A workpiece surface three-dimensional point cloud anomaly detection method based on local geometric feature fusion and cascade registration
Patent Information
- Application Number
- CN202610958049.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-30
- Publication Date
- 2026-09-08
AI Technical Summary
[0006]针对现有工件表面三维点云异常检测方法中局部微小缺陷表征不足、待检测点云与正常参考点云之间空间错位易影响异常得分计算和异常定位精度的问题，本发明提出一种基于局部几何特征融合与级联配准的工件表面三维点云异常检测方法
[0038] 1. This invention integrates the original three-dimensional coordinates, normal vector, and curvature of a point into multi-dimensional local geometric features, enabling the detection model to simultaneously perceive the spatial position of the point, the local surface direction, and the degree of local curvature, thereby improving the ability to characterize minute geometric anomalies such as missing, redundant, concave, convex, and edge damage.
Smart Images
Figure CN122714432A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of 3D vision, point cloud processing, and industrial quality inspection technology, and in particular to a method for detecting anomalies in 3D point clouds on workpiece surfaces based on local geometric feature fusion and cascade registration. This method can be used in scenarios such as precision manufacturing, automated inspection, surface defect detection of industrial parts, and 3D point cloud anomaly localization. Background Technology
[0002] With the continuous improvement of intelligent manufacturing and industrial automation, the role of workpiece surface quality inspection in the production process is becoming increasingly prominent. Anomalies such as defects, redundancy, breakage, dents, protrusions, or edge cracks on the workpiece surface not only affect the product's appearance quality but may also impact component assembly accuracy, service life, and operational reliability. Therefore, researching high-precision, automated methods for detecting workpiece surface anomalies is of great significance for improving industrial quality inspection efficiency and ensuring product quality.
[0003] Traditional industrial defect detection methods are mostly based on two-dimensional image analysis, primarily using color, texture, and grayscale changes to identify abnormal areas. While these methods are effective for detecting defects in planar textures, they are easily affected by factors such as lighting variations, metal reflections, shadows, and differences in surface texture in complex industrial environments. When defects mainly manifest as localized spatial morphological changes, relying solely on two-dimensional images is insufficient to accurately describe the true geometric structure of the workpiece surface, easily leading to false positives or false negatives.
[0004] 3D point clouds can directly record the spatial coordinates and geometric morphology information of a workpiece surface, offering advantages in representing local missing parts, redundancy, edge damage, and surface deformation. Existing anomaly detection methods based on 3D point clouds typically store normal point cloud representations in a normal sample feature memory and calculate anomaly scores through feature matching during the testing phase. However, existing methods still have the following shortcomings: On the one hand, some methods mainly rely on the original 3D coordinates or high-level semantic features, lacking full utilization of explicit local geometric information such as normal changes and curvature abrupt changes, resulting in insufficient characterization ability for minor surface anomalies; on the other hand, there may be pose deviations, local truncation, or edge misalignments between the point cloud to be detected and the normal reference point cloud, and spatial alignment errors will directly affect the anomaly score calculation and anomaly localization results.
[0005] Therefore, there is an urgent need for a three-dimensional point cloud anomaly detection method that can integrate local geometric features of point clouds and improve spatial alignment stability, so as to enhance the detection capability of small geometric anomalies on the workpiece surface and reduce the interference of point cloud misalignment on anomaly judgment. Summary of the Invention
[0006] To address the problems of insufficient characterization of local micro-defects and the potential impact of spatial misalignment between the detected point cloud and the normal reference point cloud on anomaly calculation and anomaly positioning accuracy in existing three-dimensional point cloud anomaly detection methods for workpiece surfaces, this invention proposes a three-dimensional point cloud anomaly detection method for workpiece surfaces based on local geometric feature fusion and cascade registration.
[0007] The technical solution adopted by this invention to achieve the above objectives is: a method for detecting anomalies in three-dimensional point cloud on a workpiece surface based on local geometric feature fusion and cascade registration, comprising the following steps:
[0008] During the training phase, a normal workpiece 3D point cloud is acquired and preprocessed. Multidimensional local geometric features containing spatial position, local surface orientation, and local curvature are extracted from the normal point cloud to construct a normal feature memory library.
[0009] During the testing phase, a 3D point cloud of the workpiece to be inspected is acquired and preprocessed; the 3D point cloud of the workpiece to be inspected is concatenated and registered with a normal reference point cloud to obtain a registered point cloud to be inspected; multi-dimensional local geometric features including spatial location, local surface orientation, and local curvature are extracted from the registered point cloud to be inspected and matched with the normal feature memory to calculate a point-level anomaly score; based on the point-level anomaly score, sample-level anomaly discrimination results and point-level anomaly region localization results are obtained.
[0010] The point cloud preprocessing includes the following steps:
[0011] The input point cloud is centered to translate it to a unified central position; the point cloud coordinates are scaled and normalized according to the overall scale of the point cloud; high-density point cloud is sampled to ensure that the number of points in the training and testing phases is within a preset calculation range; for any point in the point cloud, its k nearest neighbor set is constructed according to Euclidean distance as the local neighborhood of that point.
[0012] The extraction of multidimensional local geometric features, including spatial location, local surface orientation, and local curvature, from normal point clouds or registered point clouds to be detected includes the following steps:
[0013] Calculate the neighborhood center of the set of k nearest neighbors for any point in the point cloud;
[0014] Construct the covariance matrix based on the deviation between neighboring points and the neighborhood center;
[0015] Perform eigenvalue decomposition on the covariance matrix, and take the eigenvector corresponding to the smallest eigenvalue as the point. normal vector ;
[0016] Calculate the curvature at a point by the ratio of the smallest eigenvalue to the sum of all eigenvalues. .
[0017] The multidimensional local geometric features are seven-dimensional feature vectors, represented as follows:
[0018] ;
[0019] in, , , Point Three-dimensional spatial coordinates, , , Point The normal vector components, Point The local curvature.
[0020] The construction of the normal feature memory includes the following steps:
[0021] Only normal workpiece 3D point clouds are used as training samples;
[0022] Extract multidimensional local geometric features for each normal training sample;
[0023] The multidimensional local geometric features are stored in the normal feature memory. ,in Represents the features of normal points in the memory bank. Indicates the number of features in the memory;
[0024] The normal feature memory is compressed according to a preset sampling strategy to retain representative normal point features and reduce the complexity of matching calculations.
[0025] The cascaded registration includes the following steps:
[0026] Using the 3D point cloud of the workpiece to be inspected as the source point cloud Using a normal reference point cloud as the target point cloud ;
[0027] Based on source cloud and target point cloud Candidate matching point pairs are established based on local features;
[0028] The random sample consensus algorithm is used to perform coarse registration of candidate matching point pairs, filter the set of interior points, and estimate the initial rigid body transformation matrix. ;
[0029] With the initial rigid body transformation matrix As the initial pose for the iterative closest point algorithm, the source point cloud is... and target point cloud Perform precise registration;
[0030] Based on the final rigid body transformation matrix obtained from the fine registration This is used to map the three-dimensional point cloud of the workpiece to be inspected to the normal reference point cloud coordinate system, thus obtaining the registered point cloud to be inspected.
[0031] The process of matching the feature memory with normal features and calculating point-level anomaly scores includes the following steps:
[0032] For any point in the point cloud to be detected, the multidimensional local geometric features In the normal feature memory bank Search and Features of the nearest normal point;
[0033] This multidimensional local geometric feature The Euclidean distance between the feature and the nearest normal point is used as the point-level anomaly score for that point. .
[0034] The process of obtaining sample-level anomaly discrimination results and point-level anomaly region localization results based on point-level anomaly scores includes the following steps:
[0035] An anomaly response distribution is formed by combining the point-level anomaly scores of all points to generate point-level anomaly area location results; the point-level anomaly scores are aggregated to obtain sample-level anomaly scores, and the judgment result of whether the workpiece is abnormal is output based on the sample-level anomaly scores; wherein, the aggregation method is to take the maximum point-level anomaly score as the sample-level anomaly score.
[0036] Based on the comparison between the sample-level anomaly score and the preset threshold, the result of judging whether the workpiece is normal or abnormal is output; based on the point-level anomaly score and the point-level threshold or color mapping result, the result of locating the abnormal area is output.
[0037] The present invention has the following beneficial effects and advantages:
[0038] 1. This invention integrates the original three-dimensional coordinates, normal vector, and curvature of a point into multi-dimensional local geometric features, enabling the detection model to simultaneously perceive the spatial position of the point, the local surface direction, and the degree of local curvature, thereby improving the ability to characterize minute geometric anomalies such as missing, redundant, concave, convex, and edge damage.
[0039] 2. This invention estimates the normal vector and curvature through the local neighborhood covariance matrix. The features have clear physical meaning and strong interpretability, and can intuitively reflect the changes in the orientation and curvature of the workpiece surface.
[0040] 3. This invention adopts a cascaded registration strategy that combines RANSAC coarse registration with ICP fine registration, which can obtain a relatively stable initial transformation even in the presence of initial pose differences or mismatch points, and further optimizes the local spatial alignment relationship to reduce the impact of spatial misalignment on the calculation of abnormal scores.
[0041] 4. This invention only requires the construction of a normal feature memory library from the point cloud of normal workpieces, making it suitable for industrial applications where abnormal samples are scarce, defect types are varied, and annotation costs are high.
[0042] 5. This invention simultaneously outputs sample-level anomaly discrimination results and point-level anomaly location results, which can not only determine whether the entire workpiece is abnormal, but also locate the abnormal area, facilitating subsequent rework, sorting and quality traceability. Attached Figure Description
[0043] Figure 1 This is a flowchart of the overall process for the workpiece surface three-dimensional point cloud anomaly detection method based on local geometric feature fusion and cascade registration according to the present invention.
[0044] Figure 2 This is a detailed flowchart of the training and testing phases of the method of the present invention.
[0045] Figure 3 This is a schematic diagram of the local geometric feature construction process in this invention.
[0046] Figure 4 This is a schematic diagram of the cascade registration process in this invention.
[0047] Figure 5 This is a schematic diagram illustrating the abnormal region localization effect of the method of the present invention in one embodiment. Detailed Implementation
[0048] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.
[0049] like Figure 1 As shown, a method for detecting anomalies in three-dimensional point cloud on workpiece surface based on local geometric feature fusion and cascade registration includes steps such as point cloud acquisition and preprocessing, normal vector and curvature estimation, seven-dimensional local geometric feature fusion, normal feature memory construction, RANSAC+ICP cascade registration, nearest neighbor matching, anomaly score calculation, and anomaly discrimination and localization.
[0050] like Figure 2 As shown, the method of this invention can be divided into a training phase and a testing phase. The training phase is used to construct a normal feature memory library, and the testing phase is used to register the point cloud of the workpiece to be inspected, extract features, calculate anomaly scores, and output anomaly results.
[0051] Step 1: Point Cloud Acquisition and Preprocessing
[0052] Acquire 3D point clouds of normal workpieces and workpieces to be inspected. Point clouds can be acquired by structured light scanners, laser scanners, industrial 3D cameras, or other 3D imaging equipment, or they can be obtained from historical point cloud databases of the production line.
[0053] Let the input point cloud be ,in Indicating the first point cloud The three-dimensional coordinates of the points This represents the number of points in the point cloud. Since there may be differences in point count, spatial scale, and local density between different workpiece samples, this embodiment first performs centering and scale normalization on the input point cloud to ensure that different samples are compared within a relatively consistent spatial range; then, high-density point clouds are sampled to reduce the overhead of subsequent feature calculation and matching calculations; finally, for each point... Constructed based on Euclidean distance The nearest neighbor set serves as the neighborhood for local geometric feature estimation.
[0054] Step 2: Estimation of normal vector and curvature
[0055] like Figure 3 As shown, for any point in the point cloud First select its We calculate the nearest neighbor set and the center of the neighborhood points; then we construct the covariance matrix based on the deviation between the neighborhood points and the neighborhood center; next, we perform eigenvalue decomposition on the covariance matrix to obtain three eigenvalues and their corresponding eigenvectors. When the three eigenvalues satisfy... When, the smallest eigenvalue The corresponding eigenvectors are used to approximate the normal direction of the local fitting plane, denoted as point A. normal vector .
[0056] Furthermore, based on the minimum eigenvalue The point for calculating the ratio to the sum of all eigenvalues. Local curvature When the workpiece surface is relatively smooth, the change in the normal vector of adjacent points is small and the curvature is relatively stable; when the surface has defects, redundancy, depressions, protrusions, edge damage or local morphological changes, the normal vector and curvature of the abnormal area usually show significant changes.
[0057] Step 3: Multidimensional Local Geometric Feature Fusion
[0058] After obtaining the normal vector and curvature, the original 3D coordinates, normal vector, and curvature of each point are concatenated to construct a seven-dimensional local geometric feature. Therefore, input point cloud It can be represented as a set of multidimensional local geometric features. .
[0059] Compared to feature representations that only use the original three-dimensional coordinates, the aforementioned seven-dimensional local geometric features simultaneously include information on spatial location, local surface orientation, and local curvature, enabling a more comprehensive description of the local geometric state of the workpiece surface. When minute defects exist on the surface, the abnormal region will exhibit different responses in the normal vector and curvature dimensions compared to the normal region, which helps improve the ability to locate point-level anomalies.
[0060] Step 4: Construction of Normal Feature Memory
[0061] During the training phase, only normal workpiece 3D point clouds are used. Steps 1 to 3 are performed on each normal training sample to obtain the seven-dimensional local geometric features of the normal point cloud, and these seven-dimensional local geometric features are stored in the normal feature memory. ,in Represents the features of normal points in the memory bank. This indicates the number of features in the memory.
[0062] To control computational complexity, core set sampling, random sampling, voxel sampling, or other representative sampling strategies can be used to compress the normal feature memory, thereby reducing nearest neighbor matching overhead while preserving normal geometric patterns.
[0063] Step 5: Cascade Registration
[0064] like Figure 4 As shown, during the testing phase, the 3D point cloud of the workpiece to be tested is used as the source point cloud. Using a normal reference point cloud as the target point cloud The two points are then concatenated for registration. First, candidate matching point pairs are established based on local features of the source and target point clouds. Then, a random sampling consensus algorithm is used to coarsely register the candidate matching point pairs. The rigid body transformation is estimated by randomly sampling several matching point pairs, and the set of interior points is selected based on the transformed matching error to obtain the initial rigid body transformation matrix. .
[0065] Obtaining the initial rigid body transformation matrix After that, with As the initial pose for the iterative closest point algorithm, the closest point correspondence between the source point cloud and the target point cloud is established iteratively, and the distance error between corresponding points is minimized to obtain the final rigid body transformation matrix. Finally, according to The point cloud to be detected is mapped to the coordinate system of the normal reference point cloud to obtain the registered point cloud to be detected. By combining RANSAC coarse registration and ICP fine registration, the influence of initial pose differences on registration can be reduced, and the local spatial alignment accuracy can be improved.
[0066] Step 6: Anomaly Score Calculation and Result Output
[0067] Steps 2 and 3 are performed on the registered point cloud to be detected to obtain its seven-dimensional local geometric features. The features of any point in the point cloud to be detected are then analyzed. In the normal feature memory bank Search for the normal feature point with the smallest Euclidean distance, and use the smallest Euclidean distance as the point-level anomaly score for that point. If the local geometric features of a point differ significantly from those in the normal memory, the point will have a higher anomaly score, indicating that it is more likely to belong to an anomalous region.
[0068] Point-level anomaly score Represented as: ,in Indicates the first point in the cloud to be detected Seven-dimensional local geometric features of a point Represents the normal feature memory bank, Represents the first in the normal feature memory bank A normal point feature, Indicates the first Point-level anomaly scores for each point.
[0069] Point-level anomaly scores for all points Together, these constitute the abnormal response distribution, used to generate abnormal region location results. For sample-level anomaly detection, point-level anomaly scores can be assigned to all points. Aggregation is performed to obtain the sample-level anomaly score for the entire workpiece under inspection. This method uses the maximum point-level anomaly score as the sample-level anomaly score, which is expressed as: ,in Indicates the sample-level anomaly score. Indicates the first Point-level anomaly score for each point, This indicates the number of points in the point cloud to be detected.
[0070] Based on sample-level anomaly scores The comparison result with the preset threshold is used to output the judgment result of whether the workpiece is normal or abnormal; based on the point-level abnormality score... The results of point-level threshold or color mapping are used to output the abnormal region location results.
[0071] This embodiment uses a high-precision workpiece surface three-dimensional point cloud anomaly detection task as an example to illustrate the method of the present invention. This embodiment does not constitute a limitation on the scope of protection of the present invention. This embodiment includes the following steps:
[0072] Step 1: Data Preparation. Obtain the 3D point cloud of a normal workpiece and the 3D point cloud of the workpiece to be inspected. The 3D point cloud of the normal workpiece is used to construct a normal feature memory library, and the 3D point cloud of the workpiece to be inspected is used for anomaly detection and anomaly region localization.
[0073] Step 2: Point Cloud Preprocessing. The input point cloud is centered, scaled, and sampled. A local neighborhood is constructed for each point in the point cloud based on Euclidean distance. The number of neighborhood points in the local neighborhood can be set according to the point cloud density and the surface complexity of the workpiece.
[0074] Step 3: Local geometric feature extraction. A covariance matrix is constructed based on the local neighborhood of each point, and eigenvalue decomposition is performed on the covariance matrix to obtain the normal vector and curvature of the point; then the original 3D coordinates, normal vector and curvature are concatenated to form a seven-dimensional local geometric feature.
[0075] Step 4: Normal Feature Memory Construction. A normal feature memory is constructed using only normal artifact point clouds. To reduce the computational overhead of subsequent nearest neighbor matching, core set sampling, random sampling, voxel sampling, or other representative sampling strategies can be used to compress the normal feature memory.
[0076] Step 5: Cascaded Registration. Using the point cloud to be detected as the source point cloud and the normal reference point cloud as the target point cloud, coarse registration is first performed using the RANSAC algorithm, and then fine registration is performed using the initial transformation matrix obtained from the coarse registration as the initial pose of the ICP algorithm. The distance threshold, voxel size, and number of iterations used in the registration process can be set according to the point cloud scale, noise level, and detection accuracy requirements.
[0077] Step 6: Anomaly Score Calculation. Seven-dimensional local geometric features are extracted from the registered point cloud to be detected, and the nearest neighbor normal features are searched in the normal feature memory. The Euclidean distance between the two is used as the point-level anomaly score.
[0078] Step 7: Result Output. An anomaly response distribution is generated based on all point-level anomaly scores for point-level anomaly region localization. A higher point-level anomaly score indicates a greater deviation from the normal feature distribution, and a higher probability that the point belongs to an anomaly region. The maximum point-level anomaly score is used as the sample-level anomaly score, and based on the comparison between the sample-level anomaly score and a preset sample-level threshold, the judgment result of whether the workpiece to be inspected is normal or abnormal is output. Simultaneously, anomaly region localization results are generated based on the point-level anomaly scores.
[0079] In one embodiment, anomaly region localization visualization results are generated based on point-level anomaly scores, such as... Figure 5 As shown. Figure 5 This paper presents the anomaly localization results for three different geometric shapes of samples—candybar, toffee, and fish—from the Real3D-AD high-precision 3D point cloud anomaly detection dataset. The first column shows the original point cloud visualization of the sample to be detected, the second column shows the ground truth annotation of the defects provided by the dataset, and the third column shows the anomaly region predicted using the method of this invention. Figure 5 In the defect ground truth annotation and prediction results, darker gray areas represent abnormal areas or areas with strong abnormal responses, while lighter gray areas represent normal areas or areas with weak abnormal responses. Visualization results show that this method can produce a relatively obvious response to local abnormal areas in the 3D point cloud of the workpiece surface.
[0080] Experimental results show that, when validated on the Real3D-AD high-precision 3D point cloud anomaly detection dataset, the method of this invention achieves average scores of 0.8080, 0.8487, 0.7718, and 0.2313 for the four metrics of I-AUC, P-AUC, I-AP, and P-AP, respectively, which are improvements of 0.1579, 0.1717, 0.1270, and 0.1488 over the baseline method, respectively. This verifies the effectiveness of the local geometric feature fusion and cascade registration strategy for sample-level anomaly discrimination and point-level anomaly localization.
[0081] Through the above embodiments, the present invention can enhance the characterization of local small geometric anomalies by utilizing normal vectors and curvature, and reduce the impact of spatial misalignment on anomaly score calculation by using a cascaded registration strategy that combines RANSAC coarse registration and ICP fine registration, thereby realizing sample-level anomaly discrimination and point-level anomaly localization of the three-dimensional point cloud on the workpiece surface.
Claims
1. A method for detecting anomalies in three-dimensional point cloud on a workpiece surface based on local geometric feature fusion and cascade registration, characterized in that, Includes the following steps: During the training phase, a normal workpiece 3D point cloud is acquired and preprocessed. Multidimensional local geometric features containing spatial position, local surface orientation, and local curvature are extracted from the normal point cloud to construct a normal feature memory library. During the testing phase, a 3D point cloud of the workpiece to be inspected is acquired and preprocessed; the 3D point cloud of the workpiece to be inspected is concatenated and registered with a normal reference point cloud to obtain a registered point cloud to be inspected; multi-dimensional local geometric features including spatial location, local surface orientation, and local curvature are extracted from the registered point cloud to be inspected and matched with the normal feature memory to calculate a point-level anomaly score; based on the point-level anomaly score, sample-level anomaly discrimination results and point-level anomaly region localization results are obtained.
2. The method for detecting anomalies in three-dimensional point cloud on a workpiece surface based on local geometric feature fusion and cascade registration as described in claim 1, characterized in that, The point cloud preprocessing includes the following steps: The input point cloud is centered to translate it to a unified central position; the point cloud coordinates are scaled and normalized according to the overall scale of the point cloud; high-density point cloud is sampled to ensure that the number of points in the training and testing phases is within a preset calculation range; for any point in the point cloud, its k nearest neighbor set is constructed according to Euclidean distance as the local neighborhood of that point.
3. The method for detecting anomalies in three-dimensional point cloud on a workpiece surface based on local geometric feature fusion and cascade registration as described in claim 1, characterized in that, The extraction of multidimensional local geometric features, including spatial location, local surface orientation, and local curvature, from normal point clouds or registered point clouds to be detected includes the following steps: Calculate the neighborhood center of the set of k nearest neighbors for any point in the point cloud; Construct the covariance matrix based on the deviation between neighboring points and the neighborhood center; Perform eigenvalue decomposition on the covariance matrix, and take the eigenvector corresponding to the smallest eigenvalue as the point. normal vector ; Calculate the curvature at a point by the ratio of the smallest eigenvalue to the sum of all eigenvalues. .
4. A method for detecting anomalies in three-dimensional point cloud on a workpiece surface based on local geometric feature fusion and cascade registration, as described in claim 1 or 3, characterized in that, The multidimensional local geometric features are seven-dimensional feature vectors, represented as follows: ； in, , , Point Three-dimensional spatial coordinates, , , Point The normal vector components, Point The local curvature.
5. The method for detecting anomalies in three-dimensional point cloud on a workpiece surface based on local geometric feature fusion and cascade registration according to claim 1, characterized in that, The construction of the normal feature memory includes the following steps: Only normal workpiece 3D point clouds are used as training samples; Extract multidimensional local geometric features for each normal training sample; The multidimensional local geometric features are stored in the normal feature memory. ,in Represents the features of normal points in the memory bank. Indicates the number of features in the memory; The normal feature memory is compressed according to a preset sampling strategy to retain representative normal point features and reduce the complexity of matching calculations.
6. The method for detecting anomalies in three-dimensional point cloud on a workpiece surface based on local geometric feature fusion and cascade registration according to claim 1, characterized in that, The cascaded registration includes the following steps: Using the 3D point cloud of the workpiece to be inspected as the source point cloud Using a normal reference point cloud as the target point cloud ; Based on source cloud and target point cloud Candidate matching point pairs are established based on local features; The random sample consensus algorithm is used to perform coarse registration of candidate matching point pairs, filter the set of interior points, and estimate the initial rigid body transformation matrix. ; With the initial rigid body transformation matrix As the initial pose for the iterative closest point algorithm, the source point cloud is... and target point cloud Perform precise registration; Based on the final rigid body transformation matrix obtained from the fine registration This is used to map the three-dimensional point cloud of the workpiece to be inspected to the normal reference point cloud coordinate system, thus obtaining the registered point cloud to be inspected.
7. The method for detecting anomalies in three-dimensional point cloud on a workpiece surface based on local geometric feature fusion and cascade registration according to claim 1, characterized in that, The process of matching the feature memory with normal features and calculating point-level anomaly scores includes the following steps: For any point in the point cloud to be detected, the multidimensional local geometric features In the normal feature memory bank Search and Features of the nearest normal point; This multidimensional local geometric feature The Euclidean distance between the feature and the nearest normal point is used as the point-level anomaly score for that point. .
8. A method for detecting anomalies in three-dimensional point cloud on a workpiece surface based on local geometric feature fusion and cascade registration, as described in claim 1 or 7, characterized in that, The process of obtaining sample-level anomaly discrimination results and point-level anomaly region localization results based on point-level anomaly scores includes the following steps: An anomaly response distribution is formed by combining the point-level anomaly scores of all points to generate point-level anomaly area location results; the point-level anomaly scores are aggregated to obtain sample-level anomaly scores, and the judgment result of whether the workpiece is abnormal is output based on the sample-level anomaly scores; wherein, the aggregation method is to take the maximum point-level anomaly score as the sample-level anomaly score.
9. The method for detecting anomalies in three-dimensional point cloud on a workpiece surface based on local geometric feature fusion and cascade registration according to claim 8, characterized in that, Based on the comparison between the sample-level anomaly score and the preset threshold, the result of judging whether the workpiece is normal or abnormal is output; based on the point-level anomaly score and the point-level threshold or color mapping result, the result of locating the abnormal area is output.