A multi-view 3D point cloud registration method and device

Abstract

This invention discloses a multi-view 3D point cloud registration method, comprising the following steps: Step 1: acquiring multi-view point clouds to be registered and initializing the algorithm; Step 2: clustering the multi-view point clouds to be registered using the K-means clustering algorithm and updating the centroid of each cluster to obtain shape point clouds; Step 3: establishing a multi-view point cloud registration optimization model based on the KMPE loss function between the point clouds to be registered and the shape point clouds; Step 4: solving the multi-view point cloud registration optimization model based on the LM algorithm to obtain the optimal rigid body transformation between the point clouds to be registered and the shape point clouds from each viewpoint; Step 5: repeating steps 2-4 until the maximum number of iterations is reached, and outputting the optimal rigid body transformation as the registration result. This method is effective in suppressing the influence of outliers on the registration result for multi-view 3D point clouds containing outliers, improving the accuracy of multi-view point cloud registration, and the registration process is rapid and highly practical.

Description

Technical Field

[0001] This invention belongs to the field of point cloud registration technology, specifically relating to a multi-view three-dimensional point cloud registration method and apparatus. Background Technology

[0002] Point cloud registration aims to correctly register multiple point clouds to the same coordinate system to form a more complete point cloud. It is mainly widely used in 3D reconstruction, parameter evaluation, localization and pose estimation, and point cloud registration technology is also involved in emerging applications such as autonomous driving, robotics and augmented reality.

[0003] Depending on the number of point clouds to be registered, point cloud registration problems can be divided into two main categories: pairwise point cloud registration and multi-view point cloud registration. Currently, pairwise point cloud registration is more widely studied, typically employing the Iterative Closest Point (ICP) algorithm or its optimized variants. For two point cloud datasets with no completely corresponding exterior points, the ICP algorithm achieves high pairwise point cloud registration accuracy. Multi-view point cloud registration usually involves alternating registration and merging of two point clouds until all point cloud sets are registered and merged into the same model. The drawback of this method is that the registration process is prone to getting trapped in local extrema, leading to low registration accuracy. Bergevin et al. proposed using the correspondence between a point cloud and other point clouds to estimate the rigid transformation of that point cloud; however, this method requires establishing the correspondence between each point cloud and other point clouds, making the registration process very time-consuming and affecting its practicality.

[0004] Furthermore, as the number of point clouds to be registered increases, the performance of existing registration methods deteriorates significantly, making them unsuitable for multi-view 3D point cloud registration problems. Moreover, due to the variability of acquisition angles and the complexity of acquisition environments, the acquired multi-view point clouds will inevitably contain outliers. Most existing multi-view point cloud registration methods do not consider the impact of outliers on the multi-view registration accuracy, resulting in low multi-view point cloud registration accuracy. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a multi-view 3D point cloud registration method and apparatus. The technical problem to be solved by this invention is achieved through the following technical solution:

[0006] In a first aspect, the present invention provides a multi-view 3D point cloud registration method, comprising:

[0007] Step 1: Obtain multi-view point clouds to be registered and initialize the algorithm;

[0008] Step 2: Use the K-means clustering algorithm to cluster the multi-view point cloud to be registered, and update the centroid of each cluster to obtain the shape point cloud;

[0009] Step 3: Establish a multi-view point cloud registration optimization model based on the KMPE loss function between the point cloud to be registered and the shape point cloud;

[0010] Step 4: Solve the multi-view point cloud registration optimization model based on the LM algorithm to obtain the optimal rigid body transformation of the point cloud to be registered and the shape point cloud under each view.

[0011] Step 5: Repeat steps 2-4 until the maximum number of iterations is reached, and output the optimal rigid body transformation as the registration result.

[0012] In one embodiment of the present invention, step 1 includes:

[0013] Let P = {P1, P2, L, P...} be N different viewpoints containing exterior points to be registered. N}, where the point cloud at the i-th viewpoint is The total number of points in the point cloud from each viewpoint is

[0014] Given the initial rigid transformation of each point cloud The maximum number of iterations Q and the number of clusters K in the K-means clustering algorithm are set, and the current iteration number q = 1; where, and Let represent the initial rotation matrix and initial translation vector of the point cloud at the i-th viewpoint, respectively;

[0015] according to For all point clouds to be registered, P = {P1, P2, L, P} N Perform registration, and randomly select K points from the registered point cloud as the initial centroids for clustering, denoted as .

[0016] In one embodiment of the present invention, step 2 includes:

[0017] All point clouds to be registered P = {P1, P2, L, P N The 3D points in} are assigned to K clusters, and the expression is:

[0018]

[0019] in, Point p i,j The cluster number in the q-th iteration. and Let these represent the rotation matrix and translation vector of the i-th viewpoint point cloud during the (q-1)-th iteration. Let represent the centroid of the k-th cluster during the (q-1)-th iteration;

[0020] Update the centroids of the K clusters according to the following formula.

[0021]

[0022] The centroids of the updated K clusters The resulting point cloud data is referred to as a shape point cloud, denoted as...

[0023] In one embodiment of the present invention, step 3 includes:

[0024] By introducing the KMPE loss function, which is robust to external points, a multi-view point cloud registration optimization model based on the KMPE loss function is established to transform the multi-view point cloud registration problem into a minimum optimization problem of a nonlinear optimization model; wherein, the multi-view point cloud registration optimization model is expressed as:

[0025]

[0026] in, Let represent the optimal rigid body transformation of the point cloud at the q-th viewpoint in the i-th viewpoint. This indicates that the cluster number is [missing information] during the q-th iteration. The centroid of the cluster, σ represents the Gaussian kernel bandwidth, and p represents the power parameter.

[0027] In one embodiment of the present invention, step 4 includes:

[0028] The optimal rigid body transformation in the (q-1)th iteration Using the initial values ​​of the LM algorithm, the multi-view point cloud registration optimization model is optimized and solved using the LM algorithm to obtain the optimal rigid body transformation after the q-th iteration.

[0029] Secondly, the present invention provides a multi-view 3D point cloud registration device, comprising:

[0030] The initialization module is used to acquire multi-view point clouds to be registered and initialize the algorithm;

[0031] The clustering module is used to cluster the multi-view point cloud to be registered using the K-means clustering algorithm and update the centroid of each cluster to obtain the shape point cloud.

[0032] The model building module is used to establish a multi-view point cloud registration optimization model based on the KMPE loss function between the point cloud to be registered and the shape point cloud.

[0033] The calculation module is used to solve the multi-view point cloud registration optimization model based on the LM algorithm to obtain the optimal rigid body transformation of the point cloud to be registered and the shape point cloud under each view.

[0034] The output module is used to output the optimal rigid body transformation as the registration result.

[0035] The beneficial effects of this invention are:

[0036] The multi-view 3D point cloud registration method provided by this invention applies clustering algorithm and rigid body transformation estimation iteratively to multi-view point cloud registration, and establishes a robust multi-view point cloud registration optimization model, thereby transforming the multi-view point cloud registration problem into a minimum optimization problem of the optimization model. For multi-view 3D point clouds containing outliers, this method can effectively suppress the influence of outliers on the registration results, improve the accuracy of multi-view point cloud registration, and the registration process is fast and highly practical.

[0037] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the multi-view 3D point cloud registration method provided in an embodiment of the present invention;

[0039] Figure 2 This is another flowchart illustrating the multi-view 3D point cloud registration method provided in this embodiment of the invention;

[0040] Figure 3 This is a schematic diagram of the structure of the multi-view three-dimensional point cloud registration device provided in the embodiment of the present invention. Detailed Implementation

[0041] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0042] Example 1

[0043] This invention addresses the problem of multi-view point cloud registration with outliers by proposing a robust multi-view point cloud registration method based on KMPE (Kernel mean p-power error). The goal of multi-view point cloud registration is to obtain the optimal rigid body transformation between point clouds from different perspectives, which can typically be divided into multiple pairwise point cloud registration problems. This invention utilizes the K-means algorithm to cluster 3D point clouds, using the centroids of the clusters as the model point clouds for multi-view point cloud registration. Each point cloud is then sequentially rigidly registered with this model point cloud, i.e., a robust rigid body transformation optimization model is constructed by introducing a KMPE loss function robust to outliers. The Levenberg-Marquardt (LM) algorithm is then used to solve this optimization model to obtain accurate rigid body transformation values.

[0044] Please see Figure 1-2 , Figure 1 This is a flowchart illustrating the multi-view 3D point cloud registration method provided in this embodiment of the invention. Figure 2 This is another flowchart illustrating the multi-view 3D point cloud registration method provided in this embodiment of the invention. Specifically, the method includes:

[0045] Step 1: Obtain multi-view point clouds to be registered and initialize the algorithm.

[0046] First, assume that the point clouds to be registered, containing outliers, from N different viewpoints are P = {P1, P2, L, P...} N}, where the point cloud at the i-th viewpoint is The total number of points in the point cloud from each viewpoint is

[0047] Then, given the initial rigid transformation of each point cloud. The maximum number of iterations Q and the number of clusters K in the K-means clustering algorithm are set, and the current iteration number q = 1. and Let represent the initial rotation matrix and initial translation vector of the point cloud at the i-th viewpoint, respectively.

[0048] Finally, according to For all point clouds to be registered, P = {P1, P2, L, P} N Perform registration, and randomly select K points from the registered point cloud as the initial centroids for clustering, denoted as .

[0049] Step 2: Use the K-means clustering algorithm to cluster the multi-view point clouds to be registered, and update the centroid of each cluster to obtain the shape point cloud.

[0050] Specifically, all point clouds to be registered, P = {P1, P2, L, P...} N The 3D points in} are assigned to K clusters, and the expression is:

[0051]

[0052] in, Point p i,j The cluster number in the q-th iteration. and Let these represent the rotation matrix and translation vector of the i-th viewpoint point cloud during the (q-1)-th iteration. Let represent the centroid of the k-th cluster during the (q-1)-th iteration;

[0053] Update the centroids of the K clusters according to the following formula.

[0054]

[0055] The centroids of the updated K clusters The resulting point cloud data is used as a shape point cloud, i.e., the model point cloud for multi-view point cloud registration, denoted as...

[0056] Step 3: Establish a multi-view point cloud registration optimization model based on the KMPE loss function between the point cloud to be registered and the shape point cloud.

[0057] In this embodiment, for the point clouds to be registered from various viewpoints Shape point cloud The registration problem first requires solving the correspondence between the point cloud to be registered and the shape point cloud. Based on p... i,j The assigned cluster number The correspondence between the point cloud to be registered and the shape point cloud can be obtained as follows: Point cloud to be registered point p i,j Shape point cloud In For the corresponding points.

[0058] Secondly, establish the point clouds to be registered from various perspectives. Shape point cloud A robust rigid body transformation optimization model is used to obtain the point cloud P to be registered from various viewpoints by solving this optimization model. i With shape point cloud S q The optimal rigid body transformation.

[0059] Step 3 specifically includes:

[0060] By introducing the KMPE loss function, which is robust to external points, a multi-view point cloud registration optimization model based on the KMPE loss function is established to transform the multi-view point cloud registration problem into a minimum optimization problem of a nonlinear optimization model; wherein, the multi-view point cloud registration optimization model is expressed as:

[0061]

[0062] in, Let represent the optimal rigid body transformation of the point cloud at the q-th viewpoint in the i-th viewpoint. This indicates that the cluster number is [missing information] during the q-th iteration. The centroid of the cluster, σ represents the Gaussian kernel bandwidth, and p represents the power parameter.

[0063] Step 4: Solve the multi-view point cloud registration optimization model based on the LM algorithm to obtain the optimal rigid body transformation of the point cloud to be registered and the shape point cloud under each view.

[0064] Specifically, the optimal rigid body transformation of the (q-1)th iteration As initial values ​​for the LM algorithm, the LM algorithm is used to optimize the multi-view point cloud registration optimization model to obtain the optimal rigid body transformation after the qth iteration.

[0065] Step 5: Repeat steps 2-4 until the maximum number of iterations is reached, and output the optimal rigid body transformation as the registration result.

[0066] If the current iteration number q satisfies q < Q, then let q = q + 1 and go to step 2; otherwise, output the result obtained in the Qth iteration. As the optimal rigid body transformation for multi-view point cloud registration.

[0067] The multi-view 3D point cloud registration method provided by this invention applies clustering algorithm and rigid body transformation estimation iteratively to multi-view point cloud registration, and establishes a robust multi-view point cloud registration optimization model, thereby transforming the multi-view point cloud registration problem into a minimum optimization problem of the optimization model. For multi-view 3D point clouds containing outliers, this method can effectively suppress the influence of outliers on the registration results, improve the accuracy of multi-view point cloud registration, and the registration process is fast and highly practical.

[0068] Example 2

[0069] Based on Embodiment 1 above, this embodiment provides a multi-view 3D point cloud registration device. Please refer to... Figure 3 , Figure 3 This is a schematic diagram of the structure of the multi-view 3D point cloud registration device provided in an embodiment of the present invention, which includes:

[0070] Initialization module 1 is used to acquire multi-view point clouds to be registered and initialize the algorithm;

[0071] Clustering module 2 is used to cluster the multi-view point cloud to be registered using the K-means clustering algorithm and update the centroid of each cluster to obtain the shape point cloud;

[0072] Model building module 3 is used to establish a multi-view point cloud registration optimization model based on the KMPE loss function between the point cloud to be registered and the shape point cloud;

[0073] Calculation module 4 is used to solve the multi-view point cloud registration optimization model based on the LM algorithm to obtain the optimal rigid body transformation of the point cloud to be registered and the shape point cloud under each view.

[0074] Output module 5 is used to output the optimal rigid body transformation as the registration result.

[0075] The multi-view 3D point cloud registration device provided in this embodiment can realize the multi-view 3D point cloud registration method provided in Embodiment 1 above. The implementation process is the same as that in Embodiment 1 above, and will not be described in detail here.

[0076] Therefore, the multi-view 3D point cloud registration device provided in this embodiment can also effectively suppress the influence of external points on the registration results, thereby improving the accuracy of multi-view point cloud registration.

[0077] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A multi-view 3D point cloud registration method, characterized in that, include: Step 1: Obtain multi-view point clouds to be registered and initialize the algorithm, specifically including: Let P = {P1, P2, ..., Pn} be N different viewpoints containing outliers to be registered. N }, where the point cloud at the i-th viewpoint is j represents the j-th point in the point cloud of the i-th viewpoint. The total number of points in the point clouds of all viewpoints is: D i This represents the total number of points in the point cloud at the i-th viewpoint; Given the initial rigid body transformations of each point cloud The maximum number of iterations Q and the number of clusters K in the K-means clustering algorithm are set, and the current iteration number q = 1; where, and Let represent the initial rotation matrix and initial translation vector of the point cloud at the i-th viewpoint, respectively; according to For all point clouds to be registered, P = {P1, P2, ..., P} N Perform registration, and randomly select K points from the registered point cloud as the initial centroids for clustering, denoted as . Step 2: Cluster the multi-view point cloud to be registered using the K-means clustering algorithm and update the centroid of each cluster to obtain the shape point cloud. Specifically, this includes: All point clouds to be registered P = {P1, P2, ..., P N The 3D points in} are assigned to K clusters, and the expression is: in, Point p i,j The cluster number in the q-th iteration. and Let these represent the rotation matrix and translation vector of the i-th viewpoint point cloud during the (q-1)-th iteration. This represents the centroid of the k-th cluster during the (q-1)-th iteration. This indicates finding the square of the L2 norm; Update the centroids of the K clusters according to the following formula. in, Let represent the centroid of the k-th cluster during the q-th iteration; The centroids of the updated K clusters The resulting point cloud data is referred to as a shape point cloud, denoted as... Step 3: Establish a multi-view point cloud registration optimization model based on the KMPE loss function between the point cloud to be registered and the shape point cloud, specifically including: By introducing the KMPE loss function, a multi-view point cloud registration optimization model based on the KMPE loss function is established to transform the multi-view point cloud registration problem into a minimum optimization problem of a nonlinear optimization model; wherein, the multi-view point cloud registration optimization model is expressed as: Among them, (R) i ,t i ) represents the rigid body transformation of the point cloud at the i-th viewpoint. Let represent the optimal rigid body transformation of the point cloud at the q-th viewpoint in the i-th viewpoint. This indicates that the cluster number is [missing information] during the q-th iteration. The centroid of the cluster, σ represents the Gaussian kernel bandwidth, and p represents the power parameter; Step 4: Solve the multi-view point cloud registration optimization model based on the LM algorithm to obtain the optimal rigid body transformation of the point cloud to be registered and the shape point cloud under each view. Step 5: Repeat steps 2-4 until the maximum number of iterations is reached, and use the final determined optimal rigid body transformation as the registration result.

2. The multi-view 3D point cloud registration method according to claim 1, characterized in that, Step 4 includes: The optimal rigid body transformation in the (q-1)th iteration Using the initial values ​​of the LM algorithm, the multi-view point cloud registration optimization model is optimized and solved using the LM algorithm to obtain the optimal rigid body transformation after the q-th iteration.

3. A multi-view 3D point cloud registration device, characterized in that, include: The initialization module (1) is used to acquire multi-view point clouds to be registered and initialize the algorithm, specifically including: Let P = {P1, P2, ..., Pn} be N different viewpoints containing outliers to be registered. N }, where the point cloud at the i-th viewpoint is j represents the j-th point in the point cloud of the i-th viewpoint. The total number of points in the point clouds of all viewpoints is: D i This represents the total number of points in the point cloud at the i-th viewpoint; Given the initial rigid body transformations of each point cloud The maximum number of iterations Q and the number of clusters K in the K-means clustering algorithm are set, and the current iteration number q = 1; where, and Let represent the initial rotation matrix and initial translation vector of the point cloud at the i-th viewpoint, respectively; according to For all point clouds to be registered, P = {P1, P2, ..., P} N Perform registration, and randomly select K points from the registered point cloud as the initial centroids for clustering, denoted as . Clustering module (2) is used to cluster the multi-view point cloud to be registered using the K-means clustering algorithm and update the centroid of each cluster to obtain the shape point cloud. Specifically, it includes: All point clouds to be registered P = {P1, P2, ..., P N The 3D points in} are assigned to K clusters, and the expression is: in, Point p i,j The cluster number in the q-th iteration. and Let these represent the rotation matrix and translation vector of the i-th viewpoint point cloud during the (q-1)-th iteration. This represents the centroid of the k-th cluster during the (q-1)-th iteration. This indicates finding the square of the L2 norm; Update the centroids of the K clusters according to the following formula. in, Let represent the centroid of the k-th cluster during the q-th iteration; The centroids of the updated K clusters The resulting point cloud data is referred to as a shape point cloud, denoted as... The model building module (3) is used to establish a multi-view point cloud registration optimization model based on the KMPE loss function between the point cloud to be registered and the shape point cloud, specifically including: By introducing the KMPE loss function, a multi-view point cloud registration optimization model based on the KMPE loss function is established to transform the multi-view point cloud registration problem into a minimum optimization problem of a nonlinear optimization model; wherein, the multi-view point cloud registration optimization model is expressed as: Among them, (R) i ,t i ) represents the rigid body transformation of the point cloud at the i-th viewpoint. Let represent the optimal rigid body transformation of the point cloud at the q-th viewpoint in the i-th viewpoint. This indicates that the cluster number is [missing information] during the q-th iteration. The centroid of the cluster, σ represents the Gaussian kernel bandwidth, and p represents the power parameter; The calculation module (4) is used to solve the multi-view point cloud registration optimization model based on the LM algorithm to obtain the optimal rigid body transformation of the point cloud to be registered and the shape point cloud under each view. The output module (5) is used to take the final determined optimal rigid body transformation as the registration result.

Images