High-temperature target multi-view three-dimensional point cloud global registration method

Abstract

The application discloses a high-temperature target multi-view three-dimensional point cloud global registration method, calculates normal vectors of point clouds in different views respectively, adopts a preset registration method to obtain initial rigid body transformation parameters of each view, initializes covariance of points and covariance of corresponding normal vectors, then carries out iterative optimization, in each iteration, adopts a point-level Gaussian mixture model combined with geometric consistency constraint to calculate credibility of each point and its nearest neighbor in other views, and updates rigid body transformation parameters and covariance parameters based on the credibility, until an optimization end condition is reached, then adopts the optimized rigid body transformation parameters to transform point clouds of each view to a global coordinate system, so that registered point clouds are obtained. The point-level Gaussian mixture model combined with geometric consistency constraint fully utilizes relatively stable geometric features in different views, so that the quality of high-temperature target point cloud registration is improved.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional point cloud reconstruction technology, and more specifically, relates to a global registration method for three-dimensional point clouds of high-temperature targets from multiple perspectives. Background Technology

[0002] With the rapid development of modern industrial technology, especially in high-end equipment fields such as aerospace, energy and power, and precision manufacturing, increasingly higher requirements are being placed on the condition monitoring of key equipment in high-temperature environments. Three-dimensional point cloud reconstruction technology can provide accurate spatial structural information of objects, which is crucial for applications such as deformation analysis, dimensional inspection, and quality control. Due to the limitations of a single viewpoint in industrial environments, it is often necessary to acquire point cloud data from multiple angles and perform accurate registration to construct complete three-dimensional information. Therefore, multi-view point cloud registration technology is a key technology for acquiring geometric shape information of complex objects.

[0003] However, high-temperature environments present a series of unique challenges to point cloud acquisition and registration. First, there is the problem of thermal radiation interference. The surface of high-temperature targets emits strong infrared radiation, which severely interferes with depth cameras based on structured light or time-of-flight principles. The intensity of infrared radiation far exceeds that of structured light signals, causing distortion of the signals received by the sensors. Thermal radiation creates false depth information on the sensors, and these points are randomly distributed in space. The effective measurement signal is overwhelmed by thermal radiation noise, leading to a decrease in measurement accuracy.

[0004] Secondly, there is the problem of light path refraction caused by hot air flow. The air around the high-temperature target flows irregularly due to heating, forming a medium with non-uniform refractive index. When light passes through these hot air masses, it will be randomly refracted, resulting in random deviations in the position of the same physical point measured at different times. Even corresponding points in the data collected from different perspectives at different times will have systematic deviations. These deviations change with time and spatial position and are difficult to describe with a static model.

[0005] Furthermore, for equipment protection and personnel safety, high-temperature target measurements must maintain a sufficient safety distance. Since camera resolution is fixed, a greater shooting distance will reduce point cloud density and lead to a decrease in measurement quality at the edge of the field of view. At the same time, due to the uneven temperature distribution in different parts of the workpiece and the different relative positions of each viewing angle to the heat source, the point cloud quality varies significantly from different viewing angles. Traditional registration methods usually designate one point cloud as the "source point cloud" and another as the "target point cloud." This asymmetrical treatment will cause serious problems in high-temperature scenarios: if a low-quality point cloud is used as the target, the overall registration accuracy will decrease; if a low-quality point cloud is used as the source, the registration process is prone to getting trapped in local optima. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a global registration method for multi-view 3D point clouds of high-temperature targets. By combining a point-level Gaussian mixture model with geometric consistency constraints, the method makes full use of relatively stable geometric features in different viewpoints, thereby improving the quality of high-temperature target point cloud registration.

[0007] To achieve the above-mentioned objectives, the high-temperature target multi-view three-dimensional point cloud global registration method of the present invention includes the following steps:

[0008] S1: From Point cloud collection of high-temperature targets from different perspectives ,in Indicates the first The first perspective Three-dimensional data of each point, , , Indicates the first The number of points in the point cloud set of each viewpoint is calculated, and then each point is calculated separately. normal vector ;

[0009] S2: Obtain the initial rigid body transformation parameters for each viewpoint using a preset registration method. ,in Represents the initial rotation matrix. Represents the initial translation matrix;

[0010] S3: Covariance of the initialization point covariance of the normal vector ;

[0011] S4: Let the number of iterations be... ;

[0012] S5: Using a rotation matrix Translation matrix All points from each perspective Transform to common coordinate system coordinates Then, based on the coordinates of the common coordinate system, the first... Each point from each perspective In other Nearest neighbor index in a point cloud set of each viewpoint , and If point In the If a point cloud set does not have a nearest neighbor, then the point is considered to be... In the From the perspective of outliers, they are removed from the point cloud. Delete, and get the first The perspective in the first Cluster of non-outliers from various perspectives ;

[0013] S6: Calculating Outlier Clouds Based on Gaussian Mixture Model Each point in With the Nearest neighbor from each perspective Credibility :

[0014] ,

[0015] in,

[0016] ,

[0017]

[0018] in, Indicates the nearest neighbor. Common coordinate system coordinates, Indicates the first Rotation matrix for each viewpoint This indicates the search for the L2 norm. This represents the number of GMM components in a Gaussian mixture model. This represents the prior probability of an outlier.

[0019] S7: Update the rigid body transformation parameters for each viewpoint, as follows:

[0020] First calculate the... The weighted centroid of each viewpoint in the global coordinate system and nearest neighbor weighted centroid :

[0021] ,

[0022] ,

[0023] Then calculate each point using the following formula. and its nearest neighbor Decentralized coordinates in the global coordinate system , :

[0024] ,

[0025] ,

[0026] Construct the first Weighted covariance matrix of each perspective :

[0027] ,

[0028] in, Indicates the nearest neighbor. The normal vector;

[0029] For the weighted covariance matrix Perform singular value decomposition Then based on orthogonal matrices and Update the rotation matrix and translation vector:

[0030] ,

[0031] ;

[0032] Then, the two covariances are updated using the following formula:

[0033] ,

[0034] ;

[0035] S8: Determine whether the optimization termination condition has been met. If not, proceed to step S9; if it has, proceed to step S10.

[0036] S9: Let the number of iterations be... Return to step S5;

[0037] S10: Rotation matrix based on each viewpoint Translation matrix The point cloud from this perspective is transformed to the global coordinate system to obtain the registered point cloud.

[0038] This invention provides a global registration method for multi-view 3D point clouds of high-temperature targets. The method calculates the normal vectors of point clouds from different viewpoints, obtains initial rigid body transformation parameters for each viewpoint using a preset registration method, and initializes the covariance of the points and the covariance of the corresponding normal vectors. Then, iterative optimization is performed. In each iteration, a point-level Gaussian mixture model combined with geometric consistency constraints is used to calculate the confidence level of each point and its nearest neighbors in other viewpoints. Based on this confidence level, the rigid body transformation parameters and covariance parameters are updated until the optimization termination condition is met. Finally, the optimized rigid body transformation parameters are used to transform the point clouds from each viewpoint to the global coordinate system, thus obtaining the registered point clouds.

[0039] The present invention has the following beneficial effects:

[0040] 1) This invention uses a point-level Gaussian mixture model to model, which limits the impact of noise to a local area and avoids low-quality data from polluting the overall registration results. At the same time, it can handle a high proportion of outliers by explicitly modeling outliers.

[0041] 2) In high-temperature environments, normal vectors are more stable than position information. This invention introduces geometric constraints on normal vectors and makes full use of geometric features to significantly improve the accuracy of correspondence judgment and effectively avoid erroneous matching.

[0042] 3) This invention can adjust the covariance parameter and outlier ratio according to the temperature of the shooting environment of the high-temperature target, so that the algorithm can adapt to the measurement characteristics of different temperature ranges.

[0043] 4.) This invention uses a soft allocation mechanism based on posterior probability to automatically assign higher weights to high-quality correspondences and automatically suppress low-quality correspondences. The algorithm does not require pre-specifying the source and target point clouds, and all perspectives are treated equally.

[0044] 5) The point-level probabilistic modeling of the present invention enables each point to find corresponding relationships from multiple perspectives. Even if the overlap rate between two perspectives is very low, as long as the point has corresponding points in other perspectives, it can participate in the registration process, which is superior to the traditional pairwise registration method.

[0045] 6) This invention only requires the estimation of 6M+2 parameters (M represents the number of viewpoints), which is far fewer than the 10D+6M parameters (D is the number of Gaussian components) of the traditional global GMM method. It is less prone to overfitting in high-noise environments and has better robustness.

[0046] 7) This invention is based on the EM algorithm framework, which ensures the monotonically increasing likelihood function, the stable optimization process, and the absence of divergence or oscillation. It usually converges within 30-50 iterations. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the point-level Gaussian mixture model construction combined with geometric consistency constraints in this invention;

[0048] Figure 2 This is a flowchart illustrating a specific implementation of the global registration method for multi-view three-dimensional point clouds of high-temperature targets according to the present invention.

[0049] Figure 3 These are point cloud images from seven perspectives in this embodiment;

[0050] Figure 4 This is the point cloud map after point cloud registration in this embodiment. Detailed Implementation

[0051] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0052] To better illustrate the technical solution of the present invention, the point cloud probability model used in the present invention will be briefly described first.

[0053] When collecting multi-view data, for points on a real object, the observation points obtained from each viewpoint are actually noisy data of the real points due to measurement noise. The real points corresponding to these noisy observation points are actually clustered in space near the target point. In order to model the relationships between points affected by noise, an independent probability model is constructed for each observation point.

[0054] Figure 1 This is a schematic diagram illustrating the construction of a point-level Gaussian mixture model incorporating geometric consistency constraints in this invention. For example... Figure 1 As shown, the point cloud for each viewpoint is recorded. ,in Indicates the first The first perspective Three-dimensional data of each point, , , Indicates the first The number of points in a point cloud set from each viewpoint. Define the rigid body transformation function. ,in, This represents the three-dimensional data of points in a point cloud. Represents the rotation matrix. This function represents the translation matrix, indicating the transformation of a point... From the Transform the local coordinate system of a viewpoint to the global coordinate system.

[0055] For the Points from various perspectives Its position after rigid body transformation The probability of can be described by the following Gaussian mixture model:

[0056] ,

[0057] in, This indicates the number of GMM components in the Gaussian Mixture Model (GMM). This represents the covariance matrix shared by all GMM components. Represents all parameters to be estimated. Indicates the first The parameters to be estimated from each viewpoint Point In the The index of the corresponding point in the viewpoint data. and , The mean is variance is The multivariate Gaussian distribution.

[0058] In high-temperature environments, although the position of a point is greatly affected by noise, the local geometry (especially the direction of the normal vector) remains relatively stable. Under rigid body transformations, the local geometric properties of the object should remain unchanged. Let... and They are points and The unit normal vector for two truly corresponding points and They should not only be spatially similar after transformation, but their local normal vector characteristics should also remain similar after transformation. A soft constraint mechanism is used, that is, assuming that the estimation error of the normal vector follows a Gaussian distribution. , This represents the covariance matrix of the normal vectors, consistent with the probabilistic interpretation in the existing framework, and represents the observation data of the points from... Expand to The parameters to be estimated are expanded to Based on the above extensions, the probabilistic model incorporating geometric consistency constraints is constructed as follows:

[0059] .

[0060] The physical meaning expressed by the above model is that a correct correspondence should simultaneously satisfy two conditions: proximity in location (constrained by the first Gaussian distribution) and consistency in normal vector direction (constrained by the second Gaussian distribution). In high-temperature scenarios, due to high positional noise, relying solely on positional constraints can easily lead to incorrect correspondences. By introducing normal vector constraints, even if two points are relatively close in location, if their normal vector directions differ significantly, the probability of them being identified as corresponding points will be very low, thus effectively avoiding incorrect matching.

[0061] Furthermore, high-temperature point cloud measurements from various perspectives inevitably contain a large number of outliers, which have clear physical origins. These outliers include: false depth information generated by infrared radiation on the sensor; erroneous measurements caused by refraction of light through hot air masses; and diffuse reflection anomalies caused by oxide layers on high-temperature surfaces. To improve the robustness of the algorithm and address the mixed noise problem encountered in high-temperature environments, this invention employs an outlier model based on the Laplace distribution to describe the attenuation distribution pattern of outliers centered on the centroid of the point cloud, as shown in the following equation: the farther away from the centroid, the higher the probability of an outlier appearing at that location. This is because points on the surface of a real object usually cluster near the centroid, while false points generated by thermal radiation and light refraction are often distributed at the object's edge or in areas far from the main structure. Thus, the point... The probability of being an outlier :

[0062] ,

[0063] in, This represents the preset prior probability of outliers. Indicates the first Centroid of point cloud data from a single perspective This represents the distance decay scale parameter for outliers.

[0064] Based on the above analysis, and combining point-level GMM modeling, normal vector constraints, and outlier modeling, we finally obtain the following equation for each point. probability Modeling:

[0065] .

[0066] From the above analysis, it can be seen that the covariance matrix... The covariance matrix plays a crucial role in point cloud probabilistic models, not only describing the noise characteristics of the data but also influencing the convergence and robustness of the algorithm. Considering that: the measurement noise of most 3D scanning devices exhibits similar statistical characteristics in all directions, especially in the absence of significant directional deviations, the isotropic assumption can well approximate this noise characteristic; a typical 3×3 covariance matrix has 6 independent parameters (due to symmetry), while the isotropic covariance has only 1 parameter, this simplification significantly reduces the complexity of parameter estimation; and the isotropic covariance remains invariant under rotational transformations, consistent with the properties of rigid body transformations. Therefore, the covariance matrix can be expressed as: and , , Let these represent the covariances of all GMM components and the normal vector, respectively. This represents a 3×3 identity matrix. This simplification brings several advantages. First, the number of parameters for each covariance matrix is ​​reduced from 6 to 1, significantly reducing the estimation difficulty and making it less prone to overfitting in noisy environments. The simplified covariance matrix also makes subsequent matrix operations more efficient.

[0067] To apply the Expectation-Maximization (EM) algorithm framework for point cloud registration, this invention uses observation data for each point. Introducing latent variables To represent its properties. Latent variables. It is a discrete random variable, with the following specific meanings:

[0068] ,

[0069] Based on the assumptions about outliers above: each data point has... The probability of this is an outlier, and there is... The probability comes from a Gaussian distribution. Therefore, the prior distribution for the latent variables is set as follows:

[0070] ,

[0071] Observation data and latent variables combination This is called complete data, based on the assumption that the data points are independent, and outliers are removed from the optimization objective. The expected log-likelihood function of complete data is... It can be represented as:

[0072] ,

[0073] in Represents the posterior probability:

[0074] .

[0075] Substituting the probability density function of a multivariate Gaussian distribution into the expected likelihood function, ignoring constant terms irrelevant to the optimization variables, and considering the rotation matrix... It must belong to a special orthogonal group, that is... Therefore, the multi-view point cloud registration problem can be formally formulated as the following constrained optimization problem:

[0076] ,

[0077] in, , Representation matrix The determinant of the matrix. Weighted norm. .

[0078] Based on the above derivation, this invention proposes a global registration method for multi-view three-dimensional point clouds of high-temperature targets, and uses the expectation-maximization (EM) algorithm framework to iteratively optimize the registration parameters. Figure 2 This is a flowchart illustrating a specific implementation of the global registration method for multi-view 3D point clouds of high-temperature targets according to the present invention. Figure 2 As shown, the specific steps of the high-temperature target multi-view three-dimensional point cloud global registration method of the present invention include:

[0079] S201: Data Acquisition and Preprocessing

[0080] from Point cloud collection of high-temperature targets from different perspectives ,in Indicates the first The first perspective Three-dimensional data of each point, , , Indicates the first The number of points in the point cloud set for each viewpoint. Then calculate the number of points for each point separately. normal vector In this embodiment, the normal vector is obtained by fitting the local plane with neighborhood points.

[0081] S202: Initial Registration

[0082] The initial rigid body transformation parameters for each viewpoint are obtained using a preset registration method. ,in Represents the initial rotation matrix. This represents the initial translation matrix.

[0083] S203: Initialize covariance:

[0084] Covariance of initial point covariance of the normal vector .

[0085] In this embodiment, the covariance is initialized based on the point cloud characteristics of the high-temperature target. Specifically, the point cloud resolution for each viewpoint is calculated, and the average point cloud resolution is obtained by averaging the point cloud resolutions from all viewpoints. In this embodiment, the point cloud resolution for each viewpoint is defined as the average distance between each point in the point cloud set and its nearest neighbor. Then, a temperature correction coefficient is set based on the shooting environment of the high-temperature target. The covariance can typically be set to [1, 5]. Initialize the covariance. , .

[0086] S204: Set the iteration number .

[0087] S205: Correspondence Update:

[0088] Using rotation matrix Translation matrix All points from each perspective Transform to common coordinate system coordinates Then, based on the coordinates of the common coordinate system, the first... Each point from each perspective In other Nearest neighbor index in a point cloud set of each viewpoint , and If you click In the If a point cloud set does not have a nearest neighbor, then the point is considered to be... In the Each perspective is an outlier, that is... , from point cloud collection Delete, and get the first The perspective in the first Cluster of non-outliers from various perspectives .

[0089] This embodiment employs a nearest neighbor search method based on kd-trees and normal vector constraints. The specific method is as follows:

[0090] For the A point of view From another perspective Search for the set of candidate points that satisfy the normal vector constraint. :

[0091] ,

[0092] in, , They represent , The normal vector, Represents the dot product of vectors. This represents the preset threshold for the angle between the normal vectors.

[0093] Based on the common coordinate system coordinates In the candidate point set Find the closest point as the point nearest neighbor index :

[0094] .

[0095] S206: Calculate the nearest neighbor confidence level:

[0096] This step is the E-step of the Expectation Maximization (EM) algorithm framework. According to Bayes' theorem, there are latent variables. The formula for calculating the posterior probability is as follows:

[0097] ,

[0098] in, Let the marginal probability be represented, then:

[0099] ,

[0100] In the formula, the numerator has the following case for outliers (denoted as...). ):

[0101] ,

[0102] In the formula, the numerator for the normal point case (denoted as...) ):

[0103] ,

[0104] The probability density function of the 6-dimensional Gaussian distribution in the above formula can be expressed as:

[0105] .

[0106] Therefore, latent variables The posterior probability of the point represents the probability of the point. With the Nearest neighbor from each perspective Corresponding credibility In high-temperature scenarios, this confidence assessment is crucial due to noise and feature blurring. (Confidence) It can be represented as:

[0107] ,

[0108] in:

[0109] ,

[0110] ,

[0111] in, Indicates the nearest neighbor. Common coordinate system coordinates, Indicates the first Rotation matrix for each viewpoint This indicates the search for the L2 norm. This represents the number of GMM components in a Gaussian mixture model. This represents the prior probability of an outlier. Indicates the first Centroid of point cloud data from a single perspective This represents the distance decay scale parameter for outliers.

[0112] Therefore, in this invention, for non-outlier points, the confidence level between them and their nearest neighbors is calculated based on the aforementioned posterior probability. In other words, the cloud of non-outlier points is calculated based on a Gaussian mixture model. Each point in With the Nearest neighbor from each perspective Credibility :

[0113] ,

[0114] in,

[0115] ,

[0116] .

[0117] The prior probability of outliers in this embodiment and distance attenuation scale parameters Also based on the average resolution of the point cloud and temperature correction factor Settings, make , .

[0118] S208: Update parameters:

[0119] The next step is the M-step of the Expectation Maximization (EM) algorithm framework, which requires maximizing the objective function. To estimate all the parameters of the model to be estimated This indicates that although it is difficult to estimate these model parameters simultaneously, their estimation can be performed sequentially and independently. Outliers are not included in the calculation during this process.

[0120] The rigid body transformation parameters are updated for each viewpoint, as follows:

[0121] First calculate the... The weighted centroid of each viewpoint in the global coordinate system and nearest neighbor weighted centroid :

[0122] ,

[0123] ,

[0124] Then calculate each point using the following formula. and its nearest neighbor Decentralized coordinates in the global coordinate system , :

[0125] ,

[0126] ,

[0127] Construct the first Weighted covariance matrix of each perspective :

[0128] ,

[0129] in, Indicates the nearest neighbor. The normal vector.

[0130] For the weighted covariance matrix Perform singular value decomposition Then based on orthogonal matrices and Update the rotation matrix and translation vector:

[0131] ,

[0132] .

[0133] Then, by applying the objective function... Regarding covariance Taking the derivative and setting it to zero, we obtain the update formulas for the two covariances:

[0134] ,

[0135] .

[0136] S208: Determine if the optimization termination condition has been met. If not, proceed to step S209; if so, proceed to step S210. The preferred termination condition is generally set to the convergence of the transformed parameters or the reaching of the maximum number of iterations.

[0137] S209: Set the iteration count Return to step S205.

[0138] S210: Obtain point cloud registration results:

[0139] Rotation matrix based on each viewpoint Translation matrix The point cloud from this perspective is transformed to the global coordinate system to obtain the registered point cloud.

[0140] To better illustrate the technical effects of the present invention, specific examples are used to experimentally verify the present invention.

[0141] In this embodiment, a binocular camera is used to collect point clouds of the workpiece from multiple perspectives. The images are taken from seven different perspectives. Each perspective includes a rectangular plane with rounded corners and a curved workpiece. The features are sparse and highly repetitive, and the point cloud of each perspective contains approximately 80,000 points. Figure 3 These are point cloud images from seven perspectives in this embodiment. For example... Figure 3 As shown, the point clouds at each viewpoint exhibit significant spatial offset, and measurement noise and outliers exist in the edge regions.

[0142] In this embodiment, the FPFH+RANSAC method is used for initial registration. An FPFH feature descriptor is generated for each point, and then feature matching is performed between different viewpoints.

[0143] Next, we calculate the average resolution of the point cloud. For each point cloud, we calculate the average of its nearest neighbor distances, resulting in the average resolution of the point cloud across 7 viewpoints. =1.9174, set the temperature correction factor. =2, based on which the initial value of the location covariance is set. =3.8348, set the initial value of the normal vector covariance. =30.6784, the prior probability of the outlier. =4%. The EM algorithm is used for iterative optimization, with a maximum number of iterations set to 50, and the convergence threshold of the rotation matrix is... Convergence threshold of translation vector Threshold of the angle between normal vectors =20°. In this embodiment, the convergence condition is finally met on the 32nd iteration. Figure 4 This is the point cloud image after point cloud registration in this embodiment. For example... Figure 4 As shown, after transforming the multi-view point cloud to the global coordinate system, the point clouds fit together quite closely, and there is no obvious registration error visible to the naked eye. The total number of registered point clouds is approximately 550,000 points, covering the entire surface of the high-temperature target.

[0144] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A high-temperature target multi-view three-dimensional point cloud global registration method, characterized in that, Includes the following steps: S1: From Point cloud collection of high-temperature targets from different perspectives ,in Indicates the first The first perspective Three-dimensional data of each point, , , Indicates the first The number of points in the point cloud set of each viewpoint is calculated, and then each point is calculated separately. normal vector ; S2: Obtain the initial rigid body transformation parameters for each viewpoint using a preset registration method. ,in Represents the initial rotation matrix. Represents the initial translation matrix; S3: Initialize covariance of points and covariance of normal vectors ; S4: Let the iteration number ; S5: Using a rotation matrix Translation matrix All points from each perspective Transform to common coordinate system coordinates Then, based on the coordinates of the common coordinate system, the first... Each point from each perspective In other Nearest neighbor index in a point cloud set of each viewpoint , and ; If point In the If a point cloud set does not have a nearest neighbor, then the point is considered to be... In the From the perspective of outliers, we can extract them from the point cloud. Delete, and get the first The perspective in the first Cluster of non-outliers from various perspectives ; S6: Compute non-outlier point cloud set based on Gaussian mixture model each point nearest neighbor point of the first confidence : ， in, ， ， in, Indicates the nearest neighbor. Common coordinate system coordinates, Indicates the first Rotation matrix for each viewpoint This indicates the search for the L2 norm. This represents the number of GMM components in a Gaussian mixture model. This represents the prior probability of an outlier. Indicates the first Centroid of point cloud data from a single perspective This represents the outlier distance decay scale parameter; S7: Update the rigid body transformation parameters for each viewpoint, as follows: First calculate the... The weighted centroid of each viewpoint in the global coordinate system and nearest neighbor weighted centroid : ， ， Then each point is calculated by the following formula and its nearest neighbor in the global coordinate system , : ， ， Constructing a weighted covariance matrix for the first viewpoint : ， wherein represents the normal vector of the nearest neighbor point ; on the weighted covariance matrix performing singular value decomposition then updating the rotation matrix and translation vector based on the orthogonal matrix and performing singular value decomposition ， ； Then, the two covariances are updated using the following formula: ， ； S8: Determine whether the optimization termination condition has been met. If not, proceed to step S9; if it has been met, proceed to step S10. S9: Let the iteration number , return to step S5; S10: Rotation matrix based on each viewpoint Translation matrix The point cloud from this perspective is transformed to the global coordinate system to obtain the registered point cloud.

2. The high temperature target multi-view three-dimensional point cloud global registration method of claim 1, wherein, The method for initializing the covariance in step S3 is as follows: Calculate the point cloud resolution for each viewpoint separately, and then average the point cloud resolutions for all views to obtain the average point cloud resolution. Then, set the temperature correction factor according to the shooting environment of the high-temperature target. Initialize covariance , .

3. The high temperature target multi-view three-dimensional point cloud global registration method of claim 2, wherein, The point cloud resolution is the average distance between each point in the point cloud set and its nearest neighbor.

4. The high temperature target multi-view three-dimensional point cloud global registration method of claim 1, wherein, Step S5 employs a nearest neighbor search method based on kd-trees and normal vector constraints, as detailed below: For the A point of view From another perspective Search for the set of candidate points that satisfy the normal vector constraint. : ， wherein, , respectively represent normal vectors of , , denotes a dot product of vectors, denotes a preset normal vector angle threshold value; Based on public coordinate system coordinates Among the candidate point set Find the nearest point as the nearest neighbor index of point :​ 。 5. The multi-view point cloud registration method of claim 1, wherein, The prior probability , represents a temperature correction coefficient set according to the shooting environment in which the high-temperature target is located.

6. The high temperature target multi-view three-dimensional point cloud global registration method of claim 1, wherein, The outlier distance decay scale parameter , represents a temperature correction coefficient set according to a shooting environment in which the high-temperature target is located, represents point cloud average resolution of the N views.

Images