Automobile double door ring mold size deviation analysis method based on three-dimensional point cloud data

Abstract

This application relates to the field of point cloud registration technology, and in particular to a method for analyzing the dimensional deviation of an automotive double-door ring mold based on 3D point cloud data. The method includes: acquiring the full neighborhood of each measurement point in the automotive mold measurement point cloud; calculating the normal dispersion of the normal vector of each point within the full neighborhood; determining a screening threshold based on the normal dispersion of all measurement points; constructing an initial subset by screening points within the full neighborhood according to the screening threshold; obtaining a first structure ratio of the full neighborhood and a second structure ratio of the initial subset; locking the optimal subset based on the difference between the first and second structure ratios; obtaining the subset centroid and optimal structure ratio of the optimal subset; using the normalized optimal structure ratio as the registration weight, performing iterative registration with the subset centroid as the matching primitive to obtain the registered point cloud and outputting the deviation result of the automotive mold. The technical solution of this application can improve the accuracy of dimensional deviation analysis.

Description

Technical Field

[0001] This application relates to the field of point cloud registration technology, and in particular to a method for analyzing the dimensional deviation of automotive double door ring molds based on three-dimensional point cloud data. Background Technology

[0002] As the precision requirements of automobile manufacturing continue to increase, dimensional inspection of large and complex molds such as double door rings has become a crucial step in ensuring the accuracy of the car body. In mold production, 3D scanning technology is typically used to acquire point cloud data of the mold surface, which is then registered and compared with the nominal curved surface determined during the design phase to analyze mold wear, deformation, or machining deviations. For automotive molds containing numerous complex features, achieving high-precision automated alignment between the measured point cloud and the nominal curved surface is a pressing technical problem that needs to be solved.

[0003] Currently, the most common approach in the industry is to perform global registration based on the iterative nearest point algorithm. This method first performs coarse registration through feature matching, and then in the fine registration stage, each point in the measured point cloud is regarded as a primitive. By iteratively searching for the nearest point pair on the nominal surface, and constructing an objective function based on the least squares method, the rotation and translation matrix that minimizes the sum of the distances of all point pairs is solved, thereby achieving spatial alignment between the two.

[0004] However, on the one hand, automotive molds typically contain large flat areas and local complex feature areas. The equal weighting mechanism of the traditional iterative nearest point algorithm leads to the registration process being dominated by the data of flat areas with a large number of points, which severely dilutes the weight of high curvature areas with key trace features (such as corners), thus causing registration distortion. On the other hand, scanned point clouds are often accompanied by topological interference and noise at feature edges. The traditional iterative nearest point algorithm cannot adaptively remove these heterogeneous points, resulting in low registration accuracy at complex surfaces, which in turn leads to inaccurate results in mold size deviation analysis. Summary of the Invention

[0005] To address the technical problem of inaccurate results in mold size deviation analysis, this application provides a method for analyzing the size deviation of automotive double door ring molds based on three-dimensional point cloud data, which can improve the accuracy of mold size deviation analysis.

[0006] This application provides a method for analyzing the dimensional deviation of a double-door ring mold for automobiles based on three-dimensional point cloud data. The analysis method includes: acquiring a measurement point cloud and a nominal surface of the automobile mold; acquiring the full neighborhood of each measurement point in the measurement point cloud, calculating the normal dispersion of the normal vector of each point in the full neighborhood, and determining a screening threshold based on the normal dispersion of all measurement points; constructing an initial subset by screening points in the full neighborhood according to the screening threshold, and acquiring a first structure ratio of the full neighborhood and a second structure ratio of the initial subset, wherein both the first structure ratio and the second structure ratio are positively correlated with the degree of differentiation of the spatial configuration of the point cloud; locking the optimal subset based on the difference between the first structure ratio and the second structure ratio, and acquiring the subset centroid and the optimal structure ratio of the optimal subset; using the normalized optimal structure ratio as a registration weight, and using the subset centroid as a matching primitive, iteratively registering the measurement point cloud to the nominal surface using the registration weight to obtain a registered point cloud; calculating the normal deviation of the registered point cloud relative to the nominal surface, and outputting the deviation result of the automobile mold based on the normal deviation.

[0007] The optimal subset of each measurement point is selected, and the centroid and optimal structure ratio of each optimal subset are determined. The original measurement point is replaced by the centroid of the optimal subset as the matching primitive. The optimal structure ratio is used as the registration weight for iterative registration. This effectively solves the problem of key features being diluted by redundant data in flat areas in traditional registration, and improves the registration accuracy and deviation analysis accuracy of automobile molds at complex corners.

[0008] Preferably, the acquisition of the measurement point cloud and nominal surface of the automobile mold includes: performing a three-dimensional scan of the surface of the automobile mold to obtain an original point cloud; calculating the average distance and standard deviation of the nearest neighbor set for each original point in the original point cloud, removing outliers whose average distance is greater than the sum of the global mean and twice the standard deviation to obtain a denoised point cloud; dividing the three-dimensional space into multiple cubic voxel grids, retaining the centroid point inside each cubic voxel grid, and using all downsampled centroid points as the measurement point cloud.

[0009] Preferably, obtaining the full neighborhood of each measurement point in the measurement point cloud includes: constructing a spherical space with the measurement point as the center and a preset search radius, and forming the full neighborhood together with all points in the spherical space and the measurement point; constructing the covariance matrix of the full neighborhood, performing eigenvalue decomposition on the covariance matrix to obtain multiple eigenvalues ​​and corresponding eigenvectors; taking the eigenvector corresponding to the smallest eigenvalue as the center normal vector, and obtaining the neighborhood normal vectors of the remaining points in the full neighborhood.

[0010] Preferably, the calculation of the normal dispersion of the normal vector of each point in the full neighborhood includes: calculating the arccosine value of the angle between each of the neighborhood normal vectors and the central normal vector; calculating the average value of all the arccosine values ​​in the full neighborhood, and using the average value as the normal dispersion.

[0011] Preferably, determining the screening threshold based on the normal dispersion of all measurement points includes: sorting the normal dispersion of all measurement points in ascending order according to their numerical values; obtaining the preset percentile after ascending order, and using the preset percentile as the globally adaptive screening threshold.

[0012] By sorting the normal dispersion of each point in the entire neighborhood in ascending order and taking a preset percentile as the screening threshold, the threshold is adaptively generated by utilizing statistical distribution characteristics. This ensures that the threshold can dynamically adapt to the overall curvature characteristics of different mold surfaces, ensuring that there is sufficient screening strength in high curvature areas to remove heterogeneous points, while avoiding over-computation in flat areas.

[0013] Preferably, the step of filtering points in the full neighborhood based on the filtering threshold to construct an initial subset includes: retaining target points in the full neighborhood whose angle between the corresponding neighborhood normal vector and the center normal vector is less than the filtering threshold; in response to the number of target points being not less than three, the initial subset is composed of the measurement point and all the target points; in response to the number of target points being less than three, the full neighborhood is directly used as the initial subset.

[0014] Preferably, obtaining the first structure ratio of the full neighborhood and the second structure ratio of the initial subset respectively includes: constructing covariance matrices for the full neighborhood and the initial subset respectively and performing eigenvalue decomposition to obtain the first maxima eigenvalue, the first minima eigenvalue, and the sum of the first eigenvalues ​​of the full neighborhood, and the second maxima eigenvalue, the second minima eigenvalue, and the sum of the second eigenvalues ​​of the initial subset; dividing the difference between the first maxima eigenvalue and the first minima eigenvalue by the sum of the first eigenvalues ​​to obtain the first structure ratio; and dividing the difference between the second maxima eigenvalue and the second minima eigenvalue by the sum of the second eigenvalues ​​to obtain the second structure ratio.

[0015] Preferably, locking the optimal subset based on the difference between the first structure ratio and the second structure ratio, and obtaining the subset centroid and optimal structure ratio of the optimal subset includes: subtracting the first structure ratio from the second structure ratio to obtain a structure lift; in response to the structure lift being not less than zero, taking the initial subset as the optimal subset and the second structure ratio as the optimal structure ratio; in response to the structure lift being less than zero, taking the entire neighborhood as the optimal subset and the first structure ratio as the optimal structure ratio; and calculating the average of the three-dimensional coordinates of all points within the optimal subset to obtain the subset centroid.

[0016] The structure enhancement is obtained by comparing the structure ratio difference between the initial subset and the full neighborhood, and this triggers a backoff mechanism to ensure that the final subset participating in registration has a topological saliency no lower than that of the original neighborhood. This eliminates the risk of losing local geometric information due to overly aggressive normal selection, and guarantees that the system can lock the geometric primitive with the strongest representational ability under any complexity of surface.

[0017] Preferably, the step of using the normalized optimal structure ratio as the registration weight, and the subset centroid as the matching primitive, to iteratively register the measured point cloud to the nominal surface using the registration weight includes: dividing the optimal structure ratio corresponding to each measurement point by the sum of the optimal structure ratios of all measurement points to obtain the registration weight; in each iteration, searching on the nominal surface for the target surface point closest to the subset centroid, and constructing a weighted objective function with the registration weight as coefficient and the distance between the subset centroid and the target surface point as primitive; solving the weighted objective function to obtain the rotation matrix and translation vector, and updating the measured point cloud according to the rotation matrix and the translation vector until the iterative convergence condition is met.

[0018] The optimal structure ratio is normalized and used as the registration weight. A weighted objective function is constructed using the centroid of the subset as the matching primitive. This allows the registration process to be dominated by point cloud blocks with significant geometric features, which greatly suppresses random noise interference during the scanning process. This enables the measured point cloud to be aligned to the nominal surface more stably and accurately, providing a high-precision spatial reference for subsequent normal deviation calculation.

[0019] Preferably, calculating the normal deviation of the registration point cloud relative to the nominal surface and outputting the deviation result of the automobile mold based on the normal deviation includes: searching the nominal surface for the surface reference point that is closest to each registration point in the registration point cloud; calculating the signed normal deviation along the nominal normal vector direction of the surface reference point; comparing the normal deviation with a preset tolerance zone to determine whether the registration point is in an out-of-tolerance state; and outputting the deviation result based on the out-of-tolerance statistical data of each region.

[0020] In the registration point cloud, the deviation value and the deviation direction can be obtained not only by using the signed normal deviation.

[0021] The technical solution of this application has the following beneficial technical effects: This application quantifies the normal dispersion of the entire neighborhood of the point cloud and adaptively determines the screening threshold. Combined with a backoff mechanism based on the structure ratio increase, it accurately eliminates topologically heterogeneous points in complex surface regions, thus eliminating the impact of these points on matching accuracy. Furthermore, by utilizing the normalized optimal structure ratio to assign higher registration weights to high-curvature feature regions, it effectively solves the problem of key features being diluted by redundant data in flat areas in traditional registration, significantly improving the registration accuracy of automotive molds at complex corners and the reliability and accuracy of deviation analysis. Attached Figure Description

[0022] Figure 1 This is a flowchart of a method for analyzing the dimensional deviation of a car double door ring mold based on three-dimensional point cloud data, according to an embodiment of this application.

[0023] Figure 2 This is a schematic diagram of the nominal curved surface of an automotive double door ring mold according to an embodiment of this application.

[0024] Figure 3 This is a chromatogram of the dimensional deviation of the automotive double door ring mold according to an embodiment of this application. Detailed Implementation

[0025] In the automobile manufacturing process, it is necessary to inspect the geometric dimensions of the body structural components. This application provides a method for analyzing the dimensional deviation of automobile double door ring molds based on three-dimensional point cloud data. It is applicable to the quality control of mass production of automobile double door ring hot pressing molds, focusing on ensuring the registration accuracy of complex areas such as the B-pillar corner area and improving the accuracy of dimensional deviation analysis. Figure 1 This is a flowchart illustrating a method for analyzing dimensional deviations in automotive double-door ring molds based on 3D point cloud data, according to an embodiment of this application. Figure 1 As shown, the method for analyzing the dimensional deviation of the automotive double-door ring mold based on three-dimensional point cloud data includes steps S101 to S106, which are described in detail below.

[0026] S101, obtain the measurement point cloud and nominal surface of the automobile mold.

[0027] In one embodiment, to achieve high-precision dimensional deviation analysis, geometric data needs to be extracted from the physical mold and aligned with design standards. During this process, the original errors and redundant data introduced by hardware scanning must be eliminated. Obtaining the measurement point cloud and nominal surface of the automotive mold includes: performing a 3D scan of the automotive mold's surface to obtain an original point cloud; calculating the average distance and standard deviation of the nearest neighbor set for each original point in the original point cloud, removing outliers whose average distance is greater than the sum of the global mean and twice the standard deviation to obtain a denoised point cloud; dividing the 3D space into multiple cubic voxel grids, retaining the centroid points within each cubic voxel grid, and using all downsampled centroid points as the measurement point cloud.

[0028] The calculation of the average distance and standard deviation of the nearest neighbor set to remove outliers employs a well-known statistical outlier removal method, primarily used to clean up isolated scanning noise. During downsampling, the side length of the cube voxel grid is preferably 0.5 mm, specifically set to one-quarter of the minimum tolerance band width in the automotive mold drawing. By extracting the geometric centroids within each cube voxel grid, not only is the number of points compressed to an acceptable scale for subsequent registration calculations, but more importantly, the spatial distribution of the point cloud becomes more uniform.

[0029] The nominal surface is a 3D model corresponding to a standard automotive mold without dimensional deviations. Please refer to [link / reference]. Figure 2 This is a schematic diagram of the nominal curved surface of an automotive double door ring mold according to an embodiment of this application.

[0030] In this way, density deviation in overlapping areas of the scanning path and scanning noise from hardware scanning are eliminated, resulting in a spatially uniform measurement point cloud without outliers, providing a data foundation for subsequent accurate estimation of normal vectors and calculation of spatial topological features of the point cloud.

[0031] S102, obtain the full neighborhood of each measurement point in the measurement point cloud, calculate the normal dispersion of the normal vector of each point in the full neighborhood, and determine the screening threshold based on the normal dispersion of all measurement points.

[0032] In one embodiment, after acquiring the measurement point cloud, in order to accurately identify the high curvature feature area and the low curvature flat area in the mold surface, it is necessary to quantify the local topological heterogeneity of each point in the measurement point cloud, thereby providing a basis for subsequent topological consistency screening.

[0033] Obtaining the full neighborhood for each measurement point in the measurement point cloud includes: constructing a spherical space with the measurement point as the center and a preset search radius; forming the full neighborhood by combining all points within the spherical space and the measurement point; constructing the covariance matrix of the full neighborhood; performing eigenvalue decomposition on the covariance matrix to obtain multiple eigenvalues ​​and corresponding eigenvectors; using the eigenvector corresponding to the smallest eigenvalue as the center normal vector; and obtaining the neighborhood normal vectors of the remaining points within the full neighborhood in the same manner.

[0034] It should be noted that the preset search radius is determined based on the minimum fillet radius in the mold drawing, preferably five millimeters. This value is 1.5 times the minimum fillet radius, ensuring sufficient geometric information is captured in areas of abrupt feature changes. The construction of the covariance matrix and the eigenvalue decomposition to obtain the central normal vector employ the well-known method of principal component analysis for normal vector estimation, which will not be elaborated upon here. The central normal vector is directly estimated from the full neighborhood without any filtering, serving as a priori benchmark for subsequent calculations and normal vector consistency filtering.

[0035] After obtaining the center normal vector and the neighborhood normal vectors, it is necessary to further quantify the divergence of the normal vectors within the entire neighborhood. Calculating the normal vector dispersion of each point in the entire neighborhood includes: calculating the inverse cosine of the angle between each neighborhood normal vector and the center normal vector; calculating the average value of all the inverse cosine values ​​in the entire neighborhood, and using the average value as the normal vector dispersion.

[0036] Normal dispersion of measurement points Satisfying the relation:

[0037] In the formula, The normal dispersion of the measurement point, in radians; It is the set of points within the entire neighborhood; The total number of points in the entire neighborhood; is the neighborhood normal vector of the remaining points in the full neighborhood; The center normal vector of the measurement point; It is the arccosine of the angle between the neighborhood normal vector and the center normal vector, and its value ranges from zero to... between.

[0038] Understandably, a larger normal dispersion indicates a more divergent direction of the normal vector within the entire neighborhood, resulting in more topological heterogeneous points. For example, at the corner of a car's B-pillar, one side is a high-curvature arc surface, and the other side is a gently sloping surface. The angle between the normal vectors spanning the entire neighborhood on both sides is extremely large, and its normal dispersion is significantly higher than that of a flat area. Normal dispersion effectively enables the quantitative differentiation of complex geometric structures on the mold surface.

[0039] For regions with different curvature characteristics, a unified threshold needs to be determined from a global perspective to guide subsequent removal of heterogeneous points and preservation of effective topology. Determining the screening threshold based on the normal dispersion of all measurement points includes: sorting the normal dispersion of all measurement points in ascending order of numerical value; obtaining the preset percentile after ascending sorting; and using the preset percentile as the globally adaptive screening threshold.

[0040] The preset percentile can be set to the th percentile of the normal dispersion of all measurement points. Percentiles, select the first Percentiles replace median or higher quantiles to ensure the threshold is higher than the global threshold. The dispersion level of the measurement points. This allows more points in high curvature feature regions to meet the screening criteria for meaningful screening, avoiding large-scale regression due to low thresholds; at the same time, it prevents low curvature flat regions from triggering screening due to overly loose thresholds, thus introducing unnecessary computation.

[0041] In this way, each measurement point in the measurement point cloud is given a precise normal discreteness, and a screening threshold that adapts to the overall curvature distribution of the current mold is generated adaptively. This provides a reliable quantitative criterion for the subsequent accurate stripping of topological heterogeneous points and preservation of real feature structures, effectively avoiding feature confusion in subsequent steps when registering complex surfaces.

[0042] S103, the points in the entire neighborhood are filtered according to the filtering threshold to construct an initial subset, and the first structure ratio of the entire neighborhood and the second structure ratio of the initial subset are obtained respectively. The first structure ratio and the second structure ratio are both positively correlated with the degree of differentiation of the point cloud spatial configuration.

[0043] In one embodiment, after obtaining the screening threshold, topologically heterogeneous points within the entire neighborhood need to be removed to construct a subset that better represents the local true features. The effectiveness of the screening is verified by quantitatively comparing the structural features before and after screening. Screening points within the entire neighborhood to construct an initial subset based on the screening threshold includes: retaining target points in the entire neighborhood whose angle between the corresponding neighborhood normal vector and the center normal vector is less than the screening threshold; if the number of target points is not less than three, the initial subset is composed of the measured point and all the target points; if the number of target points is less than three, the entire neighborhood is directly used as the initial subset to ensure sufficient rank for subsequent matrix decomposition.

[0044] It should be noted that in flat surface areas such as the large surface of the car door sill beam, almost all points in the full neighborhood have the same normal, and the initial subset is almost equal to the full neighborhood. However, in high curvature areas such as the transition area at the B-pillar corner, the initial subset only retains points from the same side of the topology, and heterogeneous points are eliminated, achieving the goal of representing the topology as few as possible. Setting the number of points to be no less than three is to ensure that the three-dimensional spatial rank of the subsequent covariance matrix is ​​sufficient, ensuring that the subsequent covariance matrix calculation can be performed on all measurement points.

[0045] After obtaining the initial subset, the process of acquiring the first structure ratio of the full neighborhood and the second structure ratio of the initial subset includes: constructing covariance matrices for the full neighborhood and the initial subset respectively, and performing eigenvalue decomposition to obtain the first maxima eigenvalue, the first minima eigenvalue, and the sum of the first eigenvalues ​​of the full neighborhood, and the second maxima eigenvalue, the second minima eigenvalue, and the sum of the second eigenvalues ​​of the initial subset; dividing the difference between the first maxima eigenvalue and the first minima eigenvalue by the sum of the first eigenvalues ​​to obtain the first structure ratio; and dividing the difference between the second maxima eigenvalue and the second minima eigenvalue by the sum of the second eigenvalues ​​to obtain the second structure ratio.

[0046] Principal component analysis was used for covariance matrix construction and eigenvalue decomposition, which will not be elaborated here. To quantify the degree of differentiation in the spatial configuration of the subset point cloud and compare the topological saliency before and after selection, the first structure ratio... Second structure ratio Satisfying the relation:

[0047]

[0048] In the formula, The first structure ratio is dimensionless; , , These are the first maximum eigenvalue, the middle eigenvalue, and the first minimum eigenvalue of the entire neighborhood, respectively, all in square millimeters, and their sum constitutes the total first eigenvalue. The second structure ratio is dimensionless. , , These are the second maximum eigenvalue, intermediate eigenvalue, and second minimum eigenvalue of the initial subset, respectively, all in square millimeters, and their sum constitutes the total second eigenvalue. When the total second eigenvalue or the total first eigenvalue is 0, the corresponding structure ratio is directly assigned to 0.

[0049] Understandably, the structure ratio accurately measures the degree of anisotropy in the spatial distribution of a point cloud. The larger the values ​​of the first and second structure ratios, the higher the degree of differentiation in the spatial configuration of the corresponding point cloud. The point cloud tends to be distributed on a single surface, meaning the proportion of minimal eigenvalues ​​is smaller, its topological saliency is stronger, and it is less likely to be confused with points from other regions during subsequent matching. This dimensionless structure ratio calculation objectively quantifies the difference in geometric and topological expressive power between the full neighborhood and the initial subset.

[0050] In this way, heterogeneous interference points that disrupt local geometric consistency were eliminated through adaptive screening, and the spatial differentiation features of the point cloud subsets before and after screening were quantified. Geometric descriptions that can reflect the true local morphology of the mold were extracted, providing a basis for comparative verification to avoid information loss caused by aggressive screening in the future.

[0051] S104, based on the difference between the first structure ratio and the second structure ratio, the optimal subset is locked, and the subset centroid and the optimal structure ratio of the optimal subset are obtained.

[0052] In one embodiment, after obtaining the first structure ratio of the entire neighborhood and the second structure ratio of the initial subset, it is necessary to compare the two to verify whether the previous normal consistency screening has truly improved the saliency of local features, and then determine the final set of point clouds to participate in iterative registration.

[0053] The process of locking the optimal subset based on the difference between the first structure ratio and the second structure ratio, and obtaining the subset centroid and optimal structure ratio of the optimal subset, includes: subtracting the first structure ratio from the second structure ratio to obtain a structure lift; in response to the structure lift being not less than zero, using the initial subset as the optimal subset and the second structure ratio as the optimal structure ratio; in response to the structure lift being less than zero, using the entire neighborhood as the optimal subset and the first structure ratio as the optimal structure ratio to eliminate information loss caused by over-screening; and calculating the average of the three-dimensional coordinates of all points within the optimal subset to obtain the subset centroid.

[0054] Among them, structural improvement Satisfying the relation:

[0055] In the formula, The second structural ratio; This is the first structural ratio.

[0056] Understandably, if the structural improvement is not less than zero, it indicates that after removing heterogeneous points, the spatial structure differentiation of the point cloud within the initial subset has improved, and the proportion of the smallest eigenvalue has decreased. This suggests that the initial subset is more likely to lie on a single surface, and the topological saliency has indeed improved, proving that the screening operation is effective. If the structural improvement is less than zero, it indicates that this round of screening was too aggressive, resulting in the loss of structural information due to the removal of valid points. In this case, the screening is invalid, and a backoff mechanism should be triggered, using the entire neighborhood as the optimal subset and the first structure ratio as the optimal structure ratio. This self-verification mechanism ensures that the subset ultimately used for each measurement point is no worse than the entire neighborhood, eliminating the risk of quality degradation introduced by the screening operation.

[0057] After locking the optimal subset, the centroid of the subset. Satisfying the relation:

[0058] In the formula, The centroid of the optimal subset is given in millimeters. This is the optimal subset that is ultimately locked. The number of points within the optimal subset; These are the three-dimensional coordinates of each point within the optimal subset.

[0059] It should be noted that by using the centroid of the optimal subset instead of a single measurement point as the primitive for subsequent matching calculations, the centroid of the subset has a good suppression effect on single-point scanning noise within the subset, based on the variance characteristics of the mean estimation.

[0060] Thus, through the self-verification and backoff mechanism of structural features, the optimal subset with the strongest representation ability and its centroid are intelligently locked for each measurement point, fundamentally eliminating the hidden danger of feature loss caused by over-selection, and providing the optimal geometric entity for subsequent high-precision and noise-resistant weighted iterative registration.

[0061] S105, using the normalized optimal structure ratio as the registration weight, and the centroid of the subset as the matching primitive, the measured point cloud is iteratively registered to the nominal surface using the registration weight to obtain the registered point cloud.

[0062] In one embodiment, after obtaining the optimal subset corresponding to each measurement point and its centroid and structure ratio, it is necessary to use these topologically significant feature points to guide the accurate alignment of the overall point cloud, so as to avoid diluting the matching weight of the high curvature key area by a large number of similar points in the low curvature flat area.

[0063] Using the normalized optimal structure ratio as the registration weight and the centroid of the subset as the matching primitive, the measured point cloud is iteratively registered to the nominal surface using the registration weight to obtain the registered point cloud. This process includes: dividing the optimal structure ratio corresponding to each measurement point by the sum of the optimal structure ratios of all measurement points to obtain the registration weight; in each iteration, searching the nominal surface for the target surface point closest to the centroid of the subset, and constructing a weighted objective function with the registration weight as the coefficient and the distance between the centroid of the subset and the target surface point as the primitive; obtaining the rotation matrix and translation vector by solving the weighted objective function using singular value decomposition, and updating the measured point cloud according to the rotation matrix and the translation vector until the iterative convergence condition is met, thus obtaining the registered point cloud aligned to the nominal coordinate system.

[0064] In order to ensure that the weight allocation during iterative registration can truly reflect the geometric saliency of different regions, the registration weights are... Satisfying the relation:

[0065] In the formula, The registration weights are dimensionless, and the sum of all registration weights is one. The optimal structure ratio for the measurement point; This represents the total number of measurement points in the point cloud. It is the sum of the optimal structure ratios for all measurement points.

[0066] Understandably, points with a larger optimal structure ratio automatically receive higher registration weights. These points are usually located in key areas with high curvature, such as the corner of a car's B-pillar. In contrast, points in flat areas with low curvature have a relatively smaller optimal structure ratio, and their registration weights are reduced accordingly. This fundamentally solves the problem of curvature weights being diluted by the number of regional points in traditional registration schemes, allowing the registration process to be dominated by key points with significant topological features.

[0067] Weighted objective function Satisfying the relation:

[0068] In the formula, The weighted objective function; For registration weights; The centroid of the optimal subset is given in millimeters. The target surface point found on the nominal surface that is closest to the centroid of the subset, in millimeters; Let be the rotation matrix to be solved; Let be the translation vector to be solved, in millimeters.

[0069] It should be noted that, in each iteration, the composition of the optimal subset remains fixed, and only the transformation matrix of the current iteration is applied to each point to recalculate the centroid of the transformed subset, and a new target surface point is searched in the nominal surface.

[0070] During the iterative update process, it is necessary to determine whether the iterative convergence condition is met to stop the calculation. Preferably, the iterative convergence condition is that the change in the rotation matrix between two adjacent iterations is less than a preset change, which is 0.01, or the number of iterations reaches a preset upper limit, which is 50 iterations. When this convergence condition is met, the final rotation matrix and translation vector are applied to all measurement points to obtain the registered point cloud.

[0071] In this way, a larger registration weight is used to dominate the registration direction in the high curvature feature region, and a larger optimal subset is used to leverage the centroid noise reduction advantage in the low curvature flat region. This enables the measured point cloud to be aligned to the coordinate system of the nominal surface with high precision and robustness, eliminating the interference of scanning noise and redundant data in the flat region on the overall registration accuracy.

[0072] S106, calculate the normal deviation of the registered point cloud relative to the nominal surface, and output the deviation result of the automobile mold based on the normal deviation.

[0073] In one embodiment, after iterative registration, the measured point cloud has been precisely aligned to the nominal coordinate system. At this point, it is necessary to accurately quantify the actual manufacturing error of the mold surface to determine whether the mold dimensions are acceptable. Calculating the normal deviation of the registered point cloud relative to the nominal surface and outputting the deviation result of the automotive mold based on the normal deviation includes: searching the nominal surface for the surface reference point closest to each registered point in the registered point cloud; calculating the signed normal deviation along the nominal normal vector direction of the surface reference point; comparing the normal deviation with a preset tolerance zone to determine whether the registered point is in an out-of-tolerance state; and outputting the deviation result based on the out-of-tolerance statistics for each region.

[0074] Wherein, the normal deviation Satisfying the relation:

[0075] In the formula, The signified normal deviation is expressed in millimeters. To register each registration point in the point cloud; This is the closest surface reference point found on the nominal surface; Let be the nominal normal vector of the surface reference point.

[0076] Understandably, when the normal deviation is greater than zero, it indicates that the registration point is outside the nominal surface, meaning the mold surface is too large or there is material accumulation; when the normal deviation is less than zero, it indicates that it is inside, meaning the mold surface is too small or there is wear and depression.

[0077] It should be noted that when comparing the normal deviation with the preset tolerance zone, the preset tolerance zone includes the minimum tolerance and the maximum tolerance. Preferably, taking the corner area of ​​column B as an example, its preset tolerance zone is preferably -0.1 mm to 0.1 mm. When the normal deviation is between the minimum tolerance and the maximum tolerance, the registration point is judged to be in a qualified state; when the normal deviation is less than the minimum tolerance or greater than the maximum tolerance, the registration point is judged to be in an out-of-tolerance state. When outputting the deviation results, all normal deviations are mapped to warm and cold chromatograms according to their numerical values, and superimposed on the three-dimensional surface of the mold to generate a dimensional deviation chromatogram. Please refer to [link to relevant documentation]. Figure 3 The dimensional deviation chromatogram is a dimensional deviation chromatogram corresponding to the automotive double door ring mold according to the embodiments of this application. From the dimensional deviation chromatogram, it is possible to clearly determine which positions belong to the dimensionally qualified area, which positions belong to the wear area, and which positions belong to the accumulation area.

[0078] In this way, the transition from point cloud registration to intuitive quantification of three-dimensional spatial errors enables high-precision quality inspection covering the entire mold surface, improving the efficiency of mass production quality control for automotive double door ring hot pressing molds.

[0079] It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this patent application shall be determined by the appended claims.

Claims

1. A method for analyzing dimensional deviation of a double door ring mold for a vehicle based on three-dimensional point cloud data, characterized by, The analytical method includes: The method for acquiring the measurement point cloud and nominal surface of an automotive mold includes: performing a three-dimensional scan of the surface of the automotive mold to obtain an original point cloud; calculating the average distance and standard deviation of the nearest neighbor set for each original point in the original point cloud, removing outliers whose average distance is greater than the sum of the global mean and twice the standard deviation to obtain a denoised point cloud; dividing the three-dimensional space into multiple cubic voxel grids, retaining the centroid point inside each cubic voxel grid, and using all downsampled centroid points as the measurement point cloud; Obtaining the full neighborhood of each measurement point in the measurement point cloud includes: constructing a spherical space with the measurement point as the center and a preset search radius; forming the full neighborhood by combining all points in the spherical space and the measurement point; constructing the covariance matrix of the full neighborhood; performing eigenvalue decomposition on the covariance matrix to obtain multiple eigenvalues ​​and corresponding eigenvectors; using the eigenvector corresponding to the smallest eigenvalue as the center normal vector; and obtaining the neighborhood normal vectors of the remaining points in the full neighborhood. Calculate the normal dispersion of the normal vector of each point in the full neighborhood, and determine the screening threshold based on the normal dispersion of all measurement points; The points in the entire neighborhood are filtered according to the filtering threshold to construct an initial subset. The first structure ratio of the entire neighborhood and the second structure ratio of the initial subset are obtained respectively. Both the first structure ratio and the second structure ratio are positively correlated with the degree of differentiation of the point cloud spatial configuration. The optimal subset is determined based on the difference between the first structure ratio and the second structure ratio, and the subset centroid and optimal structure ratio of the optimal subset are obtained. The normalized optimal structure ratio is used as the registration weight, and the centroid of the subset is used as the matching primitive. The measured point cloud is iteratively registered to the nominal surface using the registration weight to obtain the registered point cloud. The normal deviation of the registered point cloud relative to the nominal surface is calculated, and the deviation result of the automobile mold is output based on the normal deviation.

2. The method for analyzing dimensional deviation of a double door ring mold for a vehicle based on three-dimensional point cloud data according to claim 1, characterized in that, The calculation of the normal dispersion of the normal vector of each point in the full neighborhood includes: calculating the inverse cosine value of the angle between each neighborhood normal vector and the center normal vector; Calculate the average value of all the arccosine values ​​in the full neighborhood, and use the average value as the normal dispersion.

3. The method for analyzing the dimensional deviation of automotive double-door ring molds based on three-dimensional point cloud data according to claim 1, characterized in that, The step of determining the screening threshold based on the normal dispersion of all measurement points includes: sorting the normal dispersion of all measurement points in ascending order according to their numerical values; obtaining the preset percentile after sorting in ascending order; and using the preset percentile as the globally adaptive screening threshold.

4. The method for analyzing the dimensional deviation of automotive double-door ring molds based on three-dimensional point cloud data according to claim 1, characterized in that, The step of filtering points within the entire neighborhood based on the filtering threshold to construct an initial subset includes: In the full neighborhood, target points whose angle between the corresponding neighborhood normal vector and the center normal vector is less than the filtering threshold are retained; in response to the number of target points being not less than three, the initial subset is composed of the measurement point and all the target points; in response to the number of target points being less than three, the full neighborhood is directly used as the initial subset.

5. The method for analyzing the dimensional deviation of automotive double-door ring molds based on three-dimensional point cloud data according to claim 1, characterized in that, The steps of obtaining the first structure ratio of the entire neighborhood and the second structure ratio of the initial subset respectively include: The covariance matrices are constructed for the full neighborhood and the initial subset respectively, and eigenvalue decomposition is performed to obtain the first maxima eigenvalue, the first minima eigenvalue, and the sum of the first eigenvalues ​​of the full neighborhood, and the second maxima eigenvalue, the second minima eigenvalue, and the sum of the second eigenvalues ​​of the initial subset. The difference between the first maximum eigenvalue and the first minimum eigenvalue is divided by the sum of the first eigenvalues ​​to obtain the first structure ratio; the difference between the second maximum eigenvalue and the second minimum eigenvalue is divided by the sum of the second eigenvalues ​​to obtain the second structure ratio.

6. The method for analyzing the dimensional deviation of a double-door ring mold for automobiles based on three-dimensional point cloud data according to claim 1, characterized in that, Locating the optimal subset based on the difference between the first structure ratio and the second structure ratio, and obtaining the subset centroid and optimal structure ratio of the optimal subset, includes: Subtracting the first structure ratio from the second structure ratio yields the structural improvement. In response to the structure lift being not less than zero, the initial subset is taken as the optimal subset, and the second structure ratio is taken as the optimal structure ratio; in response to the structure lift being less than zero, the entire neighborhood is taken as the optimal subset, and the first structure ratio is taken as the optimal structure ratio; the average value of the three-dimensional coordinates of all points in the optimal subset is calculated to obtain the centroid of the subset.

7. The method for analyzing the dimensional deviation of automotive double-door ring molds based on three-dimensional point cloud data according to claim 1, characterized in that, The step of using the normalized optimal structure ratio as the registration weight, and using the centroid of the subset as the matching primitive, to iteratively register the measured point cloud to the nominal surface using the registration weight includes: The optimal structure ratio corresponding to each measurement point is divided by the sum of the optimal structure ratios of all measurement points to obtain the registration weight. In each iteration, the target surface point closest to the centroid of the subset is searched on the nominal surface, and a weighted objective function is constructed with the registration weight as coefficients and the distance between the centroid of the subset and the target surface point as primitives. The rotation matrix and translation vector are obtained by solving the weighted objective function, and the measurement point cloud is updated according to the rotation matrix and the translation vector until the iteration convergence condition is met.

8. The method for analyzing the dimensional deviation of automotive double-door ring molds based on three-dimensional point cloud data according to claim 1, characterized in that, The calculation of the normal deviation of the registered point cloud relative to the nominal surface, and the output of the deviation result of the automobile mold based on the normal deviation, includes: Search the nominal surface for the surface reference point that is closest to each registration point in the registration point cloud, and calculate the signed normal deviation along the nominal normal vector direction of the surface reference point; compare the normal deviation with the preset tolerance zone to determine whether the registration point is in an out-of-tolerance state, and output the deviation result based on the out-of-tolerance statistical data of each region.

Images