for the optimal rigid transformation matrix by minimizing the sum of squared Euclidean distances between corresponding point pairs. Since the preceding steps have established high-quality point correspondences through redundancy removal, dynamic fusion, and enhanced superpoint matching, the solution to the transformation matrix achieves high accuracy. The rigid transformation matrix includes the rotation matrix. Translation vector It can fully describe the spatial transformation relationship from the preoperative source point cloud to the intraoperative target point cloud, providing accurate positioning information for subsequent surgical navigation.

[0025] Secondly, this invention provides a 3D model registration device based on hierarchical dynamic feature enhancement, comprising: a point cloud acquisition module for acquiring the target point cloud acquired intraoperatively and the source point cloud reconstructed preoperatively; a redundancy removal module for performing feature extraction and redundancy removal processing on the target point cloud and the source point cloud through a P-ScConv framework, wherein the P-ScConv framework includes an SRU submodule for group normalization and gating screening and a CRU submodule for segmentation-transformation-fusion; a downsampling module for downsampling the point cloud features after redundancy removal; a dynamic fusion module for performing channel attention weighting and spatial attention weighting fusion on the downsampled multi-scale features through P-DFF; a superpoint matching module for extracting superpoint features from the fused point cloud features and establishing superpoint correspondences through double normalization and farthest point sampling; and a transformation calculation module for propagating the superpoint correspondences to dense points and calculating a rigid transformation matrix based on the dense point correspondences.

[0026] Thirdly, the present invention provides an electronic device, the device including a processor and a memory; the memory is used to store a computer program and transmit the computer program to the processor; the processor is used to execute the hierarchical dynamic feature-enhanced 3D model registration method as described in the first aspect according to the instructions in the computer program.

[0027] Fourthly, the present invention provides a storage medium for storing computer-executable instructions, which, when executed by a computer processor, are used to perform the hierarchical dynamic feature-enhanced 3D model registration method as described in the first aspect.

[0028] The above-mentioned beneficial effects of the present invention are as follows: The present invention achieves high accuracy, high real-time performance, and high robustness in intraoperative point cloud registration through hierarchical dynamic feature enhancement technology. Compared with the prior art, the present invention has the following inventive contributions:

[0029] First, a P-ScConv redundancy removal framework is proposed. Through the group normalization and gating mechanism of the SRU module and the segmentation-transformation-fusion strategy of the CRU module, feature redundancy removal is achieved at the encoder front end. This is a technical feature that existing point cloud registration methods do not have. It effectively solves the problem of low signal-to-noise ratio of intraoperative point clouds and reduces the computational burden of subsequent processing. Secondly, a P-DFF dual dynamic attention fusion module is proposed, overcoming the limitations of traditional geometric transformers that use static feature stitching. Through a channel-space dual attention mechanism, it achieves dynamic adaptive fusion of multi-scale features, effectively suppressing noise propagation and geometric distortion, and improving the model's adaptability to different patient bone morphologies. Thirdly, an enhanced superpoint matching mechanism based on dual normalization and farthest point sampling is proposed, fundamentally solving the problem of mismatches easily occurring at similar anatomical structures in existing methods, significantly improving registration accuracy. The technical solution of this invention forms a complete hierarchical dynamic feature enhancement registration system. The modules cooperate and synergistically enhance each other, achieving real-time performance while ensuring registration accuracy, providing technical support for orthopedic surgical navigation. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the overall process of the hierarchical dynamic feature-enhanced 3D model registration method provided by the present invention. Detailed Implementation

[0031] The following description and accompanying drawings fully illustrate specific embodiments of this application to enable those skilled in the art to practice them. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operations may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. The scope of embodiments of this application includes the entire scope of the technical solutions and all available equivalents of the technical solutions. In this document, each embodiment may be referred to individually or collectively with the term "invention," which is merely for convenience and, if more than one invention is disclosed, is not intended to automatically limit the scope of the application to any single invention or inventive concept. Relational terms such as "first" and "second" are used herein only to distinguish one entity or operation from another, without requiring or implying any actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed. The various embodiments in this document are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the structures, products, etc., disclosed in the embodiments, since they correspond to the disclosed parts, the descriptions are relatively simple; relevant details can be found in the method section.

[0032] Terminology Before describing the technical solution in detail, the key terms used in this application will be explained first: Point cloud: A point cloud is a set of discrete points used to represent the surface or spatial structure of an object in a three-dimensional coordinate system. In this application, the point cloud is used to represent the three-dimensional geometry of the bone, including the target point cloud Q acquired intraoperatively and the source point cloud P reconstructed preoperatively. Each point contains three-dimensional spatial coordinate information, and the goal of point cloud registration is to find the optimal spatial transformation to align the two point clouds.

[0033] Rigid Transformation Matrix: ∈SE(3) is used to describe rigid motion in three-dimensional space, including rotation matrices. Translation vector SO(3) denotes the three-dimensional special orthogonal group, and SE(3) denotes the three-dimensional special Euclidean group. Rigid transformations preserve the shape and size of an object, changing only its spatial position and orientation.

[0034] P-ScConv Framework: P-ScConv is a point cloud sparse convolution redundancy removal framework proposed in this invention, comprising two core components: an SRU module and a CRU module. This framework is added to the front end of the encoder for redundancy removal of input point cloud features. Its overall computation is expressed as follows: .

[0035] The SRU module (Spatial Redundancy Unit) is the first component of the P-ScConv framework. It assesses and filters feature importance through group normalization and gating mechanisms. Group normalization groups features by channel and normalizes them. The gating mechanism generates weights between 0 and 1 using the sigmoid function, which are used to filter anatomical features relevant to registration and remove redundant noise.

[0036] The CRU module (Channel Redundancy Unit) is the second component of the P-ScConv framework. It employs a segmentation-transformation-fusion strategy to differentiate feature channels. This module divides channels into high-information and low-information channels using a coefficient α, applies different processing methods to each, and finally performs dynamic weighted fusion.

[0037] The P-DFF module (Point Dynamic Feature Fusion) is a multi-scale dynamic feature fusion module proposed in this invention, employing a channel-spatial dual attention mechanism. Unlike traditional static feature stitching methods, P-DFF can dynamically adjust the fusion weights based on the input features, suppressing noise propagation and compensating for geometric distortions.

[0038] Superpoints: Superpoints are representative sets of points extracted from the original point cloud using downsampling or clustering methods. Superpoints preserve the main geometric features of the original point cloud while significantly reducing the number of points, and are used to establish a rough correspondence between two point clouds.

[0039] Dual Normalization: Dual normalization is the strategy adopted in this invention for superpoint matching, which normalizes the similarity scoring matrix simultaneously along both rows and columns. Row normalization evaluates the matching degree from the perspective of the reference point cloud, while column normalization evaluates the matching degree from the perspective of the source point cloud. The two are multiplied to obtain the comprehensive analysis score.

[0040] Farthest Point Sampling (FPS): FPS is a point cloud downsampling algorithm that iteratively selects the point furthest from the selected point set to ensure a uniform spatial distribution of sampled points. In this application, FPS is used to optimize the spatial distribution of superpoint matching and avoid excessive concentration of corresponding points in local areas.

[0041] Gaussian kernel function: The Gaussian kernel function is used to convert the Euclidean distance between features into a similarity score. The Gaussian kernel function used in this application is in the form of... The smaller the distance, the higher the similarity score.

[0042] This invention provides a method for 3D model registration based on hierarchical dynamic feature enhancement. For example... Figure 1 As shown, Figure 1 This is a schematic diagram of the overall process of the hierarchical dynamic feature-enhanced 3D model registration method provided by the present invention. The method provided by the present invention can be executed by devices such as surgical navigation servers, medical image processing terminals, and robot-assisted surgical systems, which can be implemented through software and / or hardware. The method includes the following steps:

[0043] Step 101: Obtain the target point cloud Q acquired during the operation and the source point cloud P reconstructed before the operation.

[0044] In orthopedic surgical navigation scenarios, two types of point cloud data are required as input for registration: target point cloud and target point cloud. It is a point cloud of the patient's bone surface acquired in real time during surgery using a 3D scanning device (such as a structured light scanner, laser scanner, or depth camera), where M represents the number of points in the target point cloud, and each point q i It contains three-dimensional spatial coordinates (x, y, z). Source point cloud. It is a high-precision bone model point cloud reconstructed from medical images such as CT or MRI before surgery, where N represents the number of points in the source point cloud, and each point has a value of p. j It also includes three-dimensional spatial coordinates.

[0045] In practical applications, the target point cloud is often incomplete due to intraoperative environmental limitations (such as tissue occlusion, instrument interference, and limited acquisition angle), with the number of points M typically less than the number of points N in the source point cloud, and may contain noise and outliers. The source point cloud, as a high-precision model for preoperative reconstruction, usually possesses more complete skeletal geometry information and a higher point density. The overlap between the two point clouds may be limited, posing significant challenges to registration.

[0046] This step embodies the technical features of acquiring target point clouds and source point clouds in the first aspect of the technical solution, providing the input data foundation for subsequent feature extraction and registration processing.

[0047] Step 102: Perform feature extraction and redundancy removal on the target point cloud and the source point cloud respectively.

[0048] After acquiring point cloud data, the input features need to be redundant-removed using the P-ScConv framework. The P-ScConv framework is added to the front end of the encoder and consists of two core components: an SRU module and a CRU module. The overall computation is expressed as follows: , where X in It is the input feature, X out It is the output feature after redundancy removal.

[0049] For the SRU module, the input features are first grouped by channel. Assuming the number of channels in the input features is C, they are divided into G groups, each containing C / G channels. For each group of features... Perform group normalization processing and calculate the normalized feature. , where E|x g | represents the mean of this set of features, Var(x) g ) represents the variance of the set of features, ε is a small constant to prevent division by zero (usually taken as 1e-5), and γ g and β g These are the learnable scaling and offset parameters corresponding to the group. Through group normalization, features of different anatomical structures (such as trabecular orientation, articular surface curvature, cortical bone margins, etc.) can be adaptively processed within their respective channel groups, effectively improving the signal-to-noise ratio.

[0050] After group normalization, the SRU module quantifies feature importance using a gating mechanism. The formula for calculating the gating weights is as follows: σ is the sigmoid function, which maps the weighted sum to the (0,1) interval. Features with weights close to 1 are considered important anatomical features related to registration and will be preserved; features with weights close to 0 are marked as redundant features and will be suppressed in subsequent processing. This gating mechanism can effectively filter intraoperative environmental noise (such as instrument reflections, electromagnetic interference, and artifacts caused by tissue occlusion) while preserving key skeletal anatomical information.

[0051] For the CRU module processing, a segmentation-transformation-fusion strategy is adopted to differentiate the processing of different channel groups. First, the channel dimension is divided into high-information channels and low-information channels based on the coefficient α, where the number of high-information channels is... Low number of information channels For high-information channels, parallel feature extraction is performed using group convolution and point convolution. Group convolution captures local correlations between channels while maintaining computational efficiency, while point convolution (1×1 convolution) is used for information exchange between channels. For low-information channels, the dimensions are first expanded using point convolution, and then concatenated with the original features to achieve feature recombination. Finally, the features from the two branches are dynamically weighted and fused, preserving important channel information while compressing redundant dimensions.

[0052] This step embodies the technical features of feature extraction and redundancy removal through the P-ScConv framework in the first aspect of the technical solution, as well as the specific implementation methods of the SRU module and CRU module in the preferred solution.

[0053] Step 103: Downsample the point cloud features after redundancy removal.

[0054] After redundancy removal, the point cloud features need to be downsampled to obtain multi-scale feature representations. This invention uses KPConv (Kernel Point Convolution) for point cloud downsampling. This method is suitable for convolutional operations on point cloud data and can extract local geometric features while downsampling.

[0055] Downsampling is typically performed in a hierarchical manner, with each layer reducing the number of point clouds to a certain proportion (e.g., 1 / 4 or 1 / 8) while increasing the number of feature channels to maintain the amount of feature information. Through multi-layer downsampling, multi-scale feature representations from details to the global picture can be obtained: shallow features retain microscopic geometric details (such as the orientation of trabecular bone and the curvature of articular surfaces), while deep features contain macroscopic semantic information (such as the distribution of cortical bone thickness and the axial direction of the medullary cavity).

[0056] During downsampling, the point coordinates and feature information of each level are recorded simultaneously, providing a foundation for subsequent upsampling and feature fusion. The downsampled point set will then serve as superpoints in the subsequent matching process.

[0057] This step embodies the technical feature of downsampling the point cloud features after redundancy removal in the first aspect of the technical solution.

[0058] Step 104: Perform hierarchical dynamic fusion processing on the downsampled point cloud features.

[0059] After downsampling, the multi-scale features need to be dynamically fused using the P-DFF module. The P-DFF module employs a channel-space dual attention mechanism, overcoming the limitations of traditional static feature stitching methods.

[0060] For the input multi-scale features X1 and X2, a one-dimensional convolution operation is first performed to align the channel dimensions, resulting in the feature concatenation F. cat .

[0061] In the channel attention branch, for F cat Perform global average pooling to obtain F avg Then, channel attention weights are calculated using one-dimensional convolution and a sigmoid activation function. Channel attention weight W c The dimension is the same as the number of channels, and each element represents the importance weight of the corresponding channel. For F... cat Channel-weighted fusion is performed to obtain channel-weighted features. , where ⊙ represents element-wise multiplication.

[0062] In the spatial attention branch, spatial weight components are calculated by performing one-dimensional convolution on X1 and X2 respectively. and Then, spatial attention weights are obtained through the Sigmoid activation function. Spatial attention weights can capture the importance of local regions and are used to compensate for the geometric distortion of the bone surface and local coordinate deviations caused by the upsampling process.

[0063] Finally, by combining the channel attention weights and spatial attention weights, the final fused feature is output. Through this dual dynamic attention mechanism, the P-DFF module can adaptively adjust according to the differences in skeletal morphology in different cases, suppress the contamination of high-level semantic features by intraoperative noise in low-level features, and enhance the feature response of important regions.

[0064] This step embodies the technical features of the first aspect of the technical solution, which uses the P-DFF module to perform channel attention weighting and spatial attention weighting on multi-scale features, as well as the specific calculation methods of channel attention and spatial attention in the preferred solution.

[0065] Step 105: Extract and match superpoint features from the fused point cloud features.

[0066] After feature fusion is completed, an enhanced superpoint matching module is needed to establish the superpoint correspondence between the target point cloud and the source point cloud. This module includes steps such as superpoint feature extraction, similarity calculation, double normalization, Top-K selection, and farthest point sampling optimization.

[0067] First, superpoint features are extracted from the target point cloud and the source point cloud to obtain the reference skeleton superpoint features F. ref Source skeleton superpoint features F src Superpoint feature extraction can utilize geometric transformer networks such as GeoTransformer to capture the geometric relationships and semantic information between superpoints through a self-attention mechanism.

[0068] Next, a similarity score matrix is ​​calculated based on the superpoint features. A Gaussian kernel function is used to convert the Euclidean distance between features into a similarity score; the calculation formula is as follows: ,in Let S represent the Euclidean distance. The similarity score matrix S has dimensions M'×N', where M' and N' are the number of superpoints in the target point cloud and the source point cloud, respectively. Feature pairs with smaller distances receive higher similarity scores.

[0069] Then, the similarity score matrix is ​​subjected to double normalization. First, normalization is performed along the rows to obtain the reference point cloud normalized score. This represents the relative matching degree between each reference superpoint and all source superpoints from the perspective of the reference point cloud. Normalization is then performed along the columns to obtain the source point cloud normalization score. This represents the relative matching degree between each source hyperpoint and all reference hyperpoints from the perspective of the source point cloud. Finally, the two normalized scores are multiplied together to obtain the comprehensive analysis score. A high overall anatomical score can only be obtained when a matching pair has high scores from both sides, which effectively reduces mismatches caused by similar anatomical structures such as adjacent vertebrae or symmetrical joints.

[0070] Based on the comprehensive anatomical score, the Top-K method is used to select the top K candidate corresponding points with the highest matching scores as the initial matching results, including the corresponding initial point cloud index (i). ref, i src ) and matching score ρ iThe typical value for K is 512 or 1024, which needs to be adjusted according to the actual point cloud size and registration accuracy requirements.

[0071] Finally, spatial distribution optimization is performed on the candidate corresponding points using farthest point sampling. The coordinates P of the candidate superpoints in the target point cloud are then... ref Candidate superpoint coordinates P in the source point cloud src Input the farthest point sampling function respectively to obtain the sampling index fps ref and FPS src The farthest point sampling algorithm starts from the initial seed point and iteratively selects the point farthest from the selected point set to ensure that the final corresponding point pairs are evenly distributed in space and avoid over-concentration in similar anatomical regions. Considering the actual registration requirements, the final number of corresponding point pairs is usually set to 256 pairs. The final corresponding index j is output based on the sampling index. ref and j src And the corresponding matching confidence level ρ.

[0072] This step embodies the technical features of establishing superpoint correspondence through the enhanced superpoint matching module in the first aspect of the technical solution, as well as the specific implementation of double normalization and farthest point sampling in the preferred solution.

[0073] Step 106: Propagate the correspondence between superpoints to dense points, and calculate the rigid transformation matrix based on the correspondence between dense points.

[0074] After obtaining the superpoint correspondences, they need to be propagated to the dense point level, and then the optimal rigid transformation matrix is ​​calculated. The point matching module uses the superpoint correspondences as a guide to establish correspondences between dense points in the neighborhood of the superpoints. For each pair of matched superpoints (j... ref, j src Each node searches for candidate dense points within its neighborhood, establishes correspondences between dense points based on local feature similarity, and forms a set C of dense point correspondences. * .

[0075] Based on the set of correspondences of dense points C * By minimizing the objective function Solve for the optimal rigid transformation matrix T.

[0076] This optimization problem can be solved using the SVD decomposition method: First, calculate the centroid of the corresponding points in the source point cloud. The centroid of the point corresponding to the target point cloud Then calculate the covariance matrix after centroid removal. SVD decomposition of H The rotation matrix is ​​obtained. (If det(R) < 0, then it is necessary to...) Invert the last column of the matrix to get Recalculate Translation vector .

[0077] The final output is a rigid transformation matrix. ,in It is a 3×3 orthogonal rotation matrix. This is a three-dimensional translation vector. This transformation matrix is ​​used to align the preoperative source point cloud to the intraoperative target point cloud, providing accurate spatial positioning information for subsequent surgical navigation.

[0078] This step embodies the technical feature of propagating the correspondence between superpoints to dense points and calculating the rigid transformation matrix in the first aspect of the technical solution, as well as the specific implementation method of minimizing the objective function to solve the transformation matrix in the preferred solution.

[0079] Example 1: The technical solution of the present invention will be described in detail below through a specific embodiment of orthopedic surgical navigation.

[0080] In a spinal surgery navigation scenario, preoperative CT scans were used to collect lumbar spine data from the patient. A high-precision 3D model of the lumbar spine was reconstructed using Mimics software and converted into a point cloud format, resulting in a source point cloud P containing approximately 15,000 points, fully covering the complete geometry of the L3-L5 lumbar vertebrae. This model was imported into a navigation application specifically developed for HoloLens 2. During the surgery, the surgeon wore HoloLens 2 into the surgical field. The device first understood the operating room environment and achieved self-localization through its built-in SLAM (Simultaneous Localization and Mapping) system. Subsequently, the surgeon directly activated the HoloLens 2's depth sensing system via virtual interface buttons or voice commands to scan the exposed area of ​​the lumbar spine (posterior approach to the L4 vertebral body), obtaining a target point cloud Q (the acquired point cloud Q is immediately transmitted to the edge computing device). Due to limitations in surgical incision size and tissue occlusion, this point cloud only contains approximately 10,000 points, mainly covering part of the posterior structure of the L4 vertebral body.

[0081] In step 101, the system acquires the target point cloud Q (10,000 points) and the source point cloud P (15,000 points), with the overlapping area of ​​the two accounting for approximately 35% of the target point cloud. P and Q are input into HDFE-Net to obtain the transformation matrix, and then transmitted back to the HoloLens2 for registration.

[0082] In step 102, redundancy removal is performed on both point clouds using the P-ScConv framework. The number of input feature channels is set to C=64, the number of groups to G=8, and each group has 8 channels. The SRU module normalizes each group of features, and the gating mechanism identifies approximately 35% of the feature channels as primarily containing noise information (introduced by surgical light reflection and electrosurgical interference), reducing their weights. The CRU module sets α=0.6, dividing the 64 channels into 38 high-information channels and 26 low-information channels, performing differentiated processing on each before merging the output. After P-ScConv processing, the effective information density of the features is increased by approximately 40%, and the computational load is reduced by approximately 25%.

[0083] In step 103, four layers of KPConv downsampling are used to downsample the point cloud to approximately 3000, 750, 180, and 45 points, respectively. The number of feature channels in each layer is 64, 128, 256, and 512, respectively.

[0084] In step 104, the P-DFF module performs feature fusion at each layer of the decoder. Taking the first layer upsampling as an example, the input X1 is a deep feature of 45 points (512 channels), and X2 is a shallow feature of 180 points (256 channels). The channel attention mechanism calculates a 512-dimensional weight vector W. c Approximately 60% of the channels receive a weight greater than 0.5. The spatial attention mechanism generates 180 spatial weights W. s Bony landmarks (such as pedicles and articular processes) are given higher weight.

[0085] In step 105, features of 180 superpoints are extracted from each of the two point clouds. The similarity scoring matrix has a dimension of 180×180. After double normalization, the superpoints in the posterior L4 vertebral body region receive higher comprehensive anatomical scores, while the L3 and L5 regions receive lower scores due to the lack of corresponding structures in the target point cloud. The Top-K method selects the 512 candidate corresponding points with the highest scores, and after optimization by sampling the farthest point, 256 pairs of spatially uniformly distributed superpoint correspondences are output.

[0086] In step 106, the 256 pairs of superpoint correspondences are propagated to dense points, establishing approximately 8000 pairs of dense point correspondences. The rotation matrix R and translation vector t are obtained using the SVD method. Registration accuracy evaluation shows that the average relative rotation error is 0.89° and the average relative translation error is 2.3mm, meeting the sub-millimeter accuracy requirements for spinal surgery navigation. The entire registration process takes approximately 120ms, meeting the intraoperative real-time requirements.

[0087] Once registration is successful, the system directly and accurately renders the preoperative planning information (such as the pedicle screw path) holographically onto the real anatomical structure seen by the doctor through HoloLens 2, achieving true "what you see is what you treat." Doctors can move freely around the patient and observe the fusion effect of virtual navigation information and real anatomical structure from different perspectives without repeatedly checking external monitors.

[0088] Example 2: The applicability of the present invention is demonstrated below through another example of joint replacement surgery navigation.

[0089] In a total knee replacement surgery scenario, preoperative MRI scans were used to reconstruct three-dimensional models of the distal femur and proximal tibia, which were then converted into femoral point cloud P. f (40,000 points) and Tibial Origin Point Cloud P t (35,000 points). Point clouds of the joint surface were acquired intraoperatively using a structured light scanner. Due to interference from joint effusion and soft tissue, the target point cloud Q... f (12000 points) and Q t There is a lot of noise at 10,000 points.

[0090] In this embodiment, the SRU module of the P-ScConv framework sets up a dedicated channel group for articular surface curvature features, which can effectively distinguish between bony structures and soft tissue noise. The gating mechanism reduces the feature weight of the soft tissue region to below 0.2. The CRU module enhances the expression of articular surface geometric features through differential processing.

[0091] When dealing with complex curved structures like the knee joint, the P-DFF module's spatial attention mechanism effectively compensates for surface distortions introduced by upsampling. In the junction region between the femoral condyle and the tibial plateau, the channel attention mechanism adaptively enhances the weights of edge features.

[0092] The enhanced super-point matching module's dual normalization strategy effectively avoids mismatches between the medial and lateral femoral condyles, while the farthest point sampling ensures uniform coverage of the corresponding points across the entire articular surface. The final registration accuracy achieves an average relative rotation error of 0.74° and an average relative translation error of 2.16 mm, meeting the high-precision requirements of joint replacement surgery.

[0093] The present invention also provides a 3D model registration device based on hierarchical dynamic feature enhancement, the device comprising: The point cloud acquisition module is used to acquire the target point cloud collected during the operation and the source point cloud reconstructed before the operation. This module can interface with 3D acquisition devices such as optical navigation systems, structured light scanners, and depth cameras to receive target point cloud data in real time, and at the same time load the source point cloud model reconstructed before the operation from the medical imaging system.

[0094] The redundancy removal module is used to perform feature extraction and redundancy removal on the target point cloud and source point cloud through the P-ScConv framework. This module includes an SRU submodule for group normalization and gating screening and a CRU submodule for segmentation-transformation-fusion, which can effectively filter intraoperative environmental noise and retain key anatomical features.

[0095] The downsampling module is used to downsample the point cloud features after redundancy removal. This module adopts a hierarchical KPConv downsampling strategy to generate multi-scale point cloud feature representations.

[0096] The dynamic fusion module is used to perform channel attention weighting and spatial attention weighting fusion on downsampled multi-scale features through P-DFF. This module can dynamically adjust the fusion weights according to the input features, suppress noise propagation and compensate for geometric distortion.

[0097] The superpoint matching module is used to extract superpoint features from the fused point cloud features and establish superpoint correspondence through double normalization and farthest point sampling. This module can effectively reduce mismatches at similar anatomical structures and ensure the spatial uniform distribution of corresponding points.

[0098] The transformation calculation module is used to propagate the correspondence between superpoints to dense points and calculate the rigid transformation matrix based on the correspondence between dense points. This module outputs the optimal spatial transformation parameters, which are used to align the source point cloud to the target point cloud.

[0099] The specific functions and implementation methods of the above modules correspond to the corresponding steps in the aforementioned method embodiments, and will not be repeated here.

[0100] This invention also provides an electronic device, including a processor and a memory. The memory stores a computer program and transmits the computer program to the processor. The processor executes the hierarchical dynamic feature enhancement-based 3D model registration method as described in the foregoing method embodiments, according to instructions in the computer program. This electronic device can be a specialized medical device such as a surgical navigation workstation, a medical image processing server, or a robot-assisted surgical console, or it can be a general-purpose computing device equipped with a GPU accelerator card. The processor can employ a heterogeneous computing architecture combining a multi-core CPU and a GPU to meet the real-time requirements of deep learning inference. The memory should have sufficient capacity to store point cloud data and neural network model parameters.

[0101] This invention also provides a computer-readable storage medium storing computer-executable instructions. When executed by a computer processor, these instructions are used to perform the hierarchical dynamic feature-enhanced 3D model registration method as described in the foregoing method embodiments. The computer-readable storage medium can be a non-volatile storage medium, such as a hard disk, solid-state drive, optical disk, or USB flash drive, or a volatile storage medium, such as memory. The stored computer-executable instructions include program code implementing core algorithms such as the P-ScConv framework, P-DFF module, and enhanced superpoint matching module, as well as pre-trained neural network model parameters.

Claims

1. A 3D model registration method based on hierarchical dynamic feature enhancement, characterized in that, include: The target point cloud acquired during the operation and the source point cloud reconstructed before the operation are obtained, wherein the target point cloud and the source point cloud are both discrete point sets in three-dimensional space, and each point contains three-dimensional space coordinate information; Feature extraction and redundancy removal are performed on the target point cloud and the source point cloud, respectively. The input features are processed by SRU module and CRU module through the P-ScConv framework, and the point cloud features after redundancy removal are output. The redundant point cloud features are downsampled to obtain the downsampled point cloud features. The downsampled point cloud features are subjected to hierarchical dynamic fusion processing. The P-DFF module is used to perform channel attention weighting and spatial attention weighting on the multi-scale features, and the fused point cloud features are output. Superpoint feature extraction and matching are performed on the fused point cloud features, and a superpoint correspondence relationship between the target point cloud and the source point cloud is established through an enhanced superpoint matching module; The superpoint correspondence is propagated to the dense points, and the rigid transformation matrix is ​​calculated based on the dense point correspondence. The rigid transformation matrix includes a rotation matrix and a translation vector, used to align the source point cloud to the target point cloud.

2. The method according to claim 1, characterized in that, The process of processing input features using the P-ScConv framework with SRU and CRU modules includes: The P-ScConv framework is added to the front end of the encoder, and the input features are processed sequentially by the SRU module and then by the CRU module. The output features are represented as follows: , where X in As input features, X out For output features, f SRU This represents the processing function of the SRU module, f. CRU This represents the processing functions of the CRU module; In the SRU module, group normalization is used to group features, and a gating mechanism is used to quantify feature importance and filter anatomical features related to registration. In the CRU module, a segmentation-transformation-fusion strategy is used to differentiate different channel groups, and the processed features are dynamically weighted and fused.

3. The method according to claim 2, characterized in that, In the SRU module, group normalization is used to group features, and a gating mechanism is used to quantify feature importance, including: The input features are grouped by channel, and each group of features is normalized. The normalized features are then calculated. Where g is the group index, x g Let g be the eigenvector of the g-th group and N is the number of points, C is the total number of channels, G is the number of groups, L is the feature length, and E|x g | represents the mean of this set of features, Var(x) g ) represents the variance of this set of features, ε is a small constant to prevent division by zero, and γ g and β g For the learnable scaling and offset parameters corresponding to this group, ⊙ represents element-wise multiplication; The gating weight is calculated based on the normalized features. Where W is the gating weight and σ is the sigmoid activation function. This indicates summing over all groups; The feature space is filtered according to the gating weights, retaining features with weights close to 1 and marking features with weights close to 0 as redundant features.

4. The method according to claim 2, characterized in that, In the CRU module, a segmentation-transformation-fusion strategy is used to differentiate between different channel groups, including: Based on the coefficient α, the channel dimension is divided into high-information channels and low-information channels, with the number of high-information channels being... Low number of information channels α is the channel division coefficient, and its value ranges from 0 to 1, excluding itself. C total This represents the total number of channels; The high-information channels are subjected to parallel feature extraction using group convolution and point convolution; The low-information channel is expanded in dimension through point convolution and then concatenated with the original features for feature recombination; The features of the high-information channel and the low-information channel are dynamically weighted and fused to output the fused channel features.

5. The method according to claim 1, characterized in that, The process of applying channel attention weighting and spatial attention weighting to multi-scale features using the P-DFF module includes: One-dimensional convolution is performed on the input multi-scale features X1 and X2 to align the channel dimensions, resulting in the feature concatenation F. cat , where X1 is the first-scale feature and X2 is the second-scale feature; For the F cat Perform global average pooling to obtain F avg Calculate channel attention weights W c σ represents the channel attention weights, σ ​​is the Sigmoid activation function, and Conv1D represents a one-dimensional convolution operation. For the F cat Channel-weighted fusion is performed to obtain channel-weighted features. , where ⊙ denotes element-wise multiplication; One-dimensional convolutions are performed on X1 and X2 respectively to calculate the spatial weight components, thus obtaining the spatial attention weights. W s Spatial attention weights; By combining the channel attention weights and the spatial attention weights, the final fused feature is output. F fused The features are those after fusion, where C is the number of channels and w is the number of channels. c Let be the weight of the c-th channel.

6. The method according to claim 1, characterized in that, The step of establishing the superpoint correspondence between the target point cloud and the source point cloud through the enhanced superpoint matching module includes: Superpoint features are extracted from the target point cloud and the source point cloud to obtain the reference skeleton superpoint features F. ref Source skeleton superpoint features F src F ref For the superpoint features of the target point cloud, F src The superpoint features of the source point cloud; A similarity score matrix is ​​calculated based on the superpoint features. S ij The similarity score between the i-th reference superpoint and the j-th source superpoint is given by exp, which represents the exponential function. Represents Euclidean distance; The similarity scoring matrix is ​​subjected to double normalization to obtain a comprehensive anatomical score; Based on the comprehensive anatomical score, the Top-K method is used to select the top K candidate corresponding points with the highest matching scores, where K is the preset number of candidate points; Spatial distribution optimization is performed on the candidate corresponding points using farthest point sampling, and the final super-point correspondence is output.

7. The method according to claim 6, characterized in that, The process of performing double normalization on the similarity scoring matrix to obtain a comprehensive anatomical score includes: The similarity scoring matrix is ​​normalized along its rows to obtain the reference point cloud normalized score. ,in For the normalized score of the i-th reference superpoint, s i,j Let be the element in the i-th row and j-th column of the similarity score matrix. This represents the sum of all elements in the i-th row; The similarity scoring matrix is ​​normalized along its columns to obtain the normalized score of the source point cloud. ,in For the normalized score of the j-th source superpoint, This represents the sum of all elements in column j; Multiply the normalized score of the reference point cloud by the normalized score of the source point cloud to obtain the comprehensive analysis score. ,in The comprehensive anatomical score between the i-th reference superpoint and the j-th source superpoint.

8. The method according to claim 6, characterized in that, The step of optimizing the spatial distribution of the candidate corresponding points by sampling from the farthest point includes: The coordinates P of the candidate superpoint in the target point cloud ref and the coordinates P of the candidate superpoint in the source point cloud src Input the sampling function for the farthest point respectively, where P ref Let P be the set of 3D coordinates of candidate superpoints in the target point cloud. src This is the set of three-dimensional coordinates of candidate superpoints in the source point cloud; Get sampling index fps ref and FPS src This ensures that the final corresponding point pairs are evenly distributed in space, where fps ref For the sampling index of the target point cloud, fps src The sampling index for the source point cloud; The final corresponding index j is output based on the sampling index. ref and j src and the corresponding matching confidence ρ, where j ref j is the final corresponding index of the target point cloud. src ρ is the final corresponding index of the source point cloud, and ρ is the matching confidence score.

9. The method according to claim 1, characterized in that, The calculation of the rigid transformation matrix based on the correspondence of dense points includes: Based on the dense point correspondence set C * By minimizing the objective function Solve for the optimal rigid transformation matrix T, where C * Let R be the set of correspondences between dense points, and let R be the rotation matrix. SO(3) denotes the three-dimensional special orthogonal group, where t is the translation vector and R 3 p represents the three-dimensional real space. xi * Let q be the coordinates of the corresponding point in the source point cloud. yi * The coordinates of the corresponding point in the target point cloud. Represents Euclidean distance. This indicates an optimization solution for the rotation matrix and translation vector; Output the rigid transformation matrix , where SE(3) represents a three-dimensional special Euclidean group used to align the source point cloud to the target point cloud.

10. A 3D model registration device based on hierarchical dynamic feature enhancement, characterized in that, The method for 3D model registration based on hierarchical dynamic feature enhancement as described in any one of claims 1-9 includes: The point cloud acquisition module is used to acquire the target point cloud collected during the operation and the source point cloud reconstructed before the operation. A redundancy removal module is used to perform feature extraction and redundancy removal on the target point cloud and the source point cloud through the P-ScConv framework. The P-ScConv framework includes an SRU sub-module for group normalization and gating screening and a CRU sub-module for segmentation-transformation-fusion. The downsampling module is used to downsample the point cloud features after redundancy removal. The dynamic fusion module is used to perform channel attention weighting and spatial attention weighting fusion on downsampled multi-scale features using P-DFF; The superpoint matching module is used to extract superpoint features from the fused point cloud features and establish superpoint correspondence through double normalization and farthest point sampling. The transformation calculation module is used to propagate the correspondence between superpoints to dense points and calculate the rigid transformation matrix based on the correspondence between dense points.

Images