Training-free 3d point cloud anomaly detection method and system based on deformation-sensitive features

Abstract

The application discloses a kind of based on deformation sensitive feature's training-free three-dimensional point cloud anomaly detection method and system, it is related to three-dimensional point cloud anomaly detection technical field, including: the three-dimensional point cloud data to be measured is preprocessed and position coding, point cloud coordinate set is constructed local neighborhood on multiple space levels and calculates local deformation measure, generates deformation sensitive modulation weight, adjusts the geometric offset feature of local neighborhood;The position coding of adjusted geometric offset is combined with the corresponding features of center point and neighborhood point to construct multi-scale point cloud feature, and respectively map back initial point cloud resolution to obtain dense feature descriptor;In pre-set sample layered feature memory library, layered near neighbor search and distance aggregation are carried out to dense feature descriptor, and anomaly detection result is obtained.The application alleviates the technical problems that high-resolution point cloud micro-defects are easily diluted by downsampling, complex normal surface is easy to cause false alarm, and training model deployment cycle is long in prior art.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional point cloud anomaly detection technology, and in particular to a training-free three-dimensional point cloud anomaly detection method and system based on deformation-sensitive features. Background Technology

[0002] In intelligent manufacturing, aerospace component assembly, semiconductor packaging, precision mold processing, automotive parts production, and high-end equipment manufacturing, surface and structural quality inspection is a crucial step in ensuring product reliability and safety. Common defects include minor scratches, shallow dents, localized protrusions, cracks, edge burrs, and processing deformation. Traditional inspection methods typically rely on manual visual inspection or two-dimensional image inspection, which is easily affected by lighting, texture, viewing angle, and human experience, making it difficult to reliably detect defects that match the surface color but exhibit subtle three-dimensional shape variations.

[0003] With the development of 3D sensing technologies such as 3D scanning, structured light imaging, LiDAR, and industrial CT, high-resolution point clouds have gradually become an important data carrier for industrial quality inspection. 3D point clouds can directly reflect the spatial geometry and minute morphological changes of a target surface, making them suitable for detecting depth changes and surface undulations that are difficult to express in 2D images. However, point cloud data is characterized by its disorder, unstructured nature, uneven density, and large data volume, making anomaly detection directly in high-resolution point clouds a significant challenge.

[0004] On the one hand, high-resolution point clouds typically contain a huge number of normal points, while defective regions account for a very small proportion. Weak anomalous signals are easily diluted during global feature aggregation and downsampling. Existing deep learning methods often employ multi-layer downsampling or feature compression strategies to reduce computational cost, which may lead to the loss of tiny defective points in the early stages of feature extraction. On the other hand, the surfaces of actual industrial parts often contain normal high-curvature structures such as arcs, edges, corners, grooves, and complex curved surfaces. These normal structures and tiny defects may have similar spatial variation characteristics in local geometry, which can easily lead to false positives.

[0005] While existing 3D anomaly detection methods based on reconstruction, feature memory, or deep feature learning have made progress in some scenarios, they still suffer from problems such as strong dependence on training samples, large model parameter count, long deployment cycles, and insufficient cross-class generalization ability. For applications in industrial settings where product models change frequently and small or even defect-free samples are used, solutions relying on extensive anomaly labeling or long training times are not conducive to rapid deployment. Furthermore, existing training-free or non-parametric methods often employ fixed geometric matching rules, lacking adaptive adjustment capabilities for complex curved surfaces and prone to generating high anomaly responses at high-curvature normal structures. Existing non-parametric point cloud encoding methods typically focus on classification or segmentation tasks, mainly obtaining recognition results through fixed neighborhood aggregation and global feature matching, lacking curvature-guided modulation, resolution retrospective projection, and hierarchical anomaly measurement mechanisms for high-resolution point-level anomaly localization. Summary of the Invention

[0006] To address the technical problems of existing technologies, such as the easy downsampling dilution of minute defects in high-resolution point clouds, the susceptibility to false alarms caused by complex normal surfaces, and the long deployment cycle of training models, this invention provides a training-free 3D point cloud anomaly detection method and system based on deformation-sensitive features. The technical solution is as follows: On the one hand, a training-free 3D point cloud anomaly detection method based on deformation-sensitive features is provided. The method includes: performing spatial calibration and normalization on the 3D point cloud data containing the target to be detected to obtain a preprocessed point cloud coordinate set; performing position encoding on the point cloud coordinate set to obtain an initial high-dimensional geometric feature set; performing stepwise farthest point sampling on the point cloud coordinate set to construct local neighborhoods at multiple spatial levels; calculating a local deformation metric based on the geometric offset features of the local neighborhoods; generating deformation-sensitive modulation weights based on the local deformation metric; adjusting the geometric offset features of the local neighborhoods based on the deformation-sensitive modulation weights to obtain an adjusted geometric offset; and performing training-free anomaly detection on the adjusted geometric features. Offset is used for position encoding, and channel splicing and aggregation are performed by combining the center point corresponding features and the neighboring point corresponding features to construct deformation-sensitive multi-scale point cloud features. The multi-scale point cloud features are mapped back to the initial point cloud resolution and spliced ​​with the initial high-dimensional geometric feature set to obtain dense feature descriptors that correspond one-to-one with the 3D point cloud data to be tested. In the preset sample hierarchical feature memory, hierarchical nearest neighbor retrieval and distance aggregation are performed on the dense feature descriptors of the 3D point cloud data to be tested to obtain anomaly detection results. The anomaly detection results include point-level anomaly scores and object-level anomaly scores. The preset sample hierarchical feature memory is constructed based on the dense feature descriptors of normal sample reference point cloud data.

[0007] Optionally, spatial calibration and normalization are performed on the 3D point cloud data containing the target to be detected to obtain a preprocessed point cloud coordinate set. This includes: decentralizing the 3D point cloud data containing the target to be detected to obtain decentralized point cloud data; calculating the global covariance matrix of the decentralized point cloud data and performing eigenvalue decomposition to obtain multiple principal axis directions; constructing a rotation matrix based on the multiple principal axis directions; and transforming the decentralized point cloud data to the principal axis coordinate system based on the rotation matrix to obtain a transformed point cloud coordinate set; synchronously rotating the normal vector of the decentralized point cloud data based on the rotation matrix to obtain a rotated normal vector; normalizing the transformed point cloud coordinate set to a unit sphere to obtain a normalized point cloud coordinate set; and concatenating the normalized point cloud coordinate set with the rotated normal vector to obtain a preprocessed point cloud coordinate set.

[0008] Optionally, position encoding is performed on the point cloud coordinate set to obtain an initial high-dimensional geometric feature set, including: applying sine and cosine mapping to each point cloud coordinate in the point cloud coordinate set using multiple frequency bands to obtain the initial high-dimensional geometric feature set.

[0009] Optionally, the point cloud coordinate set is sampled at each level to construct local neighborhoods at multiple spatial levels, and a local deformation metric is calculated based on the geometric offset features of the local neighborhoods. This includes: sampling the point cloud coordinate set at each level to form a multi-scale point set hierarchy; in each level, selecting a center point from the point cloud coordinate set of the previous level using a farthest point sampling algorithm, and for each center point, selecting K neighboring points from the point cloud coordinate set of the previous level using K-nearest neighbor retrieval to construct a local neighborhood corresponding to each level; constructing a local covariance matrix based on the geometric offset of each neighboring point relative to the center point within the local neighborhood; performing eigenvalue decomposition on the local covariance matrix to obtain multiple local eigenvalues; calculating a local approximate curvature based on the multiple local eigenvalues, and using the local approximate curvature as a local deformation metric; wherein the local approximate curvature includes the ratio of the minimum local eigenvalue to the sum of the multiple local eigenvalues.

[0010] Optionally, the deformation-sensitive modulation weights include:

[0011] In the formula, The deformation-sensitive modulation weights, This is the measure of the local deformation; The adjusted geometric offset includes:

[0012] In the formula, The adjusted geometric offset, The geometric offset of the j-th neighboring point in the local neighborhood relative to the center point.

[0013] Optionally, the multi-scale point cloud features include aggregated features at multiple spatial levels; the adjusted geometric offset is positionally encoded, and channel splicing and aggregation are performed by combining the center point corresponding features and the neighboring point corresponding features to construct deformation-sensitive multi-scale point cloud features, including: positionally encoding the adjusted geometric offset to obtain local geometric position encoding features; channel splicing the center point corresponding features with the j-th neighboring point corresponding features to obtain extended features; and feature fusion of the extended features with the local geometric position encoding features to obtain fused features of the local neighborhood; the fused features include:

[0014] In the formula, This represents the local fused feature obtained after deformation-sensitive modulation and feature fusion of the center point and the j-th neighboring point. This represents the extended feature obtained by concatenating the features corresponding to the center point and the features corresponding to the j-th neighboring point through channels. This indicates the adjusted geometric offset. This refers to the local geometric position encoding feature obtained by position encoding the adjusted geometric offset, which matches the dimension of the extended feature. This represents element-wise multiplication; for the ... The fusion features of all neighboring points within the local neighborhood corresponding to the same center point in each spatial level are aggregated using dual-path pooling to obtain the center point in the 1st spatial level. Aggregation features at each spatial level; the dual-pooling aggregation includes adding the max pooling result to the average pooling result; the aggregation features include:

[0015] In the formula, Indicates the first The center point of each spatial level The aggregation characteristics, Indicates the first The center point of each spatial level The local fused features obtained after deformation-sensitive modulation and feature fusion with the j-th neighboring point, Indicates the first The center point of each spatial level The corresponding local neighborhood.

[0016] Optionally, the multi-scale point cloud features are mapped back to the initial point cloud resolution, and channel-wise concatenated with the initial high-dimensional geometric feature set to obtain a dense feature descriptor corresponding one-to-one with the three-dimensional point cloud data to be measured. This includes: retrieving a set of neighboring source points in the source point set at each spatial level, and calculating Gaussian weights based on the spatial distance between the target point and the source point in the set of neighboring source points; the Gaussian weights include:

[0017] In the formula, For the t-th target point p t With source point p s Gaussian weights Represents the L2 norm. The Gaussian kernel bandwidth parameter is represented; based on the Gaussian weights and using normalized weighted summation, the multi-scale point cloud features are back-interpolated and projected to obtain the recovered feature sets corresponding to each level after interpolation; the first... l Spatial hierarchical recovery feature set include:

[0018]

[0019] In the formula, Indicates the first Interpolate spatial levels back to the target point The recovery features obtained later Indicates the first Target point in each spatial level With the source Gaussian weights between them Indicates the first Source point in each spatial hierarchy Corresponding source features Indicates the target point In the The set of neighboring source points retrieved from the source point set at each spatial level. The initial point cloud contains the number of points. The initial high-dimensional geometric feature set is concatenated with the recovery feature set corresponding to each level to obtain a dense feature descriptor that corresponds one-to-one with the three-dimensional point cloud data to be tested.

[0020] Optionally, in a preset sample hierarchical feature memory, hierarchical nearest neighbor retrieval and distance aggregation are performed on the dense feature descriptors of the 3D point cloud data to be tested to obtain anomaly detection results. This includes: slicing the dense feature descriptor of each point in the 3D point cloud data to be tested according to the source spatial level to obtain the test features corresponding to each spatial level; and performing anomaly detection on the test features of each spatial level and the memory features of the corresponding spatial level in the preset sample hierarchical feature memory. Norm normalization is performed, and multiple nearest neighbor features are retrieved at the corresponding spatial level. Based on the aggregation of distance statistics of multiple nearest neighbor features at each spatial level, the point-level anomaly score of the 3D point cloud data to be tested is calculated. Based on the statistics of all point-level anomaly scores in the 3D point cloud data to be tested, the object-level anomaly score is calculated.

[0021] On the other hand, a training-free 3D point cloud anomaly detection system based on deformation-sensitive features is also provided to implement the training-free 3D point cloud anomaly detection method based on deformation-sensitive features provided in the embodiments of the present invention. The system includes: a data preprocessing module, a deformation metric calculation module, a geometric offset adjustment module, a deformation-sensitive geometric encoding module, a multi-scale parallel decoding projection module, and a hierarchical memory metric module. The data preprocessing module is used to perform spatial calibration and normalization processing on the 3D point cloud data containing the target to be detected, obtaining a preprocessed set of point cloud coordinates, and performing position encoding on the point cloud coordinate set to obtain an initial high-dimensional geometric feature set. The deformation metric calculation module is used to perform stepwise farthest point sampling on the point cloud coordinate set, constructing local neighborhoods at multiple spatial levels, and calculating local deformation metrics based on the geometric offset features of the local neighborhoods. The geometric offset adjustment module is used to generate deformation-sensitive modulation weights based on the local deformation metrics. The deformation-sensitive modulation weights are used to adjust the geometric offset features of the local neighborhood to obtain the adjusted geometric offset. The deformation-sensitive geometric encoding module is used to perform position encoding on the adjusted geometric offset and to perform channel splicing and aggregation by combining the center point corresponding features and the neighborhood point corresponding features to construct deformation-sensitive multi-scale point cloud features. The multi-scale parallel decoding projection module is used to map the multi-scale point cloud features back to the initial point cloud resolution and to perform channel splicing with the initial high-dimensional geometric feature set to obtain a dense feature descriptor that corresponds one-to-one with the test 3D point cloud data. The hierarchical memory measurement module is used to perform hierarchical nearest neighbor retrieval and distance aggregation on the dense feature descriptor of the test 3D point cloud data in the preset sample hierarchical feature memory to obtain anomaly detection results. The anomaly detection results include point-level anomaly scores and object-level anomaly scores. The preset sample hierarchical feature memory is constructed based on the dense feature descriptor of normal sample reference point cloud data.

[0022] On the other hand, an electronic device is also provided, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method provided in the embodiments of the present invention.

[0023] This invention provides a training-free 3D point cloud anomaly detection method and system based on deformation-sensitive features. The method constructs point cloud features based on explicit spatial geometric analysis, suppresses interference from normal complex surface structures through deformation-sensitive feature modulation, which is guided by local approximate curvature. It combines multi-scale parallel decoding and hierarchical feature memory to achieve point-level anomaly localization and object-level anomaly judgment. This method can effectively detect minute defects in complex 3D surfaces without the need for anomaly sample annotation and supervised training. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of a training-free 3D point cloud anomaly detection method based on deformation-sensitive features provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a training-free 3D point cloud anomaly detection system based on deformation-sensitive features provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0027] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0028] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0029] Figure 1This is a flowchart illustrating a training-free 3D point cloud anomaly detection method based on deformation-sensitive features according to an embodiment of the present invention. It can be implemented by a 3D point cloud anomaly detection device, which can be a server, industrial computer, edge computing device, or embedded detection terminal. Figure 1 As shown, the method specifically includes the following steps: Step S102: Spatial calibration and normalization processing is performed on the three-dimensional point cloud data containing the target to be detected to obtain a preprocessed set of point cloud coordinates. The point cloud coordinate set is then position-encoded to obtain an initial set of high-dimensional geometric features.

[0030] Step S104: Sample the farthest point of the point cloud coordinate set at each level, construct local neighborhoods at multiple spatial levels, and calculate the local deformation metric based on the geometric offset features of the local neighborhoods.

[0031] Step S106: Generate deformation-sensitive modulation weights based on local deformation metrics, and adjust the geometric offset features of the local neighborhood based on the deformation-sensitive modulation weights to obtain the adjusted geometric offset.

[0032] Step S108: The adjusted geometric offset is positionally encoded and then channel-joined and aggregated with the corresponding features of the center point and neighboring points to construct deformation-sensitive multi-scale point cloud features.

[0033] Step S110: Map the multi-scale point cloud features back to the initial point cloud resolution, and perform channel stitching with the initial high-dimensional geometric feature set to obtain a dense feature descriptor that corresponds one-to-one with the three-dimensional point cloud data to be measured.

[0034] Step S112: In the preset sample hierarchical feature memory, hierarchical nearest neighbor retrieval and distance aggregation are performed on the dense feature descriptors of the 3D point cloud data to be tested to obtain anomaly detection results; the anomaly detection results include point-level anomaly scores and object-level anomaly scores; the preset sample hierarchical feature memory is constructed based on the dense feature descriptors of normal sample reference point cloud data.

[0035] In an optional embodiment of the present invention, before step S102, the method further includes: acquiring three-dimensional point cloud data containing the target to be detected using a structured light scanner, a laser 3D scanner, an industrial camera depth reconstruction system, or other 3D sensors. Let the input point cloud be represented as:

[0036] in, Indicates the first The three-dimensional coordinates of each point are given, where N represents the number of points in the point cloud. Normal sample reference point cloud data is pre-collected from known defect-free, qualified workpieces and used to construct a pre-defined hierarchical feature memory library for normal samples. The three-dimensional point cloud data to be tested and the reference point cloud can have the same or different number of points; feature space matching is subsequently achieved through sampling, neighborhood retrieval, and interpolation mechanisms.

[0037] Specifically, step S102 further includes the following steps: Step S1021: Decentralize the 3D point cloud data containing the target to be detected to obtain decentralized point cloud data. Calculate the global covariance matrix of the decentralized point cloud data and perform eigenvalue decomposition to obtain multiple principal axis directions.

[0038] Specifically, to reduce the impact of random placement of industrial targets, changes in sensor viewing angle, and rigid pose variations on anomaly detection, this invention performs decentralization and principal axis alignment processing on the input point cloud. First, the centroid of the point cloud is calculated:

[0039] And translate each point to a coordinate system with the centroid as the origin:

[0040] Then, the global covariance matrix of the decentralized point cloud data is calculated:

[0041] right Eigenvalue decomposition yields three orthogonal principal axis direction vectors. , and .

[0042] Step S1022: Construct a rotation matrix based on multiple principal axis directions, and transform the decentralized point cloud data to the principal axis coordinate system based on the rotation matrix to obtain the transformed point cloud coordinate set.

[0043] To avoid alignment ambiguity caused by uncertain principal axis direction signs, this invention performs sign correction on the feature vectors based on multiple principal axis directions. Constructing the rotation matrix:

[0044] The decentralized point cloud data is then transformed to the principal axis coordinate system:

[0045] Step S1023: The normal vectors of the decentralized point cloud data are synchronously rotated based on the rotation matrix to obtain the rotated normal vectors.

[0046] If the surface normal of the point cloud is available or can be estimated, let the original normal be... Then the rotated normal vector is:

[0047] Step S1024: Perform unit sphere normalization on the transformed point cloud coordinate set to obtain the normalized point cloud coordinate set.

[0048] Specifically, the transformed point cloud coordinate set is normalized to a unit sphere so that the normalized point cloud coordinate set falls within the unit sphere. The calculation method is as follows:

[0049] Step S1025: The normalized point cloud coordinate set is concatenated with the rotated normal vector to obtain the preprocessed point cloud coordinate set.

[0050] Specifically, the preprocessed point cloud coordinate set forms a six-dimensional geometric input:

[0051] Step S1026: Perform sine and cosine mapping on each point cloud coordinate in the point cloud coordinate set using multiple frequency bands to obtain an initial high-dimensional geometric feature set.

[0052] Specifically, multi-frequency sine and cosine mappings are performed on each scalar component of the six-dimensional geometric input. Let the number of frequency bands be... For the first scalar component , ,calculate:

[0053]

[0054] in, Combining sine and cosine mappings yields the initial high-dimensional geometric features corresponding to each point cloud coordinate. f i The encoding results of all components are concatenated to obtain the initial high-dimensional geometric feature set. This invention utilizes sine and cosine mapping to expand minute spatial changes in point clouds into a higher-dimensional space, thereby enhancing the ability of subsequent anomaly detection to perceive local morphological changes.

[0055] Specifically, step S104 further includes the following steps: Step S1041: Sample the farthest point of the point cloud coordinate set step by step to form a multi-scale point set hierarchy.

[0056] Step S1042: In each level, a center point is selected from the point cloud coordinate set of the previous level using the farthest point sampling algorithm. For each center point, K neighboring points are selected from the point cloud coordinate set of the previous level using K-nearest neighbor retrieval to construct the local neighborhood corresponding to each level.

[0057] Specifically, in the first In each spatial level, the farthest point sampling algorithm is first used to select a set of center points from the point set of the previous level to obtain sparse center points that uniformly cover the input space; then, for the ... l Each center point in the spatial hierarchy Selecting points from the previous level's point set using K-nearest neighbor search. The nth neighboring point constitutes the nth... Local neighborhood at each spatial level .

[0058] Step S1043: Construct a local covariance matrix based on the geometric offset of each neighborhood point relative to the center point within the local neighborhood.

[0059] Specifically, for points in the neighborhood Calculate the spatial offset vector (i.e., geometric offset) relative to the center point:

[0060] Construct the local covariance matrix based on all neighborhood offsets:

[0061] Step S1044: Perform eigenvalue decomposition on the local covariance matrix to obtain multiple local eigenvalues.

[0062] Specifically, for Eigenvalue decomposition yields multiple local eigenvalues ​​arranged in descending order:

[0063] Step S1045: Calculate the local approximate curvature based on multiple local eigenvalues, and use the local approximate curvature as a measure of local deformation; wherein, the local approximate curvature includes the ratio of the minimum local eigenvalue to the sum of multiple local eigenvalues.

[0064] Specifically, the local approximate curvature is:

[0065] in To prevent positive numbers with a denominator of zero, this local approximate curvature reflects the degree to which the neighborhood point set deviates from the locally fitted plane. When the neighborhood approximates a flat surface, Smaller Approaching 0; when the neighborhood is located at an edge, corner, or complex curved surface region, Increase.

[0066] In an optional embodiment of the present invention, the deformation-sensitive modulation weight includes:

[0067] In the formula, For deformation-sensitive modulation weights, For local deformation measurement; The adjusted geometric offset includes:

[0068] In the formula, The adjusted geometric offset, It represents the geometric offset of the j-th neighboring point in the local neighborhood relative to the center point.

[0069] In some alternative embodiments provided by the present invention, the deformation-sensitive modulation weight can also be calculated by an exponential decay function with a scale parameter, a piecewise decay function, an inverse proportional decay function, or a normalized decay function, as long as it satisfies the monotonically decreasing relationship that the weight is smaller when the curvature is large and larger when the curvature is small.

[0070] In this embodiment of the invention, the deformation-sensitive modulation mechanism guided by local approximate curvature allows subtle offsets caused by minor defects in smooth regions to be fully preserved, while strong offsets caused by inherent surface changes in normal high-curvature structural regions are moderately suppressed, thereby reducing false alarms in normal structures.

[0071] Specifically, the multi-scale point cloud features include aggregated features across multiple spatial levels; step S108 also includes the following steps: Step S1081: The adjusted geometric offset is position-coded to obtain local geometric position-coded features.

[0072] Specifically, the adjusted geometric offset Perform sine encoding and cosine position encoding to obtain .

[0073] Step S1082: Channel concatenation is performed between the feature corresponding to the center point and the feature corresponding to the j-th neighboring point to obtain the extended feature.

[0074] Step S1083: The extended features and local geometric location coding features are fused to obtain the fused features of the local neighborhood; the fused features include:

[0075] In the formula, This represents the local fused feature obtained after deformation-sensitive modulation and feature fusion of the center point and the j-th neighboring point. This represents the extended feature obtained by concatenating the features corresponding to the center point and the features corresponding to the j-th neighboring point through channels. This indicates the adjusted geometric offset. This refers to the local geometric position encoding features obtained by position encoding the adjusted geometric offset, which match the dimensions of the extended features. This indicates element-wise multiplication.

[0076] Step S1084, for the first The fusion features of all neighboring points within the local neighborhood corresponding to the same center point in each spatial level are aggregated using dual-path pooling to obtain the center point at the th . Aggregation features at each spatial level; dual-pooling aggregation includes adding the max pooling result to the average pooling result; aggregation features include:

[0077] In the formula, Indicates the first The center point of each spatial level The aggregation characteristics, Indicates the first The center point of each spatial level The local fused features obtained after deformation-sensitive modulation and feature fusion with the j-th neighboring point, Indicates the first The center point of each spatial level The corresponding local neighborhood.

[0078] In this embodiment of the invention, max pooling is used to retain the most salient local geometric responses, while average pooling is used to maintain the overall statistical distribution of the neighborhood. The combination of the two can improve the stability of the feature representation. By repeating the above process at multiple spatial levels, multi-scale point cloud features are constructed.

[0079] Specifically, step S110 further includes the following steps: Step S1101: Retrieve the set of neighboring source points in the set of source points at each spatial level, and calculate the Gaussian weight based on the spatial distance between the target point and the source point in the set of neighboring source points.

[0080] Specifically, to perform point-level anomaly localization, it is necessary to restore the sparse hierarchical features obtained in the encoding stage to the initial point cloud resolution. For the first... l Let there be a spatial hierarchy, with its source point set denoted as . The source point feature set is For the target points in the initial 3D point cloud data to be tested... In the lSource point set at each spatial level Searching for the nearest source set And calculate the Gaussian weights based on the distance between the target point and the source point:

[0081] In the formula, For the first l The t-th target point p in each spatial hierarchy t With neighboring source points Gaussian weights Represents the L2 norm. Indicates the first l Gaussian kernel bandwidth parameters corresponding to each spatial level.

[0082] Step S1102: Based on Gaussian weights and using normalized weighted summation, perform back interpolation projection on the multi-scale point cloud features to obtain the recovered feature sets corresponding to each level after interpolation; l Spatial hierarchical recovery feature set include:

[0083]

[0084] In the formula, Indicates the first Interpolate spatial levels back to the target point The recovery features obtained later Indicates the first Target point in each spatial level With the source Gaussian weights between them Indicates the first Source point in each spatial hierarchy Corresponding source features Indicates the target point In the The set of neighboring source points retrieved from the source point set at each spatial level. This represents the number of points in the initial point cloud.

[0085] In this embodiment of the invention, the above interpolation projection is performed independently for sparse features at different spatial levels, and the interpolated features at each spatial level are mapped back to the initial number of points.

[0086] Step S1103: Channel splicing is performed between the initial high-dimensional geometric feature set and the recovery feature set corresponding to each level to obtain a dense feature descriptor that corresponds one-to-one with the three-dimensional point cloud data to be tested.

[0087] For example, using two sparse spatial levels, the dense feature descriptor includes:

[0088] This parallel decoding mechanism avoids mutual interference between features at different levels during the layer-by-layer fusion process, enabling shallow local details, mid-level neighborhood structures, and deep large-scale geometric information to be independently preserved in the channel dimension.

[0089] Specifically, step S112 further includes the following steps: Step S1121: Slice the dense feature descriptor of each point in the 3D point cloud data to be tested according to the source spatial level to obtain the test features corresponding to each spatial level.

[0090] Step S1122: Perform the following steps on the test features of each spatial level and the corresponding spatial level features in the preset sample hierarchical feature memory: Norm normalization is performed, and multiple nearest neighbor features are retrieved at the corresponding spatial level.

[0091] Step S1123: Based on the aggregation of distance statistics of multiple nearest neighbor features at each spatial level, calculate the point-level anomaly score of the 3D point cloud data to be tested.

[0092] Step S1124: Calculate the object-level outlier scores by taking the statistics of all point-level outlier scores in the data.

[0093] Specifically, firstly, normal, defect-free samples are collected as reference point cloud data, and dense feature descriptors are extracted according to the method provided in this embodiment of the invention. These descriptors are then segmented and stored according to the spatial hierarchy of feature sources to form a preset sample hierarchical feature memory library.

[0094] During online detection, for any point in the 3D point cloud data to be tested... Its dense feature descriptor is divided into hierarchical segments. ,in Before calculating the distance, the test features and memory features at each level are processed. Norm normalization:

[0095]

[0096] Retrieve from the normal sample feature memory of the corresponding level. Calculate the point-level anomaly score using the nearest neighbor features. In one feasible implementation, the point-level anomaly score is:

[0097] in Indicates the first In each spatial hierarchy memory bank and The corresponding number The feature of each point is a nearest-neighbor normal sample. The larger the point-level anomaly score, the more significantly the point deviates from the normal sample distribution in the feature space. After traversing all points in the point cloud to be tested, a point-level anomaly heatmap can be obtained.

[0098] In other alternative implementations, the point-level anomaly score can also be calculated using the mean, weighted mean, median, minimum distance, or quantile statistics of the features of multiple nearest normal samples at each level. The object-level anomaly score can be determined by the maximum value of the point-level anomaly score, the mean of the top few largest values, or the quantile statistics. In one feasible implementation, the object-level anomaly score is defined as:

[0099] Point-level anomaly scores can be compared with preset point-level thresholds to locate local defect areas, and object-level anomaly scores can be compared with preset object-level thresholds to output overall pass or fail judgment results.

[0100] The method provided in this embodiment of the invention further includes: outputting a point-level anomaly heatmap, an object-level anomaly score, and an overall qualified or abnormal determination result.

[0101] To verify the effectiveness of the method of this invention in high-resolution 3D point cloud anomaly detection tasks, the method was validated on three publicly available 3D anomaly detection datasets: MiniShift, Real3D-AD, and Anomaly-ShapeNet. These datasets cover high-resolution industrial parts, real scanned objects, and multi-class synthetic 3D shapes, respectively, and can comprehensively evaluate the performance of the method of this invention in detecting minute defects, suppressing false alarms on complex surfaces, and cross-class generalization.

[0102] In this embodiment, the area under the receiver operating characteristic (ROC) curve at the object level and the area under the receiver operating characteristic (ROC) curve at the point level are used as evaluation indicators. The object-level ROC curve, denoted as O-ROC, is used to evaluate the model's ability to determine whether the overall point cloud is abnormal; the point-level ROC curve, denoted as P-ROC, is used to evaluate the model's ability to locate local abnormal regions on the point cloud surface. For industrial quality inspection scenarios, O-ROC reflects the overall pass / fail determination capability, while P-ROC reflects the ability to locate minor defects. Combining the two allows for a more comprehensive evaluation of the practical effectiveness of the 3D point cloud anomaly detection method.

[0103] As shown in Table 1, on the MiniShift full test set, the average O-ROC and P-ROC of the method of this invention reached 76.5% and 71.2%, respectively. The MiniShift dataset has a low proportion of anomalous regions and a high point cloud resolution, making it easy for minor defects to be weakened during conventional downsampling or global feature aggregation. Compared with existing methods, this invention preserves local geometric information through spatial calibration and positional encoding, suppresses interference from normal high-curvature structures through deformation-sensitive geometric encoding, and recovers dense features at the initial resolution through multi-scale parallel decoding. Therefore, it achieves better overall performance in both object-level discrimination and point-level localization.

[0104] Table 1 Comparison of anomaly detection performance of different methods on the MiniShift dataset.

[0105] Furthermore, to verify the generalization ability of the method of the present invention on different data distributions and geometric structures, tests were conducted on the Real3D-AD and Anomaly-ShapeNet datasets. As shown in Table 2, on the Real3D-AD dataset, the method of the present invention achieved an average O-ROC of 84.2% and an average P-ROC of 83.6%; on the Anomaly-ShapeNet dataset, the method of the present invention achieved an average O-ROC of 90.3% and an average P-ROC of 93.7%. Real3D-AD is derived from real scanning scenes, and point clouds often contain uneven sampling, scanning noise, and complex surface variations; Anomaly-ShapeNet covers more categories and can reflect the anomaly detection capability in cross-category situations. Experimental results show that the method of the present invention can not only detect small anomalies in high-resolution industrial part point clouds, but also maintain relatively stable anomaly discrimination and localization capabilities in real scanned point clouds and multi-category 3D shapes.

[0106] Table 2 Performance of the method of the present invention on different 3D anomaly detection datasets

[0107] To further illustrate the role of local spatial normalization and deformation-sensitive geometric coding in this invention, ablation experiments of the core components were conducted on the MiniShift dataset. The ablation results are shown in Table 3. Without local spatial normalization and deformation-sensitive geometric coding, the baseline method's O-ROC and P-ROC were 72.5% and 62.4%, respectively. Introducing local spatial normalization alone increased the O-ROC to 74.7%, indicating that local spatial normalization can weaken the impact of macroscopic pose and overall structural differences on anomaly detection. Introducing deformation-sensitive geometric coding alone increased the P-ROC to 69.1%, indicating that the curvature-guided deformation-sensitive modulation mechanism can enhance the model's ability to locate small local deformations. When both were used simultaneously, the O-ROC and P-ROC reached 76.5% and 71.2%, respectively, achieving the best detection results, demonstrating that local spatial normalization and deformation-sensitive geometric coding have a synergistic effect in anomaly detection of complex high-resolution point clouds.

[0108] Table 3 Ablation Experiment Results of the Core Components of the Invention

[0109] In summary, the training-free 3D point cloud anomaly detection method based on deformation-sensitive features provided by this invention can effectively detect minute defects in complex 3D surfaces without the need for anomaly sample annotation and supervised training. Quantitative and ablation experiments show that spatial calibration and positional encoding, deformation-sensitive geometric encoding, multi-scale parallel decoding projection, and hierarchical feature memory measurement collectively improve the accuracy, localization capability, and engineering applicability of anomaly detection. Compared with existing technologies, this invention has at least the following advantages: (1) The online detection process of the present invention does not rely on defect sample labeling and backpropagation training process. It only needs to use normal samples to build a reference feature memory library in advance, which is suitable for rapid deployment in industrial sites and frequent product model switching scenarios.

[0110] (2) The present invention introduces a deformation-sensitive feature modulation mechanism. In one embodiment, the local approximate curvature is calculated based on the eigenvalues ​​of the local covariance matrix, and the feature response is adjusted by the local approximate curvature. This enables the feature extraction process to automatically adjust the response intensity according to the degree of local surface deformation, retaining the sensitivity to minor defects in smooth areas and reducing the risk of false alarms in normal high curvature structure areas.

[0111] (3) The present invention restores sparse hierarchical features to the initial point cloud resolution by multi-scale parallel decoding projection, avoiding feature smoothing and dilution of small anomaly information caused by layer-by-layer cascaded restoration.

[0112] (4) The present invention adopts a hierarchical feature memory and hierarchical distance measurement mechanism to enable features of different scales to be independently normalized and retrieved in their respective subspaces, thereby reducing the impact of differences in the numerical dimensions of features of different scales on anomaly scoring.

[0113] (5) The present invention can simultaneously output point-level anomaly heat map and object-level anomaly score, which can be used for local defect location and overall qualification judgment, and has good practical value for industrial testing.

[0114] Figure 2 This is a schematic diagram of a training-free 3D point cloud anomaly detection system based on deformation-sensitive features, according to an embodiment of the present invention. Figure 2 As shown, the system includes: a data preprocessing module 10, a deformation measurement calculation module 20, a geometric offset adjustment module 30, a deformation-sensitive geometric encoding module 40, a multi-scale parallel decoding projection module 50, and a hierarchical memory measurement module 60.

[0115] Specifically, the data preprocessing module 10 is used to perform spatial calibration and normalization processing on the three-dimensional point cloud data containing the target to be detected, to obtain the preprocessed point cloud coordinate set, and to perform position encoding on the point cloud coordinate set to obtain the initial high-dimensional geometric feature set. The deformation measurement calculation module 20 is used to sample the farthest point of the point cloud coordinate set step by step, construct local neighborhoods at multiple spatial levels, and calculate local deformation measurement based on the geometric offset features of the local neighborhoods. The geometric offset adjustment module 30 is used to generate deformation-sensitive modulation weights based on local deformation metrics, and adjust the geometric offset features of the local neighborhood based on the deformation-sensitive modulation weights to obtain the adjusted geometric offset. The deformation-sensitive geometric coding module 40 is used to encode the adjusted geometric offset and perform channel splicing and aggregation with the corresponding features of the center point and neighboring points to construct deformation-sensitive multi-scale point cloud features. The multi-scale parallel decoding projection module 50 is used to map the multi-scale point cloud features back to the initial point cloud resolution, and perform channel stitching with the initial high-dimensional geometric feature set to obtain a dense feature descriptor that corresponds one-to-one with the three-dimensional point cloud data to be measured. The hierarchical memory measurement module 60 is used to perform hierarchical nearest neighbor retrieval and distance aggregation on the dense feature descriptors of the 3D point cloud data to be tested in the preset sample hierarchical feature memory to obtain anomaly detection results; the anomaly detection results include point-level anomaly scores and object-level anomaly scores; the preset sample hierarchical feature memory is constructed based on the dense feature descriptors of normal sample reference point cloud data.

[0116] The present invention also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method provided in the embodiments of the present invention.

[0117] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0118] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0119] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0120] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0121] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0122] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0123] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A training-free 3D point cloud anomaly detection method based on deformation-sensitive features, characterized in that, The method includes: Spatial calibration and normalization are performed on the three-dimensional point cloud data containing the target to be detected to obtain a preprocessed set of point cloud coordinates. The point cloud coordinate set is then position-encoded to obtain an initial set of high-dimensional geometric features. The point cloud coordinate set is sampled level by level for the farthest point, local neighborhoods are constructed at multiple spatial levels, and local deformation measures are calculated based on the geometric offset features of the local neighborhoods. Deformation-sensitive modulation weights are generated based on the local deformation metric, and the geometric offset features of the local neighborhood are adjusted based on the deformation-sensitive modulation weights to obtain the adjusted geometric offset. The adjusted geometric offset is positionally encoded, and channel splicing and aggregation are performed by combining the center point corresponding features and the neighborhood point corresponding features to construct deformation-sensitive multi-scale point cloud features. The multi-scale point cloud features are mapped back to the initial point cloud resolution, and channel-separated with the initial high-dimensional geometric feature set to obtain a dense feature descriptor that corresponds one-to-one with the three-dimensional point cloud data to be measured. In a pre-defined sample hierarchical feature memory, hierarchical nearest neighbor retrieval and distance aggregation are performed on the dense feature descriptors of the three-dimensional point cloud data to be tested to obtain anomaly detection results; the anomaly detection results include point-level anomaly scores and object-level anomaly scores; the pre-defined sample hierarchical feature memory is constructed based on the dense feature descriptors of normal sample reference point cloud data.

2. The method according to claim 1, characterized in that, Spatial calibration and normalization are performed on the 3D point cloud data containing the target to be detected to obtain a preprocessed set of point cloud coordinates, including: The 3D point cloud data containing the target to be detected is decentralized to obtain decentralized point cloud data. The global covariance matrix of the decentralized point cloud data is calculated and eigenvalue decomposition is performed to obtain multiple principal axis directions. A rotation matrix is ​​constructed based on the multiple principal axis directions, and the decentralized point cloud data is transformed to the principal axis coordinate system based on the rotation matrix to obtain the transformed point cloud coordinate set. The normal vectors of the decentralized point cloud data are synchronously rotated based on the rotation matrix to obtain the rotated normal vectors; The transformed point cloud coordinate set is normalized by a unit sphere to obtain a normalized point cloud coordinate set. The normalized point cloud coordinate set is concatenated with the rotated normal vector to obtain the preprocessed point cloud coordinate set.

3. The method according to claim 1, characterized in that, Position encoding is performed on the point cloud coordinate set to obtain an initial high-dimensional geometric feature set, including: applying sine and cosine mapping to each point cloud coordinate in the point cloud coordinate set using multiple frequency bands to obtain the initial high-dimensional geometric feature set.

4. The method according to claim 1, characterized in that, The point cloud coordinate set is sampled level by level for farthest points to construct local neighborhoods at multiple spatial levels, and local deformation metrics are calculated based on the geometric offset features of the local neighborhoods, including: The point cloud coordinate set is sampled level by level for the farthest point to form a multi-scale point set hierarchy; In each level, the center point is selected from the point cloud coordinate set of the previous level by the farthest point sampling algorithm. For each center point, K neighbor points are selected from the point cloud coordinate set of the previous level by K nearest neighbor retrieval to construct the local neighborhood corresponding to each level. Based on the geometric offset of each neighboring point relative to the center point within the local neighborhood, a local covariance matrix is ​​constructed. The local covariance matrix is ​​decomposed into eigenvalues ​​to obtain multiple local eigenvalues; The local approximate curvature is calculated based on the multiple local eigenvalues, and the local approximate curvature is used as a measure of local deformation; wherein, the local approximate curvature includes the ratio of the minimum local eigenvalue to the sum of the multiple local eigenvalues.

5. The method according to claim 1, characterized in that, The deformation-sensitive modulation weights include: In the formula, The deformation-sensitive modulation weights, This is the measure of the local deformation; The adjusted geometric offset includes: In the formula, The adjusted geometric offset, The geometric offset of the j-th neighboring point in the local neighborhood relative to the center point.

6. The method according to claim 1, characterized in that, The multi-scale point cloud features include aggregated features at multiple spatial levels; The adjusted geometric offset is positionally encoded, and channel stitching and aggregation are performed by combining the center point corresponding features and the neighboring point corresponding features to construct deformation-sensitive multi-scale point cloud features, including: The adjusted geometric offset is position-encoded to obtain local geometric position-encoded features; The extended feature is obtained by concatenating the features corresponding to the center point and the features corresponding to the j-th neighboring point through channels. The extended features are fused with the local geometric location encoding features to obtain fused features of the local neighborhood; the fused features include: In the formula, This represents the local fused feature obtained after deformation-sensitive modulation and feature fusion of the center point and the j-th neighboring point. This represents the extended feature obtained by concatenating the features corresponding to the center point and the features corresponding to the j-th neighboring point through channels. This indicates the adjusted geometric offset. This refers to the local geometric position encoding feature obtained by position encoding the adjusted geometric offset, which matches the dimension of the extended feature. This represents element-wise multiplication; For the The fusion features of all neighboring points within the local neighborhood corresponding to the same center point in each spatial level are aggregated using dual-path pooling to obtain the center point in the 1st spatial level. Aggregation features at each spatial level; the dual-pooling aggregation includes adding the max pooling result to the average pooling result; the aggregation features include: In the formula, Indicates the first The center point of each spatial level The aggregation characteristics, Indicates the first The center point of each spatial level The local fused features obtained after deformation-sensitive modulation and feature fusion with the j-th neighboring point, Indicates the first The center point of each spatial level The corresponding local neighborhood.

7. The method according to claim 1, characterized in that, The multi-scale point cloud features are mapped back to the initial point cloud resolution, and channel concatenation is performed with the initial high-dimensional geometric feature set to obtain a dense feature descriptor corresponding one-to-one with the three-dimensional point cloud data to be measured, including: Retrieve neighboring source point sets from the source point sets at each spatial level, and calculate Gaussian weights based on the spatial distances between the target point and the source point in the neighboring source point sets; the Gaussian weights include: In the formula, For the t-th target point p t With source point p s Gaussian weights Represents the L2 norm. This represents the Gaussian kernel bandwidth parameter; Based on the Gaussian weights and using normalized weighted summation, the multi-scale point cloud features are back-interpolated and projected to obtain the recovered feature sets corresponding to each level after interpolation; l Spatial hierarchical recovery feature set include: In the formula, Indicates the first Interpolate spatial levels back to the target point The recovery features obtained later Indicates the first Target point in each spatial level With the source Gaussian weights between them Indicates the first Source point in each spatial hierarchy Corresponding source features Indicates the target point In the The set of neighboring source points retrieved from the source point set at each spatial level. This represents the number of points in the initial point cloud; The initial high-dimensional geometric feature set and the corresponding recovery feature set of each level are concatenated to obtain a dense feature descriptor that corresponds one-to-one with the three-dimensional point cloud data to be tested.

8. The method according to claim 1, characterized in that, In a pre-defined hierarchical feature memory, hierarchical nearest neighbor retrieval and distance aggregation are performed on the dense feature descriptors of the 3D point cloud data to be tested to obtain anomaly detection results, including: The dense feature descriptor of each point in the three-dimensional point cloud data to be tested is sliced ​​according to the source spatial level to obtain the test feature corresponding to each spatial level. For each spatial level, the test features and the corresponding spatial level features in the preset sample hierarchical feature memory are executed. Norm normalization is performed, and multiple nearest neighbor features are retrieved at the corresponding spatial level; The point-level anomaly score of the three-dimensional point cloud data under test is calculated by aggregating distance statistics of multiple nearest neighbor features at each spatial level. Based on the statistics of all point-level anomaly scores in the three-dimensional point cloud data to be tested, the object-level anomaly score is calculated.

9. A training-free 3D point cloud anomaly detection system based on deformation-sensitive features, characterized in that, This system is used to implement the training-free 3D point cloud anomaly detection method based on deformation-sensitive features as described in any one of claims 1-8; the system includes: a data preprocessing module, a deformation metric calculation module, a geometric offset adjustment module, a deformation-sensitive geometric encoding module, a multi-scale parallel decoding projection module, and a hierarchical memory metric module; wherein... The data preprocessing module is used to perform spatial calibration and normalization on the three-dimensional point cloud data containing the target to be detected, to obtain a preprocessed set of point cloud coordinates, and to perform position encoding on the set of point cloud coordinates to obtain an initial set of high-dimensional geometric features. The deformation measurement calculation module is used to sample the farthest point of the point cloud coordinate set step by step, construct local neighborhoods at multiple spatial levels, and calculate local deformation measurement based on the geometric offset features of the local neighborhoods. The geometric offset adjustment module is used to generate deformation-sensitive modulation weights based on the local deformation metric, and adjust the geometric offset features of the local neighborhood based on the deformation-sensitive modulation weights to obtain the adjusted geometric offset. The deformation-sensitive geometric coding module is used to perform position coding on the adjusted geometric offset, and to perform channel splicing and aggregation by combining the center point corresponding features and the neighborhood point corresponding features to construct deformation-sensitive multi-scale point cloud features. The multi-scale parallel decoding projection module is used to map the multi-scale point cloud features back to the initial point cloud resolution, and perform channel stitching with the initial high-dimensional geometric feature set to obtain a dense feature descriptor that corresponds one-to-one with the three-dimensional point cloud data to be measured. The hierarchical memory measurement module is used to perform hierarchical nearest neighbor retrieval and distance aggregation on the dense feature descriptors of the three-dimensional point cloud data to be tested in the preset sample hierarchical feature memory to obtain anomaly detection results; the anomaly detection results include point-level anomaly scores and object-level anomaly scores; the preset sample hierarchical feature memory is constructed based on the dense feature descriptors of normal sample reference point cloud data.

10. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as claimed in any one of claims 1-8.

Images