Rapid three-dimensional reconstruction method for large converter

Abstract

The invention discloses a rapid three-dimensional reconstruction method for a large converter, and belongs to the technical field of point cloud three-dimensional reconstruction. The method comprises the following steps: S1, data acquisition: scanning different parts of a buggy ladle, acquiring point cloud data of each part of the furnace top, the furnace body and the furnace bottom of the converter, and performing data preprocessing; s2, feature point detection: carrying out feature point detection on the converter point cloud data preprocessed in the step S1; s3, point cloud segmentation and point cloud registration are carried out on the point cloud data obtained through preprocessing; and S4, performing three-dimensional reconstruction: performing curved surface reconstruction on the converter point cloud by using a three-dimensional reconstruction algorithm to obtain a converter three-dimensional model. According to the method, the features of point cloud three-dimensional reconstruction are utilized, more comprehensive and accurate overall three-dimensional information of the converter is obtained, and detailed geometric information of the converter can be provided.

Description

Technical Field

[0001] The invention belongs to the technical field of point cloud three-dimensional reconstruction, and more specifically, relates to a fast three-dimensional reconstruction method for a large converter. Background Art

[0002] Large converters are usually used in industrial production. Due to the large size of the converter, it is generally rarely disassembled regularly after installation. Therefore, from the beginning to the end of use, the internal structure and detailed geometric information of the converter cannot be observed throughout the process, resulting in engineers being unable to monitor the wear, deformation and structural damage of the equipment, and thus unable to formulate effective maintenance and overhaul plans. In addition, since the three-dimensional reconstruction of the internal space and process of the converter cannot be directly observed from the outside, the flow characteristics during the metal smelting process cannot be analyzed, and data support cannot be provided for the optimization of process parameters, thereby failing to effectively improve production efficiency and product quality.

[0003] After searching, patent CN114758333A discloses a method and system for identifying the detachment of a ladle from a hook when it is hoisted by a casting crane. The method includes receiving point cloud data collected by a laser radar; receiving image data collected by a camera; preprocessing the point cloud data to remove clutter and noise to obtain preprocessed point cloud data; performing three-dimensional reconstruction based on the preprocessed point cloud data and image data to obtain fused point cloud data and establish a three-dimensional model of the ladle and hook. However, the patent is for the three-dimensional reconstruction of the ladle and hook, and the solution does not involve the construction of a curved surface model, so it cannot be directly used in the model construction of the converter in this application.

[0004] Patents US20160125226A1 and WO2015039375A1 both disclose methods for automatically optimizing the quality of point cloud data, which use a Possion surface reconstruction method on the obtained initial discrete point cloud to obtain a Possion reconstructed point cloud model. Patent CN106504332A discloses using a moving least squares algorithm to smoothly resample point cloud data; spatially downsample the point cloud data; and use a Poisson surface reconstruction algorithm to triangulate the point cloud data.

[0005] Patent CN110111430A discloses a method for extracting quadratic surfaces from three-dimensional point clouds. First, the surface corresponding to each point is fitted using the K-neighborhood point set of each point and the corresponding surface parameter standard deviation is calculated; secondly, the point set is arranged in ascending order according to the surface parameter standard deviation; then, each point is taken out from the sorted point set in sequence, and when the surface parameter standard corresponding to the taken point is less than the set threshold and is not marked as a grown point, it is regarded as an initial seed point; regional growth is performed from the initial seed point, and the growth points belonging to the surface are marked with the sequence number of the current region, and the characteristic parameters of the surface are dynamically fitted until a surface is grown; the steps are repeated to take out points from the point set and perform dynamic fitting until the surface parameter standard deviation is greater than the threshold, and finally the corresponding surface is extracted according to the marked sequence number in the point cloud.

[0006] However, the accuracy of the surface reconstructed using the above patent is low, which affects the monitoring of the internal structure of the converter. Summary of the invention

[0007] 1. Problem to be solved

[0008] In view of the problem of low accuracy of existing surface reconstruction, the present invention provides a method for rapid three-dimensional reconstruction of a large converter. The surface reconstructed by this method has high accuracy and can achieve accurate monitoring of the internal structure of the converter.

[0009] 2. Technical solution

[0010] In order to solve the above problems, the technical solution adopted by the present invention is as follows:

[0011] The present invention discloses a method for rapid three-dimensional reconstruction of a large converter, comprising the following steps:

[0012] Step S1, collecting data: collecting point cloud data of the top, body and bottom of the converter, and then preprocessing the point cloud data and estimating the point cloud normal vector;

[0013] Step S2, identifying feature points: after preprocessing the collected converter point cloud data, feature point detection is performed on the converter point cloud data to identify feature points;

[0014] Step S3, performing point cloud segmentation and point cloud registration on the pre-processed point cloud data;

[0015] Step S4, three-dimensional reconstruction: After acquiring the converter point cloud data through the point cloud acquisition device, the complete point cloud data can be obtained by using point cloud segmentation, point cloud registration and other processing methods, and then the three-dimensional reconstruction algorithm is used to complete the surface reconstruction of the converter point cloud to obtain the converter three-dimensional model.

[0016] The step one specifically includes the following steps:

[0017] Step S101, collecting data: using a non-contact point cloud data acquisition method to respectively acquire point cloud data of the converter top, furnace body and furnace bottom, the non-contact point cloud data acquisition method is selected from an optical measurement method, an acoustic measurement method or an electromagnetic measurement method, and converting the acquired point cloud data file into a PCD file to facilitate subsequent data processing, preferably a laser scanner, and calibrating the internal parameters of the laser scanner before scanning.

[0018] Step S102, data preprocessing: The specific steps of preprocessing are: setting filtering conditions, filtering out data that are not within the required range to achieve conditional filtering, and then deleting the converter point cloud data that does not meet the given conditions, wherein the set filtering conditions can be set by the user according to actual conditions.

[0019] Among them, conditional filtering sets the filtering conditions through key member functions, and points that do not meet the filtering conditions will be filtered out; in the process of using member function settings, it can be considered whether the points deleted after filtering need to be retained, so as to maintain the orderliness of the point cloud;

[0020] Furthermore, to use conditional filtering on the initial converter point cloud, it is necessary to first determine the value range of the comparison operator of the conditional filtering on the fields in the x, y, and z directions of the point cloud data.

[0021] Step 103, estimate the point cloud normal vector: the point cloud normal vector involves the extraction, segmentation, registration and three-dimensional reconstruction of the converter point cloud and other links. Therefore, the point cloud normal vector needs to be accurately estimated for the converter point cloud data. The principal component analysis method is used to estimate the point cloud normal vector, which is calculated by the local plane fitting method.

[0022] The process of estimating the point cloud normal vector specifically includes:

[0023] For any point A(x,y,z) in the point cloud model, the K-neighborhood point set of point A is N p , use KD tree structure to store point cloud data to improve the search efficiency of spatial points, for N p The least squares method is used to fit the local plane, and the formula is:

[0024]

[0025] Where: n represents the plane normal vector, satisfying ||n||2=1, d is the distance from the origin to the plane, A i is any point A(x,y,z) in the point cloud model, and A(n,d) is N p Fitted local plane expression for a set of points.

[0026] The value of d is calculated according to the following method:

[0027] The formula for the plane equation is:

[0028] ax+by+cz+d=0

[0029] Right now:

[0030] d=-(ax+by+cz)

[0031] Where a, b, and c are plane parameters.

[0032] The covariance matrix C of point A in the converter point cloud is:

[0033]

[0034] in k is a constant with a value of 1-k. The eigenvalue λ of the covariance matrix C can be calculated by singular value decomposition i (i=0,1,2), satisfying λ2≥λ1≥λ0. The eigenvalues ​​λ2, λ1, and λ0 correspond to eigenvectors v2, v1, and v0 respectively. The eigenvector v0 corresponding to the minimum eigenvalue λ0 is the normal vector of point A, that is, the plane parameters a, b, and c are the component values ​​of vector v0 respectively, and the calculation formula of parameter d is:

[0035]

[0036] in:

[0037] are the weighted averages of x, y, and z respectively.

[0038] The second step is to detect feature points on the pre-processed converter point cloud data, which specifically includes the following steps:

[0039] Harris operator is a common feature detection operator proposed by Harris et al. It can extract edge points and corner points in the model. 3D Harris feature point detection is an extension of the Harris detection operator in three dimensions.

[0040] Step 201: In the converter three-dimensional point cloud space, take any point M as the center, find the neighborhood point set of point M within the radius r, perform principal component analysis (PCA) on the points in this set, and fit the quadratic surface by the least squares method. The expression of the surface is:

[0041]

[0042] Where x, y and z are the coordinates of the surface, m1, m2, m3, m4, m5 and m6 are the neighborhood points of point M within the radius r.

[0043] The partial derivatives of the surface z in the x and y directions can be denoted by f x(x,y) and f y (x,y), where:

[0044] f x (x,y)=m1x+m2y+m4

[0045] f y (x,y)=m2x+m3y+m5

[0046] Step 202: Integrate the partial derivative using a continuous Gaussian function. The specific formula is as follows:

[0047]

[0048] Where: σ is a constant, the support of the Gaussian function is defined by it, and R is a constant in a given interval;

[0049] Step 203, calculate A, B, C, the specific formula is:

[0050] A=m4 2 +2m1 2 +2m2 2

[0051] B=m5 2 +2m2 2 +2m3 2

[0052] C=m4m5+2m1m2+2m2m3

[0053] The symmetric matrix E can be defined as:

[0054]

[0055] Step 204: define the 3D Harris detection operator V(M) of point M:

[0056] V(M)=det(E)-a×tr(E) 2

[0057] Where: det(E) and tr(E) represent the determinant and trace of the matrix E, respectively, and a is the parameter that needs to be adjusted in the experiment;

[0058] Step 205: Finally, the non-maximum suppression method (NMS) is used to regard the points with a threshold value T less than V(M) as feature points.

[0059] Among them, in step 3, the point cloud data obtained by preprocessing is segmented and registered, and the process is as follows:

[0060] Step 301: point cloud segmentation adopts point cloud region growing segmentation algorithm, and its basic process is as follows:

[0061] a. Select the initial seed point from the preprocessed point cloud data and add it to the seed list;

[0062] b. Find the neighboring points of the seed point and use the angle between the normal vectors of the point cloud as the growth criterion. If the angle between the normal vectors of the seed point and its neighboring points is less than the given threshold ε, it means that they belong to the same surface, and the points in the neighboring points whose curvature is less than the threshold τ are used as new seed points to continue growing outward.

[0063] c. Similarly, points with similar attributes are merged into the same area until all points are divided into a certain area and the growth stops.

[0064] Step 302: A feature matching registration method is used for the converter point cloud registration to achieve a rough registration of the converter point cloud. The specific steps are:

[0065] a. Rough registration of point cloud: The rough registration of converter point cloud is realized by using the registration method based on feature matching. The specific method is as follows: the feature descriptors are matched by using the improved nearest neighbor distance ratio method, and the mismatched point pairs are removed, so as to optimize the correspondence between the feature points in the matching pairs. Finally, the transformation matrix is ​​calculated by the quaternion method to complete the rough registration of point cloud.

[0066] b. Point cloud fine registration: The improved ICP fine registration algorithm is used to optimize the rough registration results, and finally obtain the complete converter point cloud data. The improved ICP fine registration algorithm is: by iteratively finding the corresponding point pairs between the two point clouds and minimizing the residual registration error, the rigid transformation matrix is ​​optimized. In the process of finding the nearest point, the KD tree is used to improve the search efficiency of the nearest neighbor point, and the sampling strategy from coarse to fine is used to eliminate the point pairs whose distance is greater than the adaptive distance threshold in the iteration to improve the registration accuracy. Finally, the rigid body transformation relationship is calculated by the SVD algorithm, so as to optimize the ICP fine registration algorithm.

[0067] In step 4, the 3D reconstruction algorithm is used to complete the surface reconstruction of the converter point cloud to obtain the 3D model of the converter. The surface reconstruction process is as follows:

[0068] Poisson surface reconstruction based on Poisson's equation is a triangular mesh reconstruction method based on implicit functions. The approximate surface is obtained by performing optimal interpolation processing on the converter point cloud data.

[0069] Poisson surface reconstruction is an implicit surface reconstruction method that transforms surface reconstruction into a problem of solving the Poisson equation.

[0070] For the actual surface Ω, any point m i (x i ,y i ,z i)∈Ω, if F(x,y,z) is the reconstructed surface equation, then the surface model can be described by the following formula:

[0071]

[0072] At the same time:

[0073]

[0074] Compared with most converter equipment that cannot frequently conduct effective equipment inspection and structural observation, the present invention utilizes the characteristics of point cloud three-dimensional reconstruction to not only obtain more comprehensive three-dimensional information of the converter as a whole, but also provide detailed geometric information of the converter, helping engineers monitor the wear, deformation and structural damage of the equipment, so as to formulate effective maintenance and overhaul plans; secondly, in terms of process optimization, through the three-dimensional reconstruction of the internal space and process of the converter, the flow characteristics in the metal smelting process can be analyzed, providing data support for the optimization of process parameters, improving production efficiency and product quality; in addition, in terms of safety assessment, three-dimensional reconstruction can help identify potential safety hazards, such as high-temperature areas and molten metal flow paths, thereby strengthening safety management and reducing accident risks. Finally, the experiment shows that the method of the present invention still has good performance in other environments, can well solve the problem of needing to perform three-dimensional analysis on large converter equipment in reality, and has good generalization ability.

[0075] 3. Beneficial effects

[0076] Compared with the prior art, the present invention has the following beneficial effects:

[0077] (1) The present invention performs point cloud segmentation and point cloud registration on point cloud data, and uses a three-dimensional reconstruction algorithm to reconstruct the surface of the converter point cloud to obtain a three-dimensional model of the converter. In the process of surface reconstruction, the least squares method is used to fit the local plane to Np, and the quadratic surface is fit by the least squares method. The corresponding quadratic surface is extracted, and the point cloud data is registered to achieve reconstruction of the three-dimensional model of the converter, which is beneficial to improving the accuracy of the surface and can deeply analyze the fluid flow, heat distribution and reaction process inside the converter. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] The technical solution of the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments, but it should be understood that these drawings are designed only for explanation purposes and are not intended to limit the scope of the present invention. In addition, unless otherwise specified, these drawings are intended only to conceptually illustrate the structural configurations described herein and are not necessarily drawn to scale.

[0079] Figure 1 The present invention is a flow chart of a method for rapid three-dimensional reconstruction of a large converter. DETAILED DESCRIPTION

[0080] The following detailed description of exemplary embodiments of the present invention refers to the accompanying drawings, which form a part of the description, and in which exemplary embodiments of the present invention that can be implemented are shown as examples. Although these exemplary embodiments are described in sufficient detail to enable those skilled in the art to implement the present invention, it should be understood that other embodiments can be implemented and various changes can be made to the present invention without departing from the spirit and scope of the present invention. The following more detailed description of the embodiments of the present invention is not intended to limit the scope of the claimed invention, but is only for the purpose of illustrating and not limiting the description of the characteristics and features of the present invention, so as to propose the best mode for performing the present invention and to enable those skilled in the art to implement the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

[0081] It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention. As used herein, the terms "parallel" and "perpendicular" are not limited to their strict geometric definitions, but include tolerances for reasonable and inconsistent machining or human errors.

[0082] Example

[0083] The present invention discloses a method for rapid three-dimensional reconstruction of a large converter, comprising the following steps:

[0084] Step S1, collecting data: collecting point cloud data of the top, body and bottom of the converter, and then preprocessing the point cloud data and estimating the point cloud normal vector;

[0085] Step S2, identifying feature points: after preprocessing the collected converter point cloud data, feature point detection is performed on the converter point cloud data to identify feature points;

[0086] Step S3, performing point cloud segmentation and point cloud registration on the pre-processed point cloud data;

[0087] Step S4, three-dimensional reconstruction: After acquiring the converter point cloud data through the point cloud acquisition device, the complete point cloud data can be obtained by using point cloud segmentation, point cloud registration and other processing methods, and then the three-dimensional reconstruction algorithm is used to complete the surface reconstruction of the converter point cloud to obtain the converter three-dimensional model.

[0088] The step one specifically includes the following steps:

[0089] Step S101, collecting data: using a non-contact point cloud data acquisition method to respectively acquire point cloud data of the converter top, furnace body and furnace bottom, the non-contact point cloud data acquisition method is selected from an optical measurement method, an acoustic measurement method or an electromagnetic measurement method, and converting the acquired point cloud data file into a PCD file to facilitate subsequent data processing, preferably a laser scanner, and calibrating the internal parameters of the laser scanner before scanning.

[0090] Step S102, data preprocessing: The initially acquired converter point cloud data contains external points, noise and invalid data, so the PCL platform is used to preprocess the original converter point cloud data. The specific steps of preprocessing are as follows:

[0091] Set the filtering conditions, filter out the data that is not within the required range to achieve conditional filtering, and then delete the converter point cloud data that does not meet the given conditions. The set filtering conditions can be set by the user according to actual conditions.

[0092] Among them, conditional filtering sets the filtering conditions through key member functions, and points that do not meet the filtering conditions will be filtered out; in the process of using member function settings, it can be considered whether the points deleted after filtering need to be retained, so as to maintain the orderliness of the point cloud;

[0093] Furthermore, to use conditional filtering on the initial converter point cloud, it is necessary to first determine the value range of the comparison operator of the conditional filtering on the fields in the x, y, and z directions of the point cloud data.

[0094] Step 103, estimating the point cloud normal vector: The point cloud normal vector involves the extraction, segmentation, registration and 3D reconstruction of the converter point cloud, so the converter point cloud data also needs to accurately estimate the point cloud normal vector. The point cloud normal vector is estimated using the principal component analysis method, which is calculated by the local plane fitting method. The process of estimating the point cloud normal vector specifically includes:

[0095] For any point A(x,y,z) in the point cloud model, the K-neighborhood point set of point A is N p , use KD tree structure to store point cloud data to improve the search efficiency of spatial points, for N p The least squares method is used to fit the local plane, and the formula is:

[0096]

[0097] Where: n represents the plane normal vector, satisfying ||n||2=1, d is the distance from the origin to the plane, A i is any point A(x,y,z) in the point cloud model, and A(n,d) is N p Fitted local plane expression for a set of points.

[0098] The value of d is calculated according to the following method:

[0099] The formula for the plane equation is:

[0100] ax+by+cz+d=0

[0101] Right now:

[0102] d=-(ax+by+cz)

[0103] Where a, b, and c are plane parameters.

[0104] The covariance matrix C of point A in the converter point cloud is:

[0105]

[0106] in k is a constant with a value of 1-k. The eigenvalue λ of the covariance matrix C can be calculated by singular value decomposition i (i=0,1,2), satisfying λ2≥λ1≥λ0. The eigenvalues ​​λ2, λ1, and λ0 correspond to eigenvectors v2, v1, and v0 respectively. The eigenvector v0 corresponding to the minimum eigenvalue λ0 is the normal vector of point A, that is, the plane parameters a, b, and c are the component values ​​of vector v0, which are the projections of vector v0 on the Cartesian axis. The calculation formula of parameter d is:

[0107]

[0108] in:

[0109] are the weighted averages of x, y, and z respectively.

[0110] The second step is to detect feature points on the pre-processed converter point cloud data, which specifically includes the following steps:

[0111] Harris operator is a common feature detection operator proposed by Harris et al. It can extract edge points and corner points in the model. 3D Harris feature point detection is an extension of the Harris detection operator in three dimensions.

[0112] Step 201: In the converter three-dimensional point cloud space, take any point M as the center, find the neighborhood point set of point M within the radius r, perform principal component analysis (PCA) on the points in this set, and fit the quadratic surface by the least squares method. The expression of the surface is:

[0113]

[0114] Where x, y and z are the coordinates of the surface, m1, m2, m3, m4, m5 and m6 are the neighborhood points of point M within the radius r.

[0115] The partial derivatives of the surface z in the x and y directions can be denoted by f x (x,y) and f y (x,y), where:

[0116] f x (x,y)=m1x+m2y+m4

[0117] f y (x,y)=m2x+m3y+m5

[0118] Step 202: Integrate the partial derivative using a continuous Gaussian function. The specific formula is as follows:

[0119]

[0120] Where: σ is a constant, and the support of the Gaussian function is defined by it. R is a constant for a given interval, reflecting the range of the interval.

[0121] Step 203, calculate A, B, C, the specific formula is:

[0122] A=m4 2 +2m1 2 +2m2 2

[0123] B=m5 2 +2m2 2 +2m3 2

[0124] C=m4m5+2m1m2+2m2m3

[0125] The symmetric matrix E can be defined as:

[0126]

[0127] Step 204: define the 3D Harris detection operator V(M) of point M:

[0128] V(M)=det(E)-a×tr(E) 2

[0129] Where: det(E) and tr(E) represent the determinant and trace of the matrix E, respectively, and a is the parameter that needs to be adjusted in the experiment;

[0130] Step 205: Finally, the non-maximum suppression method (NMS) is used to regard the points with a threshold value T less than V(M) as feature points.

[0131] The principle of its non-maximum suppression method is:

[0132]

[0133] In the process of target detection, many detection frames (A, B, C, D, E, F, etc.) are generated, many of which detect the same target, but in the end each target only needs one detection frame. NMS selects the detection frame with the highest score (i.e. overlap) (assuming it is C), and then calculates the corresponding IOU value between C and the remaining frames. When the IOU value exceeds the set threshold N t , that is, for t The suppression method is to set the score of the detection box to 0. After one round, continue to look for the highest score in the remaining detection boxes, and then suppress the boxes whose IOU exceeds the threshold, until almost no overlapping boxes are retained. In this way, basically only one detection box is left for each target.

[0134] Among them, in step 3, the point cloud data obtained by preprocessing is segmented and registered, and the process is as follows:

[0135] Step 301: point cloud segmentation adopts point cloud region growing segmentation algorithm, and its basic process is as follows:

[0136] a. Select the initial seed point from the preprocessed point cloud data and add it to the seed list;

[0137] b. Find the neighboring points of the seed point and use the angle between the normal vectors of the point cloud as the growth criterion. If the angle between the normal vectors of the seed point and its neighboring points is less than the given threshold ε, it means that they belong to the same surface, and the points in the neighboring points whose curvature is less than the threshold τ are used as new seed points to continue growing outward.

[0138] c. Similarly, points with similar attributes are merged into the same area until all points are divided into a certain area and the growth stops.

[0139] Step 302: A feature matching registration method is used for the converter point cloud registration to achieve a rough registration of the converter point cloud. The specific steps are:

[0140] a. Rough registration of point clouds: A registration method based on feature matching is used to achieve rough registration of converter point clouds. The specific method is as follows: In step S1 and step S2, normal vector estimation and feature point detection are performed on the source point cloud and the target point cloud, and the covariance matrix descriptor is used to describe the feature points; the source point cloud is the collected point cloud data, and the target point cloud is the point cloud data at the target position. In this step, the improved nearest neighbor distance ratio method is used to perform feature matching on the feature descriptor, and the mismatched point pairs are removed, thereby optimizing the correspondence between the feature points in the matching pairs.

[0141] Where: Assume that the point cloud S after feature point detection in step S2 i With S j The feature descriptor sets extracted from and

[0142] The main idea of ​​the nearest neighbor distance ratio feature matching is to i Any feature descriptor In F j Find the closest feature descriptor and the next closest descriptor If the formula is satisfied: Where ε1 is the set threshold, then For the matching feature descriptor, the specific steps of feature matching are as follows:

[0143] Ⅰ. For F i Any feature descriptor Use the nearest neighbor distance ratio method to find the matching feature descriptor

[0144] II. Feature Descriptors According to the idea of ​​the nearest neighbor distance ratio method, find the corresponding feature descriptor

[0145] III. If and Corresponding to the same feature point, and are matching feature descriptors, and their corresponding feature points form feature point pairs

[0146] IV. Through iteration, the feature descriptor set F i and F j Perform feature matching to obtain a series of corresponding point pairs

[0147] V. Calculate the transformation matrix by quaternion method: The registration between point sets using quaternion method mainly calculates the rigid transformation T = [R, t], where R is the rotation matrix and t is the translation vector. Given two corresponding point sets P = {p i ,i=1,2,...,N} and Q={q i ,i=1,2,...,N}, N is the number of points in the point set, and the p with the same subscript i and q i For the corresponding feature point pairs, complete the point cloud coarse registration;

[0148] b. Point cloud fine registration: The improved ICP fine registration algorithm is used to optimize the rough registration results, and finally obtain the complete converter point cloud data. The improved ICP fine registration algorithm is: by iteratively finding the corresponding point pairs between the two point clouds and minimizing the residual registration error, the rigid transformation matrix is ​​optimized. In the process of finding the nearest point, the KD tree is used to improve the search efficiency of the nearest neighbor point, and the sampling strategy from coarse to fine is used to eliminate the point pairs whose distance is greater than the adaptive distance threshold in the iteration to improve the registration accuracy. Finally, the rigid body transformation relationship is calculated by the SVD algorithm, so as to optimize the ICP fine registration algorithm.

[0149] Among them: The steps of calculating the rotation matrix and translation vector by SVD algorithm are:

[0150] Ⅰ. Given a set of points to be registered P = {p i} and Q = {q i}, i = 1, 2, ..., N, P i and q i are corresponding point pairs, and the centers of P and Q are μ P, μ Q , assuming that the rotation matrix and translation vector are R and T respectively, then:

[0151] q i =Rp i +t

[0152] Essential Matrix:

[0153] Ⅱ. Centroid the point sets P and Q respectively to form a new point cloud set in

[0154] III. Substitute the new point cloud set into the essential matrix E(R,t) above and minimize it:

[0155]

[0156] IV. Calculation and The correlation coefficient matrix M between them is as follows:

[0157]

[0158] Ⅴ. Perform singular value decomposition on M, the formula is: M = UΛV T ,

[0159] Where U is an m-order orthogonal matrix, V is an n-order orthogonal matrix, m and n represent the rows and columns of the matrix respectively, V T is the transposed matrix of V.

[0160] VI. The rotation matrix R = VSU that minimizes the error function T ,

[0161] Where S is a diagonal matrix, and its expression is:

[0162] S = diag(1,1,det(U)det(V)), optimal translation vector t = μ Q -Rμ P ,

[0163] Where det(U) is the determinant of U, and det(V) is the determinant of V.

[0164] In step 4, the 3D reconstruction algorithm is used to complete the surface reconstruction of the converter point cloud to obtain the 3D model of the converter. The surface reconstruction process is as follows:

[0165] Poisson surface reconstruction based on Poisson's equation is a triangular mesh reconstruction method based on implicit functions. The approximate surface is obtained by performing optimal interpolation processing on the converter point cloud data.

[0166] Poisson surface reconstruction is an implicit surface reconstruction method that transforms surface reconstruction into a problem of solving the Poisson equation.

[0167] The optimal interpolation can be processed by nearest neighbor interpolation and linear interpolation.

[0168] For the actual surface Ω, any point m i (x i ,y i ,z i )∈Ω, if F(x,y,z) is the reconstructed surface equation, then the surface model can be described by the following formula:

[0169]

[0170] At the same time:

[0171]

[0172] M i (x i ,y i ,z i ) into F(x,y,z),

[0173] Where (x i ,y i ,z i ) are the coordinates obtained by solving, which constitute the reconstructed surface equation.

[0174] The present invention achieves reconstruction of the three-dimensional model of the converter by improving the accuracy of the curved surface, and can deeply analyze the fluid flow, heat distribution and reaction process inside the converter; this information can be used to optimize the smelting process, improve production efficiency and product quality; it can identify potential safety hazards, such as high-temperature areas and dangerous parts, and take timely measures to prevent accidents, thereby improving overall operational safety; combined with precise three-dimensional data analysis, it can evaluate the environmental impact of the converter, such as waste gas emissions, so as to take measures to reduce environmental pollution and improve energy efficiency.

[0175] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope involved in the present application is not limited to the technical solution formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the inventive concept, such as a technical solution formed by replacing the above-mentioned features with (but not limited to) technical features with similar functions disclosed in this application.

[0176] Except for the technical features described in the specification, the remaining technical features are known technologies to those skilled in the art. In order to highlight the innovative features of the present invention, the remaining technical features will not be described here in detail.

Claims

1. A method for rapid three-dimensional reconstruction of a large converter, characterized in that: The following steps are involved: Step S1, data collection: scanning different parts of the ladle car, collecting point cloud data of the top, body and bottom of the converter, and performing data preprocessing; Step S2, feature point detection: performing feature point detection on the converter point cloud data obtained by preprocessing; Step S3, performing point cloud segmentation and point cloud registration on the pre-processed point cloud data; Step S4, three-dimensional reconstruction: using a three-dimensional reconstruction algorithm to perform surface reconstruction on the converter point cloud to obtain a three-dimensional model of the converter.

2. The method according to claim 1, characterized in that The method for estimating the point cloud normal vector is a principal component analysis method, which specifically includes: for any point A(x, y, z) in the point cloud model, the K neighborhood point set of point A is N p , for N p The least squares method is used to fit the local plane, and the plane formula is: In the formula, n represents the plane normal vector, satisfying ||n||2=1, d is the distance from the origin to the plane, and the result is solved using the principal component analysis method, k is a constant with a value of 1-k, and A i is any point A(x,y,z) in the point cloud model.

3. The method according to claim 2, characterized in that The numerical calculation process of d is: d=-(ax+by+cz) In the formula, a, b, c are plane parameters, which are calculated according to the covariance matrix C of point A in the point cloud. The formula is: In the formula, A i For any point A(x,y,z) in the point cloud model, k is a constant with a value of 1-k. The eigenvalue λ of the covariance matrix C is calculated by singular value decomposition. i (i=0,1,2), satisfying λ2≥λ1≥λ0, the eigenvalues ​​λ2, λ1, λ0 correspond to the eigenvectors v2, v1, v0 respectively, the eigenvector v0 corresponding to the minimum eigenvalue λ0 is the normal vector of point A, and the plane parameters a, b, c are the component values ​​of vector v0 respectively.

4. The method according to claim 3, characterized in that The specific steps of feature point detection in step S2 include: Step 201: In the converter three-dimensional point cloud space, take any point M as the center, find the neighborhood point set of point M within the radius r, perform principal component analysis on the points in this set, and fit a quadratic surface by the least squares method. The expression of the surface is: Where x, y and z are the coordinates of the surface, m1, m2, m3, m4, m5 and m6 are the neighborhood points of point M within the radius r. Step 202: Integrate the partial derivative using a continuous Gaussian function. The specific formula is as follows: Where: σ is a constant, R is a constant in a given interval; Step 203: A, B, and C are calculated using calculus. The specific formula is: A=m4 2 +2m1 2 +2m2 2 B=m5 2 +2m2 2 +2m3 2 C=m4m5+2m1m2+2m2m3 The symmetric matrix E is defined as: Step 204: define the 3D Harris detection operator V(M) of point M: V(M)=det(E)-a×tr(E) 2 In the formula, det(E) and tr(E) represent the determinant and trace of the matrix E, respectively, and a is the parameter that needs to be adjusted in the experiment; Step 205: Use the non-maximum suppression method to regard the points with a threshold value T less than V(M) as feature points.

5. The method according to claim 4, characterized in that The point cloud segmentation in step S3 adopts the point cloud region growing segmentation algorithm, and its basic process is: a. Select the initial seed point in the converter point cloud and add it to the seed list; b. Find the neighboring points of the seed point and use the angle between the normal vectors of the point cloud as the growth criterion. If the angle between the normal vectors of the seed point and its neighboring points is less than the given threshold ε, it can be said that they belong to the same surface, and the points with curvature less than the threshold τ in the neighboring points are used as new seed points to continue growing outward; c. Similarly, points with similar attributes are merged into the same area until all points are divided into a certain area and the growth stops.

6. The method according to claim 5, characterized in that The specific steps of point cloud registration in step S3 include: a. Use the feature matching-based registration method to achieve coarse registration of the converter point cloud; b. Use the improved ICP fine registration algorithm to optimize the coarse registration results and finally obtain the complete converter point cloud data.

7. The method according to claim 6, characterized in that The nearest neighbor distance ratio method is used to match the feature descriptors, remove the mismatched point pairs, optimize the correspondence between the matching feature points, and finally calculate the transformation matrix through the quaternion method to complete the point cloud coarse registration.

8. The method according to claim 7, characterized in that The improved ICP precise registration algorithm comprises the steps of: optimizing the rigid transformation matrix by iteratively searching for corresponding point pairs between point clouds and minimizing the residual registration error.

9. The method according to claim 8, characterized in that Step S4 uses the Poisson surface reconstruction method, which is specifically as follows: for the actual surface Ω, any point m i (x i ,y i ,z i )∈Ω, if F(x,y,z) is the reconstructed surface equation, then the surface model can be described by the following formula: At the same time:

10. The method according to claim 1, characterized in that Step S1: The pre-processing steps include: Conditional filtering is achieved by setting filtering conditions and filtering out data that is not within the required range, and then deleting converter point cloud data that does not meet the given conditions.

Images