A single-view point cloud feature extraction-based weld joint recognition method and system

Abstract

The application discloses a kind of based on single view point cloud feature extraction of weld seam identification method and system, it is related to welding identification technical field, including steps: obtaining the workpiece three-dimensional cleaning point cloud dataset under current view point;Obtain the spherical neighborhood point cloud set corresponding to each single point cloud, obtain the single point cloud geometry feature corresponding to single point cloud;Using clustering algorithm, the feature vectors in feature vector set are classified clustering;Obtain view point vector and shape vector;According to the included angle between view point vector and shape vector, identify non-weld seam convex point and weld seam concave point.The method provided by the application can identify real weld seam point cloud without initial weld seam fitting, uses the same concept for real weld seam point identification for straight-line weld seam and curve weld seam, and has strong versatility;Meanwhile, the technical problem that the present art lacks distinction of real weld seam point cloud after identifying potential weld seam area, resulting in low weld seam identification precision, is solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent welding technology, and in particular to a weld seam recognition method and system based on single-view point cloud feature extraction. Background Technology

[0002] Welding, as a widely used material joining technology, plays a crucial role in important industrial sectors such as automotive manufacturing, construction, transportation, and aerospace. With the rapid development of industrial automation technology, the application scope of welding robots is constantly expanding, significantly improving the production efficiency and quality consistency of the welding process. Welding robots perceive their working environment through a diverse range of sensor systems. While different sensors have their advantages in specific application scenarios, visual sensing technology, with its non-contact detection characteristics, strong information acquisition capabilities, high detection accuracy, and strong scene adaptability, has become the industry's preferred sensing technology. In complex real-world application scenarios, welding robots need to acquire images or point cloud information of the workpiece through visual sensors and combine this with specific algorithms developed by engineers to accurately extract weld features from the visual information, thereby performing the welding task. However, existing image-based processing methods have significant limitations. Point cloud-based processing methods can obtain global three-dimensional information of the workpiece, providing welding robots with richer and more complete workpiece characterization data, thus enabling effective identification of complex spatial welds. Therefore, point cloud-based processing methods are more widely used in welding robot systems, significantly improving the accuracy, stability, and efficiency of robots in welding operations.

[0003] Currently, methods for weld identification and extraction based on weld point cloud information to construct geometric features are susceptible to noise interference, have high computational complexity, and are time-consuming, making it difficult to meet the application requirements of real-time welding operations. In addition, since the weld locations of steel structure workpieces are mainly distributed in concave areas, existing methods generally lack the ability to accurately distinguish between concave and convex areas (the actual weld areas) after identifying potential weld areas, affecting the accuracy of weld identification. Furthermore, existing methods are usually designed separately for specific workpiece types (curved welds or straight welds), resulting in poor adaptability to different types of workpieces and severely restricting the versatility and practicality of the methods.

[0004] Therefore, it is necessary to propose a weld seam identification method and system based on single-view point cloud feature extraction to solve or at least alleviate some of the above-mentioned defects.

[0005] In existing technologies, vision-based intelligent welding technology is mainly applied to relatively simple scenarios, making it difficult to cope with the diversity of welding scenarios, the uncertainty of the number of welds, and the complexity of welded workpieces. Therefore, improving the adaptability of welding robots in complex environments and developing more intelligent welding systems are particularly important. It is necessary to propose a weld recognition method and system based on single-view point cloud feature extraction to solve or at least alleviate some of the above-mentioned shortcomings. Summary of the Invention

[0006] The main objective of this invention is to provide a weld seam recognition method and system based on single-view point cloud feature extraction, which aims to solve the technical problems in the prior art, such as the lack of differentiation between the actual point cloud of the weld seam after identifying potential weld seam areas, resulting in low weld seam recognition accuracy, and the inability of the prior art to simultaneously satisfy the recognition of straight weld seams and curved weld seams, resulting in low applicability of weld seam recognition.

[0007] To achieve the above objectives, the present invention provides a weld seam recognition method based on single-viewpoint point cloud feature extraction, comprising the following steps: S10, acquiring a three-dimensional cleaning point cloud dataset of the workpiece under the current viewpoint. The workpiece 3D cleaning point cloud dataset consists of multiple single points The cloud formations correspond to the viewpoint coordinates of the current viewpoint.

[0008] S20, Obtain each single point cloud based on the first preset search radius r. The corresponding spherical neighborhood point cloud set Median principal component analysis was used to analyze the point cloud set in the neighborhood of the sphere. Analyze the point cloud in the data to obtain each individual point cloud. Corresponding single-point cloud geometric features ;

[0009] S30, which combines all single-point cloud geometric features Combine them to form a set of feature vectors A clustering algorithm is used to perform binary classification clustering on the feature vectors in the feature vector set, and each single point cloud is then classified. Feature point cloud obtained by marking welding labels The welding label is either the first label or the second label;

[0010] S40, acquire feature point cloud From the perspective of the center of mass with viewpoint coordinates Constructing a pointing viewpoint view vector Obtain feature point cloud Locally weighted centroid With shape-weighted centroid Constructing the shape-weighted centroid Shape vector ;

[0011] Among them, based on feature point cloud The corresponding spherical neighborhood point cloud set Calculate and obtain the centroid of the viewpoint and local weighted centroid Based on feature point cloud The shape-weighted centroid is obtained by calculating the corresponding spherical ring point cloud set. The inner diameter of the spherical ring point cloud set is the first preset search radius, and the outer diameter of the spherical ring point cloud set is the second preset search radius, which is greater than the first preset search radius.

[0012] S50, if the viewpoint vector With shape vector If the included angle between them is greater than 90 degrees, then the feature point cloud is identified. For non-weld seam protrusions; if the viewpoint vector With shape vector If the included angle between them is no greater than 90 degrees, then the feature point cloud is identified. This is a weld seam depression.

[0013] Furthermore, it also includes the step: S60, which involves processing the identified feature point cloud. Update to target point cloud The feature fusion method is used to integrate the target point cloud with initial concavity / convexity attribute labels. Perform semi-supervised optimization and update the current target point cloud. Weld markings.

[0014] Furthermore, step S60 specifically includes:

[0015] S61, the labeled feature point cloud Update to target point cloud ;

[0016] S62, for all target point clouds Construct a KD-tree data structure;

[0017] S63, acquires point cloud for each target based on preset search distance. The corresponding search neighborhood point cloud set;

[0018] S64, obtain the current target point cloud in each search neighborhood point cloud set. With each neighboring point cloud Distance weight ;

[0019] S65, based on the target point cloud in the search neighborhood point cloud set The labels are used to calculate the total weight of convex points within each search neighborhood point cloud set. The total weight of the concave points ;

[0020] S66, if the total weight of the convex points is greater than the total weight of the concave points, then the current target point cloud is corrected to convex points; if the total weight of the convex points is less than the total weight of the concave points, then the current target point cloud is corrected to concave points; if the total weight of the convex points is equal to the total weight of the concave points, then the current target point cloud is not corrected.

[0021] S67. After correction, the point cloud category weights within each search neighborhood are reacquired. The point cloud category weights include the total weights of convex points and concave points. If the point cloud category weights within the search neighborhood are not less than the first threshold, then the target point cloud corresponding to the search neighborhood is determined. A point is considered a high-confidence point; if the category weight of the point cloud within the search neighborhood is less than the first threshold, then the target point cloud corresponding to the search neighborhood is determined. This is a low confidence point;

[0022] S681, For low confidence points, expand and update the preset search distance based on the preset step size, and proceed to step S63; For high confidence points, determine the weld identifier of the target point cloud based on the high confidence points;

[0023] Repeat steps S63 to S681 until the preset termination condition is met, progressively assimilate the attributes of the points with ambiguous attributes, and update the target point cloud. Weld markings.

[0024] Furthermore, the preset termination condition is reaching a preset number of iterations or when the proportion of attribute correction points in a single iteration to the total target points is less than a preset percentage.

[0025] Furthermore, using the formula The distance weights are calculated, where, Distance weights For the target point cloud, For adjacent point clouds, For target point cloud With one of the neighboring point clouds (neighboring point clouds) The Euclidean distance between them It is the adjustment coefficient. Using formula The total weight of convex points within the search neighborhood is statistically analyzed. , Point cloud representing convex points The distance weight is calculated using the formula. The total weight of concave points within the search neighborhood is statistically analyzed. , Target point cloud representing concave points Distance weights. Preferably, .

[0026] Furthermore, median principal component analysis was used to analyze the point cloud set in the neighborhood of the sphere. The point cloud was analyzed to obtain three eigenvalues. , , ,in, ; Use formula Perform calculations to obtain each single point cloud. Single-point cloud geometric features .

[0027] Furthermore, using the formula Calculate and obtain the centroid of the computational viewpoint ,in, These are weighting coefficients. , and These are the current feature point clouds. of The eigenvalue of the j-th point in the neighborhood. , For the current feature point cloud of The total number of points in the neighborhood, For the current feature point cloud of The set of points in the neighborhood, For the current feature point cloud The coordinates of the j-th point within the spherical neighborhood;

[0028] Using formula Calculate and obtain the local weighted centroid ,in, These are the centroid weighting coefficients. , For the current feature point cloud and k-th point in the neighborhood Euclidean distance, For the current feature point cloud The coordinates of the k-th point within the spherical neighborhood;

[0029] Calculate using formula Calculate and obtain the shape-weighted centroid ,in, These are shape weighting coefficients. , For the current feature point cloud and The nth point in the neighborhood Euclidean distance, , For the current feature point cloud of The total number of points in the neighborhood; For the current feature point cloud of The coordinates of the nth point in the neighborhood, For the current feature point cloud of The set of points in the neighborhood, For the current feature point cloud of The coordinates of the nth point in the neighborhood,

[0030] The first preset search radius, The second preset search radius.

[0031] Furthermore, using the formula For feature point clouds Perform initial weld marking.

[0032] Furthermore, the steps for obtaining the workpiece 3D cleaning point cloud dataset from the current viewpoint specifically include:

[0033] Obtain the workpiece 3D point cloud dataset from the current viewpoint; perform voxel downsampling on the point cloud data in the workpiece 3D point cloud dataset to obtain the workpiece 3D cleaning point cloud dataset from the current viewpoint. The workpiece 3D cleaning point cloud dataset is formed by combining multiple single point clouds.

[0034] Furthermore, it also includes the step of fitting all weld seam depressions to obtain weld seam identification lines.

[0035] The present invention also provides a weld seam recognition system based on single-view point cloud feature extraction, including an coded structured light sensor and a processing device. The coded structured light sensor is used to acquire a three-dimensional cleaning point cloud dataset of the workpiece, and the processing device is used to implement the steps of the above-mentioned weld seam recognition method based on single-view point cloud feature extraction.

[0036] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described weld seam identification method based on single-view point cloud feature extraction.

[0037] Compared with existing technologies, the weld seam identification method based on single-view point cloud feature extraction provided by this invention has the following beneficial effects:

[0038] The weld seam recognition method based on single-view point cloud feature extraction provided by this invention first obtains a 3D cleaning point cloud dataset corresponding to the target workpiece, which is formed by combining multiple single point clouds; then, it obtains the single point clouds. The corresponding spherical neighborhood point cloud set Median principal component analysis was used to analyze the point cloud set in the neighborhood of the sphere. Analyze the point cloud in the data to obtain each individual point cloud. Corresponding single-point cloud geometric features Next, based on the feature vector set, binary classification and welding labeling are performed on the single point cloud in the workpiece 3D cleaning point cloud dataset to obtain the feature point cloud. The binary classification involves classifying the point cloud into either a first label or a second label. Then, based on the single point cloud... The corresponding spherical neighborhood point cloud set Calculate and obtain the centroid of the viewpoint Local weighted centroid and shape-weighted centroid Obtain the coordinates from the centroid of the viewpoint to the viewpoint. view vector Obtain the locally weighted centroid Pointing to shape-weighted centroid Shape Vector Finally, based on viewpoint vectors With shape vector The included angle between the points is used to identify non-weld seam protrusions or weld seam depressions in each individual point cloud of the workpiece's 3D cleaning point cloud dataset. The method of this invention, in acquiring the individual point clouds of the target workpiece... Then, it can directly perform further identification on a single point cloud to determine the single point cloud. Whether it is a real weld point, where a real weld point is a weld concave point, does not require a single point cloud in the workpiece 3D cleaning point cloud dataset. The actual weld points can be identified by performing initial weld fitting. The same approach is used for both straight and curved welds, making it highly adaptable to different workpiece types and possessing versatility and practicality. Furthermore, it is based on viewpoint vectors. With shape vector Angle identifier feature point cloud between Whether it is a weld concave point or not, it solves the technical problem in the prior art that the lack of differentiation of the actual point cloud of the weld after identifying the potential weld area leads to low weld recognition accuracy. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0040] Figure 1 This is a flowchart illustrating a weld seam identification method based on single-view point cloud feature extraction in one embodiment of the present invention.

[0041] Figure 2 This is a flowchart illustrating the feature point cloud based on the angle between the viewpoint vector and the shape vector in one embodiment of the present invention. In this flowchart, 2a is a schematic diagram of the clustering results of a cylindrical workpiece, and 2b is a schematic diagram of the classification of concave and convex points of a cylindrical workpiece.

[0042] Figure 3 This is a schematic diagram of the principle of weld seam recognition for rectangular workpieces in another embodiment of the present invention, wherein 3a is a schematic diagram of the original point cloud of the rectangular workpiece, 3b is a schematic diagram of the initial feature point cloud differentiation of the rectangular workpiece, and 3c is a schematic diagram of the result of semi-supervised optimization of concave and convex points using the feature fusion method.

[0043] Figure 4 This is a schematic diagram of the principle of weld seam recognition for a round tube workpiece in one embodiment of the present invention, wherein 4a is a schematic diagram of the original point cloud of the round tube workpiece, 4b is a schematic diagram of the initial feature point cloud differentiation of the round tube workpiece, and 4c is a schematic diagram of the result of semi-supervised optimization of concave and convex points using the feature fusion method.

[0044] Figure 5 This is a schematic diagram illustrating the principle of identifying non-weld seam protrusions and weld seam concave points using the angle between the viewpoint vector and the shape vector in one embodiment of the present invention.

[0045] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0046] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0048] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0049] Furthermore, the use of terms such as "first" and "second" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. When the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0050] Research reveals that existing point cloud-based weld seam recognition technologies primarily employ mathematical modeling methods. These methods involve pre-fitting the point cloud of the target workpiece into a plane or cylinder to obtain the intersection lines as weld seam features. However, this approach has limited applicability, requiring separate algorithm design for planar and cylindrical models. Furthermore, the obtained intersection lines often contain both convex (false weld seams) and concave (real weld seams) regions, necessitating manual secondary screening. While existing point cloud descriptor-based methods offer richer feature representation, they suffer from high computational complexity and long processing times, making them unsuitable for the industrial applications requiring real-time robotic welding. Finally, existing point cloud-based weld seam recognition methods lack the ability to effectively distinguish between convex and concave regions, exhibiting significant technical limitations.

[0051] Please refer to the appendix. Figure 1 , Figure 2 , Figure 3 , Figure 4 as well as Figure 5 This invention provides a weld seam recognition method based on single-viewpoint point cloud feature extraction, comprising the following steps: S10, acquiring a 3D cleaning point cloud dataset of the workpiece under the current viewpoint. The workpiece 3D cleaning point cloud dataset consists of multiple single points The cloud formations correspond to the viewpoint coordinates of the current viewpoint.

[0052] S20, Obtain each single point cloud based on the first preset search radius r. The corresponding spherical neighborhood point cloud set Median principal component analysis was used to analyze the point cloud set in the neighborhood of the sphere. Analyze the point cloud in the data to obtain each individual point cloud. Corresponding single-point cloud geometric features ;

[0053] S30, which combines all single-point cloud geometric features Combine them to form a set of feature vectors A clustering algorithm is used to perform binary classification clustering on the feature vectors in the feature vector set, and each single point cloud is then classified. Feature point cloud obtained by marking welding labels The welding label is either the first label or the second label;

[0054] S40, acquire feature point cloud From the perspective of the center of mass with viewpoint coordinates Constructing a pointing viewpoint view vector Obtain feature point cloud Locally weighted centroid With shape-weighted centroid Constructing the shape-weighted centroid Shape vector ;

[0055] Among them, based on feature point cloud The corresponding spherical neighborhood point cloud set Calculate and obtain the centroid of the viewpoint and local weighted centroid Based on feature point cloud The shape-weighted centroid is obtained by calculating the corresponding spherical ring point cloud set. The inner diameter of the spherical ring point cloud set is the first preset search radius, and the outer diameter of the spherical ring point cloud set is the second preset search radius, which is greater than the first preset search radius.

[0056] S50, if the viewpoint vector With shape vector If the included angle between them is greater than 90 degrees, then the feature point cloud is identified. For non-weld seam protrusions; if the viewpoint vector With shape vector If the included angle between them is no greater than 90 degrees, then the feature point cloud is identified. This is a weld seam depression.

[0057] The weld seam recognition method based on single-view point cloud feature extraction provided by this invention first obtains a three-dimensional cleaning point cloud dataset of the target workpiece, which is formed by combining multiple single point clouds; then, it obtains the single point clouds. The corresponding spherical neighborhood point cloud set Median principal component analysis was used to analyze the point cloud set in the neighborhood of the sphere. Analyze the point cloud in the data to obtain each individual point cloud. Corresponding single-point cloud geometric features Next, based on the feature vector set, binary classification and welding labeling are performed on the single point cloud in the workpiece 3D cleaning point cloud dataset to obtain the feature point cloud. The binary classification involves classifying the point cloud into either a first label or a second label. Then, based on the single point cloud... The corresponding spherical neighborhood point cloud set Calculate and obtain the centroid of the viewpoint Local weighted centroid and shape-weighted centroid Obtain the coordinates from the centroid of the viewpoint to the viewpoint. view vector Obtain the locally weighted centroid Pointing to shape-weighted centroid Shape Vector Finally, based on viewpoint vectors With shape vector The included angle between the points is used to identify non-weld seam protrusions or weld seam depressions in each individual point cloud of the workpiece's 3D cleaning point cloud dataset. The method of this invention, in acquiring the individual point clouds of the target workpiece... Then, it can directly perform further identification on a single point cloud to determine the single point cloud. Whether it is a real weld point, where a real weld point is a weld concave point, does not require a single point cloud in the workpiece 3D cleaning point cloud dataset. The actual weld points can be identified by performing initial weld fitting. The same approach is used for both straight and curved welds, making it highly adaptable to different workpiece types and possessing versatility and practicality. Furthermore, it is based on viewpoint vectors. With shape vector Angle identifier feature point cloud between Whether it is a weld concave point or not, it solves the technical problem in the prior art that the lack of differentiation of the actual point cloud of the weld after identifying the potential weld area leads to low weld recognition accuracy.

[0058] Understandably, in the solution of the present invention, the three-dimensional cleaning point cloud dataset of the target workpiece to be welded at the current viewpoint is obtained. ,in, , For 3D cleaning point cloud dataset The i-th single point cloud in the diagram represents the target workpiece to be welded. This workpiece can be a straight weld seam workpiece, a curved weld seam workpiece, or a combination of straight and curved weld seams. In specific embodiments of this invention, circular tube workpieces and rectangular workpieces are used for detailed explanation. It is understood that in the scheme of this invention, a clustering algorithm is used to analyze the feature vector set. The feature vectors in the cloud are used for binary classification and clustering to classify each single point cloud. The welding label is marked as either the first label or the second label, where, , For the i-th single-point cloud geometric feature, the first label can be a potential weld label or a non-weld label, and the first label and the second label are different category labels.

[0059] Optionally, the second preset search radius is larger than the first search radius, in the sphere's neighborhood point cloud set. A spherical ring point cloud set is formed on the periphery. The second search radius can be a multiple of 1.2 times, 1.5 times, 2.2 times, 2.5 times, etc. of the first search radius. No specific limitation is made in the scheme of this invention.

[0060] Furthermore, it also includes the step: S60, which involves processing the identified feature point cloud. Update to target point cloud The feature fusion method is used to integrate the target point cloud with initial concavity / convexity attribute labels. Perform semi-supervised optimization and update the current target point cloud. Weld seam identification. Understandably, in a preferred embodiment of the present invention, a semi-supervised optimization is performed using a feature fusion method based on an improved DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm.

[0061] Furthermore, S61, the labeled feature point cloud Update to target point cloud S62, for all target point clouds Construct a KD-tree data structure; S63, obtain the point cloud of each target based on a preset search distance. The corresponding search neighborhood point cloud set; S64, obtain the current target point cloud in each search neighborhood point cloud set. With each neighboring point cloud Distance weight S65, based on the target point cloud in the search neighborhood point cloud set. The labels are used to calculate the total weight of convex points within each search neighborhood point cloud set. The total weight of the concave points S66, If the total weight of the convex points is greater than the total weight of the concave points, then the current target point cloud is corrected to convex points; if the total weight of the convex points is less than the total weight of the concave points, then the current target point cloud is corrected to concave points; if the total weight of the convex points is equal to the total weight of the concave points, then the current target point cloud is not corrected; S67, After correction, the point cloud category weights in each search neighborhood are re-acquired. The point cloud category weights include the total weights of the convex points and the total weights of the concave points. If the point cloud category weights in the search neighborhood are not less than the first threshold, then the target point cloud corresponding to the search neighborhood is determined. A point is considered a high-confidence point; if the category weight of the point cloud within the search neighborhood is less than the first threshold, then the target point cloud corresponding to the search neighborhood is determined. For low-confidence points; S681, for low-confidence points, expand and update the preset search distance based on the preset step size, and proceed to step S63; for high-confidence points, determine the weld identifier of the target point cloud based on the high-confidence points; repeat steps S63 to S681 until the preset termination condition is reached, perform attribute assimilation on the points with ambiguous attributes, and update the target point cloud. Weld markings.

[0062] Further, the preset termination condition is reaching a preset number of iterations or when the proportion of attribute-corrected points in a single iteration to the total target points is less than a preset percentage. Optionally, the preset percentage can be any value from 0.5% to 10%. In another preferred embodiment of the present invention, in order to reduce the number of iterations, the preset percentage is preferably 1%.

[0063] Furthermore, using the formula The distance weights are calculated, where, Distance weights For the target point cloud, This refers to the neighboring point cloud (neighborhood point cloud). For target point cloud With one of the adjacent point clouds The Euclidean distance between them It is the adjustment coefficient. Using formula The total weight of convex points within the search neighborhood is statistically analyzed. , Point cloud representing convex points The distance weight is calculated using the formula. The total weight of concave points within the search neighborhood is statistically analyzed. , Target point cloud representing concave points Distance weights.

[0064] Furthermore, median principal component analysis was used to analyze the point cloud set in the neighborhood of the sphere. The point cloud was analyzed to obtain three eigenvalues. , , ,in, ; Use formula Perform calculations to obtain each single point cloud. Single-point cloud geometric features .

[0065] Furthermore, using the formula Calculate and obtain the centroid of the computational viewpoint ,in, These are weighting coefficients. , and These are the current feature point clouds. of The eigenvalue of the j-th point in the neighborhood. , For the current feature point cloud of The total number of points in the neighborhood, For the current feature point cloud of The set of points in the neighborhood, For the current feature point cloud The coordinates of the j-th point within the spherical neighborhood;

[0066] Using formula Calculate and obtain the local weighted centroid ,in, These are the centroid weighting coefficients. , For the current feature point cloud and k-th point in the neighborhood Euclidean distance, For the current feature point cloud The coordinates of the k-th point within the spherical neighborhood;

[0067] Calculate using formula Calculate and obtain the shape-weighted centroid ,in, These are shape weighting coefficients. , For the current feature point cloud and The nth point in the neighborhood Euclidean distance, , For the current feature point cloud of The total number of points in the neighborhood; For the current feature point cloud of The coordinates of the nth point in the neighborhood, For the current feature point cloud of The set of points in the neighborhood, For the current feature point cloud of The coordinates of the nth point in the neighborhood,

[0068] The first preset search radius, The second preset search radius is defined as follows. It can be understood that, in one specific embodiment of the present invention, the second preset search radius is twice the first preset search radius.

[0069] Furthermore, using the formula For feature point clouds Perform initial weld marking.

[0070] In specific implementation, S10 includes: S11, acquiring the workpiece 3D point cloud dataset from the current viewpoint. , , For the i-th 3D point cloud, record the viewpoint coordinate information corresponding to the current viewpoint; S12, perform voxel downsampling on the point cloud data in the workpiece 3D point cloud dataset to obtain the workpiece 3D cleaning point cloud dataset under the current viewpoint. The workpiece 3D cleaning point cloud dataset is formed by combining multiple single point clouds. .

[0071] Furthermore, in order to accurately identify the weld line to be welded, the process also includes the step of fitting all the weld pits to obtain the weld identification line.

[0072] This invention provides an optional implementation method for viewpoint-based dual-neighborhood weld seam recognition, including feature extraction and binary classification of a single-view workpiece point cloud; geometric attribute recognition of the single-view workpiece point cloud to determine non-weld seam protrusions and weld seam concave points; and feature fusion based on an improved DBSCAN algorithm to gradually assimilate the feature point cloud. The surrounding attributes are ambiguous.

[0073] Firstly, the feature extraction and binary classification of single-view workpiece point clouds specifically include:

[0074] Workpiece point cloud data acquisition utilizes an coded structured light camera as the 3D data acquisition device for the target workpiece. This is achieved by scanning the target workpiece at a preset viewpoint to obtain a 3D point cloud dataset of the workpiece at the current viewpoint. Simultaneously record the viewpoint coordinates corresponding to the current viewpoint. ;

[0075] Workpiece point cloud preprocessing and feature calculation: First, the original workpiece 3D point cloud dataset is processed. The point cloud data in the image is downsampled using voxel mesh to obtain the downsampled 3D cleaning point cloud dataset of the workpiece. Based on the downsampled workpiece 3D cleaning point cloud dataset Performing analysis and processing helps reduce data redundancy and improve subsequent processing efficiency; then, for the workpiece 3D cleaning point cloud dataset... Each single point cloud in Set the first preset search radius Furthermore, by constructing a KD-Tree data structure to accelerate the neighborhood point search process, it is possible to achieve single-point cloud... Centered on, with radius Points within the spherical neighborhood are extracted to obtain a set of all spherical neighborhood point clouds. Next, median principal component analysis (median PCA) was used to analyze the point cloud set in the neighborhood of each sphere. The analysis yielded three eigenvalues. , , ,in, Finally, based on eigenvalues , , Using formula Calculate a single point cloud , where, For a single point cloud after downsampling Geometric features;

[0076] Feature region segmentation based on k-means clustering, single point cloud All the features Combine to form feature vectors The k-means clustering algorithm is used to cluster the feature vectors. Perform binary clustering to divide the point cloud region into a first label or a second label.

[0077] Secondly, the geometric attribute identification of non-weld seam protrusions and weld seam concave points from single-view workpiece point clouds specifically includes:

[0078] For each feature point cloud selected by k-means clustering With it as the center, and a radius of Within a spherical neighborhood, select all feature point clouds and calculate the view centroid. For each feature point cloud In a radius of Within a spherical neighborhood, select all point cloud data and calculate the local weighted centroid. Perform bilateral neighborhood calculations for feature point clouds. With the center and radius in to Calculate the shape-weighted centroid of the point cloud in the neighborhood of the spherical ring within the range. ;

[0079] The feature point cloud selected by the k-means clustering algorithm is updated to the target point cloud, and its concavity / convexity properties are expressed by the formula. Make a judgment.

[0080] Thirdly, after identifying the concavity and convexity attributes of the workpiece point cloud from a single perspective, due to factors such as noise interference, local misjudgments may occur in some areas. The traditional DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm is a classic density-based clustering method. Its core principle is to form clusters by identifying densely connected points in space and it can effectively handle noisy data. However, the traditional DBSCAN algorithm is mainly suitable for unsupervised clustering scenarios and cannot directly handle the semi-supervised optimization problem with "initial concavity and convexity attribute labels" in this invention. To solve this problem, this invention proposes a feature fusion method based on an improved DBSCAN algorithm to gradually assimilate the feature point cloud. The surrounding attribute fuzzy points specifically include:

[0081] Construct a KD-Tree data structure for the target point cloud that has completed geometric concavity and convexity attribute identification;

[0082] For each target point The algorithm searches for a set of neighboring points within a radius epsilon (a preset search distance) to obtain valid convex or concave points within the neighborhood. For each point in the neighborhood... Calculate distance weights ,in, For point and The Euclidean distance between them It is an adjustment coefficient used to control the degree to which distance affects the weight;

[0083] The total weight of convex points within the statistical neighborhood ( The total weight of convex and concave points (m is a concave point). If , correct For convex points; if , correct It is a concave point.

[0084] Iterative optimization: After the initial correction, points where the majority attribute accounts for more than 70% of the total weight of all neighborhoods are recorded as "high-confidence points"; points where the majority attribute accounts for between 50% and 70% of the total weight of neighborhoods are called "low-confidence points". For low-confidence points, the search radius is increased by 2 times to search for high-confidence points in the neighborhood, and a second correction is performed based on the majority attribute of the high-confidence points; this process is repeated 1-2 times to allow the majority attribute to gradually assimilate the surrounding ambiguous attribute points.

[0085] Early termination condition: The iteration terminates when the proportion of attribute correction points to the total target points is less than 1% in a single iteration.

[0086] The beneficial effects of the weld seam recognition method based on single-view point cloud feature extraction provided by this invention include: the single-view workpiece point cloud concavity / convexity attribute recognition method adopts the dual-neighborhood discrimination idea, has a fast calculation speed, and can quickly obtain the geometric attributes of feature point clouds, making it suitable for industrial applications of real-time welding; the feature fusion method based on the improved DBSCAN algorithm can effectively reduce misjudgments caused by point cloud noise, making the welding robot's point cloud recognition accuracy higher, its anti-interference ability stronger, and its robustness higher; it does not require a CAD model of the welding workpiece, is suitable for extracting straight and curved weld seams, and has good versatility. This method has a fast calculation speed, strong robustness to point cloud noise, and can adapt to different workpiece types, achieving the goal of one algorithm adapting to multiple workpieces, thus meeting the intelligent operation requirements of welding robots.

[0087] This invention also provides a weld seam recognition system based on single-view point cloud feature extraction.

[0088] It includes an coded structured light sensor and a processing device. The coded structured light sensor is used to acquire a three-dimensional cleaning point cloud dataset of the workpiece, and the processing device is used to implement the steps of the weld recognition method based on single-view point cloud feature extraction described above.

[0089] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described weld seam identification method based on single-view point cloud feature extraction.

[0090] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A weld seam identification method based on single-view point cloud feature extraction, characterized in that, Includes the following steps: S10, Obtain the workpiece 3D cleaning point cloud dataset under the current viewpoint. The workpiece 3D cleaning point cloud dataset is formed by combining multiple single point clouds. The current viewpoint has corresponding viewpoint coordinates. S20, based on the first preset search radius, obtain the spherical neighborhood point cloud set corresponding to each single point cloud, and use the median principal component analysis method to analyze the point cloud in the spherical neighborhood point cloud set to obtain the single point cloud geometric features corresponding to each single point cloud; S30, combine all the geometric features of the single point cloud to form a feature vector set, use a clustering algorithm to perform binary classification clustering on the feature vector set, and label each single point cloud with a welding label to obtain a feature point cloud, wherein the welding label is a first label or a second label; S40, obtain the viewpoint vector of the feature point cloud, which is composed of the viewpoint centroid and the viewpoint coordinates, pointing to the viewpoint; obtain the shape vector of the feature point cloud, which is composed of the local weighted centroid and the shape weighted centroid, pointing to the shape weighted centroid. Specifically, the viewpoint centroid and the local weighted centroid are calculated and obtained based on the spherical neighborhood point cloud set corresponding to the feature point cloud, and the shape weighted centroid is calculated and obtained based on the spherical ring point cloud set corresponding to the feature point cloud. The inner diameter of the spherical ring point cloud set is a first preset search radius, and the outer diameter of the spherical ring point cloud set is a second preset search radius. The second preset search radius is greater than the first preset search radius. S50, if the angle between the viewpoint vector and the shape vector is greater than 90 degrees, the feature point cloud is identified as a non-weld seam protrusion; if the angle between the viewpoint vector and the shape vector is not greater than 90 degrees, the feature point cloud is identified as a weld seam concave point.

2. The weld seam identification method based on single-view point cloud feature extraction according to claim 1, characterized in that, It also includes the following steps: S60, update the identified feature point cloud to the target point cloud; use a feature fusion method to perform semi-supervised optimization on the target point cloud with initial concavity / convexity attribute labels, and update the weld identification of the current target point cloud.

3. The weld seam identification method based on single-view point cloud feature extraction according to claim 2, characterized in that, The S60 specifically includes: S61, update the identified feature point cloud to the target point cloud; S62, construct a KD tree data structure for all the target point clouds; S63, obtain the search neighborhood point cloud set corresponding to each target point cloud based on the preset search distance; S64, obtain the distance weights between the target point cloud and each neighboring point cloud in the search neighborhood point cloud set; S65, based on the label identifier of the target point cloud in the search neighborhood point cloud set, calculate and obtain the total weight of the convex points and the total weight of the concave points in each search neighborhood point cloud set; S66, if the total weight of the convex points is greater than the total weight of the concave points, then the current target point cloud is corrected to convex points; if the total weight of the convex points is less than the total weight of the concave points, then the current target point cloud is corrected to concave points; if the total weight of the convex points is equal to the total weight of the concave points, then the current target point cloud is not corrected. S67, After correction, the point cloud category weights in each search neighborhood are re-acquired. The point cloud category weights include the total weight of convex points and the total weight of concave points. If the point cloud category weights in the search neighborhood are not less than a first threshold, the target point cloud corresponding to the search neighborhood is determined to be a high-confidence point; if the point cloud category weights in the search neighborhood are less than the first threshold, the target point cloud corresponding to the search neighborhood is determined to be a low-confidence point. S68, For the low confidence point, the preset search distance is increased and updated based on the preset step size, and then proceed to step S63; Repeat steps S63 to S68 until the preset termination condition is met, gradually assimilate the attributes of the ambiguous points, and update the weld identification of the target point cloud.

4. The weld seam identification method based on single-view point cloud feature extraction according to claim 3, characterized in that, The preset termination condition is reaching a preset number of iterations or when the proportion of attribute correction points in a single iteration to the total target points is less than a preset percentage.

5. The weld seam identification method based on single-view point cloud feature extraction according to claim 3, characterized in that, Using formula The distance weights are calculated, where, Distance weights For the target point cloud, For neighboring point clouds, For target point cloud With one of the said neighborhood point clouds The Euclidean distance between them It is the adjustment coefficient. , Using formula Calculate the total weight of convex points within the search neighborhood. , The target point cloud representing the convex points The distance weight is calculated using the formula. Calculate the total weight of concave points within the search neighborhood. , The target point cloud representing the concave point Distance weights.

6. The weld seam identification method based on single-view point cloud feature extraction according to any one of claims 1 to 3, characterized in that, Median principal component analysis was used to analyze the point cloud in the neighborhood of the sphere to obtain eigenvalues. , , ,in, ; Use formula Calculations are performed to obtain the geometric features of each individual point cloud.

7. The weld seam identification method based on single-view point cloud feature extraction according to any one of claims 1 to 3, characterized in that, Using formula Calculate and obtain the centroid of the computational viewpoint ,in, These are weighting coefficients. , and These are the current feature point clouds. of The eigenvalue of the j-th point in the neighborhood. , For the current feature point cloud of The total number of points in the neighborhood, For the current feature point cloud of The set of points in the neighborhood, For the current feature point cloud The coordinates of the j-th point within the spherical neighborhood; Using formula Calculate and obtain the local weighted centroid ,in, These are the centroid weighting coefficients. , For the current feature point cloud and k-th point in the neighborhood Euclidean distance, For the current feature point cloud The coordinates of the k-th point within the spherical neighborhood; Calculate using formula Calculate and obtain the shape-weighted centroid ,in, These are shape weighting coefficients. , For the current feature point cloud and The nth point in the neighborhood Euclidean distance, , For the current feature point cloud of The total number of points in the neighborhood; For the current feature point cloud of The coordinates of the nth point in the neighborhood, For the current feature point cloud of The set of points in the neighborhood, For the current feature point cloud of The coordinates of the nth point in the neighborhood, The first preset search radius, The second preset search radius.

8. The weld seam identification method based on single-view point cloud feature extraction according to any one of claims 1 to 3, characterized in that, The specific steps for obtaining the 3D cleaning point cloud dataset of the workpiece from the current viewpoint include: Obtain the workpiece 3D point cloud dataset from the current viewpoint; perform voxel downsampling on the point cloud data in the workpiece 3D point cloud dataset to obtain the workpiece 3D cleaning point cloud dataset from the current viewpoint, wherein the workpiece 3D cleaning point cloud dataset is formed by combining multiple single point clouds.

9. The weld seam identification method based on single-view point cloud feature extraction according to claim 2, characterized in that, It also includes the step of fitting all the weld seam depressions to obtain weld seam identification lines.

10. A weld seam recognition system based on single-view point cloud feature extraction, characterized in that, It includes an coded structured light sensor and a processing device, wherein the coded structured light sensor is used to acquire a three-dimensional cleaning point cloud dataset of the workpiece. The processing device is used to implement the steps of the weld identification method based on single-view point cloud feature extraction as described in any one of claims 1 to 9.

Images