Point cloud data registration method, electronic device, and storage medium

Abstract

The application provides a point cloud data registration method, an electronic device and a storage medium, and relates to the technical field of point cloud registration. The point cloud data registration method comprises the following steps: performing coarse registration based on an iterative closest point algorithm; obtaining target point clouds of a stable region as regional target point clouds, and obtaining updated source point clouds of the stable region as regional source point clouds; wherein the stable region is a region of nodes at both ends of a main beam and a region of nodes at the bottom of a pier; establishing a Gaussian mixture model according to the regional target point clouds and the regional source point clouds; establishing a target function based on a regularized deformation field, optimizing deformation parameters of non-rigid deformation by using an expectation maximization algorithm, and obtaining target deformation parameters; and registering the updated source point clouds and the target point clouds based on the target deformation parameters. The application performs coarse registration as a foundation, focuses on key stable regions, and effectively improves registration accuracy; a Gaussian mixture model is used for modeling, deep matching of point clouds in a feature level is realized, and high-precision registration of point clouds is realized.

Description

Technical Field

[0001] This invention relates to the field of point cloud registration technology, and in particular to a point cloud data registration method, electronic device, and storage medium. Background Technology

[0002] In the application of laser scanning technology for three-dimensional deformation monitoring of bridge structures, it is typically necessary to conduct multi-phase three-dimensional digital modeling of the target bridge to obtain a time-series point cloud dataset. After acquiring this dataset, spatial registration operations are performed on the multi-temporal point cloud data (the goal of point cloud registration is to minimize the error between the transformed point cloud and the target point cloud). This allows for the extraction of spatial coordinate differences between corresponding feature points, and these numerical features characterize the three-dimensional deformation of the bridge within the monitoring period. In this technical process, the computational accuracy of the spatial registration algorithm directly impacts the reliability of the bridge deformation monitoring results. Especially in long-span bridge engineering, when faced with the processing requirements of millions of point cloud data, effectively balancing algorithm accuracy and computational efficiency has become a key challenge in the application of this technology, as the rationality of the registration scheme directly determines the accuracy of the structural health assessment conclusions.

[0003] In existing technologies, rigid registration and non-rigid registration are commonly used for spatial registration. Rigid registration refers to the situation where the geometric properties and shape of an object remain unchanged under affine transformations such as translation and rotation, while non-rigid registration involves accurately determining the deformation field of the source point cloud. The registration of two bridge point clouds from multiple sources and time periods is a non-rigid registration. However, due to the large amount of data in the two bridge point clouds, the presence of various degrees of deformation, noise, outliers, and local incompleteness, the registration effect is poor, affecting the accuracy of bridge deformation monitoring. Summary of the Invention

[0004] This invention provides a point cloud data registration method, electronic device, and storage medium to solve the problem that existing non-rigid registration methods, when applied to the registration of point clouds of two-phase bridges, result in poor registration performance and affect the accuracy of bridge deformation monitoring.

[0005] In a first aspect, embodiments of the present invention provide a point cloud data registration method, including: The source point cloud and the target point cloud are coarsely registered based on the iterative nearest point algorithm to obtain the updated source point cloud. The target point cloud of the stable region is obtained as the regional target point cloud, and the updated source point cloud of the stable region is obtained as the regional source point cloud; wherein, the stable region is the node region at both ends of the main beam and the node region at the bottom of the pier. A Gaussian mixture model is established based on the regional target point cloud and the regional source point cloud. The probability density characteristics of the Gaussian mixture model are obtained by fitting the regional target point cloud, and the point cloud of the regional source point cloud after non-rigid deformation is used as the Gaussian components of the Gaussian mixture model. An objective function is established based on a regularized deformation field, and the deformation parameters of non-rigid deformation are optimized using the expectation-maximization algorithm to obtain the target deformation parameters. The updated source and target point clouds are registered based on the target deformation parameters.

[0006] In a second aspect, embodiments of the present invention provide an electronic device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the point cloud data registration method as described in the first aspect or any possible implementation of the first aspect.

[0007] Thirdly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the point cloud data registration method as described in the first aspect or any possible implementation thereof.

[0008] This invention provides a point cloud data registration method, an electronic device, and a storage medium. The point cloud data registration method includes: coarsely registering a source point cloud and a target point cloud based on an iterative nearest-point algorithm to obtain an updated source point cloud; acquiring a target point cloud in a stable region as a regional target point cloud, and acquiring the updated source point cloud in the stable region as a regional source point cloud; wherein the stable region is the node region at both ends of the main beam and the node region at the bottom of the pier; establishing a Gaussian mixture model based on the regional target point cloud and the regional source point cloud; wherein the probability density characteristics of the Gaussian mixture model are obtained by fitting the regional target point cloud, and the point cloud of the regional source point cloud after non-rigid deformation is used as each Gaussian component of the Gaussian mixture model; establishing an objective function based on a regularized deformation field, and optimizing the deformation parameters of the non-rigid deformation using an expectation-maximization algorithm to obtain target deformation parameters; and registering the updated source point cloud and the target point cloud based on the target deformation parameters. This application first completes coarse registration using an iterative nearest-point algorithm, pre-aligning the overall spatial positions of the source and target point clouds to lay the foundation for subsequent precise local registration. Then, it constructs an objective function using a regularized deformation field to optimize non-rigid deformation parameters, effectively constraining the rationality of point cloud deformation, suppressing registration offset problems caused by point cloud noise and measurement errors, and improving the accuracy of the registration results. At the same time, it focuses on stable region point clouds to avoid interference from unstable region point clouds in the registration process, improving registration accuracy, while reducing the amount of point cloud data involved in complex calculations and reducing computational difficulty. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the implementation of a point cloud data registration method provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the point cloud data registration device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0010] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0011] See Figure 1 The diagram illustrates a flowchart of a point cloud data registration method provided in an embodiment of the present invention, which is described in detail below: The above point cloud data registration method includes: S101: The source point cloud and the target point cloud are coarsely registered based on the iterative nearest point algorithm to obtain the updated source point cloud; A 3D laser scanner can be used to obtain bridge point clouds from two different periods: the source point cloud (first phase bridge point cloud) and the target point cloud (second phase bridge point cloud).

[0012] This application quickly eliminates the significant positional deviation caused by differences in scanning perspective and scanning position between two phases of bridge point clouds through coarse registration, keeping the point cloud alignment error within a reasonable range; it eliminates the need for pre-extraction of point cloud features, adapts to point cloud data of complex bridge structures, and has a wider range of applications; it has a fast iterative convergence speed, which can meet the high-efficiency processing requirements of bridge point clouds and save computational costs for subsequent fine registration.

[0013] This application employs the ICP algorithm to eliminate the initial positional deviation between the source point cloud and the target point cloud. Essentially, it is an optimization problem based on the least squares criterion. By iteratively updating the rotation matrix and translation vector, the sum of the Euclidean distances between corresponding point pairs in the source point cloud and the target point cloud is minimized.

[0014] In one possible implementation, S101 may include: S1011: For any point in the source point cloud of the current iteration, transform the point using the optimal transformation parameters of the previous iteration to obtain the transformed point; find the point in the target point cloud with the smallest Euclidean distance to the transformed point of the current point, and form a point pair with the transformed point; where the source point cloud of the first iteration is the source point cloud. Assuming that the closest points in space are likely to be points in the same physical location, we can use the spatial metric property of Euclidean distance to select the closest points as corresponding point pairs. The calculation formula is as follows:

[0015] in, For the points in the source point cloud of the current iteration The corresponding transformation point, In the target point cloud The point closest to Euclidean distance, that is, the point closest to... The corresponding points thus establish a pair of points {( , )}.

[0016] S1012: Determine the optimal transformation parameters for the current iteration by taking the minimum mean square error of all point pairs in the current iteration as the objective function; Minimizing the mean square error is the core criterion for solving the optimal transformation parameters. Essentially, it ensures the best overall alignment between the source point cloud and the target point cloud after transformation by minimizing the sum of squares of the coordinate deviations of corresponding points.

[0017] The mean square error of all point pairs is expressed as:

[0018] in, The number of point pairs, For the first The points in the source point cloud of the current iteration in a pair of points No. Points in the target point cloud of a point pair It is a translation vector. It is a rotation matrix.

[0019] Considering the rigidity of bridge point clouds, the transformation parameters only include two types of parameters: rotation (controlling attitude deviation) and translation (controlling position deviation). Minimizing the mean square error of all point pairs is a least squares optimization problem. This application achieves parameter solution through centroid alignment and covariance matrix decomposition.

[0020] In one possible implementation, S1012 may include: 1. Calculate the centroid of the source point cloud in all point pairs in the current iteration as the first centroid, and determine the centroid-free coordinates of each point in the source point cloud in all point pairs in the current iteration relative to the first centroid; 2. Calculate the centroid of the target point cloud in all point pairs in the current iteration as the second centroid, and determine the centroid-free coordinates of each point in the target point cloud in all point pairs in the current iteration relative to the second centroid. Calculate the first centroid and the second mass center The calculation formula is as follows:

[0021]

[0022] in, For the first The coordinates of the source point in the nth point pair, the th For the first The coordinates of the target point in the point pair.

[0023] Then, based on the first and second centroids, calculate the centroid-free coordinates of each point using the following formula:

[0024] in, for centroid coordinates for The centroid coordinates.

[0025] 3. Based on the centroid-free coordinates of the points in the source point cloud of all point pairs in the current iteration, and the centroid-free coordinates of the points in the target point cloud of all point pairs in the current iteration, construct the covariance matrix; Construct a 3x3 covariance matrix The calculation formula is as follows:

[0026] in, yes (3x1 column vector) and The outer product of (1x3 row vectors) yields a 3x3 matrix.

[0027] 4. Perform SVD decomposition on the covariance matrix to obtain two orthogonal matrices; For covariance matrix The SVD decomposition is performed using the following formula:

[0028] in, and Given two 3x3 orthogonal matrices, It is a 3x3 diagonal matrix containing non-negative singular values ​​(σ1, σ2, σ3, usually in descending order).

[0029] 5. Based on the two orthogonal matrices, obtain the rotation matrix for the current iteration, and determine the translation vector for the current iteration based on the rotation matrix for the current iteration; Calculate the rotation matrix based on two orthogonal matrices. The calculation formula is as follows:

[0030] 6. The rotation matrix and translation vector of the current iteration form the optimal transformation parameters for the current iteration.

[0031] Translation vector The translation vector can be obtained by aligning the centroids. The first center of mass after rotation should be such that Equal to the second mass center The calculation formula is as follows:

[0032] In one possible implementation, obtaining the rotation matrix for the current iteration based on two orthogonal matrices, and determining the translation vector for the current iteration based on the rotation matrix, may include: (1) Based on the two orthogonal matrices, and combined with the first formula, the rotation matrix for the current iteration is obtained; (2) Determine the translation vector for the current iteration based on the rotation matrix of the current iteration and the second formula; The first formula may include:

[0033] The second formula may include:

[0034] in, Let be the rotation matrix for the current iteration. Let be the translation vector for the current iteration; and These are the first and second centroids, respectively. and Let them be two orthogonal matrices.

[0035] The rotation matrix can be calculated from two orthogonal matrices, namely: .

[0036] like If it is a valid rotation matrix, then ;like If it is a reflection matrix, then Therefore, to ensure that the calculated result is a valid rotation matrix, it is necessary to determine... : like ,but ; like Then, the last column of the matrix is ​​negative, and a new matrix is ​​constructed as the rotation matrix:

[0037] The matrix satisfies , which is an effective rotation matrix.

[0038] The translation vector in the current iteration should make the rotated first centroid equal to the second centroid, that is:

[0039] This leads to the formula for calculating the translation vector.

[0040] This application eliminates the influence of the overall point cloud position on parameter solving through centroid removal, focusing the solution on attitude (rotation) deviation. The covariance matrix reflects the spatial distribution correlation of the point clouds in two phases, and its SVD decomposition yields the optimal rotation direction. Combined with the centroid coordinate difference, the translation vector is solved, ensuring the accuracy and stability of the transformation parameter solution. Specifically, SVD decomposition of the rotation matrix avoids the singularity problem of direct solution, improving the stability of parameter solving. Using the mean square error as the objective function further suppresses the influence of individual outliers on the overall transformation effect, improving the robustness of coarse registration.

[0041] S1013: Determine whether the convergence condition is met; S1014: If so, the source point cloud is transformed using the optimal transformation parameters of the current iteration to obtain the updated source point cloud; If the convergence condition is met, the source point cloud is updated using the rotation matrix and translation vector of the current iteration, which serves as the basis for subsequent data processing.

[0042] S1015: If not, then use the optimal transformation parameters of the current iteration to transform the source point cloud of the current iteration to obtain the source point cloud of the next iteration.

[0043] If the convergence condition is not met, continue iteratively to solve the problem.

[0044] Convergence criteria are set to balance registration accuracy and computational efficiency. Convergence criteria may include the following: 1. The error variation is small; for example, the error can be the mean square error or its square root.

[0045] 2. The optimal transformation parameters change little; 3. Reach the maximum number of iterations.

[0046] The convergence condition can be any one of the above, or iterative can stop when one of the three conditions is met. The specific condition can be set according to the actual application requirements.

[0047] This application gradually corrects the deviation of transformation parameters through multiple rounds of iteration, thereby continuously optimizing the coarse registration result. The final updated source point cloud and target point cloud have a significantly reduced overall deviation, which reduces the optimization pressure for the subsequent fine registration process and improves the overall registration efficiency.

[0048] S102: Obtain the target point cloud of the stable region as the region target point cloud, and obtain the updated source point cloud of the stable region as the region source point cloud; wherein, the stable region is the node region at both ends of the main beam and the node region at the bottom of the pier. The spatial displacement of bridge structures in actual operation exhibits a specific pattern: the displacement in the mid-span region of the main girder (e.g., deflection) is more significant than that in the girder end region; and the displacement in the pier top region (e.g., horizontal displacement) is more significant than that in the pier bottom region. Based on the above displacement distribution characteristics, this application provides differentiated treatment for different components.

[0049] First, the structure is skeletonized by extracting the outer contours of each component at preset intervals and defining control nodes. Then, the connection nodes between components (such as beam-beam connections and beam-pier connections) are identified, and the main beams and piers are divided accordingly. Finally, stable regions are selected: the node regions near the two ends are used for registration of the main beam components; the node regions near the bottom are used for registration of the pier components.

[0050] Accurate selection of stable regions provides high-quality data for subsequent Gaussian mixture model construction, reducing interference from invalid data (point clouds in deformed regions); compared to full-area point cloud processing, it reduces the amount of data to be processed subsequently and improves registration efficiency; the selection of stable regions conforms to the mechanical characteristics of bridge structures, ensuring the reliability of the registration benchmark.

[0051] S103: Based on the regional target point cloud and the regional source point cloud, establish a Gaussian mixture model; wherein, the probability density characteristics of the Gaussian mixture model are obtained by fitting the regional target point cloud, and the point cloud of the regional source point cloud after non-rigid deformation is used as each Gaussian component of the Gaussian mixture model. Gaussian mixture models are probabilistic models composed of linear combinations of multiple Gaussian distributions. The core principle is to fit the probability density distribution of complex data by superimposing multiple simple Gaussian distributions.

[0052] The probability density function of the Gaussian mixture model is expressed as:

[0053] in, This indicates the number of Gaussian distributions, i.e., the number of mixture components; The weights of the k-th Gaussian distribution satisfy... And 0≤ ≤1 reflects the proportion of each Gaussian distribution in the mixture model; It is the first The probability density function of a Gaussian distribution. The mean vector of this Gaussian distribution determines the location of the distribution's center. It is the covariance matrix, which determines the shape and scale of the distribution. For example, when When it is two-dimensional data, , It is a 2×2 matrix whose elements determine the distribution characteristics of the data in two dimensions.

[0054] In this method, the probability density distribution of the regional target point cloud is used as the "baseline distribution," and the regional source point cloud is subjected to non-rigid deformation. Each point cluster thereafter is considered as a Gaussian component, i.e. .in, It is a function dependent on the parameter θ, which can describe the complex deformation of the source point cloud, allowing the source point cloud to better match the distribution of the target point cloud. For example, for a 3D point cloud, non-rigid transformations may include translation, rotation, scaling, and more complex nonlinear deformations, which can be achieved by adjusting... The parameters can realize various deformations of the source point cloud in space, so that the Gaussian distribution centered on the deformed source point cloud can accurately fit the distribution of the target point cloud, and achieve fine alignment of the point cloud through probability distribution matching.

[0055] The advantage of Gaussian mixture models lies in their strong fitting ability. Theoretically, as long as... A sufficiently large Gaussian mixture model can approximate any continuous probability distribution. This makes it excellent at handling complex data distributions, capturing various features and patterns within the data. In point cloud registration, the distribution of target point clouds is often complex and diverse, potentially including noise, outliers, and the effects of various nonlinear deformations. Gaussian mixture models can effectively model such complex distributions, providing a solid foundation for subsequent registration calculations.

[0056] S104: Based on the regularized deformation field, an objective function is established, and the expectation-maximization algorithm is used to optimize the deformation parameters of non-rigid deformation to obtain the target deformation parameters; The core challenge of non-rigid deformation is to achieve precise alignment of point clouds while avoiding structural distortion caused by excessive deformation. The introduction of a regularized deformation field, by adding a "smoothing constraint term" (regularization term), limits drastic changes in the deformation field, ensuring that the deformation conforms to the physical characteristics of the bridge structure. The EM algorithm is used to solve the parameter estimation problem of Gaussian mixture models, ensuring the consistency and reliability of the deformation parameter optimization results.

[0057] In one possible implementation, S104 may include: S1041: Calculate the posterior probability of each point in the target point cloud of the region; The above describes the E-step (expectation step) in the EM algorithm.

[0058] In one possible implementation, the formula for calculating the posterior probability can be:

[0059]

[0060] in, For the first Deformation parameters at the next iteration For the first point in the regional target point cloud One point; Deformation parameters Down, Belongs to the The posterior probability of each Gaussian component; Deformation parameters Next The likelihood probability of each Gaussian component; For the first Prior probabilities of Gaussian components , The total number of Gaussian components; This represents the proportion of outliers.

[0061] Posterior probability Reflects the point The degree of matching between the target point and each point after transformation of the source point cloud (i.e., the center of the Gaussian distribution) allows for the establishment of soft correspondences between point clouds. This means that each target point corresponds to multiple points after transformation of the source point cloud with a certain probability, rather than searching for a unique hard correspondence pair as in traditional rigid registration.

[0062] S1042: Based on the objective function, update the deformation parameters to obtain the updated deformation parameters; The above describes the M-step (maximization step) in the EM algorithm: using the posterior probability calculated in the E-step, the parameters and deformation parameters of the Gaussian mixture model are updated to maximize the likelihood function of the target point cloud.

[0063] In one possible implementation, S1042 may include: 1. Find the gradient of the deformation parameters with respect to the objective function; 2. The deformation parameters are iteratively updated using the gradient descent method until the objective function converges, thus obtaining the updated deformation parameters.

[0064] In another possible implementation, the formula for updating the deformation parameters can be obtained by taking the partial derivative of the likelihood function with respect to these parameters and setting it to zero.

[0065] In point cloud registration, non-rigid transformations can cause unreasonable distortions and deformations in the source point cloud, rendering the registration result meaningless or inaccurate. To avoid this, a regularized deformation field is introduced to constrain the deformation of the source point cloud. Regularized deformation fields are typically constructed based on some prior geometric or physical knowledge. The core idea is to limit certain properties of the deformation field by adding a regularization term to the objective function. For example, a common regularization term based on smoothness constraints aims to make the deformation field as smooth as possible in space, avoiding abrupt and discontinuous deformations. Mathematically, we assume the deformation field can be represented as a function... It describes the points in the source point cloud. The displacement, based on the regularization term of the thin-plate spline function, can be expressed as:

[0066] That is, in one possible implementation, the objective function can be:

[0067]

[0068] in, For the first Deformation parameters of the next iteration The likelihood function under the given conditions, The regularization coefficient is . For regularization terms, Deformation parameters Non-rigid deformation field; For the Laplace operator, For the dimensions of the point cloud, Non-rigid deformation field The norm of the second derivative.

[0069] This is used to punish drastic changes in the deformation field, ensuring that the deformation field remains smooth in space. When When the value is small, it indicates that the deformation field changes relatively smoothly, and the deformation of the source point cloud is more reasonable; conversely, if... A larger value indicates a significant abrupt change in the deformation field, which may lead to unreasonable deformation of the source point cloud.

[0070] Used to balance the relative importance of data fitting and regularization constraints. When When the value is large, the regularization constraint has a stronger effect and the deformation field is smoother, but it may sacrifice some data fitting accuracy; when When the value is small, the data fitting effect is stronger, which may lead to an uneven deformation field and unreasonable deformation. Therefore, it is important to choose an appropriate value. The value of is crucial for obtaining accurate and reasonable registration results. Specifically, the optimal value may need to be determined experimentally. The value is adjusted to adapt to different point cloud data and registration requirements. Through the above objective function, the regularized deformation field can effectively constrain the non-rigid transformation of the source point cloud while ensuring the fit between the source and target point cloud data, thus achieving high-quality non-rigid point cloud registration.

[0071] S1043: Determine the change in the updated deformation parameters compared to the original deformation parameters; S1044: If the change is less than the preset change, the updated deformation parameter will be used as the target deformation parameter. S1045: Otherwise, perform non-rigid deformation on the source point cloud of the region using the updated deformation parameters, take each point after deformation as a new Gaussian component, and jump to the step of calculating the posterior probability of each point in the target point cloud of the region to continue execution.

[0072] Repeat the E-step and M-step until convergence is achieved, and the target deformation parameters are obtained.

[0073] The convergence of the EM algorithm is based on the monotonically increasing likelihood function. In each iteration, the M-step guarantees that the likelihood function reaches a local maximum under the posterior probability calculated in the current E-step, while the E-step recalculates the posterior probability based on the updated parameters, providing a more accurate estimate for the next M-step. Through continuous iteration, the likelihood function gradually increases, eventually converging to a local optimum. Although the EM algorithm cannot guarantee convergence to the global optimum, for this application, the locally optimal solution it converges to is sufficient.

[0074] S105: Register the updated source point cloud and target point cloud based on the target deformation parameters.

[0075] The target deformation parameters obtained in S104 (reflecting the non-rigid deformation law of the stable region) are extended to the entire updated source point cloud. Through non-rigid transformation, the position of each point in the source point cloud is adjusted according to the optimal deformation law, ultimately achieving precise alignment of the entire source point cloud and the target point cloud. The core logic is to utilize the deformation continuity of the bridge structure; the deformation law of the stable region can indirectly reflect the deformation trend of the entire bridge, ensuring the rationality of the full-area registration.

[0076] This application can achieve full-area accurate alignment of two phases of bridge point clouds, with registration error reduced to the millimeter level, fully meeting the high-precision requirements for monitoring minor bridge deformations. The registration results provide high-quality aligned point cloud data for subsequent bridge structural deformation analysis and health status assessment, greatly improving the reliability of subsequent analysis results.

[0077] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0078] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0079] Figure 3 A schematic diagram of the point cloud data registration device provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below: like Figure 3 As shown, the point cloud data registration device includes: The first registration module 21 is used to perform coarse registration of the source point cloud and the target point cloud based on the iterative nearest point algorithm to obtain the updated source point cloud. The stable point cloud acquisition module 22 is used to acquire the target point cloud of the stable region as the region target point cloud, and to acquire the updated source point cloud of the stable region as the region source point cloud; wherein, the stable region is the node region at both ends of the main beam and the node region at the bottom of the pier. The model building module 23 is used to build a Gaussian mixture model based on the regional target point cloud and the regional source point cloud; wherein, the probability density characteristics of the Gaussian mixture model are obtained by fitting the regional target point cloud, and the point cloud of the regional source point cloud after non-rigid deformation is used as each Gaussian component of the Gaussian mixture model. The parameter optimization module 24 is used to establish an objective function based on the regularized deformation field, and to optimize the deformation parameters of non-rigid deformation using the expectation-maximization algorithm to obtain the target deformation parameters. The second registration module 25 is used to register the updated source point cloud and target point cloud based on the target deformation parameters.

[0080] In one possible implementation, the parameter optimization module 24 may include: The posterior probability calculation unit is used to calculate the posterior probability of each point in the regional target point cloud. The parameter update unit is used to update the deformation parameters based on the objective function to obtain the updated deformation parameters; The change determination unit is used to determine the change in the updated deformation parameters compared to the original deformation parameters. The parameter output unit is used to take the updated deformation parameter as the target deformation parameter if the change is less than the preset change. The first loop jump unit is used otherwise to perform non-rigid deformation on the source point cloud of the region using the updated deformation parameters, take each point after deformation as a new Gaussian component, and jump to the step of calculating the posterior probability of each point in the target point cloud of the region to continue execution.

[0081] In one possible implementation, the formula for calculating the posterior probability can be:

[0082]

[0083] in, For the first Deformation parameters at the next iteration For the first point in the regional target point cloud One point; Deformation parameters Down, Belongs to the The posterior probability of each Gaussian component; Deformation parameters Next The likelihood probability of each Gaussian component; For the first Prior probabilities of Gaussian components , The total number of Gaussian components; This represents the proportion of outliers.

[0084] In one possible implementation, the objective function can be:

[0085]

[0086] in, For the first Deformation parameters of the next iteration The likelihood function under the given conditions, The regularization coefficient is . For regularization terms, Deformation parameters Non-rigid deformation field; For the Laplace operator, For the dimensions of the point cloud, Non-rigid deformation field The norm of the second derivative.

[0087] In one possible implementation, the parameter update unit can be specifically used for: 1. Find the gradient of the deformation parameters with respect to the objective function; 2. The deformation parameters are iteratively updated using the gradient descent method until the objective function converges, thus obtaining the updated deformation parameters.

[0088] In one possible implementation, the first registration module 21 may include: The point-to-point matching unit is used to transform any point in the source point cloud of the current iteration using the optimal transformation parameters of the previous iteration to obtain the transformed point of that point; and to find the point in the target point cloud with the smallest Euclidean distance to the transformed point of that point, and to form a point pair with that point; wherein, the source point cloud of the first iteration is the source point cloud; The parameter optimization unit is used to determine the optimal transformation parameters for the current iteration with the objective function of minimizing the mean square error of all point pairs in the current iteration. The convergence judgment unit is used to determine whether the convergence condition is met; The first point cloud update unit is used to transform the source point cloud using the optimal transformation parameters of the current iteration if the condition is met, thus obtaining the updated source point cloud. The second point cloud update unit is used to transform the source point cloud of the current iteration using the optimal transformation parameters of the current iteration if no, to obtain the source point cloud of the next iteration.

[0089] In one possible implementation, the parameter optimization unit can be specifically used for: 1. Calculate the centroid of the source point cloud in all point pairs in the current iteration as the first centroid, and determine the centroid-free coordinates of each point in the source point cloud in all point pairs in the current iteration relative to the first centroid; 2. Calculate the centroid of the target point cloud in all point pairs in the current iteration as the second centroid, and determine the centroid-free coordinates of each point in the target point cloud in all point pairs in the current iteration relative to the second centroid. 3. Based on the centroid-free coordinates of the points in the source point cloud of all point pairs in the current iteration, and the centroid-free coordinates of the points in the target point cloud of all point pairs in the current iteration, construct the covariance matrix; 4. Perform SVD decomposition on the covariance matrix to obtain two orthogonal matrices; 5. Based on the two orthogonal matrices, obtain the rotation matrix for the current iteration, and determine the translation vector for the current iteration based on the rotation matrix for the current iteration; 6. The rotation matrix and translation vector of the current iteration form the optimal transformation parameters for the current iteration.

[0090] In one possible implementation, the rotation matrix for the current iteration is obtained based on two orthogonal matrices, and the translation vector for the current iteration is determined based on the rotation matrix for the current iteration, including: (1) Based on the two orthogonal matrices, and combined with the first formula, the rotation matrix for the current iteration is obtained; (2) Determine the translation vector for the current iteration based on the rotation matrix of the current iteration and the second formula; The first formula may include:

[0091] The second formula may include:

[0092] in, Let be the rotation matrix for the current iteration. Let be the translation vector for the current iteration; and These are the first and second centroids, respectively. and Let them be two orthogonal matrices.

[0093] Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. For example... Figure 3 As shown, the electronic device 3 of this embodiment includes a processor 30 and a memory 31. The memory 31 stores a computer program 32. When the processor 30 executes the computer program 32, it implements the steps in the various method embodiments described above. Alternatively, when the processor 30 executes the computer program 32, it implements the functions of each module / unit in the various device embodiments described above.

[0094] For example, computer program 32 may be divided into one or more modules / units, which are stored in memory 31 and executed by processor 30 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 32 in electronic device 3.

[0095] Electronic device 3 may include, but is not limited to, processor 30 and memory 31. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 3 may also include input / output devices, network access devices, buses, etc.

[0096] The processor 30 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0097] The memory 31 can be an internal storage unit of the electronic device 3, such as a hard disk or memory of the electronic device 3. The memory 31 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 3. Furthermore, the memory 31 can include both internal and external storage units of the electronic device 3. The memory 31 is used to store the computer program 32 and other programs and data required by the electronic device 3. The memory 31 can also be used to temporarily store data that has been output or will be output.

[0098] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.

[0099] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.

[0100] This invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.

[0101] Computer programs include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0102] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0103] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A point cloud data registration method, characterized in that, include: The source point cloud and the target point cloud are coarsely registered based on the iterative nearest point algorithm to obtain the updated source point cloud. The target point cloud of the stable region is obtained as the regional target point cloud, and the updated source point cloud of the stable region is obtained as the regional source point cloud; wherein, the stable region is the node region at both ends of the main beam and the node region at the bottom of the pier. A Gaussian mixture model is established based on the target point cloud and the source point cloud of the region; wherein, the probability density features of the Gaussian mixture model are obtained by fitting the target point cloud of the region, and the point cloud of the source point cloud after non-rigid deformation is used as each Gaussian component of the Gaussian mixture model. An objective function is established based on a regularized deformation field, and the deformation parameters of the non-rigid deformation are optimized using the expectation-maximization algorithm to obtain the target deformation parameters. The updated source point cloud and the target point cloud are registered based on the target deformation parameters.

2. The point cloud data registration method according to claim 1, characterized in that, The optimization of the deformation parameters of the non-rigid deformation using the expectation-maximization algorithm yields the target deformation parameters, including: Calculate the posterior probability of each point in the target point cloud of the region; Based on the objective function, the deformation parameters are updated to obtain the updated deformation parameters; Determine the amount of change between the updated deformation parameters and the original deformation parameters; If the change is less than the preset change, then the updated deformation parameter is used as the target deformation parameter; Otherwise, the updated deformation parameters are used to perform non-rigid deformation on the source point cloud of the region, and each point after deformation is used as a new Gaussian component. Then, the process jumps to the step of calculating the posterior probability of each point in the target point cloud of the region and continues.

3. The point cloud data registration method according to claim 2, characterized in that, The formula for calculating the posterior probability is: in, For the first Deformation parameters at the next iteration For the target point cloud in the region One point; Deformation parameters Down, Belongs to the The posterior probability of each Gaussian component; Deformation parameters Next The likelihood probability of each Gaussian component; For the first Prior probabilities of Gaussian components , The total number of Gaussian components; This represents the proportion of outliers.

4. The point cloud data registration method according to claim 1, characterized in that, The objective function is: in, For the first Deformation parameters of the next iteration The likelihood function under the given conditions, The regularization coefficient is . For regularization terms, Deformation parameters Non-rigid deformation field; For the Laplace operator, For the dimensions of the point cloud, Non-rigid deformation field The norm of the second derivative.

5. The point cloud data registration method according to claim 4, characterized in that, The step of updating the deformation parameters based on the objective function includes: For the objective function, find the gradient of the deformation parameter; The deformation parameters are iteratively updated using the gradient descent method until the objective function converges, thus obtaining the updated deformation parameters.

6. The point cloud data registration method according to any one of claims 1 to 5, characterized in that, The coarse registration of the source point cloud and the target point cloud based on the iterative nearest point algorithm to obtain the updated source point cloud includes: For any point in the source point cloud of the current iteration, the point is transformed using the optimal transformation parameters of the previous iteration to obtain the transformed point; in the target point cloud, the point with the smallest Euclidean distance to the transformed point is found, and a point pair is formed with the transformed point; wherein, the source point cloud of the first iteration is the source point cloud. The optimal transformation parameters for the current iteration are determined by taking the minimum mean square error of all point pairs in the current iteration as the objective function. Determine whether the convergence condition is met; If so, the source point cloud is transformed using the optimal transformation parameters of the current iteration to obtain the updated source point cloud; If not, the source point cloud of the current iteration is transformed using the optimal transformation parameters of the current iteration to obtain the source point cloud of the next iteration.

7. The point cloud data registration method according to claim 6, characterized in that, The process of determining the optimal transformation parameters for the current iteration, with the objective function being the minimum mean square error of all point pairs in the current iteration, includes: Calculate the centroid of the source point cloud in all point pairs in the current iteration as the first centroid, and determine the decentroid coordinates of each point in the source point cloud in all point pairs in the current iteration relative to the first centroid. Calculate the centroid of the target point cloud in all point pairs in the current iteration as the second centroid, and determine the centroid-free coordinates of each point in the target point cloud in all point pairs in the current iteration relative to the second centroid; Based on the decentroid coordinates of the points in the source point cloud of all point pairs in the current iteration, and the decentroid coordinates of the points in the target point cloud of all point pairs in the current iteration, a covariance matrix is ​​constructed. The covariance matrix is ​​decomposed by SVD to obtain two orthogonal matrices; Based on the two orthogonal matrices, the rotation matrix for the current iteration is obtained, and the translation vector for the current iteration is determined based on the rotation matrix for the current iteration. The rotation matrix and translation vector of the current iteration form the optimal transformation parameters for the current iteration.

8. The point cloud data registration method according to claim 7, characterized in that, The step of obtaining the rotation matrix for the current iteration based on the two orthogonal matrices, and determining the translation vector for the current iteration based on the rotation matrix for the current iteration, includes: Based on the two orthogonal matrices, the rotation matrix for the current iteration is obtained using the first formula; The translation vector for the current iteration is determined based on the rotation matrix of the current iteration and the second formula. The first formula includes: The second formula includes: in, Let be the rotation matrix of the current iteration. Let be the translation vector for the current iteration; and These are the first centroid and the second centroid, respectively. and Let be the two orthogonal matrices.

9. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the point cloud data registration method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the point cloud data registration method as described in any one of claims 1 to 8.

Images