Automobile part registration detection method based on 3D laser sensor

Abstract

The invention discloses an automobile part registration detection method based on a 3D laser sensor, and relates to the technical field of automobile part detection, and the method comprises the following steps: S1, collecting 3D point cloud data, and carrying out the non-contact scanning of a to-be-detected automobile part through a scanning device based on the 3D laser sensor; s2, registering the point cloud data; s3, estimating and adjusting the attitude; s4, matching and difference detection; s5, carrying out difference visualization; s6, multi-view point cloud merging; and S7, generating a result. According to the automobile part registration detection method based on the 3D laser sensor disclosed by the invention, the 3D laser sensor scans the surfaces of the automobile parts to obtain the three-dimensional point cloud data of the parts, the accurate detection of the automobile parts is realized by comparing the difference of the point cloud data, and the difference parts are highlighted through a visual means, so that the accuracy of the registration detection of the automobile parts is improved. And manufacturing and quality inspection personnel are helped to judge the quality of the spare and accessory parts more intuitively.

Description

Technical Field

[0001] The present invention relates to the technical field of automobile parts detection, and in particular to an automobile parts registration detection method based on a 3D laser sensor. Background Art

[0002] Auto parts are the various units that make up the whole of a car and the products that serve the car. They are of various types, including engine assemblies, filters, cylinders and components. The auto parts market is expanding with the increase in people's consumption of cars. In recent years, auto parts manufacturers have also developed rapidly. During the use of the car, some auto parts need to be replaced regularly, such as air filters, oil filters, gasoline filters, etc. These are called auto spare parts. Some parts that are not often replaced, such as exhaust pipes, oil pans, headlights, etc., belong to the category of auto parts.

[0003] With the continuous development of the automobile manufacturing industry, the precision requirements for automobile parts are becoming higher and higher during the production and manufacturing of automobile parts. In order to ensure the production quality of automobile parts, it is usually necessary to accurately detect their geometric shape, size and alignment.

[0004] Traditional detection methods mainly rely on contact measuring tools (such as calipers, micrometers, etc.), but this method is not only time-consuming and labor-intensive, but also easily affected by human errors, and it is difficult to meet the requirements of the modern automotive industry for high-precision and high-efficiency detection.

[0005] In view of this, the present invention proposes an automobile parts registration and detection method based on a 3D laser sensor. Summary of the invention

[0006] The invention discloses an automobile parts registration detection method based on a 3D laser sensor, aiming to solve the technical problems in the background technology.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A 3D laser sensor-based auto parts registration detection method comprises the following steps:

[0009] S1. Collection of 3D point cloud data: non-contact scanning of the automotive parts to be inspected by a scanning device based on a 3D laser sensor;

[0010] Equipment configuration: The 3D laser sensor is synchronized with the assembly line to obtain the 3D point cloud data of the object surface through laser scanning to generate the initial point cloud model P = {p1, p2, ..., pn}, where p i =(x i ,y i ,z i) is each 3D point in the point cloud;

[0011] Multi-angle scanning: For parts with complex geometric shapes, multi-angle scanning is used to ensure that all point cloud data on the surface of the part is fully captured to reduce blind spots caused by occlusion or complex surfaces;

[0012] During installation, the laser direction of the laser sensor is perpendicular to the movement direction. During the movement, the laser sensor is triggered and data is collected to obtain grid data. At the same time, the grid data is supplemented according to the trigger spacing in the movement direction and the point spacing of the laser to obtain 3D point cloud data.

[0013] Correction and determination of laser posture: Align and compare 3D point cloud data with the 3D model of the measured product;

[0014] Data preprocessing: The collected point cloud data needs to be preprocessed;

[0015] S2. Registration of point cloud data: Based on the preprocessed point cloud, the point cloud of the part to be inspected is aligned with the point cloud of the standard part through the point cloud registration algorithm for subsequent geometric analysis and comparison. The registration method includes the following stages:

[0016] (1) Preliminary alignment;

[0017] Since the orientations of objects on the assembly line are inconsistent, the collected point cloud is quite different from the standard point cloud, so preliminary alignment is required first;

[0018] (2) Feature matching;

[0019] A common preliminary alignment method is to perform a coarse registration by extracting global geometric features of the point cloud; the following steps are used:

[0020] Calculate point cloud feature descriptors: such as FPFH (Fast Point Feature Histograms) or SHOT (Signature of Histograms of OrienTations) features, which can describe the local geometric features of each point in the point cloud;

[0021] Feature matching: Match the feature descriptors of the collected point cloud with those of the standard point cloud, find similar point pairs, and use these point pairs to estimate the rough alignment transformation of the two sets of point clouds;

[0022] (3) Shape-based alignment;

[0023] If the object has obvious geometric features (such as symmetry or known specific geometric shapes), estimate the orientation of the object through shape matching and perform preliminary alignment;

[0024] (4) Precise alignment;

[0025] After rough registration, the alignment of the point cloud is usually still not accurate enough; in order to further improve the accuracy of the alignment, it is necessary to use the iterative closest point algorithm (ICP) for fine registration;

[0026] S3, attitude estimation and adjustment;

[0027] Since the objects on the assembly line are placed in different directions, the posture of the objects is further corrected through posture estimation based on preliminary alignment and precise registration:

[0028] (1) Estimate the direction through the geometric center and principal axis of the point cloud:

[0029] 1) Determine the location of the object by calculating the object's center of mass;

[0030] The center of mass C is the geometric center of the object, which represents the overall position of the object;

[0031] 2) Use the principal axis direction estimation to infer the orientation of the object;

[0032] The main axis direction of an object reflects the main orientation of the object in space, and the main axis direction of the object is estimated by principal component analysis (PCA);

[0033] (2) Posture adjustment-rotation transformation;

[0034] 1) Rotation Matrix

[0035] Assume that the main axis of the object is v1, and the main axis of the standard object is vQ1. The rotation matrix R is calculated to rotate v1 to vQ1. This is achieved by the Axis-Angle Representation.

[0036] 2) Apply a rotation transformation

[0037] Once we have the rotation matrix R, we apply it to each point P' of the object i , to achieve posture adjustment:

[0038] ”'

[0039] p i =Rp i

[0040] In this way, the orientation of the object will be aligned with the standard object, ensuring that both are matched in the same coordinate system;

[0041] S4. Matching and Difference Detection

[0042] After alignment, matching and difference detection are performed; the following geometric distance calculation methods are used to determine whether the object meets the standards:

[0043] Hausdorff distance: Calculates the maximum and minimum distance between two sets of point clouds to determine the geometric difference between them; this distance is more sensitive to large local differences; this distance reflects the distance between the points with the largest geometric deviation in the two point clouds; Hausdorff distance measures the maximum and minimum distance between two point sets;

[0044] Chamfer distance: used to measure the average distance from each point to the nearest point, suitable for processing local shape deviations; Chamfer distance better handles partial occlusion;

[0045] Point-to-point Euclidean distance: Calculate the Euclidean distance between the two sets of point clouds after registration point by point, mark the areas with large errors as possible abnormal areas; areas with large differences will be displayed as highlighted areas; for each pair of registered points, calculate the Euclidean distance between them, and use histograms or colors to display local geometric differences; this method is suitable for highly aligned point clouds:

[0046] S5. Difference visualization

[0047] After matching and difference detection, the error heat map or color mapping is used to highlight the areas with large differences, helping operators to quickly determine whether there are defects or deviations in the objects;

[0048] Difference heat map: After alignment, the difference between the object point cloud and the standard point cloud is converted into color values ​​for 3D rendering;

[0049] Difference area annotation: Mark points where the error exceeds the threshold as abnormal areas, and further analyze or detect these areas;

[0050] S6, multi-view point cloud merging;

[0051] Generate a complete point cloud model by merging multi-view point clouds; the core idea of ​​merging multi-view point clouds is to align and fuse point clouds from multiple different viewpoints to form a unified 3D model;

[0052] Result generation: Generate a complete point cloud model containing all view information;

[0053] S7, result generation: generating a complete point cloud model containing all viewing angle information.

[0054] In a preferred solution, the scanning device based on the 3D laser sensor in step S1 is installed above the assembly line, and the 3D point cloud number of the object's surface is collected in real time based on the 3D laser sensor as the object passes through the assembly line.

[0055] In a preferred solution, the data preprocessing method in step S1 is as follows:

[0056] (1) Denoising: Use statistical filtering or radius filtering to remove noise from the scan;

[0057] (2) Downsampling: The density of the point cloud is reduced by downsampling the voxel grid, which reduces the amount of data while retaining the geometric features of the point cloud.

[0058] In a preferred solution, the iterative closest point algorithm (ICP) in step S2 is as follows:

[0059] The core principle of the ICP algorithm: ICP iteratively finds the closest point pair in two point clouds, calculates the error between the point pairs, and then continuously optimizes the rotation matrix R and the translation matrix T to minimize the following objective function:

[0060]

[0061] where p i is a point in the standard point cloud, q i It is to collect points in the point cloud, and gradually align the two sets of point clouds by iteratively optimizing R and T.

[0062] ICP improvement: In order to avoid falling into the local optimal solution, different heuristic strategies or multi-resolution ICP schemes are combined to make the alignment more robust.

[0063] In a preferred solution, in step S3, the centroid is calculated as follows:

[0064] Given a point cloud P = {p1, p2, ..., pN} of an object, where each point p i =(x i ,y i ,z i ) is the 3D coordinate in the point cloud, the centroid C P The calculation formula is:

[0065]

[0066] The centroid represents the position of the object in three-dimensional space; in order to align the object point cloud with the standard point cloud, the centroids of the two need to be aligned through translation transformation.

[0067] In a preferred embodiment, in step S3,

[0068] By performing eigenvalue decomposition on the covariance matrix Σ, the eigenvalues ​​λ1,λ2,λ3 and the corresponding eigenvectors v1,v2,v3 are obtained;

[0069] The eigenvector v1 corresponds to the maximum eigenvalue λ1, which represents the direction of the main axis of the object;

[0070] The eigenvectors v2 and v3 correspond to the secondary principal axis directions and are usually used for the adjustment of the secondary directions;

[0071] The main direction of the object is determined by the feature vector v1; if it is inconsistent with the main axis direction of the standard object, the posture is adjusted through the rotation matrix.

[0072] In a preferred solution, in step S3, the calculation process of the axis-angle method is as follows:

[0073] Rotation axis: The rotation axis r is the cross product of vectors v1 and vQ1.

[0074] r=v1×v Q1

[0075] Rotation angle: The rotation angle θ is the angle between v1 and vQ1 and is calculated using the dot product formula:

[0076]

[0077] Among them, ||v1|| and ||v Q1 || are the moduli of v1 and vQ1 respectively.

[0078] Rodriguez rotation formula: The rotation matrix RRR is calculated using the Rodriguez rotation formula:

[0079]

[0080] where I is the identity matrix, [r] X is the antisymmetric matrix of the rotation axis r, defined as:

[0081]

[0082] In a preferred embodiment, in step S7,

[0083] Finally, the system will output the following:

[0084] Comparison report of differences between finished objects and standard objects;

[0085] Visualized point cloud models and difference heat maps are used to help inspectors make decisions;

[0086] Areas where the difference exceeds the threshold are marked, indicating the presence of defects.

[0087] As can be seen from the above, the present invention provides a registration detection method for automobile parts based on a 3D laser sensor, which scans the surface of automobile parts through a 3D laser sensor to obtain the three-dimensional point cloud data of the parts, and matches the point cloud data with the point cloud data of the standard parts to analyze whether there are deviations or defects in their geometric shapes. The method realizes accurate detection of automobile parts by comparing the differences in point cloud data, and highlights the difference parts through visualization means, helping manufacturing and quality inspection personnel to more intuitively judge the quality of parts. At the same time, the present invention provides a registration detection method for automobile parts based on a 3D laser sensor, which uses the 3D model of the measured product and the collected laser data for double alignment and comparison, on the one hand, the laser emission position of the 3D laser sensor is obtained relative to the laser posture of the product to achieve calibration, and on the other hand, the position of the automobile parts can be adjusted, which effectively reduces the time and processing cycle of the existing specially designed calibration block, reduces the production cost, and can achieve calibration by adjusting the attitude parameters of the laser and adjusting the conveying position of the automobile parts, saving the time of laser adjustment parameters, and effectively improving the accuracy of the point cloud data. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 This is an overall flow chart of an automobile parts registration detection method based on 3D laser sensors proposed by the present invention.

[0089] Figure 2 The present invention proposes a 3D laser sensor roadmap for an automobile parts registration detection method based on a 3D laser sensor. DETAILED DESCRIPTION

[0090] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0091] Reference Figure 1 , a registration detection method for automobile parts based on 3D laser sensor, comprising the following steps:

[0092] S1. Collection of 3D point cloud data: non-contact scanning of the automotive parts to be inspected by a scanning device based on a 3D laser sensor;

[0093] Equipment configuration: The 3D laser sensor is synchronized with the assembly line to obtain the 3D point cloud data of the object surface through laser scanning to generate the initial point cloud model P = {p1, p2, ..., pn}, where p i =(x i ,y i ,z i ) is each 3D point in the point cloud;

[0094] Multi-angle scanning: For parts with complex geometric shapes, multi-angle scanning is used to fully capture all point cloud data on the surface of the part to reduce blind spots caused by occlusion or complex curved surfaces;

[0095] In a preferred solution, the scanning device based on the 3D laser sensor in step S1 is installed above the assembly line, and the 3D point cloud data of the surface of the object is collected in real time based on the 3D laser sensor as the object passes through the assembly line.

[0096] Every time an object on the assembly line moves 0.1mm, the laser is automatically triggered and data is collected.

[0097] During installation, the laser direction of the laser sensor is perpendicular to the direction of movement. During the movement, the laser sensor is triggered and data is collected to obtain grid-shaped data.

[0098] At the same time, the grid data is supplemented according to the trigger spacing in the direction of motion and the point spacing of the laser to obtain 3D point cloud data.

[0099] The present invention uses a 3D laser sensor to perform non-contact point cloud scanning to obtain three-dimensional surface information of automotive parts in real time. This technology can be used in an assembly line operation environment, has high-speed and accurate point cloud acquisition capabilities, and ensures that the geometric features of the parts can be fully captured.

[0100] Among them, the 3D laser sensor parameters are as follows:

[0101] Use a line laser 3D vision JK_3D20_200B for scanning;

[0102] Z-axis accuracy: 0.0241mm

[0103] X-axis detection range (axis where the line laser is located): the middle width is 200mm, which meets the scanning range

[0104] X-axis detection resolution estimation: 0.1mm

[0105] Y-axis detection resolution estimation: 0.1mm

[0106] A material with a length of 400 mm needs 2 seconds to be scanned.

[0107] Scanning speed: 200mm / s

[0108]

[0109] Data preprocessing: The collected point cloud data needs to be preprocessed;

[0110] In a preferred solution, the data preprocessing method in step S1 is as follows:

[0111] (1) Denoising: Use statistical filtering or radius filtering to remove noise from the scan;

[0112] (2) Downsampling: The density of the point cloud is reduced by downsampling the voxel grid, which reduces the amount of data while retaining the geometric features of the point cloud.

[0113] S2. Registration of point cloud data: Based on the preprocessed point cloud, the point cloud of the part to be inspected is aligned with the point cloud of the standard part through the point cloud registration algorithm for subsequent geometric analysis and comparison. The registration method includes the following stages:

[0114] (1) Calibrate and determine the laser posture: align and compare the 3D point cloud data with the 3D model of the measured product;

[0115] It is important to note that when collecting 3D point cloud data of automotive parts, the 3D point cloud data is aligned and compared with the 3D model of the measured product to obtain the deviation value. By adjusting the attitude parameters of the 3D laser sensor, the parameters with the smallest deviation value are used as the final laser attitude.

[0116] The alignment steps include:

[0117] Select N points on the 3D point cloud data and the 3D model respectively, N>3, and the N points in the 3D point cloud data and the N points in the 3D model are not coplanar;

[0118] An affine matrix based on N points is obtained by matrix transformation calculation, and the obtained affine matrix is ​​used as the initial value for ICP alignment;

[0119] The affine matrix of the 3D point cloud data and the 3D model is calculated by ICP alignment, and the laser posture relative to the product is obtained by aligning the calculated affine matrix;

[0120] Specifically, the present application utilizes the 3D model of the measured product and the collected laser data for dual alignment and comparison. On the one hand, the calibration is achieved by obtaining the laser emission position of the 3D laser sensor relative to the laser posture of the product. On the other hand, the conveying position of automotive parts can be adjusted, thereby effectively reducing the time and processing cycle required for the special design of calibration blocks and reducing production costs. Calibration can also be achieved by adjusting the laser posture parameters and adjusting the conveying position of automotive parts, thereby saving the time for laser parameter adjustment and effectively improving the accuracy of point cloud data.

[0121] (2) Preliminary alignment;

[0122] Since the orientations of objects on the assembly line are inconsistent, the collected point cloud is quite different from the standard point cloud, so preliminary alignment is required first;

[0123] (3) Feature matching;

[0124] A common preliminary alignment method is to perform a coarse registration by extracting global geometric features of the point cloud; the following steps are used:

[0125] Calculate point cloud feature descriptors: such as FPFH (Fast Point Feature Histograms) or SHOT (Signature of Histograms of OrienTations) features, which can describe the local geometric features of each point in the point cloud;

[0126] Feature matching: Match the feature descriptors of the collected point cloud with those of the standard point cloud, find similar point pairs, and use these point pairs to estimate the rough alignment transformation of the two sets of point clouds;

[0127] (4) Shape-based alignment;

[0128] If the object has obvious geometric features (such as symmetry or known specific geometric shapes), estimate the orientation of the object through shape matching and perform preliminary alignment;

[0129] (5) Precise alignment;

[0130] After rough registration, the point cloud alignment is usually still not accurate enough. In order to further improve the accuracy of the alignment, it is necessary to use the iterative closest point algorithm (ICP) for fine registration;

[0131] In a preferred solution, the iterative closest point algorithm (ICP) in step S2 is as follows:

[0132] The core principle of the ICP algorithm: ICP iteratively finds the closest point pair in two point clouds, calculates the error between the point pairs, and then continuously optimizes the rotation matrix R and the translation matrix T to minimize the following objective function:

[0133]

[0134] where p i is a point in the standard point cloud, q i It is to collect points in the point cloud, and gradually align the two sets of point clouds by iteratively optimizing R and T.

[0135] ICP improvement: In order to avoid falling into the local optimal solution, different heuristic strategies or multi-resolution ICP schemes are combined to make the alignment more robust.

[0136] S3, attitude estimation and adjustment;

[0137] Since the orientation of objects on the assembly line may be inconsistent, the posture of the objects is further corrected through posture estimation based on preliminary alignment and fine registration:

[0138] (1) Estimate the direction through the geometric center and principal axis of the point cloud:

[0139] 1) Determine the location of the object by calculating the object's center of mass;

[0140] The center of mass C is the geometric center of the object, which represents the overall position of the object;

[0141] In a preferred solution, in step S3, the centroid is calculated as follows:

[0142] Given a point cloud P = {p1, p2, ..., pN} of an object, where each point p i =(x i ,y i ,z i ) is the 3D coordinate in the point cloud, the centroid C P The calculation formula is:

[0143]

[0144] The centroid represents the position of the object in three-dimensional space; in order to align the object point cloud with the standard point cloud, the centroids of the two need to be aligned through translation transformation.

[0145] 2) Use the principal axis direction estimation to infer the orientation of the object;

[0146] The main axis direction of an object reflects the main orientation of the object in space, and the main axis direction of the object is estimated by principal component analysis (PCA);

[0147] First, the point cloud data is decentralized to obtain the decentralized point cloud P′, that is, each point minus the centroid C p :

[0148] p' i =p i -C P

[0149] Next, calculate the covariance matrix Σ of the point cloud:

[0150]

[0151] The covariance matrix Σ is a 3x3 matrix that reflects the variance of the point cloud in all directions;

[0152] In a preferred embodiment, in step S3,

[0153] By performing eigenvalue decomposition on the covariance matrix Σ, the eigenvalues ​​λ1,λ2,λ3 and the corresponding eigenvectors v1,v2,v3 are obtained;

[0154] The eigenvector v1 corresponds to the maximum eigenvalue λ1, which represents the direction of the main axis of the object;

[0155] The eigenvectors v2 and v3 correspond to the secondary principal axis directions and are usually used for the adjustment of the secondary directions;

[0156] The main direction of the object is determined by the feature vector v1; if it is inconsistent with the main axis direction of the standard object, the posture is adjusted through the rotation matrix.

[0157] (2) Posture adjustment:

[0158] If the orientation of the object is inconsistent with the orientation of the standard object, use the rotation matrix to adjust the object to a uniform posture;

[0159] 1) Rotation Matrix

[0160] Assume that the main axis of the object is v1, and the main axis of the standard object is vQ1. The rotation matrix R is calculated to rotate v1 to vQ1. This is achieved by the Axis-Angle Representation.

[0161] In a preferred solution, in step S3, the calculation process of the axis-angle method is as follows:

[0162] Rotation axis: The rotation axis r is the cross product of vectors v1 and vQ1.

[0163] r=v1×v Q1

[0164] Rotation angle: The rotation angle θ is the angle between v1 and vQ1 and is calculated using the dot product formula:

[0165]

[0166] Among them, ||v1|| and ||v Q1 || are the moduli of v1 and vQ1 respectively.

[0167] Rodriguez rotation formula: The rotation matrix RRR is calculated using the Rodriguez rotation formula:

[0168]

[0169] where I is the identity matrix, [r] X is the antisymmetric matrix of the rotation axis r, defined as:

[0170]

[0171] 2) Apply a rotation transformation

[0172] Once we have the rotation matrix R, we apply it to each point P' of the object i , to achieve posture adjustment:

[0173] ”'

[0174] p i =Rp i

[0175] In this way, the orientation of the object will be aligned with the standard object, ensuring that both are matched in the same coordinate system;

[0176] Among them, the present invention introduces the technology of point cloud registration and posture estimation. First, rough alignment is performed through preliminary registration (such as based on global feature matching), and then precise alignment is performed using a fine registration algorithm (such as ICP). At the same time, the posture estimation and adjustment technology ensures that the object to be detected and the standard part can still be automatically aligned even if the objects are placed in inconsistent directions.

[0177] S4. Matching and Difference Detection

[0178] After alignment, the next step is to perform matching and difference detection; the following geometric distance calculation methods are used to determine whether the object meets the standard:

[0179] Hausdorff distance: Calculate the maximum and minimum distance between two sets of point clouds to determine the geometric difference between them. This distance is more sensitive to large local differences; this distance reflects the distance between the points with the largest geometric deviation in the two point clouds; Hausdorff distance measures the maximum and minimum distance between two point sets; given point clouds P and Q, Hausdorff distance is defined as:

[0180]

[0181] Chamfer distance: used to measure the average distance from each point to the nearest point, suitable for dealing with local shape deviations; Chamfer distance better handles partial occlusion; Chamfer distance is a relaxed form of Hausdorff distance, measuring the average distance from each point to the nearest point:

[0182]

[0183] Point-to-point Euclidean distance: Calculate the Euclidean distance between the two groups of point clouds after registration point by point, mark the areas with large errors as possible abnormal areas; areas with large differences will be displayed as highlighted areas; for each pair of registered points, calculate the Euclidean distance between them, and use histograms or colors to display local geometric differences; this method is suitable for highly aligned point clouds;

[0184] The geometric difference between the object to be detected and the standard part is analyzed by calculating the geometric distance (such as Hausdorff distance, Chamfer distance, etc.). The present invention can accurately identify the geometric errors and potential defects of the parts through point-by-point comparison between point cloud data.

[0185] S5. Difference visualization

[0186] After matching and difference detection, the error heat map or color mapping is used to highlight the areas with large differences, helping operators to quickly determine whether there are defects or deviations in the objects;

[0187] Difference heat map: After alignment, the difference between the object point cloud and the standard point cloud is converted into color values ​​for 3D rendering;

[0188] Difference area annotation: Mark points where the error exceeds the threshold as abnormal areas, and further analyze or detect these areas;

[0189] The present invention uses 3D difference visualization technology to generate an error heat map, which intuitively displays the geometric differences between the object and the standard part in a color-highlighted manner. This visualization method greatly improves the readability and intuitiveness of the test results, making it easier to quickly find problem areas.

[0190] S6. Multi-view point cloud merging

[0191] If the shape of the object is complex, a single-view scan may not be able to capture the complete geometric shape of the object. In order to improve the matching accuracy, a complete point cloud model is generated by merging multi-view point clouds. The core idea of ​​multi-view point cloud merging is to align and fuse point clouds from multiple different perspectives to form a unified 3D model.

[0192] The main steps of multi-view merging:

[0193] Point cloud acquisition: Get point cloud data of objects from multiple angles P1, P2, ..., P k ;

[0194] Global alignment: Use a registration algorithm (such as ICP or feature matching) to align point clouds from different perspectives to ensure that all point clouds are fused in the same coordinate system;

[0195] Point cloud fusion: Use voxel grid filtering to fuse point clouds from different perspectives to ensure uniform point cloud density and remove redundant points; use voxel weighted averaging to smooth the fused point cloud and reduce errors;

[0196] Among them, the present invention provides a multi-view point cloud data merging technology, which generates a more complete 3D model by scanning automotive parts from multiple angles, and is particularly suitable for detecting parts with complex shapes or occluded areas.

[0197] S7, result generation: generating a complete point cloud model containing all viewing angle information.

[0198] In a preferred embodiment, in step S7,

[0199] Finally, the system will output the following:

[0200] Comparison report of differences between finished objects and standard objects;

[0201] Visualized point cloud models and difference heat maps are used to help inspectors make decisions;

[0202] Areas where the difference exceeds the threshold are marked, indicating possible defects.

[0203] The system of the present invention can automatically output a test report, including information such as geometric differences and matching errors, and provide a basis for quality judgment to help quality inspectors make decisions.

[0204] In summary, the present invention can be integrated into the assembly line operation to achieve fully automatic detection of auto parts. The system obtains the three-dimensional point cloud data of the parts through the combination of 3D laser sensors and point cloud data processing, and matches the point cloud data with the point cloud data of standard parts to analyze whether there are deviations or defects in their geometric shapes. This method achieves accurate detection of auto parts by comparing the differences in point cloud data, and highlights the difference parts through visualization, helping manufacturing and quality inspection personnel to more intuitively judge the quality of parts. In addition, the auto parts registration detection method based on 3D laser sensors proposed in the present invention can automatically collect, register, detect and output results, reduce human intervention, and improve detection efficiency. At the same time, the laser emission posture of the 3D laser sensor and the delivery position of the auto parts are calibrated by using a double alignment comparison method, which effectively improves the accuracy of the point cloud data and reduces errors.

[0205] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A registration detection method for automobile parts based on 3D laser sensor, characterized in that: The following steps are involved: S1. Collection of 3D point cloud data: non-contact scanning of the automotive parts to be inspected by a scanning device based on a 3D laser sensor; Equipment configuration: The 3D laser sensor is synchronized with the assembly line to obtain the 3D point cloud data of the object’s surface through laser scanning and generate the initial point cloud model; Multi-angle scanning: For parts with complex geometric shapes, multi-angle scanning is used; Data preprocessing: The collected point cloud data needs to be preprocessed; S2. Registration of point cloud data: Based on the preprocessed point cloud, the point cloud of the part to be inspected is aligned with the point cloud of the standard part through the point cloud registration algorithm. The registration method includes the following stages: (1) Calibrate and determine the laser posture: align and compare the 3D point cloud data with the 3D model of the measured product; The steps of alignment include: Selecting N points on the 3D point cloud data and the 3D model respectively, N>3, and the N points in the 3D point cloud data and the N points in the 3D model are not coplanar; An affine matrix based on N points is obtained by matrix transformation calculation, and the obtained affine matrix is ​​used as an initial value for ICP alignment; The affine matrix of the 3D point cloud data and the 3D model is calculated by ICP alignment, and the laser posture of the laser relative to the product is obtained by aligning the calculated affine matrix; (2) Preliminary alignment; (3) Feature matching; Use the following steps: Calculate point cloud feature descriptors; Feature matching: Match the feature descriptors of the collected point cloud with those of the standard point cloud, find similar point pairs, and use these point pairs to estimate the rough alignment transformation of the two sets of point clouds; (4) Shape-based alignment; Estimate the orientation of objects through shape matching and perform preliminary alignment; (5) Precise alignment; Use the iterative closest point algorithm (ICP) for precise registration; S3, attitude estimation and adjustment; (2) Posture adjustment: Use rotation matrix to adjust objects to a uniform posture; 1) Rotation matrix; 2) Apply a rotation transformation; S4. Matching and Difference Detection Detection is done through the following geometric distance calculation methods: Hausdorff distance: Calculate the maximum and minimum distance between two sets of point clouds to determine the geometric differences between them; Chamfer distance: used to measure the average distance from each point to the nearest point; Point-to-point Euclidean distance; S5. Difference visualization Difference heat map: After alignment, the difference between the object point cloud and the standard point cloud is converted into color values ​​for 3D rendering; Difference area annotation: Mark points where the error exceeds the threshold as abnormal areas and analyze or detect these areas; S6, multi-view point cloud merging; S7, result generation: generating a complete point cloud model containing all viewing angle information.

2. The method for detecting automobile parts registration based on 3D laser sensor according to claim 1, characterized in that: The scanning device based on the 3D laser sensor in step S1 is installed above the assembly line, and the 3D point cloud data of the surface of the object is collected in real time as the object passes through the assembly line.

3. The automobile parts registration detection method based on 3D laser sensor according to claim 1, characterized in that: The data preprocessing method in step S1 is as follows: (1) Denoising: Use statistical filtering or radius filtering to remove noise from the scan; (2) Downsampling: The density of the point cloud is reduced by downsampling the voxel grid, which reduces the amount of data while retaining the geometric features of the point cloud.

4. The method for detecting automobile parts registration based on 3D laser sensor according to claim 1, characterized in that: The iterative closest point algorithm (ICP) in step S2 is as follows: The core principle of the ICP algorithm: ICP iteratively finds the closest point pair in two point clouds, calculates the error between the point pairs, and then continuously optimizes the rotation matrix R and the translation matrix T to minimize the following objective function: where p i is a point in the standard point cloud, q i It is to collect points in the point cloud, and gradually align the two sets of point clouds by iteratively optimizing R and T. ICP improvement: In order to avoid falling into the local optimal solution, different heuristic strategies or multi-resolution ICP schemes are combined to make the alignment more robust.

5. The automobile parts registration detection method based on 3D laser sensor according to claim 1, characterized in that: In step S3, the centroid is calculated as follows: Given a point cloud P = {p1, p2, ..., pN} of an object, where each point p i =(x i ,y i ,z i ) is the 3D coordinate in the point cloud, the centroid C P The calculation formula is: The centroid represents the position of the object in three-dimensional space; in order to align the object point cloud with the standard point cloud, the centroids of the two need to be aligned through translation transformation.

6. The automobile parts registration detection method based on 3D laser sensor according to claim 1, characterized in that: In step S3, By performing eigenvalue decomposition on the covariance matrix Σ, the eigenvalues ​​λ1,λ2,λ3 and the corresponding eigenvectors v1,v2,v3 are obtained; The eigenvector v1 corresponds to the maximum eigenvalue λ1, which represents the direction of the main axis of the object; The eigenvectors v2 and v3 correspond to the secondary principal axis directions and are usually used for the adjustment of the secondary directions; The main direction of the object is determined by the feature vector v1; if it is inconsistent with the main axis direction of the standard object, the posture is adjusted through the rotation matrix.

7. The method for detecting automobile parts registration based on 3D laser sensor according to claim 1, characterized in that: In step S3, the calculation process of the axis-angle method is as follows: Rotation axis: The rotation axis r is the cross product of vectors v1 and vQ1 r=v1×v Q1 Rotation angle: The rotation angle θ is the angle between v1 and vQ1 and is calculated using the dot product formula: Among them, ||v1|| and ||v Q1 || are the moduli of v1 and vQ1 respectively. Rodriguez rotation formula: The rotation matrix RRR is calculated using the Rodriguez rotation formula: where I is the identity matrix, [r] X is the antisymmetric matrix of the rotation axis r, defined as:

8. The automobile parts registration detection method based on 3D laser sensor according to claim 1, characterized in that: In step S7, Finally, the system will output the following: Comparison report of differences between finished objects and standard objects; Visualized point cloud models and difference heat maps are used to help inspectors make decisions; Areas where the difference exceeds the threshold are marked, indicating the presence of defects.

Images