A 3D Vision Inspection Method for Motorcycle Tires

Abstract

This invention belongs to the field of 3D modeling technology, and particularly relates to a 3D visual inspection method for motorcycle tires. The method involves fixing the tire and scanning it through multi-view point clouds, then using an improved SAC-IA algorithm for registration and fusion to construct a 3D point cloud model. Based on tire geometric priors, an adaptive voxel mesh filter is designed for noise reduction. After smoothing with an improved bilateral filter, a standardized point cloud model is obtained by unifying the coordinate system through principal component analysis and least squares fitting. Four core parameters—thickness uniformity, sidewall curvature, etc.—are extracted and input into a hybrid dynamic weighted scoring network to output a quality score. This invention meets the high-precision and high-efficiency quality inspection requirements of motorcycle tires in industrial environments.

Description

Technical Field

[0001] This invention belongs to the field of 3D modeling technology, and in particular relates to a 3D visual inspection method for motorcycle tires. Background Technology

[0002] With the increasing demands for product quality in the motorcycle industry, traditional tire inspection methods are no longer sufficient to meet the requirements of high-precision and high-efficiency quality inspection. There is an urgent need to develop non-contact, intelligent inspection technologies to achieve accurate extraction and objective evaluation of key tire quality parameters. Current motorcycle tire inspection methods mostly employ contact measurement or basic visual inspection, which have significant drawbacks: contact measurement easily causes wear on the tire surface, has low inspection efficiency, and cannot achieve omnidirectional inspection; traditional visual inspection is limited by two-dimensional image information, making it difficult to accurately characterize the three-dimensional geometric features of the tire. Even some non-contact point cloud inspection technologies suffer from insufficient multi-view point cloud registration accuracy, poor noise removal, and inconsistent parameter extraction benchmarks, resulting in low reliability and weak versatility of inspection results, making it difficult to meet the practical application requirements of large-scale industrial quality inspection. Summary of the Invention

[0003] In view of the technical problems existing in the background art, the present invention proposes a three-dimensional vision inspection method for motorcycle tires.

[0004] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0005] S1. Fix the motorcycle tire to be inspected onto the inspection platform, perform an all-round scan using a multi-view point cloud scanning device, and use the improved SAC-IA algorithm to accurately register and fuse the multi-view point cloud data to construct a three-dimensional point cloud model.

[0006] S2. The constructed 3D point cloud model is processed. An adaptive voxel grid filtering algorithm is designed based on the prior knowledge of the tire geometry to intelligently identify and remove outlier noise points. An improved bilateral filtering method is used to smooth the point cloud. The direction of the tire's main axis is determined by principal component analysis and least squares fitting. The point cloud coordinate system is aligned to the standard posture to obtain a standardized point cloud model.

[0007] S3. For the standardized point cloud model, extract parameters such as uniform thickness of motorcycle tire, sidewall profile curvature, bead roundness error and tire radial runout.

[0008] S4. Input the extracted parameters into the preset scoring model to obtain the output motorcycle tire quality score.

[0009] Preferably, the specific implementation of step S1, which uses the improved SAC-IA algorithm to accurately register and fuse multi-view point cloud data to construct a 3D point cloud model, includes:

[0010] S11. First, obtain the number of original point cloud data collected from each viewpoint. For each point, calculate the local curvature of the 3D point cloud, the feature vector of the fast point feature histogram, and the normal vector.

[0011] S12. Sample the original point cloud data, generating a pair of nonlinear collinear point sets for each sampling. ,in, For the source viewpoint set, For the target viewpoint set, the non-collinearity determination condition is that the area of ​​the triangle formed by the three points is greater than a set threshold;

[0012] S13. Apply probability weighting to the samples, defining the sampling weight as follows: It is obtained by fusing local curvature, fast point feature histogram, and normal vector. The calculation method is as follows: ,in, These are the weighting coefficients, and This is used to weigh the contribution of the three factors to the sampling weight. Let be the local curvature at the k-th point. For the global maximum curvature, For fast point feature histogram feature vectors, Let be the magnitude of the feature vector of the fast point feature histogram of the k-th point. This represents the maximum magnitude of the feature vector of the global maximum fast point feature histogram. Let be the normal vector of the k-th point. Let be the mean of the normal vectors of all points in the neighborhood of the k-th point. It is the absolute value of the dot product of the normal vector of the k-th point and the mean of the normal vectors of the neighborhood, reflecting the consistency of the directions of the two vectors;

[0013] S14. Sampling terminates when the cumulative number of samples reaches M, resulting in a set of point pairs; each point pair in the set... Construct a multi-constraint transformation matrix solution model to solve the source point set. to target point set initial rotation matrix Translation vector It satisfies the following formula:

[0014] ,

[0015] in, Let the centroid of the source point set be... Let the centroid of the target point set be... Let be the covariance matrix of the source point set and the target point set. Singular value decomposition is performed on the covariance matrix H, resulting in three matrices: U (left singular matrix). Singular value matrix and The transpose of a right singular matrix;

[0016] S15. Calculate the pairwise distances within the source point set. The distance between corresponding points after transformation using the obtained transformation matrix If the absolute value of the distance difference between the two is greater than the set distance threshold, the transformation matrix is ​​discarded and the filtered valid transformation matrix is ​​retained.

[0017] S16. Calculate all points from the source viewpoint for each valid transformation matrix. Transformed reprojection error Thresholds are adaptively adjusted based on local density of the point cloud: ,in, Based on the threshold, For dynamic thresholds, Let L be the local density of the k-th point, and L be the total number of source point clouds;

[0018] S17. Calculate the consistency scoring function, combining the internal point ratio and constraint satisfaction: ,in, To meet the reprojection error The number of interior points, To constrain the weights, To satisfy the distance constraint, To determine the number of points participating in the distance constraint verification, the transformation matrix that maximizes the consistency scoring function is selected as the final transformation matrix for solving the problem. A complete 3D point cloud model is then constructed based on this transformation matrix.

[0019] Preferably, in step S2, the constructed 3D point cloud model is processed, and an adaptive voxel mesh filtering algorithm is designed based on prior knowledge of the tire geometry to intelligently identify and remove outlier noise points. The specific implementation is as follows:

[0020] S211. Based on the design specifications of motorcycle tires and the initial contour features of the 3D point cloud model, extract the core geometric prior parameters of the tire, including the tire's nominal outer diameter D, nominal inner diameter d, tread width W, and tire rotation center O. Establish a local cylindrical coordinate system with the tire rotation center O as the origin. Where r is the radial distance, z is the circumferential angle, and z is the axial distance;

[0021] S212. Perform radial mesh generation, dividing the radial range according to the tire's nominal inner diameter d and nominal outer diameter D. Divided into mg sub-intervals, each sub-interval has a corresponding voxel radial side length. satisfy: Where jm are the indices of mg sub-intervals, This is the attenuation coefficient, and the circumferential angle range is also considered. The area is uniformly divided into ng circumferential sub-intervals, and the circumferential angle of the voxel corresponding to each sub-interval is . Additionally, based on the tread width W, the axial range is... Divide into p axial sub-intervals, and the axial side length of the voxel corresponding to each sub-interval is... ;

[0022] S213. Convert all point cloud data in the 3D point cloud model from the Cartesian coordinate system to... The coordinates are then categorized into the corresponding adaptive voxel grids to obtain the number of point clouds within each voxel grid. For each adaptive voxel grid, if the number of point clouds within that grid... If the value is less than the set threshold, all point clouds within that grid are identified as outlier noise points and are removed.

[0023] Preferably, in step S2, after smoothing the point cloud using an improved bilateral filtering method, the tire axis direction is determined by principal component analysis and least squares fitting, and the point cloud coordinate system is aligned to the standard orientation to obtain the standardized point cloud model. The specific implementation of this method is as follows:

[0024] S221. Based on the geometric features of a tire's annular rotation, an improved bilateral filtering algorithm integrating radial distance weights and local curvature weights is designed to smooth the point cloud while preserving the tire's contour edge features. The target point coordinates after filtering are... Satisfying the formula: ,in, Let P be the set of neighboring points. Normalization factor Geometric distance weights For local curvature constraint weights;

[0025] S222. Using the improved bilaterally filtered point cloud as input, calculate the mean point of all point clouds. Convert each point cloud coordinate into a centralized coordinate system. Construct the covariance matrix of the centralized point cloud, solve for the eigenvalues ​​and corresponding eigenvectors of the covariance matrix, and take the eigenvector with the largest eigenvalue as the initial principal axis direction;

[0026] S223. Establish a Cartesian standard coordinate system with the tire rotation center O as the origin. The z-axis coincides with the direction of the principal axis. The x-axis lies in the radial plane of the tire and is taken as the radial direction at the midpoint of the tire tread width. The y-axis is determined based on the right-hand rule to ensure the consistency of the right-hand screw of the coordinate system.

[0027] S224. Transform the filtered point cloud coordinates to the standard coordinate system to complete the construction of the standardized point cloud model.

[0028] Preferably, the geometric distance weight in step S221 ,in, These are the radial distances between the target point and its neighboring points, respectively, taken from cylindrical coordinates. The standard deviation of radial distance. The standard deviation of spatial distance is taken as 1.5 times the average point spacing of the point cloud; local curvature constraint weight. ,in, Let P be the local curvature. This is the curvature threshold.

[0029] Preferably, in step S3, a standardized point cloud unified coordinate reference is used to ensure the consistency of the normal direction calculation, which is characterized by calculating the standard deviation of the normal distance from each point in the tread area to the inner wall of the tire; the bead roundness error is calculated by projecting the inner edge point cloud of the bead onto a reference plane perpendicular to the Z-axis, fitting elliptic parameters, and calculating the ellipticity deviation and the maximum radial deviation; the sidewall profile curvature is characterized by uniformly extracting multiple profile lines along the circumference of the sidewall based on the standardized coordinate system Z-axis, and calculating the average absolute value of their second derivatives; the tire radial runout is obtained by simulating the rotation state based on the Z-axis of the standardized model and measuring the displacement amplitude of the tread centerline in the radial direction.

[0030] Preferably, the scoring model in step S4 employs a hybrid dynamic weight scoring network, which includes the following processing units:

[0031] The parameter standardization unit is used to normalize each extracted parameter to the [0,1] interval;

[0032] The feature nonlinear transformation unit is used to perform nonlinear mapping of each parameter through a learnable piecewise function;

[0033] The dynamic weight allocation unit is used to automatically adjust the weights of each parameter according to tire specifications and usage scenarios.

[0034] The comprehensive scoring calculation unit is used to integrate the converted values ​​of various parameters and their corresponding weights to output a tire quality score.

[0035] Preferably, the feature nonlinear transformation unit, used to perform nonlinear mapping on each parameter through a learnable piecewise function, specifically includes: dividing each extracted parameter into multiple segmented intervals according to the parameter value range; for the c-th parameter... Its nonlinear mapping function is: Where kc is the segment index and KC is the total number of segment intervals. This is the amplitude coefficient of the kc-th segment of the c-th parameter, controlling the contribution of this segment to the mapping result. Let be the slope coefficient of the c-th parameter in the kc-th segment, controlling the rate of change of the mapping function within that segment. Let the segment breakpoint position of the c-th parameter in the kc-th segment define the range of parameter values ​​for that segment. The function is a differentiable smooth transition function, using the Sigmoid function.

[0036] Compared with existing technologies, the advantages and positive effects of this invention are as follows: it adopts non-contact multi-view point cloud scanning, avoiding the tire wear problem of contact measurement, and the detection is more comprehensive and efficient; the improved SAC-IA algorithm enhances the registration accuracy of multi-view point clouds, and combined with adaptive voxel grid filtering based on tire geometric priors and improved bilateral filtering, it retains contour edge features while efficiently removing noise; through principal component analysis and standard coordinate system construction, the parameter extraction benchmark is unified to ensure detection consistency; four core quality parameters are accurately extracted, and combined with a hybrid dynamic weight scoring network, the weights can be adaptively adjusted according to tire specifications and usage scenarios to achieve objective and intelligent quality evaluation, effectively solving the problems of insufficient three-dimensional representation, unstable registration, and weak versatility of traditional detection, and meeting the needs of large-scale, high-precision quality inspection in industrial applications. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of the structural process of a three-dimensional vision inspection method for motorcycle tires. Detailed Implementation

[0039] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0040] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0041] In this embodiment, considering the problems of insufficient detection accuracy, large noise interference, unstable multi-view registration, inconsistent detection benchmarks, and difficulty in achieving intelligent comprehensive evaluation in existing technologies, a 3D visual inspection method for motorcycle tires based on multi-view 3D visual perception, high-precision point cloud modeling, and intelligent scoring fusion is proposed. The specific implementation process is as follows: Figure 1 As shown.

[0042] To address the issues of severe occlusion, incomplete information, and large registration errors in traditional single-view detection, which hinder the construction of high-precision 3D tire models, this method fixes the motorcycle tire to be inspected on a detection platform and performs omnidirectional scanning using a multi-view point cloud scanning device. An improved SAC-IA algorithm is then employed to accurately register and fuse the multi-view point cloud data, constructing a 3D point cloud model. Specifically, the motorcycle tire is first fixed to the detection platform, allowing it to rotate controllably around its own axis of rotation. Simultaneously, multiple sets of structured light or laser 3D scanners collect point cloud data from different directions, forming raw point cloud sets from multiple perspectives. For each point in the point cloud from each perspective, its local curvature, Fast Point Feature Histogram (FPFH) feature vector, and normal vector information are calculated to construct a highly discriminative feature descriptor that considers both geometric shape and local texture distribution. This enhances the matching probability of source points from different perspectives, improving the reliability of the initial correspondence. The raw point cloud data is then sampled, with each sampling generating a pair of three nonlinear collinear point sets. ,in, For the source viewpoint set, For the target viewpoint set, the non-collinearity criterion is that the area of ​​the triangle formed by the three points is greater than a set threshold. In the SAC-IA sampling process, a probability weighting strategy based on local curvature, FPFH feature magnitude, and normal consistency is introduced to assign higher sampling weights to points with significant features and stable structures. The sampling is probabilistically weighted, and the sampling weight is defined as follows: It is obtained by fusing local curvature, fast point feature histogram, and normal vector. The calculation method is as follows: ,in, These are the weighting coefficients, and This is used to weigh the contribution of the three factors to the sampling weight. Let be the local curvature at the k-th point. For the global maximum curvature, For fast point feature histogram feature vectors, Let be the magnitude of the feature vector of the fast point feature histogram of the k-th point. This represents the maximum magnitude of the feature vector of the global maximum fast point feature histogram. Let be the normal vector of the k-th point. Let be the mean of the normal vectors of all points in the neighborhood of the k-th point. The absolute value of the dot product of the normal vector of the k-th point and the mean of the normal vectors of its neighborhood reflects the directional consistency of the two vectors. Sampling terminates when the cumulative number of samples reaches M, resulting in a set of point pairs; each point pair in the set... Construct a multi-constraint transformation matrix solution model to solve the source point set. to target point set initial rotation matrix Translation vector It satisfies the following formula:

[0043] ,

[0044] in, Let the centroid of the source point set be... Let the centroid of the target point set be... Let be the covariance matrix of the source point set and the target point set. Singular value decomposition is performed on the covariance matrix H, resulting in three matrices: U (left singular matrix). Singular value matrix and The transpose of the right singular matrix. Then calculate the pairwise distances within the source point set. The distance between corresponding points after transformation using the obtained transformation matrix If the absolute value of the distance difference between the two is greater than the set distance threshold, the transformation matrix is ​​discarded, and the filtered valid transformation matrices are retained; for each valid transformation matrix, all points from the source viewpoint are calculated. Transformed reprojection error Thresholds are adaptively adjusted based on local density of the point cloud: ,in, Based on the threshold, For dynamic thresholds, Let L be the local density of the k-th point and L be the total number of source point clouds; calculate the consistency scoring function, combining the proportion of interior points and the constraint satisfaction: ,in, To meet the reprojection error The number of interior points, To constrain the weights, To satisfy the distance constraint, To determine the number of points participating in the distance constraint verification, the transformation matrix that maximizes the consistency scoring function is selected as the final transformation matrix for solving the problem. A complete 3D point cloud model is then constructed based on this transformation matrix.

[0045] To address the issues of high outlier noise and the potential for uniform filtering to damage tire structural features in tire point clouds, the following steps first process the constructed 3D point cloud model. An adaptive voxel mesh filtering algorithm is designed based on prior knowledge of the tire's geometry to intelligently identify and remove outlier noise points. Specifically, based on motorcycle tire design specifications and the initial contour features of the 3D point cloud model, core geometric prior parameters of the tire are extracted, including the tire's nominal outer diameter D, nominal inner diameter d, tread width W, and tire rotation center O. A local cylindrical coordinate system is then established with the tire rotation center O as the origin. Where r is the radial distance, Let z be the circumferential angle and z be the axial distance; perform radial mesh generation, dividing the radial range according to the tire's nominal inner diameter d and nominal outer diameter D. Divided into mg sub-intervals, each sub-interval has a corresponding voxel radial side length. satisfy: Where jm are the indices of mg sub-intervals, This is the attenuation coefficient, and the circumferential angle range is also considered. The area is uniformly divided into ng circumferential sub-intervals, and the circumferential angle of the voxel corresponding to each sub-interval is . Additionally, based on the tread width W, the axial range is... Divide into p axial sub-intervals, and the axial side length of the voxel corresponding to each sub-interval is... Transform all point cloud data in the 3D point cloud model from the Cartesian coordinate system to... The coordinates are then categorized into the corresponding adaptive voxel grids to obtain the number of point clouds within each voxel grid. For each adaptive voxel grid, if the number of point clouds within that grid... Points within the grid that are less than a set threshold are considered outliers and removed. Then, an improved bilateral filtering method is used to smooth the point cloud. Principal component analysis and least squares fitting are then used to determine the tire's principal axis direction, aligning the point cloud coordinate system to a standard orientation, resulting in a standardized point cloud model. This is achieved by designing an improved bilateral filtering algorithm that integrates radial distance and local curvature weights based on the tire's toroidal rotational geometry. This smooths the point cloud while preserving the tire's contour edge features. The filtered target point coordinates... Satisfying the formula: ,in, Let P be the set of neighboring points. Normalization factor Geometric distance weights The local curvature constraint weights are used; the mean point of all point clouds is calculated using the improved bilaterally filtered point cloud as input. Convert each point cloud coordinate into a centralized coordinate system. The process involves constructing the covariance matrix of the centered point cloud, solving for the eigenvalues ​​and corresponding eigenvectors of the covariance matrix, and using the eigenvector with the largest eigenvalue as the initial principal axis direction. A Cartesian standard coordinate system is established with the tire's rotation center O as the origin, the z-axis coinciding with the principal axis direction, the x-axis lying in the tire's radial plane and taken as the radial direction at the midpoint of the tire's tread width, and the y-axis determined based on the right-hand rule to ensure consistency of the right-hand screw in the coordinate system. The filtered point cloud coordinates are then transformed to the standard coordinate system, completing the construction of the standardized point cloud model. Specifically, to suppress random noise while preserving the key contour edge features of the tire during point cloud smoothing, a local neighborhood search structure is first established for the point cloud after outlier removal, based on the geometric characteristics of the tire's toroidal rotation. Then, the spatial relationship and curvature difference between the target point and its neighbors are jointly modeled. On the one hand, geometric distance weights are constructed based on the radial distance difference between the target point and its neighbors in cylindrical coordinates and the three-dimensional Euclidean distance. ,in, These are the radial distances between the target point and its neighboring points, respectively, taken from cylindrical coordinates. The standard deviation of radial distance. The standard deviation of spatial distance is set to 1.5 times the average point spacing in the point cloud, ensuring that points closer to each other contribute more to the update of the target point. On the other hand, the local curvature difference between the target point and its neighbors is calculated, and local curvature weights and local curvature constraint weights are constructed. ,in, Let P be the local curvature. A curvature threshold is used. Points with similar curvature are given higher weights, while the weights at points of abrupt curvature change decay rapidly, thus avoiding smoothing across the tire edge structure. Subsequently, the two types of weights are multiplied to form a comprehensive weight, and the coordinates of neighboring points are weighted and summed. After normalization, the new position of the target point is updated. By performing this filtering operation sequentially on all points, and setting multiple iterations until convergence, overall smoothing of the point cloud is achieved. This implementation effectively protects high-curvature areas such as the tire bead and sidewall edges, while significantly reducing local vibrations caused by noise. To obtain the tire's principal axis direction as a basis for attitude standardization, the improved bilaterally filtered point cloud is first used as input. All point cloud data are traversed, and the mean point of its three-dimensional coordinates is calculated as the geometric center of the entire point cloud. Then, the coordinates of each point are subtracted from this mean point to obtain centered point cloud data, eliminating the influence of translation on subsequent analysis. Next, a three-dimensional covariance matrix is ​​constructed based on the centered point cloud coordinates. This matrix reflects the dispersion of the point cloud along each coordinate axis. The covariance matrix is ​​decomposed into eigenvalues, yielding three eigenvalues ​​and their corresponding eigenvectors. The eigenvector corresponding to the largest eigenvalue indicates the most dispersed distribution of the point cloud in that direction, representing the principal extension direction of the tire's overall geometry. Since the tire is a typical body of revolution, its axial length is significantly greater than its radial and circumferential thickness. Therefore, this largest eigenvector stably corresponds to the tire's principal axis of rotation. Finally, this eigenvector is used as the initial principal axis direction output for the tire, providing a reliable directional reference for subsequent standard coordinate system construction. This implementation allows for adaptive identification of the tire's principal axis direction without prior installation attitude information. To unify the coordinate reference for different tire samples and different detection attitudes, a Cartesian coordinate system is first constructed using the known tire rotation center O from the previous steps as the origin. Then, the obtained principal axis direction vector is normalized and set as the Z-axis direction of the standard coordinate system, ensuring that the Z-axis strictly coincides with the actual rotation axis of the tire. Next, in the radial plane perpendicular to the Z-axis, based on tire geometric priors, a radially outward direction is selected at the midpoint of the tread width as the X-axis direction, and this direction vector is also normalized. Subsequently, according to the right-hand rule, the Y-axis direction is determined by the cross product of the Z-axis and X-axis, ensuring that the established coordinate system satisfies orthogonality and the consistency of the right-hand screw. Finally, the three-axis direction vectors are orthogonally normalized to form a complete rotation matrix. The standard coordinate system constructed through this process is not only highly consistent with the tire's physical structure but also possesses uniqueness and repeatability, thus ensuring that subsequent alignment and comparative analysis of different tire samples are performed under the same coordinate reference, providing a unified reference framework for parameter extraction. Finally, a rotation matrix R is constructed based on the obtained three-axis unit vectors, with the tire's rotation center O serving as the translation reference.Then, a coordinate transformation is performed on each point in the point cloud: first, the coordinates are subtracted from the rotation center O to achieve translational alignment; then, multiplied by the rotation matrix R, the point cloud is rotated and mapped from the original coordinate system to the standard coordinate system; if necessary, scale normalization is performed to ensure that tires of different sizes are analyzed under a unified scale reference. By performing the above rigid transformation on all point clouds, a standardized point cloud model can be obtained. This standardization process eliminates the influence of arbitrary tire placement postures and installation errors on the geometric analysis on the testing platform.

[0046] To address the challenge of accurately acquiring multiple three-dimensional tire quality parameters simultaneously using traditional inspection methods, this study extracts parameters such as uniform thickness, sidewall profile curvature, bead roundness error, and radial runout from a standardized point cloud model. This enables high-precision and objective quantitative characterization of the tire's multi-dimensional quality features. The implementation is based on a unified coordinate reference using a standardized point cloud, ensuring consistent calculations in the normal direction. Uniform thickness is characterized by calculating the standard deviation of the normal distance from each point in the tread region to the tire's inner wall. Bead roundness error is achieved by projecting the inner edge point cloud of the bead onto a reference plane perpendicular to the Z-axis, fitting ellipse parameters, and calculating the ellipticity deviation and maximum radial deviation. Sidewall profile curvature is characterized by uniformly extracting multiple profile lines along the circumference of the sidewall using the standardized coordinate system Z-axis, and calculating the average absolute value of their second derivatives. Radial runout is obtained by simulating rotation using the Z-axis of the standardized model and measuring the displacement amplitude of the tread centerline in the radial direction. Specifically, based on a standardized point cloud model, the tread region point cloud is automatically segmented according to radial position and curvature characteristics under a unified coordinate reference. This segment consists of points located near the tire's outer diameter and whose axial position is within the tread width. Then, normal vector estimation is performed on the tread region point cloud. A neighborhood plane fitting method can be used to calculate the outward normal direction for each point. The normal orientation is corrected by ensuring consistency with the radial direction of the standardized coordinate system, so that all normal directions point uniformly to the outer side of the tire, thus guaranteeing the consistency of the normal direction calculation. Next, an inner wall reference surface is constructed based on the tire's inner wall point cloud. A continuous inner wall model can be obtained through least-squares surface fitting or radial interpolation. Then, for each point within the tread region, a ray search is performed inward along its normal direction, intersecting with the inner wall surface to obtain the corresponding normal thickness distance. The standard deviation of the normal distances obtained for all tread points is calculated as a uniformity thickness index. In the standardized point cloud model, point cloud data located in the inner edge region of the tire bead is selected based on axial position and radial range. This segment consists of points close to the tire's inner diameter and located at both edges. Then, the point set is projected onto a reference plane perpendicular to the Z-axis, i.e., the XOY plane, in the standard coordinate system, thus transforming the three-dimensional problem into a two-dimensional planar contour analysis problem. Next, the least squares ellipse fitting algorithm is used to solve for the ellipse parameters of the projected two-dimensional point set, obtaining the center position, major and minor axis lengths, and directions of the ellipse. By using this ellipse as an approximation of the actual shape of the bead, its ellipticity deviation is further calculated, such as the ratio of the difference between the major and minor axes to the average radius, to reflect the overall roundness deviation. Simultaneously, the radial distance from each projected point to the ellipse fitting curve is calculated, and the maximum radial deviation is statistically analyzed as a local maximum error index. Finally, the ellipticity deviation and the maximum radial deviation can be combined or separately output as bead roundness error parameters. Based on the standardized point cloud model, the lateral region is determined in the axial direction, i.e., the point cloud region located between the tread edge and the bead.Then, using the Z-axis of the standard coordinate system as the rotation axis, the sidewall region is uniformly divided into several angular intervals in the circumferential direction. For example, a cross-sectional line parallel to the Z-axis is extracted at fixed angles. For the point cloud within each angular interval, it is sorted according to the axial coordinate, and a continuous sidewall profile curve is constructed using interpolation or fitting. Next, each profile curve is smoothed, and its second derivative is calculated based on the curve parameter expression to reflect the degree of curve curvature. The average of the absolute values ​​of the second derivatives within the curve range is taken as the local curvature index of that cross-section. Finally, the local curvature indices obtained from all circumferential cross-sections are averaged to obtain the overall sidewall profile curvature parameters. Based on the standardized point cloud model, the tread centerline is extracted in the tread region, that is, a set of points representing the outer surface of the tread is obtained along the circumferential direction near the axial midpoint. Then, using the Z-axis of the standard coordinate system as the theoretical rotation axis of the tire, this centerline point set is regarded as a full circumferential sampling of the tire in a stationary state. Next, the process of the tire rotating around the Z-axis is simulated. This involves sorting the centerline points according to their circumferential angles and calculating the radial distance from each point to the Z-axis, forming a sequence of radial distances that vary with the angle. Then, the difference between the maximum and minimum radial distances in this sequence is calculated as the maximum radial displacement amplitude of the tire tread within one rotation cycle, i.e., the radial runout.

[0047] Finally, the extracted parameters are input into a preset scoring model to obtain the output motorcycle tire quality score. The preset scoring model includes the following processing units: a parameter standardization unit, used to normalize each extracted parameter to the [0,1] interval; a feature nonlinear transformation unit, used to perform nonlinear mapping on each parameter using a learnable piecewise function; a dynamic weight allocation unit, used to automatically adjust the weights of each parameter according to tire specifications and usage scenarios; and a comprehensive score calculation unit, used to fuse the transformed values ​​of each parameter and their corresponding weights to output the tire quality score. Specifically, after obtaining the original parameter values ​​of the tire sample, a linear normalization method is used to map them to the [0,1] interval. Then, in the feature nonlinear transformation unit, each extracted parameter is divided into multiple segmented intervals according to its value range. For the c-th parameter... Its nonlinear mapping function is: Where kc is the segment index and KC is the total number of segment intervals. This is the amplitude coefficient of the kc-th segment of the c-th parameter, controlling the contribution of this segment to the mapping result. Let be the slope coefficient of the c-th parameter in the kc-th segment, controlling the rate of change of the mapping function within that segment. Let the segment breakpoint position of the c-th parameter in the kc-th segment define the range of parameter values ​​for that segment. The function is a differentiable smooth transition function, employing the sigmoid function. Subsequently, a tire specification and application scenario description vector is established in the dynamic weight allocation unit, including information such as tire size, load rating, speed rating, and usage environment category, and encoded as input features for the weight allocation unit. Then, a weight prediction model is constructed using a lightweight neural network, trained with historical sample data, and outputs weights. In practical applications, the current tire specification and scenario information are input, and the model outputs the weight coefficients corresponding to each parameter, which are normalized to ensure the sum of all weights is 1. Finally, the parameter mapping values ​​from the feature nonlinear transformation unit are received to form a quality feature vector, along with the weight vector output from the dynamic weight allocation unit. Then, a weighted summation method is used to calculate the comprehensive score, ensuring that each parameter contributes to the final result according to its importance, resulting in a quality score.

[0048] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A three-dimensional vision inspection method for motorcycle tires, characterized in that, Includes the following steps: S1. Fix the motorcycle tire to be inspected onto the inspection platform, perform an all-round scan using a multi-view point cloud scanning device, and use the improved SAC-IA algorithm to accurately register and fuse the multi-view point cloud data to construct a three-dimensional point cloud model. S2. The constructed 3D point cloud model is processed. An adaptive voxel grid filtering algorithm is designed based on the prior knowledge of the tire geometry to intelligently identify and remove outlier noise points. An improved bilateral filtering method is used to smooth the point cloud. The direction of the tire's main axis is determined by principal component analysis and least squares fitting. The point cloud coordinate system is aligned to the standard posture to obtain a standardized point cloud model. S3. For the standardized point cloud model, extract parameters such as uniform thickness of motorcycle tire, sidewall profile curvature, bead roundness error and tire radial runout. S4. Input the extracted parameters into the preset scoring model to obtain the output motorcycle tire quality score. The specific implementation of step S1, which uses the improved SAC-IA algorithm to accurately register and fuse multi-view point cloud data to construct a 3D point cloud model, includes: S11. First, obtain the number of original point cloud data collected from each viewpoint. For each point, calculate the local curvature of the 3D point cloud, the feature vector of the fast point feature histogram, and the normal vector. S12. Sample the original point cloud data, generating a pair of nonlinear collinear point sets for each sampling. ,in, For the source viewpoint set, For the target viewpoint set, the non-collinearity determination condition is that the area of ​​the triangle formed by the three points is greater than a set threshold; S13. Apply probability weighting to the samples, defining the sampling weight as follows: It is obtained by fusing local curvature, fast point feature histogram, and normal vector. The calculation method is as follows: ,in, These are the weighting coefficients, and This is used to weigh the contribution of the three factors to the sampling weight. Let be the local curvature at the k-th point. For the global maximum curvature, For fast point feature histogram feature vectors, Let be the magnitude of the feature vector of the fast point feature histogram of the k-th point. This represents the maximum magnitude of the feature vector of the global maximum fast point feature histogram. Let be the normal vector of the k-th point. Let be the mean of the normal vectors of all points in the neighborhood of the k-th point. It is the absolute value of the dot product of the normal vector of the k-th point and the mean of the normal vectors of the neighborhood, reflecting the consistency of the directions of the two vectors; S14. Sampling terminates when the cumulative number of samples reaches M, resulting in a set of point pairs; each point pair in the set... Construct a multi-constraint transformation matrix solution model to solve the source point set. to target point set initial rotation matrix Translation vector It satisfies the following formula: ， in, Let the centroid of the source point set be... Let the centroid of the target point set be... Let be the covariance matrix of the source point set and the target point set. Singular value decomposition is performed on the covariance matrix H, resulting in three matrices: U (left singular matrix). Singular value matrix and The transpose of a right singular matrix; S15. Calculate the pairwise distances within the source point set. The distance between corresponding points after transformation using the obtained transformation matrix If the absolute value of the distance difference between the two is greater than the set distance threshold, the transformation matrix is ​​discarded and the filtered valid transformation matrix is ​​retained. S16. Calculate all points from the source viewpoint for each valid transformation matrix. Transformed reprojection error Thresholds are adaptively adjusted based on local density of the point cloud: ,in, Based on the threshold, For dynamic thresholds, Let L be the local density of the k-th point, and L be the total number of source point clouds; S17. Calculate the consistency scoring function, combining the internal point ratio and constraint satisfaction: ,in, To meet the reprojection error The number of interior points, To constrain the weights, To satisfy the distance constraint, To determine the number of points participating in the distance constraint verification, the transformation matrix that maximizes the consistency score function is selected as the final transformation matrix for the solution. A complete 3D point cloud model is then constructed based on this transformation matrix. In step S2, after smoothing the point cloud using an improved bilateral filtering method, the tire's main axis direction is determined through principal component analysis and least squares fitting. The point cloud coordinate system is then aligned to a standard orientation, resulting in a standardized point cloud model. S221. Based on the geometric features of a tire's annular rotation, an improved bilateral filtering algorithm integrating radial distance weights and local curvature weights is designed to smooth the point cloud while preserving the tire's contour edge features. The target point coordinates after filtering are... Satisfying the formula: ,in, Let P be the set of neighboring points. Normalization factor, Geometric distance weights For local curvature constraint weights; S222. Using the improved bilaterally filtered point cloud as input, calculate the mean point of all point clouds. Convert each point cloud coordinate into a centralized coordinate system. Construct the covariance matrix of the centralized point cloud, solve for the eigenvalues ​​and corresponding eigenvectors of the covariance matrix, and take the eigenvector with the largest eigenvalue as the initial principal axis direction; S223. Establish a Cartesian standard coordinate system with the tire rotation center O as the origin. The z-axis coincides with the direction of the principal axis. The x-axis lies in the radial plane of the tire and is taken as the radial direction at the midpoint of the tire tread width. The y-axis is determined based on the right-hand rule to ensure the consistency of the right-hand screw of the coordinate system. S224. Transform the filtered point cloud coordinates to the standard coordinate system to complete the construction of the standardized point cloud model.

2. The three-dimensional vision inspection method for motorcycle tires according to claim 1, characterized in that, In step S2, the constructed 3D point cloud model is processed. An adaptive voxel mesh filtering algorithm is designed based on prior knowledge of the tire geometry to intelligently identify and remove outlier noise points. The specific implementation is as follows: S211. Based on the design specifications of motorcycle tires and the initial contour features of the 3D point cloud model, extract the core geometric prior parameters of the tire, including the tire's nominal outer diameter D, nominal inner diameter d, tread width W, and tire rotation center O. Establish a local cylindrical coordinate system with the tire rotation center O as the origin. Where r is the radial distance, z is the circumferential angle, and z is the axial distance; S212. Perform radial mesh generation, dividing the radial range according to the tire's nominal inner diameter d and nominal outer diameter D. Divided into mg sub-intervals, each sub-interval has a corresponding voxel radial side length. satisfy: Where jm are the indices of mg sub-intervals, This is the attenuation coefficient, and the circumferential angle range is also considered. The area is uniformly divided into ng circumferential sub-intervals, and the circumferential angle of the voxel corresponding to each sub-interval is . Additionally, based on the tread width W, the axial range is... Divide into p axial sub-intervals, and the axial side length of the voxel corresponding to each sub-interval is... ; S213. Convert all point cloud data in the 3D point cloud model from the Cartesian coordinate system to... The coordinates are then categorized into the corresponding adaptive voxel grids to obtain the number of point clouds within each voxel grid. For each adaptive voxel grid, if the number of point clouds within that grid... If the value is less than the set threshold, all point clouds within that grid are identified as outlier noise points and are removed.

3. The three-dimensional vision inspection method for motorcycle tires according to claim 1, characterized in that, Geometric distance weight in step S221 ,in, These are the radial distances between the target point and its neighboring points, respectively, taken from cylindrical coordinates. The standard deviation of radial distance. The standard deviation of spatial distance is taken as 1.5 times the average point spacing of the point cloud; local curvature constraint weight. ,in, Let P be the local curvature. This is the curvature threshold.

4. The three-dimensional vision inspection method for motorcycle tires according to claim 1, characterized in that, In step S3, the uniformity of the normal direction calculation is ensured by using a standardized point cloud as the unified coordinate reference. The uniformity thickness is characterized by calculating the standard deviation of the normal distance from each point in the tread area to the inner wall of the tire. The bead roundness error is calculated by projecting the inner edge point cloud of the bead onto a reference plane perpendicular to the Z-axis, fitting elliptic parameters, and calculating the ellipticity deviation and the maximum radial deviation. The sidewall profile curvature is characterized by extracting multiple profile lines evenly along the circumference of the sidewall using the standardized coordinate system Z-axis, and calculating the average absolute value of their second derivatives. The tire radial runout is obtained by simulating the rotation state using the Z-axis of the standardized model and measuring the displacement amplitude of the tread centerline in the radial direction.

5. The three-dimensional vision inspection method for motorcycle tires according to claim 1, characterized in that, The scoring model in step S4 employs a hybrid dynamic weight scoring network, which includes the following processing units: The parameter standardization unit is used to normalize each extracted parameter to the [0,1] interval; The feature nonlinear transformation unit is used to perform nonlinear mapping of each parameter through a learnable piecewise function; The dynamic weight allocation unit is used to automatically adjust the weights of each parameter according to tire specifications and usage scenarios. The comprehensive scoring calculation unit is used to integrate the converted values ​​of various parameters and their corresponding weights to output a tire quality score.

6. The three-dimensional vision inspection method for motorcycle tires according to claim 5, characterized in that, The feature nonlinear transformation unit, used to perform nonlinear mapping on each parameter through a learnable piecewise function, specifically includes: dividing each extracted parameter into multiple segmented intervals according to the parameter value range; for the c-th parameter... Its nonlinear mapping function is: Where kc is the segment index and KC is the total number of segment intervals. This is the amplitude coefficient of the kc-th segment of the c-th parameter, controlling the contribution of this segment to the mapping result. Let be the slope coefficient of the c-th parameter in the kc-th segment, controlling the rate of change of the mapping function within that segment. Let the segment breakpoint position of the c-th parameter in the kc-th segment define the range of parameter values ​​for that segment. The function is a differentiable smooth transition function, using the Sigmoid function.

Images