A point set registration method based on clustering

Through the cluster-based point set registration method, combined with Gaussian hybrid model and local linear embedding technology, the problem that the point set registration algorithm is prone to fall into local minimum values ​​under non-rigid transformation is solved, and more accurate point set registration and topological structure are achieved Keep.

CN113902004BActive Publication Date: 2024-11-19BEIJING YOUCHUANG YOUNENG TECHNOLOGY CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202111161681.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-09-30
Publication Date
2024-11-19
Estimated Expiration
2041-09-30

AI Technical Summary

Technical Problem

Existing point set registration algorithms are prone to local minimums in non-rigid transformations and are difficult to accurately ensure the global and local topology of point sets, especially in the presence of noise and outliers.

Method used

Using cluster-based point set registration method, the cluster center is used to perform preliminary registration through Gaussian hybrid model and local linear embedding technology, and combined with variation method and maximum expected algorithm, the local topology is integrated to avoid local best. Appear.

Benefits of technology

It improves the accuracy and efficiency of point set registration, reduces the probability of local optimal points, and can more accurately keep the overall and local structure of point set unchanged under non-rigid transformation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113902004B_ABST
    Figure CN113902004B_ABST
Patent Text Reader

Abstract

The invention discloses a point set registration method based on clustering, comprising: obtaining a reference point set and a target point set, clustering the reference point set and the target point set respectively to obtain a reference point set cluster center and a target point set cluster center, and a cluster correspondence matrix, expanding each cluster cluster pair in the cluster correspondence matrix into a point set pair of each cluster cluster pair corresponding to a point set to obtain a first objective function; fusing a displacement function model with the first objective function to obtain a second objective function, using a local linear trapping method to calculate the reference point set to obtain a cost function, fusing the cost function with the second objective function, and then using a maximum expectation algorithm to iterate the fused objective function to complete parameter optimization to obtain a final objective function. The method can accurately perform point set registration.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of image processing, and in particular relates to a point set registration method based on clustering. Background Art

[0002] The point set registration problem is a widely studied problem in computer vision, and the development of accurate point set registration algorithms has always been a research hotspot in the field of pattern recognition. The goal of point set registration is to find the correspondence between two sets of related point sets, or to restore the transformation relationship between two sets of point sets, so as to establish a one-to-one mapping between the two point sets. According to the different transformation methods, the registration problem can be divided into two categories: rigid registration problem and non-rigid registration problem. The transformation method of rigid registration only considers translation, rotation and scaling. The transformation method of non-rigid registration is usually complex, difficult to model, and contains more transformation parameters. In the real world, many problems will involve more non-rigid transformations, that is, they face the problem of non-rigid registration. Non-rigid registration is more challenging because of the unknown transformation model and the existence of a large number of transformation parameters.

[0003] At present, the research on rigid point set registration has achieved certain results. The iterative closest point (ICP) algorithm and its variants are one of the most widely used methods for rigid registration problems. The algorithm iteratively establishes the correspondence between the nearest neighbor point pairs and minimizes the distance between the point pairs. However, due to its assumption that the nearest point pairs between the target point set and the source point set correspond to each other, the algorithm has limitations. The convergence and accuracy of the algorithm largely depend on the similarity of the given point sets and are prone to fall into local minima. The robust point matching (RPM) algorithm used to determine the pose and correspondence of two point sets in space can prevent falling into local minima. The algorithm applies a deterministic annealing method to convexify their energy functions. However, the RPM algorithm is limited to affine and piecewise affine mappings. The thin plate spline robust point matching (TPS-RPM) algorithm attempts to solve the non-rigid mapping problem inherent in the RPM and ICP algorithms, and proposes a general framework to solve the non-rigid registration problem with outliers and noise. The coherent point drift (CPD) algorithm that emerged later is a probabilistic method that can align point sets under both rigid and non-rigid transformations. The algorithm optimizes the alignment of two point sets by estimating the probability density function, finds the optimal transformation parameters by maximizing the likelihood function, and iteratively fits the GMM centroid of the source point set with the data of the target point set. At the same time, in order to maintain the global topological structure of the point set, the center point is moved as a whole. However, there is a weight parameter w in the CPD algorithm, which is used to predict the noise level and the number of outliers in the GMM, and is manually selected under unclear constraints. To solve this problem, some researchers introduced an algorithm to automatically select the optimal value of w. In addition, in order to solve the limitation of the CPD algorithm that only considers the global structure of the point set, the global and local distance thin plate spline (GLMDTPS) non-rigid point set registration algorithm came into being. The algorithm evaluates the correspondence relationship based on global and local features, and the global and local features are fused into hybrid features through deterministic annealing technology. In the global-local topology preservation (GLTP) algorithm, a local descriptor called local linear embedding (LLE) is added to maintain the local structure between two point sets. A common problem in probabilistic non-rigid registration algorithms is that inaccurate correspondence estimates can easily cause the optimization process to fall into a local optimum.The present invention uses cluster centers representing local data information instead of the to-be-registered point set for coarse registration, thereby reducing the probability of local optimum points appearing in the registration, thereby effectively avoiding the optimization process from falling into the local optimum. Summary of the invention

[0004] The invention provides a point set registration method based on clustering, which can accurately perform point set registration.

[0005] A point set registration method based on clustering, comprising:

[0006] Obtain a reference point set and a target point set, cluster the reference point set and the target point set respectively to obtain the reference point set cluster center and the target point set cluster center, use the Gaussian mixture model point set registration method to obtain the cluster correspondence matrix of the reference point set cluster center and the target point set cluster center, expand each cluster cluster pair in the cluster correspondence matrix to a point set pair of corresponding point sets of each cluster cluster pair to obtain the first objective function, and complete the preliminary registration of the reference point set and the target point set;

[0007] The variational method is used to construct a displacement function model of the reference point set moving toward the target point set. The displacement function model is fused with the first objective function to obtain the second objective function, so that the overall structure of the reference point set remains unchanged during the registration process between the reference point set and the target point set. The local linear trapping method is used to calculate the reference point set to obtain the cost function, which is fused with the second objective function. The maximum expectation algorithm is then used to iterate the fused objective function to obtain the final objective function after parameter optimization. The reference point set completes the final registration to the target point set through the final objective function.

[0008] The present invention first uses clustering to cluster data point sets with larger dimensions, and then performs alignment to improve computational efficiency while preventing the occurrence of local optimal phenomena. After the initial alignment, local topology technology is used to maintain accurate alignment of the initial alignment, so that the reference point set can be moved while keeping the overall structure and local structure as unchanged as possible, thereby completing accurate alignment.

[0009] The K-means clustering method is used to cluster the reference point set and the target point set respectively. The target point set is clustered to generate b clusters, and the reference point set is clustered to generate a cluster.

[0010] The point set registration method of Gaussian mixture model is used to obtain the cluster correspondence matrix of the reference point set cluster center and the target point set cluster center, including:

[0011] The reference point cluster center is used as the mean of the Gaussian model, and the target point cluster center is used as the data point generated by the Gaussian model. The probability density function of the mixed Gaussian model is constructed as follows:

[0012]

[0013] Among them, x k is the cluster center of the k clusters of the target point set, ω is the parameter used, a is the number of clusters generated by the target point set clustering, b is the number of clusters generated by the reference point set clustering, l is the index of the reference point set cluster, based on the probability density function p(x k ), the maximum expectation algorithm is used to obtain the cluster correspondence matrix between the reference point set cluster center and the target point set cluster center.

[0014] Each cluster pair in the cluster correspondence matrix is ​​expanded to a point set pair of each cluster pair corresponding to a point set to obtain the first objective function, including:

[0015] Each cluster pair q in the cluster correspondence matrix l,k Expand to get the point set pair p i,j , point set pair p i,j With q l,k The relationship is:

[0016]

[0017] where p i,j It represents the correspondence between the i-th point of the k-th cluster of the target point set and the j-th point of the l-th cluster of the reference point set. are k clusters formed by clustering the target points. is the l clusters formed by the reference point set, and multiple point sets are clustered into p i,j A point set matrix P is constructed, and based on the point set matrix P, the first objective function is obtained using the Gaussian model.

[0018] The displacement function model is:

[0019]

[0020] Among them, v is the displacement function, G is the Gaussian kernel function, is the Fourier transform function of G, is the Fourier transform of v, s represents the frequency variable in the Fourier domain, Indicates that the data is D-dimensional.

[0021] The cost function is:

[0022]

[0023] Among them, G(m1,·) represents the m1th row of the Gaussian kernel function, W is the weight parameter, and L miis the weight matrix, that is, the domain information of the reference point set, m is the mth point in the reference point set, M is the number of points in the reference point set, K is the K nearest neighbor points of each point in the reference point set, and i is the index of the K nearest neighbor points.

[0024] The third objective function is partially derived based on variables, namely, conversion parameters and weight parameters to obtain conversion parameter partial derivatives and weight parameter partial derivatives, and the conversion parameter partial derivatives and weight parameter partial derivatives are set to zero to obtain the initially optimized conversion parameters and weight parameters, and then the maximum expectation algorithm is used to complete parameter optimization to obtain the final objective function.

[0025] The final objective function is:

[0026]

[0027] Among them, σ 2 is the conversion parameter, P old (·) is the posterior probability function, x n is the nth point in the target point set, y m is the mth point in the reference point set, λ and α are the trade-off parameters of the displacement function model and the cost function, respectively.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The present invention replaces the correspondence between the two point sets with the correspondence between the cluster centers of the two point sets, thereby replacing the point correspondence projection problem with the cluster center correspondence projection problem. Finally, the cluster correspondence projection result is converted back to point correspondence, which greatly improves the solution efficiency and reduces the probability of the local optimum in the registration, thereby effectively avoiding the optimization process from falling into the local optimum. In the point set registration framework based on the Gaussian mixture model, the local structure topology is added to maintain a more accurate registration of the two point sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A point set registration method process based on clustering is provided for a specific implementation method;

[0031] Figure 2 A registration process using a clustering-based point set registration method provided for a specific implementation;

[0032] Figure 3 A diagram showing the registration results of different deformation data sets of a cat model using a clustering-based point set registration method provided in a specific implementation manner;

[0033] Figure 4 The registration results of different deformation data sets of the lion model completed by using the clustering-based point set registration method provided in the specific implementation method;

[0034] Figure 5 A point set registration result of a cat model with outliers completed by a clustering-based point set registration method provided in a specific implementation manner;

[0035] Figure 6 The 3D monkey head data set registration result completed by using the clustering-based point set registration method provided in the specific implementation method;

[0036] Figure 7 A 3D head model registration result obtained by using a clustering-based point set registration method according to a specific implementation method;

[0037] Figure 8 The human body data set registration result completed by using the clustering-based point set registration method provided for the specific implementation manner. DETAILED DESCRIPTION

[0038] 1 Clustering-based probabilistic registration algorithm

[0039] The key to the study of clustering-based probabilistic registration algorithms is to integrate clustering projection and local topology preservation into the probabilistic registration framework based on Gaussian mixture models, thereby solving the problem of insufficient accuracy in point set registration algorithms and the problem that the optimization process is prone to fall into local optimality. First, the two given point sets are clustered separately; the cluster centers are registered based on the point set registration framework of the Gaussian mixture model, and then the corresponding relationship of the cluster centers is projected into the point set, and finally the registration is performed to preserve the local structure. In the following content, the general point set registration framework based on the Gaussian mixture model is first introduced; on this basis, the cluster correspondence projection method is given; and finally, the local structure topology preservation is introduced.

[0040] 1.1 Point set registration framework based on Gaussian mixture model

[0041] The point set registration framework based on Gaussian mixture model regards the alignment problem of two point sets as a probability density estimation problem, where one point set represents the mean of each sub-model of the Gaussian mixture model, and the other point set is represented as data points, and registration is performed by continuously resetting the position of the center of mass.

[0042] Given two point sets, the target point set X N×D =(x1,...,x n ,...,x N ) T , reference point set Y M×D =(y1,...,y m ,...,y M ) T, D is the dimension of the points in the point set, N is the number of points in the point set X, and M is the number of points in the point set Y. The points in the point set Y are used as the mean of the GMM, and the points in the point set X are used as the data points generated by the GMM. The probability density of each GMM sub-model is

[0043]

[0044] Among them, x n is the nth point of the X point set, y m is the mth point in the y point set, M is the number of the Y point set, assuming that all Gaussian components are independently distributed and have the same variance, and the probability of each component is the same. Considering the existence of noise and outliers in the data, an additional uniform distribution is introduced in the mixed model At this time, click x n The GMM probability density function is

[0045]

[0046] Combining the two probability distributions, the probability density function of the mixed model after weighting the two distributions using the parameter ω (0≤ω≤1) is

[0047]

[0048] Point set registration uses the method of minimum negative log-likelihood function to continuously reset the mean of GMM. At this time, the variance also changes with the change of the mean position. Let θ be the conversion parameter of GMM mean change, then the negative log-likelihood function can be written as

[0049]

[0050] The required parameters (θ,σ) are calculated by the maximum expectation (EM) algorithm. 2 ), so that the negative log-likelihood function is minimized.

[0051] y m and x n The correlation coefficient between two points defines the posterior probability

[0052]

[0053] Solving for P old (m|x n ) is the E-step of the EM algorithm, and the M-step uses the obtained posterior probability P old (m|x n ), by minimizing the negative log-likelihood function Q constructed by Jensen's inequality and Bayes' formula based on the above negative log-likelihood function, the new parameter value (θ,σ 2 ).

[0054]

[0055] The EM algorithm alternates between E-step and M-step until convergence.

[0056] According to the probability density function of the weighted mixture model above, and ignoring the sum parameters (θ,σ 2 ) is irrelevant, the previous function Q can be rewritten into a form related to the parameters to obtain the first objective function, that is,

[0057]

[0058] in τ(y m ,θ) refers to the transformation parameter θ, m The transformation implemented, the posterior probability density function is

[0059]

[0060] The posterior probability matrix P M×N It is the correspondence matrix, the matrix P M×N Each element in represents y m and x n The correlation coefficient between two points.

[0061] 1.2 Clustering and cluster corresponding projection

[0062] When the cardinality of the two point sets increases and the dimension of the point correspondence matrix increases, the efficiency of solving the point correspondence relationship will decrease. Therefore, we first seek an approximate solution to the problem through clustering. The main idea is to cluster the two point sets separately, and replace the correspondence relationship between the two point sets with the correspondence relationship between the cluster centers of the two point sets, thereby replacing the point correspondence projection problem with the cluster center correspondence projection problem. Finally, the cluster correspondence projection result is converted back to point correspondence.

[0063] Through experiments, among various clustering algorithms, subspace K-means clustering shows better registration accuracy. Therefore, the subspace K-means clustering algorithm is used to cluster the X point set into a clusters, which can be expressed as Cluster the Y point set into b clusters, expressed as In addition, use x k Indicates the cluster center of k clusters in the X point concentration, with y l represents the cluster center of cluster l in the Y point set. After clustering the two point sets, the cluster centers are aligned based on the Gaussian mixture model. At this time, the probability density function is

[0064]

[0065] After registration, we get a matrix Q l×k To represent the correspondence between the cluster centers of two point sets. l,k is the element in the matrix Q, which means the concentration of point X Concentrated with point Y The corresponding relationship.

[0066]

[0067] Copy the corresponding relationship in the cluster center correspondence Q to construct the point correspondence P, that is, To achieve the effect of projecting the cluster correspondence matrix Q to the point correspondence matrix P, where p i,j The i-th point of the k-th cluster of the X-point set is represented by clustering the X-point set into a cluster, and the j-th point of the l-th cluster of the Y-point set is represented by clustering the Y-point set into b clusters.

[0068] 1.3 Topology Preservation

[0069] After obtaining the correspondence between point sets by cluster correspondence projection, local structure topology preservation is added to the point set registration framework based on Gaussian mixture model to more accurately align the two point sets.

[0070] For the registration problem of non-rigid point set data, the transformation relationship T(Y,v) is defined as follows:

[0071] T(Y,v)=Y+v(Y)

[0072] Where Y represents the initial position of the Y point set and v represents the displacement function. In the case of non-rigid transformation, registration usually faces ill-posed problems due to the lack of proper regularization. Motion Coherence Theory (MCT) believes that adjacent points tend to move coherently, so the displacement function between point sets should be smooth. The above displacement function v(Y) is estimated using the variational method.

[0073] Define a rule item

[0074]

[0075] Among them, v is the displacement function, G is the Gaussian kernel function, is the Fourier transform function of G, is the Fourier transform of v, s represents the frequency variable in the Fourier domain, Indicates that the data is D-dimensional.

[0076] Add the defined transformation and regularization terms to the objective function to obtain the second objective function:

[0077]

[0078] In the above formula, the displacement function v(y m ) is defined as the Fourier transform form, that is, in is the Fourier transform of v, s represents the frequency variable in the Fourier domain, R D Indicates that the data is D-dimensional and j is an imaginary unit.

[0079] To find the minimum value of this objective function, first find the The partial derivative of , and then set the partial derivative equal to zero, we get:

[0080]

[0081] where a m,n =P(m|x n )[x n -(y m +v(y m ))].

[0082] Define a new coefficient Substitute into the above formula and rearrange it to get the following formula:

[0083]

[0084] Perform Fourier transform on both sides of the above equation and get the following equation:

[0085]

[0086] Adding a term ψ(z) in the null space to the above equation gives the general representation of the optimal solution as shown below, (what does Z represent)

[0087]

[0088] After obtaining the expression of the smooth function v, the problem of solving v becomes solving the coefficient w m .

[0089] Define a coefficient w m The matrix W formed M×D =(w1,...,w M ) T , and define a The kernel matrix G M×M Therefore, the transformation function T can be re-expressed as follows:

[0090] T=T(Y,W)=Y+GW

[0091] At this time, the displacement function v can also be re-expressed as shown below:

[0092]

[0093] So the second objective function can be re-expressed as follows:

[0094]

[0095] Where tr(·) represents the trace of the matrix. To find the minimum value of this objective function, first find the partial derivative with respect to the variable W, then set the partial derivative equal to zero, and after sorting and calculating, we can get:

[0096] (G+λσ 2 d -1 (P1))W=d -1 (P1)PX-Y

[0097] where d -1 (·) represents a diagonal matrix, 1 represents a column vector of all 1s, and W is calculated from this formula to obtain the coefficient w m .

[0098] Similarly, the parameter σ is required 2 We need to find the objective function with respect to the variable σ 2 The partial derivative of , and setting the partial derivative result to zero, we get:

[0099]

[0100] At this point, the corresponding transformation relationship between the two point sets has been found. Next, the aforementioned EM algorithm is used to iterate the parameters and find the optimal value.

[0101] The application of the above-mentioned theory of interferometric motion enables all points to move together as a whole, maintaining the global topology and preserving the connectivity and integration of different parts.

[0102] But it may not be enough to handle detailed local deformations. Therefore, it is necessary to maintain the local neighborhood structure and adapt to different local deformations while maintaining the global structure. The Local Linear Entanglement (LLE) algorithm is used as a nonlinear dimensionality reduction method to maintain the local neighborhood structure in a low-dimensional latent space. The local structure is represented as the spatial relationship between each vertex and its pre-selected nearest neighbor. In the mapping from high dimension to low dimension, the local structure is maintained by sharing the same set of local linear reconstruction coefficients.

[0103] During registration, we hope that the local structure of the Y point set can be preserved after non-rigid transformation. The LLE idea is applied to perform non-rigid registration. First, the K nearest neighbors of each point in the Y point set are calculated based on the Euclidean distance; second, each point in the Y point set is represented by a weighted linear combination of its neighboring points, and the weight can be obtained by minimizing the cost function, which is

[0104]

[0105] The weight matrix L mi , contains the neighborhood information of each point in the Y point set, where m∈(1···M), i∈(1···K).

[0106] Finally, by Y tran =Y+GW After the non-rigid transformation controlled by the weight matrix W defined in the formula, the total reconstruction error of all points with corresponding weights is calculated. To obtain the optimal W, minimize the cost function:

[0107]

[0108] G(m,·) represents the m-th row of G.

[0109] By leveraging complementary global and local topological regularization terms, both local structures and global connectivity can be preserved during complex non-rigid transformations involving various local deformations.

[0110] So far, the point set registration problem is regarded as a regularized density estimation problem based on a Gaussian mixture model. According to the general formula of the objective function and the two regularization terms that maintain the global local topology, the final objective function is

[0111]

[0112] Among them, σ 2 is the conversion parameter, P old (·) is the posterior probability function, x n is the nth point in the target point set, y m is the mth point in the reference point set, λ and α are the trade-off parameters of the displacement function model and the cost function respectively. Figure 2 shown.

[0113] 1.4 Algorithm

[0114] Given a set of X points and a set of Y points, cluster them, and then first align the cluster centers, find the corresponding relationship between the cluster centers, and copy the corresponding relationship to the corresponding relationship of the point sets. Based on this corresponding relationship, update the position of the Y point set to make the initial state between the X point set and the Y point set closer; then align them in combination with local topology preservation. The algorithm flow is as follows Figure 1 shown.

[0115] Experiments were conducted on four types of data: Chui-Rangarajan synthetic dataset, Stanford synthetic dataset, deformable 3D object dataset and 3D landscape human dataset: 1) 3D animal point set; 2) 3D head point set; 3) 3D human point set; and the experimental results were compared and analyzed with the existing advanced non-rigid registration algorithm GLTP. The software environment for all experiments is MATLAB R2016a, and the hardware environment is 16GB memory and AMD Ryzen 5 2600 six-core processor.

[0116] To demonstrate the registration process, we first conduct experiments on the fish dataset in the Chui-Rangarajan synthetic data, which is used by most point set registration algorithms to test performance. The template point set and the target point set of the point set shape model have the same number of points. First, the template point set X and the target point set Y are clustered to obtain cluster a of the X point set and cluster b of the Y point set, and the cluster centers of each cluster are registered. Then, the correspondence of the cluster centers is copied to the correspondence between the point sets, and finally the correspondence of the point sets is further registered to complete the registration.

[0117] Animal Point Set

[0118] First, a clustering-based point set registration method is tested on a 3D animal dataset, where the points are extracted from 3D mesh models. The dataset contains several animal models: lion, containing 5000 points, cat, containing 7207 points, horse, containing 8431 points, etc., and the number of points in the template point set and the target point set of each shape model is the same. Taking the cat model and the lion model as examples, a clustering-based point set registration method is experimentally analyzed. Figure 3 and Figure 4 The registration results of the two models under different deformation degrees are shown. As shown in the figure, for two datasets with different deformations, a clustering-based point set registration method provides good registration results, especially in the highly variable limbs and tail parts.

[0119] The Euclidean distance is used as the registration error to evaluate the performance of the registration algorithm. Table 1 shows Figure 3 Table 2 shows the comparison of the registration errors of GLTP and a clustering-based point set registration method under three different deformations of the Chinese cat model. Figure 4 The registration errors of GLTP and a clustering-based point set registration method are compared under three different deformations of the Chinese cat model. The results show that the clustering-based point set registration method has more advantages.

[0120] Table 1 Comparison of registration errors of different deformation datasets of 7207-point cat model

[0121]

[0122] Table 2 Comparison of registration errors of 5000-point lion model with different deformation datasets

[0123]

[0124] In addition, the performance of clustering-based point set registration methods is evaluated in matching incomplete shapes (missing some data) and with outliers. Figure 5 The results of deleting some points on the tail of the cat model and randomly adding some outliers to the dataset are shown. As shown in the figure, the clustering-based point set registration method still has good registration results in the presence of outliers.

[0125] 3D head point set

[0126] We also conducted experiments on animal head datasets and human head datasets, which are from the Stanford synthetic dataset. The number of target points and template points in the monkey head dataset is the same, both 7958 points. Figure 6 The registration results of the 3D monkey head dataset in the algorithm are shown.

[0127] Figure 6 The monkey head model shows good registration results, where the registration error in the GLTP algorithm is 5.55E-05, and the registration error in the clustering-based point set registration method is 5.51E-05. The results show that the clustering-based point set registration method has more advantages.

[0128] The number of target points and template points in the human head dataset is the same, both are 15941 points. The registration results are as follows: Figure 7 As shown in the figure, the head dataset contains subtle facial expressions, which places higher requirements on the registration algorithm. The results show that the clustering-based point set registration method for the head and facial point sets also has good registration results.

[0129] 3D human body point set

[0130] Finally, experiments were conducted on a more challenging 3D human body dataset, in which the number of target points is 12,500 and the number of template points is 2,150. The template point set is manually labeled and has significant shape and pose differences compared to the target point set. Figure 8 The following figure shows the registration results of the algorithm for 3D human body datasets at different deformation levels. Figure 8 As shown in Figure 2, for two human body datasets with different deformations, the clustering-based point set registration method provides good registration results.

[0131] Table 4 shows Figure 8The registration errors of GLTP and clustering-based point set registration methods are compared under different deformations of the human body model. The results show that the clustering-based point set registration method has more advantages.

[0132] Table 4 Comparison of registration errors of different deformation datasets of human body models

[0133]

Claims

1. A point set registration method based on clustering, It is characterized in that include: Obtain a reference point set and a target point set, cluster the reference point set and the target point set respectively to obtain the reference point set cluster center and the target point set cluster center, use the Gaussian mixture model point set registration method to obtain the cluster correspondence matrix of the reference point set cluster center and the target point set cluster center, expand each cluster cluster pair in the cluster correspondence matrix to a point set pair of corresponding point sets of each cluster cluster pair to obtain the first objective function, and complete the preliminary registration of the reference point set and the target point set; The variational method is used to construct a displacement function model of the reference point set moving toward the target point set. The displacement function model is fused with the first objective function to obtain the second objective function, so that the overall structure of the reference point set remains unchanged during the registration process between the reference point set and the target point set. The local linear trapping method is used to calculate the reference point set to obtain the cost function, which is fused with the second objective function. The maximum expectation algorithm is then used to iterate the fused objective function to obtain the final objective function after parameter optimization. The reference point set completes the final registration to the target point set through the final objective function. The point set registration method of Gaussian mixture model is used to obtain the cluster correspondence matrix of the reference point set cluster center and the target point set cluster center, including: The reference point cluster center is used as the mean of the Gaussian model, and the target point cluster center is used as the data point generated by the Gaussian model. The probability density function of the mixed Gaussian model is constructed as follows: Among them, x k is the cluster center of the k clusters of the target point set, ω is the parameter used, a is the number of clusters generated by the target point set clustering, b is the number of clusters generated by the reference point set clustering, l is the index of the reference point set cluster, based on the probability density function p(x k ), the maximum expectation algorithm is used to obtain the clustering correspondence matrix between the reference point set cluster center and the target point set cluster center; Each cluster pair in the cluster correspondence matrix is ​​expanded to a point set pair of each cluster pair corresponding to a point set to obtain the first objective function, including: Each cluster pair q in the cluster correspondence matrix l,k Expand to get the point set pair p i,j , point set pair p i,j With q l,k The relationship is: where p i,j It represents the correspondence between the i-th point of the k-th cluster of the target point set and the j-th point of the l-th cluster of the reference point set. are k clusters formed by clustering the target points. is the l clusters formed by the reference point set, and multiple point sets are clustered into p i,j A point set matrix P is constructed, and based on the point set matrix P, the first objective function is obtained using the Gaussian model; The displacement function model is: Among them, v is the displacement function, G is the Gaussian kernel function, is the Fourier transform function of G, is the Fourier transform of v, s represents the frequency variable in the Fourier domain, Indicates that the data is D-dimensional; The cost function is: Among them, G(m1,·) represents the m1th row of the Gaussian kernel function, W is the weight parameter, and L mi is the weight matrix, that is, the domain information of the reference point set, m is the mth point in the reference point set, M is the number of points in the reference point set, K is the K nearest neighbor points of each point in the reference point set, and i is the index of the K nearest neighbor points; The final objective function is: Among them, σ 2 is the conversion parameter, P old (·) is the posterior probability function, x n is the nth point in the target point set, y m is the mth point in the reference point set, λ and α are the trade-off parameters of the displacement function model and the cost function, respectively.

2. The point set registration method based on clustering according to claim 1, It is characterized in that The K-means clustering method is used to cluster the reference point set and the target point set respectively. The target point set is clustered to generate b clusters, and the reference point set is clustered to generate a cluster.

3. The clustering point set registration method according to claim 1, It is characterized in that The third objective function is partially derived based on variables, namely, conversion parameters and weight parameters to obtain conversion parameter partial derivatives and weight parameter partial derivatives, and the conversion parameter partial derivatives and weight parameter partial derivatives are set to zero to obtain the initially optimized conversion parameters and weight parameters, and then the maximum expectation algorithm is used to complete parameter optimization to obtain the final objective function.

Citation Information

Patent Citations

  • Recovering objects by using shape priors for level set representations

    AU2003207796A1

  • Non-rigid point set registration method for reserving partial structure

    CN106485741A

  • Spectral clustering for multi-type relational data

    US20080294686A1

  • Image Registration

    US20120328203A1