Novel vehicle body part detection algorithm
By combining contour recognition methods based on normal vectors and nearest neighbors, along with maximal clique point cloud registration and the PointNN model, the problems of low recognition accuracy and inaccurate assembly evaluation in traditional vehicle body parts inspection are solved, achieving efficient and robust vehicle body parts recognition and assembly quality evaluation.
Patent Information
- Application Number
- CN202510704355.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-28
- Publication Date
- 2025-10-31
AI Technical Summary
Existing methods for inspecting body parts rely on manual labor or two-dimensional image processing, resulting in low recognition accuracy, poor efficiency, and susceptibility to subjective factors. Quality assessment based on historical data faces the challenge of point cloud registration. Traditional point cloud registration algorithms are prone to getting stuck in local optima when initial position errors are large or data is missing, affecting the accuracy of assembly quality assessment.
A vehicle body contour recognition method based on normal vectors and nearest neighbors is adopted, which combines maximal clique point cloud registration and PointNN nonparametric point cloud matching model. Through data preprocessing, point cloud registration, feature segmentation and quality assessment, efficient recognition of vehicle body parts and accurate assembly quality assessment are achieved.
It significantly improves the accuracy of body part identification and the precision of assembly quality assessment, maintains robustness under complex environments and initial error conditions, supports efficient identification of various part types and sizes, and provides fine-grained assembly quality analysis.
Smart Images
Figure CN120876352A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent manufacturing, computer vision, and computer graphics, specifically to a novel algorithm for detecting vehicle body parts. Background Technology
[0002] With the continuous development of the automobile manufacturing industry, the manufacturing precision and assembly quality of body parts have an increasingly significant impact on the performance of the whole vehicle. Traditional body part inspection methods usually rely on manual visual inspection or inspection technology based on two-dimensional images. Although these methods are simple to operate, they generally have the disadvantages of low recognition accuracy, insufficient efficiency, and great influence from human subjective factors, making it difficult to meet the higher requirements of modern manufacturing industry for precision and automation. In recent years, with the rapid development of 3D scanning technology and computer vision technology, body part recognition and assembly quality assessment based on point cloud data have gradually become the mainstream. Point cloud data can comprehensively and accurately reflect the spatial geometric features of body parts, which helps to improve the accuracy and automation of part recognition and quality assessment. However, point cloud data itself also has problems such as uneven distribution, data redundancy and noise interference, which directly affect the effectiveness of subsequent part contour recognition and assembly quality assessment algorithms. Therefore, developing effective point cloud data preprocessing methods and efficient feature extraction technology has become an important issue that needs to be addressed.
[0003] Currently, the mainstream methods for identifying body parts typically include traditional geometric feature matching algorithms and machine learning-based classification and recognition algorithms. Although these methods are intuitive and easy to understand, they are not well-suited for complex and diverse body parts. While machine learning-based algorithms can effectively improve recognition accuracy, they require a large amount of labeled data for training, resulting in strong data dependence and insufficient generalization ability. In terms of parts assembly quality assessment, point cloud registration (such as the ICP algorithm) is often used to compare the point cloud data of the part to be inspected with that of a standard part to determine assembly quality. However, the ICP algorithm is prone to getting trapped in local optima and reduced matching accuracy when faced with large initial position errors, missing point cloud data, or significant interference, which seriously affects the accuracy and reliability of assembly assessment. Summary of the Invention
[0004] This invention provides a novel vehicle body part detection algorithm that effectively addresses the problems mentioned in the background art, such as the reliance on manual or two-dimensional image processing methods for traditional vehicle part detection, which suffers from low accuracy, poor efficiency, and susceptibility to subjective factors. Furthermore, in the quality assessment process based on historical data, there is often a problem of misalignment between historical point clouds and target point clouds. Without considering point cloud registration, the quality of point cloud matching will be significantly reduced. Traditional point cloud registration and matching algorithms are prone to getting trapped in local optima when faced with large initial position errors, missing data, or noise interference, thus affecting the accuracy of assembly quality assessment.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a novel vehicle body part detection algorithm, comprising the steps of: A. identifying the outline of the vehicle body part, and B. evaluating the assembly quality of the part;
[0006] The A includes the following steps: A1 vehicle body contour recognition based on normal vectors and nearest neighbor points, and A2 part size information analysis.
[0007] The A1 includes the following steps:
[0008] A101, Data Preprocessing;
[0009] A102. Preliminary contour recognition based on normal vectors;
[0010] A103. Contour recognition based on nearest neighbor points;
[0011] A104, Results Fusion;
[0012] The A2 includes the following steps:
[0013] A201. Obtain the outline of each point cloud;
[0014] A202, Remove outliers;
[0015] A203. Obtain the main direction of the current point cloud;
[0016] A204. Solve for various shapes;
[0017] A205. Output and save the information;
[0018] The B includes the following steps: B1 vehicle body point cloud registration and point cloud matching algorithm, B2 point cloud feature segmentation and correspondence, and B3 quality assessment.
[0019] The B1 includes the following steps:
[0020] B101, Point cloud registration based on maximal cliques;
[0021] B102, PointNN nonparametric point cloud matching model;
[0022] The B2 includes the following steps:
[0023] B201. Segmentation based on part inspection results;
[0024] B202, Local correspondence after segmentation;
[0025] The B3 includes the following steps:
[0026] B301, Deviation Calculation;
[0027] B302. Assembly tolerance judgment;
[0028] B303, Automated assessment and report generation.
[0029] According to the above technical solution, in the A2, the shape inside the part is an important feature of the part in the vehicle body point cloud. It is necessary to extract fine-grained features inside the part. The shape features include round holes, square holes, and oblong holes. In addition to extracting the shape, it is also necessary to extract the radius and length and width size features to facilitate the subsequent identification and evaluation of the point cloud parts.
[0030] According to the above technical solution, in B1, the part assembly quality assessment needs to find the point cloud data that is closest to the target point cloud in the historical database. In actual industrial scenarios, the similar point clouds obtained often have problems such as inconsistent measurement coordinate systems and different poses. In order to better compare and analyze the historical data or CAD model with the currently measured point cloud, it is necessary to perform registration first so that the two point clouds are aligned under the same coordinate reference system.
[0031] According to the above technical solution, after completing point cloud alignment and matching, B2 enters the body part detection and feature segmentation stage. The goal is to divide the overall point cloud into several local regions with consistent geometric or functional meaning based on the matching information obtained in the previous step, and then establish a finer-grained correspondence with the target point cloud and CAD model.
[0032] According to the above technical solution, B3 is the quality assessment stage after point cloud registration and matching are completed and each part or local area is accurately matched. It is necessary to make a detailed comparison between the measured point cloud and the target data to obtain the error range and deviation distribution of each part or assembly part in order to determine the qualification status of the assembly.
[0033] According to the above technical solution, before processing the vehicle body point cloud, A101 needs to perform coordinate system 1, noise removal and outlier detection, and use statistical filtering or radius filtering methods to remove noise and outliers to obtain purer and more representative vehicle body point cloud data.
[0034] A102, the normal vector is an important indicator for measuring the local features of a point cloud. To calculate the normal vector of each point, the nearest neighbor query of the point cloud is used to calculate and obtain the direction of the normal vector. Point cloud nearest neighbors: k-neighborhood, radius neighborhood;
[0035] For a point p in the point cloud, select several neighborhood point sets, calculate the covariance matrix and perform eigenvalue decomposition to obtain the principal direction vector, which is the normal vector of point p.
[0036] The specific calculation process is as follows: First, calculate the mean and covariance of each point in sequence. Then, perform eigenvalue decomposition based on the covariance, and use the final eigenvector as the local normal vector of that point.
[0037]
[0038] In the formula, For p i The k nearest neighbors are used to determine the vehicle body boundary region based on the variation law of the normal vector after obtaining the normal vector;
[0039] Rapid changes in the direction and magnitude of the normal vector reflect the discontinuity of the object's surface;
[0040] After completing the initial recognition based on normal vectors, A103 optimizes contour extraction using features based on the distribution of nearest neighbors, achieved in the following way:
[0041] Neighborhood density analysis: For certain regions on the vehicle body with significant geometric changes, the point cloud density is often reduced or the distribution shows obvious local differences. The number of neighboring points or density distribution characteristics of each point can be calculated.
[0042] When the distance and distribution between a point and its neighborhood exhibit abrupt changes, the region containing that point lies on the outer boundary. This is addressed by using the distance from points within the neighborhood to point p. i The density is represented by the reciprocal of the average distance. The smaller the distance, the more compact the local distribution, and therefore the greater the density. That is:
[0043]
[0044] The A104 combines the recognition results based on normal vectors and those based on nearest neighbors, using the following strategy:
[0045] Joint confidence assessment: Assign confidence to each candidate contour segment, including assigning higher confidence to segments in which both normal vector jumps and neighborhood density abrupt changes are detected;
[0046] Weighted fusion: The results of different recognition methods are weighted. Specifically, if a region has a significant jump in normal vector, the result of normal vector recognition is given more weight. If a region has a large change in neighborhood density or curvature, the result based on nearest neighbor points is given more weight.
[0047] Post-processing optimization includes connecting contour segments, repairing cracks, and removing redundant boundaries to make the contours smoother and more complete.
[0048] According to the above technical solution, A201 uses a density-based clustering algorithm to divide the point cloud contours to obtain the contours of each point cloud block.
[0049] A202 uses a geometric rule-based method to remove outliers, with the geometric rule being: k nearest neighbors;
[0050] The A203 uses SVD decomposition to decompose the decentralized point matrix to obtain the main direction of the current point cloud:
[0051] P=UΣV T ;
[0052] Where P is the centered point matrix, the principal direction of the point cloud and the corresponding plane equation are obtained by solving for the eigenvalues and eigenvectors of Σ: Finally, all points on the 3D point cloud are mapped to the corresponding 2D plane for specific shape recognition;
[0053] The A204, after obtaining the 2D scatter plot, uses a constrained quadratic programming algorithm to solve for various shapes and obtains the corresponding shape's size parameters;
[0054] The A205 will output and save the extracted body parts and the information on the holes in the parts.
[0055] According to the above technical solution, B101 infers the optimal rigid body transformation by searching for maximal cliques in the feature correspondence between two point clouds. The rigid body transformation includes rotation and translation. To ensure the efficiency of searching for maximal cliques, geometric features are extracted from the point cloud, including key points and local descriptors. Feature matching pairs are then established. The specific process is as follows:
[0056] First, a compatibility graph is constructed to reflect the relationship between the initial corresponding point clouds;
[0057] Then, search for maximal cliques in the graph. Each maximal clique represents a set of consensus. Through node-guided clique selection, select the maximal clique with the largest graph weight.
[0058] Next, during the search for maximal clusters, the total number of maximal clusters is limited to prevent the search time from becoming too long;
[0059] Finally, the transformation hypothesis of the selected clique is calculated using the SVD algorithm, the best hypothesis is selected for registration, and appropriate pruning and pre-screening strategies are implemented to avoid a surge in computational load on large-scale point clouds and to ensure a balance between efficiency and accuracy.
[0060] The B102 model is used to perform point cloud matching tasks after registering two point clouds. The Point-NN consists of a non-parametric encoder for 3D feature extraction and a point cloud memory for task-specific recognition.
[0061] The nonparametric encoder adopts a multi-stage structural design, using farthest point sampling, k-nearest neighbor, trigonometric functions and pooling to gradually aggregate local geometry, generating a high-dimensional global feature for the point cloud;
[0062] Its model includes a nonparametric encoder and a point memory. For the input point cloud, it uses EncNP to extract its global features and uses PoM feature similarity matching to output the classification result.
[0063] f G =EncNP(P); logits = PoM(f G );
[0064] The similarity between point clouds is determined by using the vector embeddings output by PointNN.
[0065] According to the above technical solution, B201 is used when there are different parts or components on the vehicle body, including doors, engine hoods, and fenders. With the help of the previous matching or prior knowledge, the point cloud is divided into different segments by point cloud coordinate distribution, normal vector changes, and existing detection results.
[0066] For specific segmentation methods, the methods used include: region-based growth and clustering-based methods. Since the spatial location or shape features of the parts in the vehicle body are known, the segmentation efficiency is accelerated by combining detection boxes or key feature points.
[0067] After obtaining the point cloud of each sub-part, B202 performs a more detailed correspondence with the target point cloud. It uses local geometric descriptors or features to determine the alignment error of each local part. When there are internal fine-grained features, including holes, connectors, and welds, the position and orientation of these geometric elements can be further calculated in the segmented point cloud and compared with the design value of the target part, providing rich deviation information in the later evaluation.
[0068] According to the above technical solution, B301 uses the simplest strategy to calculate the Euclidean distance ||p for each corresponding point. i -q i ||, calculate the average, maximum, and root mean square of the overall error;
[0069] In engineering applications, more attention is paid to deviations in key features, including key holes, connecting surfaces, and bent edges. During the judgment process, multiple features are compared based on round holes, rectangular holes, and weld points in the part. The features have been extracted and have accurate information. Therefore, the target point cloud part is judged to meet the standard through two aspects of judgment.
[0070] The B302 sets the corresponding allowable deviation range according to the design tolerance of the whole vehicle or part. If the local error exceeds a certain threshold, it is marked as unqualified or needs to be reworked. It can also further analyze different assembly links of welding and sheet metal forming to determine which stage causes the maximum error in order to improve the process flow.
[0071] The B303 integrates error distribution, deviation statistics, and pass rate into the quality assessment system, automatically generates assembly quality reports, and compares and analyzes historical data. It can continuously track the quality stability of the production line and provide early warnings of potential risks.
[0072] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0073] 1. By combining contour recognition based on normal vectors and contour recognition based on nearest neighbor information, the contours obtained by the two methods are combined and optimized to effectively remove outliers and improve the accuracy of recognition results. Specific optimizations have been made to address the failure of geometric rule-based methods in point cloud data of different sizes, significantly improving the robustness and adaptability of the algorithm. This ensures efficient and stable recognition under various part types and sizes. Furthermore, it supports the recognition of various geometric shapes in point clouds, such as circular holes, square holes, oblong holes, and polygonal holes, demonstrating strong versatility and practicality. Through the design of part detection from coarse to fine granularity, compared to traditional part detection algorithms, the algorithm has more and more practical functions, enabling fine-grained disassembly and analysis of automotive parts. Moreover, this detection algorithm is independent of the size of the point cloud and unaffected by outliers, making it more stable in real-world scenarios.
[0074] 2. By constructing a high-quality point cloud database of historical car body parts, a large amount of rigorously verified and qualified point cloud data of car body parts is stored. The historical data covers feature information of various part types and different manufacturing conditions, providing rich reference for subsequent registration and matching algorithms. The cutting-edge point cloud registration algorithm based on maximal cliques is adopted to align the point cloud of the target part with the point cloud of the parts in the historical database. By adopting the optimized registration method, the matching error caused by initial position error, data missing or external environmental interference can be effectively reduced, ensuring that the registration process is more stable and accurate. After the initial alignment is completed, the point cloud matching algorithm based on PointNN is further used for processing to accurately find the historical part data most similar to the target part.
[0075] 3. After determining the most similar historical point cloud data, this invention uses a feature calculation method to divide the point clouds of the target part and historical parts into multiple feature regions, and compares the point clouds one by one. This part-by-part feature comparison method can not only comprehensively capture the minor deviations and anomalies that may exist in the part assembly process, but also quickly identify the problems in the part assembly process by calculating the one-to-one correspondence of each feature part, thereby significantly improving the accuracy and detail of automobile assembly quality assessment. By deeply integrating historical data with advanced point cloud technology, a more accurate, efficient and reliable automobile part assembly quality assessment is achieved.
[0076] In summary, by constructing a more robust contour recognition algorithm, the challenge of automated identification of complex parts can be effectively solved. At the same time, by introducing historical part quality data and adopting an improved algorithm of point cloud matching and registration, the accuracy and stability of part assembly quality assessment can be effectively improved to meet the high-precision, high-efficiency and intelligent development needs of modern automobile manufacturing. Compared with traditional methods, it exhibits higher robustness in the face of complex environmental interference and initial errors, and can effectively improve the accuracy of assembly quality control. Attached Figure Description
[0077] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.
[0078] In the attached diagram:
[0079] Figure 1 This is a schematic diagram illustrating the specific content of the algorithm of this invention;
[0080] Figure 2 This is a flowchart of the steps of the algorithm of the present invention. Detailed Implementation
[0081] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0082] Example: Figure 1-2 As shown, the present invention provides a technical solution, a novel vehicle body part detection algorithm, including the steps of: A. identifying the outline of the vehicle body part, and B. evaluating the assembly quality of the part.
[0083] A includes the following steps: A1 Vehicle body contour recognition based on normal vectors and nearest neighbor points, and A2 Part size information analysis;
[0084] A1 includes the following steps:
[0085] A101, Data Preprocessing;
[0086] A102. Preliminary contour recognition based on normal vectors;
[0087] A103. Contour recognition based on nearest neighbor points;
[0088] A104, Results Fusion;
[0089] A2 includes the following steps:
[0090] A201. Obtain the outline of each point cloud;
[0091] A202, Remove outliers;
[0092] A203. Obtain the main direction of the current point cloud;
[0093] A204. Solve for various shapes;
[0094] A205. Output and save the information;
[0095] B includes the following steps: B1 vehicle body point cloud registration and point cloud matching algorithm, B2 point cloud feature segmentation and correspondence, and B3 quality assessment.
[0096] B1 includes the following steps:
[0097] B101, Point cloud registration based on maximal cliques;
[0098] B102, PointNN nonparametric point cloud matching model;
[0099] B2 includes the following steps:
[0100] B201. Segmentation based on part inspection results;
[0101] B202, Local correspondence after segmentation;
[0102] B3 includes the following steps:
[0103] B301, Deviation Calculation;
[0104] B302. Assembly tolerance judgment;
[0105] B303, Automated assessment and report generation.
[0106] Based on the above technical solution, A2, in the vehicle body point cloud, the shape inside the part is an important feature of the part. It is necessary to extract fine-grained features inside the part. The shape features include round holes, square holes, and oblong holes. In addition to extracting the shape, it is also necessary to extract the radius and length and width size features to facilitate the subsequent identification and evaluation of the point cloud parts.
[0107] Based on the above technical solution, B1, the quality assessment of part assembly requires finding the point cloud data that is closest to the target point cloud in the historical database. In actual industrial scenarios, the similar point clouds obtained often have problems such as inconsistent measurement coordinate systems and different poses. In order to better compare and analyze the historical data or CAD model with the currently measured point cloud, it is necessary to perform registration first so that the two point clouds are aligned under the same coordinate reference system.
[0108] Based on the above technical solution, after completing point cloud alignment and matching, B2 enters the body part detection and feature segmentation stage. The goal is to segment the overall point cloud into several local regions with consistent geometric or functional meaning based on the matching information obtained in the previous step, and then establish a finer-grained correspondence with the target point cloud and CAD model to facilitate subsequent quality assessment.
[0109] Based on the above technical solution, B3 is to enter the quality assessment stage after completing point cloud registration and matching and accurately corresponding each part or local area. It is necessary to make a detailed comparison between the measured point cloud and the target data to obtain the error range and deviation distribution of each part or assembly part in order to determine the qualification status of the assembly.
[0110] Based on the above technical solution, before processing the vehicle body point cloud, A101 needs to perform coordinate system 1, noise removal and outlier detection. Statistical filtering or radius filtering methods are used to remove noise and outliers to obtain cleaner and more representative vehicle body point cloud data. The quality of this step directly affects the accuracy of subsequent normal vector calculation and contour recognition.
[0111] A102, the normal vector is an important indicator for measuring the local features of a point cloud. To calculate the normal vector of each point, the nearest neighbor query of the point cloud is used to calculate and obtain the direction of the normal vector. Point cloud nearest neighbors: k-neighborhood, radius neighborhood;
[0112] For a point p in the point cloud, select several neighborhood point sets, calculate the covariance matrix and perform eigenvalue decomposition to obtain the principal direction vector, which is the normal vector of point p.
[0113] The specific calculation process is as follows: First, calculate the mean and covariance of each point in sequence. Then, perform eigenvalue decomposition based on the covariance, and use the final eigenvector as the local normal vector of that point.
[0114]
[0115] In the formula, For p i The k nearest neighbors are used to determine the vehicle body boundary region based on the variation law of the normal vector after obtaining the normal vector;
[0116] Rapid changes in the direction and magnitude of the normal vector reflect the discontinuity of the object's surface. When the angle between the normal vectors of adjacent regions increases significantly, or when the normal vectors change dramatically, it indicates the edges, broken lines, or openings of the vehicle's shape. Combined with threshold settings, the abruptly changed regions can be initially marked as candidate contour regions.
[0117] A103, after completing the initial recognition based on normal vectors, optimizes contour extraction using features based on the distribution of nearest neighbors, achieved in the following way:
[0118] Neighborhood density analysis: For certain regions on the vehicle body with significant geometric changes, the point cloud density is often reduced or the distribution shows obvious local differences. The number of neighboring points or density distribution characteristics of each point can be calculated.
[0119] When the distance and distribution between a point and its neighborhood exhibit abrupt changes, the region containing that point lies on the outer boundary. This is addressed by using the distance from points within the neighborhood to point p. i The density is represented by the reciprocal of the average distance. The smaller the distance, the more compact the local distribution, and therefore the greater the density. That is:
[0120]
[0121] In addition, when the point cloud density is relatively uniform, a contour point recognition method based on the number of neighboring points and average density is developed. The contour information obtained by the nearest neighbor method is more inclined to supplement local features, while the result obtained by the normal vector method provides a more overall and continuous contour direction. The two corroborate each other and improve the integrity and accuracy of the contour line.
[0122] A104 combines the recognition results based on normal vectors and those based on nearest neighbors, employing the following strategy:
[0123] Joint confidence assessment: Assign confidence to each candidate contour segment, including assigning higher confidence to segments in which both normal vector jumps and neighborhood density abrupt changes are detected;
[0124] Weighted fusion: The results of different recognition methods are weighted. Specifically, if a region has a significant jump in normal vector, the result of normal vector recognition is given more weight. If a region has a large change in neighborhood density or curvature, the result based on nearest neighbor points is given more weight.
[0125] Post-processing optimization includes connecting contour segments, repairing cracks, and removing redundant boundaries to make the contours smoother and more complete.
[0126] Based on the above technical solution, A201 uses the density-based clustering algorithm DBSCAN to divide the point cloud contours and obtain the contours of each point cloud block.
[0127] A202 uses a geometric rule-based method to remove outliers. The geometric rule is: k nearest neighbors.
[0128] A203 uses SVD decomposition to decompose the decentralized point matrix to obtain the main direction of the current point cloud:
[0129] P=UΣV T ;
[0130] Where P is the centered point matrix, the principal direction of the point cloud and the corresponding plane equation are obtained by solving for the eigenvalues and eigenvectors of Σ: Finally, all points on the 3D point cloud are mapped to the corresponding 2D plane for specific shape recognition;
[0131] A204: After obtaining the 2D scatter plot, a constrained quadratic programming algorithm is used to solve for various shapes and obtain the corresponding shape's size parameters.
[0132] A205 will output and save the extracted body parts and the information on the holes in the parts.
[0133] Based on the above technical solution, B101 infers the optimal rigid body transformation by searching for maximal cliques in the feature correspondence between two point clouds. Rigid body transformations include rotation and translation. To ensure the efficiency of maximal clique search, geometric features are extracted from the point cloud, including keypoints and local descriptors. Feature matching pairs are then established. The specific process is as follows:
[0134] First, a compatibility graph is constructed to reflect the relationship between the initial corresponding point clouds;
[0135] Then, search for maximal cliques in the graph. Each maximal clique represents a set of consensus. Through node-guided clique selection, select the maximal clique with the largest graph weight.
[0136] Next, during the search for maximal clusters, the total number of maximal clusters is limited to prevent the search time from becoming too long;
[0137] Finally, the transformation hypothesis of the selected clique is calculated using the SVD algorithm, the best hypothesis is selected for registration, and appropriate pruning and pre-screening strategies are implemented to avoid a surge in computational load on large-scale point clouds and to ensure a balance between efficiency and accuracy.
[0138] B102 uses the cutting-edge PointNN model to perform point cloud matching tasks after registering two point clouds. Point-NN consists of a non-parametric encoder for 3D feature extraction and a point cloud memory bank for task-specific recognition.
[0139] The nonparametric encoder adopts a multi-stage structural design, using farthest point sampling (FPS), k-nearest neighbor (k-NN), trigonometric functions, and pooling to gradually aggregate local geometry and generate a high-dimensional global feature for the point cloud.
[0140] Its model includes a non-parametric encoder, EncNP, and a point-memory bank, PoM. For the input point cloud, EncNP is used to extract its global features, and the classification result is output through feature similarity matching using PoM.
[0141] f G =EncNP(P); logits = PoM(f G );
[0142] The similarity between point clouds is determined by using the vector embeddings output by PointNN.
[0143] Based on the above technical solution, B201 is used when there are different parts or components on the vehicle body, including doors, engine hoods, and fenders. With the help of the previous matching or prior knowledge, the point cloud is divided into different segments by point cloud coordinate distribution, normal vector changes, and existing detection results.
[0144] For specific segmentation methods, the methods used include: region-based growth and clustering-based methods. Since the spatial location or shape features of the parts in the vehicle body are known, the segmentation efficiency is accelerated by combining detection boxes or key feature points.
[0145] B202 After obtaining the point cloud of each sub-part, it performs a more detailed correspondence with the target point cloud. It uses local geometric descriptors or features to determine the alignment error of each local part. When there are internal fine-grained features, including holes, connectors, and welds, the position and orientation of these geometric elements can be further calculated in the segmented point cloud and compared with the design value of the target part, providing rich deviation information in the later evaluation.
[0146] By employing a detection-segmentation-correspondence approach, the vehicle body point cloud is subdivided layer by layer, enabling clearer observation and inspection of local quality or deviations while preserving the overall assembly relationship.
[0147] Based on the above technical solution, B301 uses the simplest strategy to calculate the Euclidean distance ||p for each corresponding point. i -q i ||, Statistical analysis of the overall error: average, maximum, and root mean square (RMS).
[0148] In engineering applications, more attention is paid to deviations in key features, including key holes, connecting surfaces, and bent edges. During the judgment process, multiple features are compared based on round holes, rectangular holes, and weld points in the part. The features have been extracted and have accurate information. Therefore, the target point cloud part is judged to meet the standard through two aspects of judgment.
[0149] B302 sets the corresponding allowable deviation range based on the design tolerance of the whole vehicle or part. If the local error exceeds a certain threshold, it is marked as unqualified or needs to be reworked. Further analysis can be conducted on different assembly links of welding and sheet metal forming to determine which stage causes the maximum error in order to improve the process flow.
[0150] B303 integrates error distribution, deviation statistics, and pass rate into the quality assessment system, automatically generates assembly quality reports, and compares and analyzes historical data to continuously track the quality stability of the production line and provide early warnings of potential risks.
[0151] Through the above deviation calculation, feature evaluation, and tolerance judgment, the automated assembly quality assessment at the vehicle or part level is completed.
[0152] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A novel vehicle body part detection algorithm, characterized in that: The steps include: A. Identifying the outlines of body parts; B. Assessing the assembly quality of the parts. The A includes the following steps: A1 vehicle body contour recognition based on normal vectors and nearest neighbor points, and A2 part size information analysis. The A1 includes the following steps: A101, Data Preprocessing; A102. Preliminary contour recognition based on normal vectors; A103. Contour recognition based on nearest neighbor points; A104, Results Fusion; The A2 includes the following steps: A201. Obtain the outline of each point cloud; A202, Remove outliers; A203. Obtain the main direction of the current point cloud; A204. Solve for various shapes; A205. Output and save the information; The B includes the following steps: B1 vehicle body point cloud registration and point cloud matching algorithm, B2 point cloud feature segmentation and correspondence, and B3 quality assessment. The B1 includes the following steps: B101, Point cloud registration based on maximal cliques; B102, PointNN nonparametric point cloud matching model; The B2 includes the following steps: B201. Segmentation based on part inspection results; B202, Local correspondence after segmentation; The B3 includes the following steps: B301, Deviation Calculation; B302. Assembly tolerance judgment; B303, Automated assessment and report generation.
2. The novel vehicle body part detection algorithm according to claim 1, characterized in that: In the A2 section, the shape inside the part is an important feature of the part in the vehicle body point cloud. Fine-grained feature extraction is required for the part's interior. Shape features include round holes, square holes, and oblong holes. In addition to extracting the shape, the radius and length and width dimension features also need to be extracted to facilitate the subsequent identification and evaluation of the point cloud part.
3. The novel vehicle body part detection algorithm according to claim 1, characterized in that: In B1, the quality assessment of part assembly requires finding the point cloud data that is closest to the target point cloud in the historical database. In actual industrial scenarios, the similar point clouds obtained often have problems such as inconsistent measurement coordinate systems and different poses. In order to better compare and analyze the historical data or CAD model with the currently measured point cloud, it is necessary to perform registration first so that the two point clouds are aligned under the same coordinate reference system.
4. The novel vehicle body part detection algorithm according to claim 1, characterized in that: After completing point cloud alignment and matching, B2 enters the vehicle body part detection and feature segmentation stage. The goal is to segment the overall point cloud into several local regions with consistent geometric or functional meaning based on the matching information obtained in the previous step, and then establish a finer-grained correspondence with the target point cloud and CAD model.
5. The novel vehicle body part detection algorithm according to claim 1, characterized in that: The B3 stage, after point cloud registration and matching are completed and each part or local area is accurately matched, is the quality assessment stage. It requires a detailed comparison between the measured point cloud and the target data to determine the error range and deviation distribution of each part or assembly part in order to determine the qualification status of the assembly.
6. The novel vehicle body part detection algorithm according to claim 1, characterized in that: Before processing the vehicle body point cloud, the A101 needs to perform coordinate system 1, noise removal and outlier detection. Statistical filtering or radius filtering methods are used to remove noise and outliers to obtain cleaner and more representative vehicle body point cloud data. A102, the normal vector is an important indicator for measuring the local features of a point cloud. To calculate the normal vector of each point, the nearest neighbor query of the point cloud is used to calculate and obtain the direction of the normal vector. Point cloud nearest neighbors: k-neighborhood, radius neighborhood; For a point p in the point cloud, select several neighborhood point sets, calculate the covariance matrix and perform eigenvalue decomposition to obtain the principal direction vector, which is the normal vector of point p. The specific calculation process is as follows: First, calculate the mean and covariance of each point in sequence. Then, perform eigenvalue decomposition based on the covariance, and use the final eigenvector as the local normal vector of that point. In the formula, For p i The k nearest neighbors are used to determine the vehicle body boundary region based on the variation law of the normal vector after obtaining the normal vector; Rapid changes in the direction and magnitude of the normal vector reflect the discontinuity of the object's surface; After completing the initial recognition based on normal vectors, A103 optimizes contour extraction using features based on the distribution of nearest neighbors, achieved in the following way: Neighborhood density analysis: For certain regions on the vehicle body with significant geometric changes, the point cloud density is often reduced or the distribution shows obvious local differences. The number of neighboring points or density distribution characteristics of each point can be calculated. When the distance and distribution between a point and its neighborhood exhibit abrupt changes, the region containing that point lies on the outer boundary. This is addressed by using the distance from points within the neighborhood to point p. i The density is represented by the reciprocal of the average distance. The smaller the distance, the more compact the local distribution, and therefore the greater the density. That is: The A104 combines the recognition results based on normal vectors and those based on nearest neighbors, using the following strategy: Joint confidence assessment: Assign confidence to each candidate contour segment, including assigning higher confidence to segments in which both normal vector jumps and neighborhood density abrupt changes are detected; Weighted fusion: The results of different recognition methods are weighted. Specifically, if a region has a significant jump in normal vector, the result of normal vector recognition is given more weight. If a region has a large change in neighborhood density or curvature, the result based on nearest neighbor points is given more weight. Post-processing optimization includes connecting contour segments, repairing cracks, and removing redundant boundaries to make the contours smoother and more complete.
7. A novel vehicle body part detection algorithm according to claim 2, characterized in that: The A201 uses a density-based clustering algorithm to divide the point cloud contours to obtain the contours of each point cloud block. A202 uses a geometric rule-based method to remove outliers, with the geometric rule being: k nearest neighbors; The A203 uses SVD decomposition to decompose the decentralized point matrix to obtain the main direction of the current point cloud: Where P is the centered point matrix, the principal direction of the point cloud and the corresponding plane equation are obtained by solving for the eigenvalues and eigenvectors of Σ: Finally, all points on the 3D point cloud are mapped to the corresponding 2D plane for specific shape recognition; The A204, after obtaining the 2D scatter plot, uses a constrained quadratic programming algorithm to solve for various shapes and obtains the corresponding shape's size parameters; The A205 will output and save the extracted body parts and the information on the holes in the parts.
8. A novel vehicle body part detection algorithm according to claim 3, characterized in that: The B101 method infers the optimal rigid body transformation by searching for maximal cliques in the feature correspondence between two point clouds. Rigid body transformations include rotation and translation. To ensure efficiency in searching for maximal cliques, geometric features are extracted from the point clouds, including keypoints and local descriptors. Feature matching pairs are then established. The specific process is as follows: First, a compatibility graph is constructed to reflect the relationship between the initial corresponding point clouds; Then, search for maximal cliques in the graph. Each maximal clique represents a set of consensus. Through node-guided clique selection, select the maximal clique with the largest graph weight. Next, during the search for maximal clusters, the total number of maximal clusters is limited to prevent the search time from becoming too long; Finally, the transformation hypothesis of the selected clique is calculated using the SVD algorithm, the best hypothesis is selected for registration, and appropriate pruning and pre-screening strategies are implemented to avoid a surge in computational load on large-scale point clouds and to ensure a balance between efficiency and accuracy. The B102 model is used to perform point cloud matching tasks after registering two point clouds. The Point-NN consists of a non-parametric encoder for 3D feature extraction and a point cloud memory for task-specific recognition. The nonparametric encoder adopts a multi-stage structural design, using farthest point sampling, k-nearest neighbor, trigonometric functions and pooling to gradually aggregate local geometry, generating a high-dimensional global feature for the point cloud; Its model includes a nonparametric encoder and a point memory. For the input point cloud, it uses EncNP to extract its global features and uses PoM feature similarity matching to output the classification result. f G =EncNP(P);logits=PoM(f G ); The similarity between point clouds is determined by using the vector embeddings output by PointNN.
9. A novel vehicle body part detection algorithm according to claim 4, characterized in that: The B201 refers to the process of dividing the point cloud into different segments by means of point cloud coordinate distribution, normal vector change and existing detection results when there are different parts or components on the vehicle body, including doors, engine hood and fenders, with the help of the previous matching or prior knowledge. For specific segmentation methods, the methods used include: region-based growth and clustering-based methods. Since the spatial location or shape features of the parts in the vehicle body are known, the segmentation efficiency is accelerated by combining detection boxes or key feature points. After obtaining the point cloud of each sub-part, B202 performs a more detailed correspondence with the target point cloud. It uses local geometric descriptors or features to determine the alignment error of each local part. When there are internal fine-grained features, including holes, connectors, and welds, the position and orientation of these geometric elements can be further calculated in the segmented point cloud and compared with the design value of the target part, providing rich deviation information in the later evaluation.
10. A novel vehicle body part detection algorithm according to claim 5, characterized in that: B301 uses the simplest strategy to calculate the Euclidean distance ||p for each corresponding point. i -q i ||, calculate the average, maximum, and root mean square of the overall error; In engineering applications, more attention is paid to deviations in key features, including key holes, connecting surfaces, and bent edges. During the judgment process, multiple features are compared based on round holes, rectangular holes, and weld points in the part. The features have been extracted and have accurate information. Therefore, the target point cloud part is judged to meet the standard through two aspects of judgment. The B302 sets the corresponding allowable deviation range according to the design tolerance of the whole vehicle or part. If the local error exceeds a certain threshold, it is marked as unqualified or needs to be reworked. It can also further analyze different assembly links of welding and sheet metal forming to determine which stage causes the maximum error in order to improve the process flow. The B303 integrates error distribution, deviation statistics, and pass rate into the quality assessment system, automatically generates assembly quality reports, and compares and analyzes historical data. It can continuously track the quality stability of the production line and provide early warnings of potential risks.