Weld point cloud region segmentation method and system based on multi-feature dynamic threshold analysis

Abstract

The application belongs to the technical field of three-dimensional measurement, and particularly discloses a welding seam point cloud region segmentation method and system based on multi-feature dynamic threshold analysis, which comprises the following steps: obtaining three-dimensional point cloud data containing a welding seam region, selecting an arbitrary point in the welding seam as an initial seed point for region growth; calculating a seed point normal vector divergence feature and a tangent plane deviation feature, and performing three-dimensional point cloud feature fusion; modeling prior feature information as a Gaussian process to realize dynamic threshold updating of the region growth process; adopting the dynamic threshold updating method to determine whether the seed point belongs to the welding seam region; if yes, the K-neighborhood of the welding seam feature point is included in the seed set as a seed point set for subsequent determination of whether it belongs to the welding seam region; if not, it is determined whether the seed point set is empty; if not, the seed point is updated; and if empty, it is determined that complete welding seam region segmentation has been completed. The application can effectively extract the welding seam point cloud of a partial weak curvature feature region of a welding seam surface.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of three-dimensional measurement, and more particularly relates to a welding seam point cloud region segmentation method and system based on multi-feature dynamic threshold analysis. BACKGROUND

[0002] Robot welding technology has almost achieved full automation, and real-time evaluation of welding results has become one of the core issues in the welding field. Weld extraction is an important step in welding result evaluation and is also a prerequisite for subsequent surface quality processing. The principle of the weld extraction process is to extract the weld surface information and contour information to achieve semantic segmentation of the weld and different regions of the workpiece. However, for welds with weak geometric features, or when there is no obvious distinction between the weld and the workpiece surface features, a reasonable evaluation method and a more robust algorithm are needed to achieve the extraction process of the weld under different working conditions.

[0003] Weld extraction algorithms mainly include two-dimensional and three-dimensional weld extraction. Among them, two-dimensional extraction relies on visual images, and common methods such as structured light centerline algorithm and laser stripe gray distribution extraction algorithm have real-time performance, but have some shortcomings such as complex extraction process, low precision, and great limitations. Three-dimensional extraction is often processed on point clouds, and common methods such as region clustering and feature extraction algorithm, and region growing algorithm based on point cloud similarity, etc. The clustering and extraction idea has a great error in the extraction of target features in reality, and the feature extraction process of the region growing algorithm highly depends on experience and is not universal for different types of welds. With the development of machine learning, some people have begun to try to use machine learning to extract welds, using learned models to extract features of new welds, but such methods require a large amount of training data and high computing resources.

[0004] Weld region features change significantly, and geometric features at different positions have certain differences, which can cause the following problems: single feature description for the weld region is difficult to completely cover the weld surface, the selection of feature threshold highly depends on experience, and noise information on the workpiece base material surface can have a great impact on feature region segmentation. There is no suitable solution to the above problems of weld extraction. SUMMARY

[0005] In view of the above defects or improvement needs of the prior art, the present application provides a welding seam point cloud region segmentation method and system based on multi-feature dynamic threshold analysis, which significantly improves the accuracy and robustness of three-dimensional point cloud welding seam region segmentation through multi-feature fusion and dynamic threshold analysis. This method uses the combination of point cloud normal vector divergence features and tangent plane deviation features to effectively identify the welding seam region, including those weak curvature feature regions, while avoiding false extraction caused by workpiece base material surface noise. In addition, by modeling the welding seam region features as a Gaussian process and dynamically updating the threshold, this method can adapt to different working conditions, reducing the dependence on human experience, thereby improving the application range and flexibility of the welding seam extraction algorithm.

[0006] To achieve the above object, according to one aspect of the present application, a welding seam point cloud region segmentation method based on multi-feature dynamic threshold analysis is proposed, comprising the following steps:

[0007] Step one, obtain three-dimensional point cloud data containing the welding seam region, and select an arbitrary point in the welding seam as the initial seed point for region growing;

[0008] Step two, calculate the normal vector divergence feature and tangent plane deviation feature of the seed point, and perform three-dimensional point cloud feature fusion based on the same to form a point cloud fusion feature;

[0009] Step three, according to the workpiece point cloud fusion feature, use a dynamic threshold updating method to determine whether the seed point belongs to the welding seam region, if yes, include the K-neighborhood of the welding seam feature point into the seed set as the seed point set for subsequent judgment of whether it belongs to the welding seam region, if not, judge whether the seed point set is empty, if not, update the seed point and return to step two, if empty, judge that the complete welding seam region segmentation has been completed, and enter step four;

[0010] Step four, extract the welding seam point cloud and contour information after segmentation to complete the welding seam region segmentation of three-dimensional point cloud.

[0011] As a further optimization, in step one, a line laser scanner is used to scan the welding seam region to obtain a narrow point cloud containing the welding seam region, and the narrow point cloud is taken as the three-dimensional point cloud data.

[0012] As a further optimization, in step two, the calculation of the normal vector divergence feature of the seed point includes:

[0013] (211) Calculate the K-neighborhood of each point in the three-dimensional point cloud data and calculate the L1 median point;

[0014] (212) Based on the L1 median point, estimate the normal vector of each point in the three-dimensional point cloud data using the PCA method;

[0015] (213) calculating a local normal vector divergence feature factor of the feature point based on the normal vector.

[0016] As a further preferred, in step (211), set the three-dimensional point cloud data point set P k ={p1, p2, …, p j , …, p k}, where p k ∈R 3 is the k different points near p j , β i is the weight of p j , for the point q ∈ R j in the neighborhood of p 3 , C(q) is the weighted sum of the distance of q and p1, p2, …, p k , the equation of C(q) includes:

[0017]

[0018] where d i (q) = ||q-p i || represents the Euclidean distance of q and p i ,

[0019] Find a point q ∈ R 3 to minimize the equation of C(q):

[0020] M = M(p1, …, p k ; β1, … β k ) = argmin {C(q); q ∈ R 3}

[0021] Where the solution of M is the L1 median point in the k neighborhood;

[0022] Preferably, step (212) includes:

[0023] Construct the covariance matrix of C 3×3 :

[0024]

[0025] Where K is the number of neighborhood points;

[0026] Calculate the eigenvalue of the covariance matrix, and the eigenvector corresponding to the minimum eigenvalue is the normal vector;

[0027] Preferably, step (213) includes:

[0028] Define the local normal vector mean:

[0029]

[0030] According to the local normal vector mean, a local divergence feature factor calculation model is constructed:

[0031]

[0032] Wherein, n ij is the unit normal vector of each point in the K neighborhood, K is the number of neighborhood points, α i is the local divergence feature factor.

[0033] As a further preferred, in step two, the calculation of the seed point tangent plane deviation feature includes:

[0034] (221) Slice the feature points along the scanning line direction, and calculate the projection of each point on the tangent plane to form a two-dimensional coordinate;

[0035] (222) Curve fitting is carried out on the two-dimensional coordinates by using a sliding average filter;

[0036] (223) Construct a deviation feature factor to calculate the tangent plane deviation feature, and the deviation feature factor includes:

[0037]

[0038] In the formula, d i is the deviation of the original point of the two-dimensional coordinate and its corresponding point on the parent curve.

[0039] As a further preferred, step (221) includes:

[0040] Let the feature point p A = [x A , y A , z A ] T , the distance between point p j and the tangent plane C is:

[0041] d j = |[p A -p j ] T ·n C |

[0042] Wherein, n C is the normal vector of the tangent plane, when the distance d j is less than the threshold d thre , it is considered that the point p j is located on the tangent plane;

[0043] A tangent plane coordinate system is constructed by using a tangent plane normal vector, a scanning line direction vector and a direction vector perpendicular to the two vectors, a homogeneous transformation matrix T between the tangent plane coordinate system and the space coordinate system is calculated, and three-dimensional point cloud coordinates are converted into two-dimensional coordinates in the tangent plane coordinate system.

[0044] As a further preferred, in step (222), the formula of the moving average filter includes:

[0045]

[0046] Wherein, N represents the length of the filter, p M-i represents the value within the filter range;

[0047] The deviation d of the original point of the two-dimensional coordinate and its corresponding point on the parent curve is calculated by the following formula i :

[0048]

[0049] Wherein, [x pi ,y pi ] represents the corresponding coordinate point after smoothing, [x i ,y i ] represents the initial coordinate before smoothing.

[0050] As a further preferred, in step two, the three-dimensional point cloud feature fusion includes: combining the normal vector divergence feature and the tangent plane deviation feature information to form a new weld feature factor:

[0051] θ i =ω1α i +ω2β i

[0052] Wherein, ω1 and ω2 are weight values.

[0053] As a further preferred, step three includes the following steps:

[0054] (311) Collect and prepare data samples, combine part of the acquired prior weld feature information to form a weld region feature set.

[0055] (312) Train the Gaussian process model, input the feature information into the Gaussian process model, set a suitable kernel function k(x, x') to model the correlation of the input data. By training the Gaussian process model, the mean function μ(x) and the covariance function σ(x) of the output are obtained;

[0056] (313) Predict the probability distribution of the new point, for a new feature point x new , use the Gaussian process model to predict its mean μ(x new) and variance σ(x new ), P(X>Threshold) is calculated by using the cumulative probability function F(z) of the normal distribution;

[0057] (314) calculating the threshold value of the cumulative probability reaching 99%;

[0058] (315) judging whether the seed point belongs to the weld feature point according to the threshold value, if yes, the point is included in the prior data sample, which is used for subsequent threshold value updating, if not, the threshold value is unchanged, judging whether the seed point set is empty, if not, the seed point is updated, and returning to step two, if empty, judging that the complete weld region segmentation is completed, and entering step four.

[0059] According to another aspect of the present application, a weld point cloud region segmentation system based on multi-feature dynamic threshold analysis is also provided, comprising:

[0060] A first main control module is used for acquiring three-dimensional point cloud data containing a weld region, and selecting an arbitrary point in the weld as an initial seed point for region growing;

[0061] A second main control module is used for calculating the divergence feature of the normal vector of the seed point and the deviation feature of the tangent plane, and performing three-dimensional point cloud feature fusion according to the features to form a point cloud fusion feature;

[0062] A third main control module is used for judging whether the seed point belongs to the weld region by using a dynamic threshold updating method according to the workpiece point cloud fusion feature, if yes, the K-neighborhood of the weld feature point is included in the seed set as a seed point set for subsequent judgment of whether it belongs to the weld region, if not, judging whether the seed point set is empty, if not, updating the seed point, and inputting the updated seed point into the second main control module, if empty, judging that the complete weld region segmentation is completed;

[0063] A fourth main control module is used for extracting the weld point cloud and contour information after the segmentation is completed after the third main control module judges that the complete weld region segmentation is completed, and completing the three-dimensional point cloud weld region segmentation.

[0064] Overall, compared with the prior art, the above technical solutions conceived by the present application mainly have the following technical advantages:

[0065] 1.The application constructs a multi-feature dynamic threshold analysis algorithm framework, extracts multiple features from the pre-processed complete point cloud data, outputs the feature values of the feature point cloud to be judged, and then inputs the obtained prior feature information into the dynamic threshold updating algorithm framework to accurately determine the region to which the feature points belong, thereby completing the weld point cloud region extraction work.

[0066] 2.The application combines the point cloud normal vector divergence feature and the tangent plane deviation feature to effectively extract the weld point cloud of the weak curvature feature region of the weld surface, and improves the situation that the weld point cloud may be missed in the single feature extraction process.

[0067] 3.The application incorporates the tangent plane deviation feature into the feature item, which can effectively avoid the situation that the local high curvature point cloud of the workpiece base material surface part region is incorrectly extracted due to noise point cloud, thereby improving the accuracy of the weld point cloud extraction.

[0068] 4.The application models the weld region feature as a Gaussian process to realize dynamic threshold updating of the region growing process, avoids the inapplicability of manual threshold setting to different working conditions, and improves the application range of the weld extraction algorithm, so that it can adapt to different welding conditions and conditions.

[0069] 5.The application uses the point with the minimum weighted distance as the neighborhood center point to estimate the point cloud normal vector, uses the L1 median as the data center to highlight the differences between the feature points in the adjacent points, and is more resistant to the influence of the feature points in the neighborhood than the average value, and is closer to the center of the data body. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 is a flowchart of a weld point cloud region segmentation method based on multi-feature dynamic threshold analysis according to an embodiment of the application;

[0071] Figure 2 is a point cloud feature value calculation flowchart according to an embodiment of the application;

[0072] Figure 3 is a cloud feature judgment threshold acquisition flowchart according to an embodiment of the application;

[0073] Figure 4 is a weld point cloud extraction effect diagram according to an embodiment of the application. DETAILED DESCRIPTION

[0074] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application. In addition, the technical features involved in the various embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0075] As shown in Figure 1 , the welding seam point cloud region segmentation method based on multi-feature dynamic threshold analysis provided by the embodiment of the present application, the welding seam region segmentation process of the three-dimensional point cloud includes:

[0076] (1) Since the welding seam volume will cause certain interference to the line laser, the excessive area of the welding seam and the workpiece base material surface is blocked, resulting in the loss of part of the point cloud details, so it is necessary to control the relative pose of the scanner and the welding workpiece to collect complete welding seam point cloud data. The welding seam has a certain length feature in its main extension direction, and the line laser scanner will generate a long point cloud containing the welding seam region after single scanning of the welding seam region.

[0077] (2) After collection is completed, the normal vector of the acquired point cloud needs to be estimated in order to facilitate the subsequent feature extraction process. Due to the exposure problem in the process of line structured light scanning the metal surface and the precision of the camera device, the collected point cloud data will introduce noise and outliers. Therefore, the collected point cloud needs to be preprocessed, such as point cloud filtering, uniform downsampling, and the like, and the unnecessary profile point cloud data in the thickness direction is removed and cleaned.

[0078] Specifically, in the present application, the feature extraction module includes point cloud normal vector divergence feature and tangent plane deviation feature, and a multi-feature fusion method is used to form a seed point feature value, which is used to judge whether the point cloud belongs to the welding seam region in the subsequent process. The complete feature extraction process is shown in Figure 2 .

[0079] (3) The divergence feature adopted in the present application estimates the normal vector of the feature point and the points in its K-neighborhood, and constructs the divergence feature value according to the dispersion degree of the normal vector.

[0080] ① Step one, calculate the median of the point cloud and its K-neighborhood L1. Let the point set P k ={p1,p2,…,p j ,…,p k} where p k ∈R 3 is the k different points near p j , and β i is the weight of p j . For p jq∈R 3 , C(q) is the weighted sum of distances of q to p1, p2, …, p k . The equation of C(q) is shown as follows:

[0081]

[0082] where d i (q) = ||q-p i || represents the Euclidean distance of q and p i .

[0083] The L1 median in k-neighborhood is the centroid of the region formed by the point set p i . Find a point q∈R 3 to minimize equation (1). This process can be represented as follows:

[0084] M = M(p1, …, p k ; β1, … β k ) = arg min{C(q); q∈R 3} (2)

[0085] The solution of M is the L1 median in k-neighborhood.

[0086] ②Step two, PCA method estimates point cloud normal vector. The L1 median in equation (2) is used to optimize the principal component analysis of the point cloud. By subtracting the L1 median from the neighboring points of p j , the covariance matrix of C 3×3 is as follows:

[0087]

[0088] The eigenvalue λ1 < λ2 < λ3 of the covariance matrix, and the eigenvector corresponding to the minimum eigenvalue is the normal vector.

[0089] ③Step three, calculate the divergence feature of the normal vector. Define the local normal vector mean as follows:

[0090]

[0091] Then the local divergence feature factor can be defined by the local normal vector mean as follows:

[0092]

[0093] where n ij is the unit normal vector of each point in the K-neighborhood, K is the number of neighborhood points, and α i is the local divergence feature factor.

[0094] (4) The deviation feature adopted in the patent is obtained by point cloud slicing processing on the feature points, extracting the projection of the adjacent point cloud on the slicing plane, smoothing the convex profile of the local weld area by using the moving average filter method, and constructing the deviation feature value according to the deviation degree between the original projection point and the point after smoothing processing.

[0095] ① Step one, slicing the feature points along the scanning line direction, and calculating the projection of each point on the slicing plane to form a two-dimensional coordinate. Let the feature point p A =[x A ,y A ,z A ] T , the distance between the point p j and the slicing plane C is:

[0096] d j =|[p A -p j ] T ·n C | (6)

[0097] Wherein, n C is the normal vector of the slicing plane. When the distance d j is less than the threshold value d thre , it can be considered that the point p j is located on the slicing plane.

[0098] Then, the slicing plane coordinate system is constructed with the slicing plane normal vector, the scanning line direction vector and the direction vector perpendicular to the two, the homogeneous transformation matrix T of the slicing plane coordinate system and the space coordinate system is calculated, and the three-dimensional point cloud coordinates are converted into two-dimensional coordinates in the slicing plane coordinate system.

[0099] ② Step two, curve fitting is carried out by using a moving average filter. The sampling data set is adjusted by creating a series of average values of different subsets. With the increase of the length of the filter, the smoothness of the output increases, and the sharp disturbance in the data can be obviously eliminated. The formula of the sliding average filter can be expressed by the following formula.

[0100]

[0101] Wherein, N represents the length of the filter, and p M-i represents the value within the filter range.

[0102] The moving average filter is adopted to obtain the fitting point of the mother curve, and the deviation between the original point and the corresponding point on the mother curve is calculated by the following formula.

[0103]

[0104] Wherein, [x pi ,y pi[x i ,y i ] represents the initial coordinates before smoothing.

[0105] ③ Step three, calculate the tangent plane deviation feature. The patent adopts the following formula to construct the deviation feature factor:

[0106]

[0107] (5) The two feature information is combined by weighting to form a new weld feature factor for subsequent feature judgment.

[0108] θ i = ω1α i + ω2β i (10)

[0109] Where ω1 and ω2 are the weights of the two features.

[0110] The point cloud feature information obtained in steps (3), (4) and (5) will be used in the feature point threshold judgment process of the subsequent region growing process, which will be introduced in the following. Figure 3

[0111] (5) The prior acquired part of the weld feature information is used in the dynamic threshold updating process in the embodiment of the present application, and the interval with a cumulative probability of 99% is taken as the feature threshold for judging whether the subsequent feature point belongs to the weld point cloud. By continuously updating the prior information, the weld features belonging to the same distribution can be extracted, and the feature threshold is continuously updated to realize the adaptive updating of the threshold.

[0112] ① Step one, collect and prepare data samples. Part of the acquired prior weld feature information is combined to form a weld region feature set.

[0113] ② Step two, train the Gaussian process model. The feature information is input into the Gaussian process model, and a suitable kernel function k(x, x') is set to model the correlation of the input data. By training the Gaussian process model, the mean function μ(x) and the covariance function σ(x) are obtained.

[0114] ③ Step three, predict the probability distribution of the new point. For a new feature point x new , the mean μ(x new ) and variance σ(x new ) are predicted by using the Gaussian process model. Since the output of the Gaussian process obeys the normal distribution, the cumulative probability function F(z) of the normal distribution can be used to calculate P(X>Threshold).

[0115] ​Step four, calculate the threshold value of cumulative probability reaching 99%. For normal distribution N(μ,σ 2 ), the threshold value of cumulative probability of 0.99 can be expressed as:

[0116] θ threshold = μ(x new )-z 0.99 ×σ(x new ) (11)

[0117] Wherein, z 0.99 is the quantile under the standard normal distribution.

[0118] Step five, use the threshold value obtained in step four to determine whether the seed point belongs to the weld feature point. If yes, the point is included in the prior data sample for subsequent threshold updating. If not, the threshold value remains unchanged and is the threshold value obtained in step four.

[0119] (6) The weld feature judgment threshold value in the region growing process is dynamically updated by the above method. When the seed point set is empty in the region growing process, it is determined that the complete weld region segmentation is completed. The weld point cloud and contour information after segmentation are extracted, and the weld region segmentation of the three-dimensional point cloud is accurately completed, as shown in Figure 4 .

[0120] According to another aspect of the present application, a weld point cloud region segmentation system based on multi-feature dynamic threshold analysis is also provided, which is used to realize the method of any embodiment or combination of multiple embodiments, comprising:

[0121] A first master control module is used to obtain three-dimensional point cloud data containing a weld region, and select an arbitrary point in the weld as an initial seed point for region growing;

[0122] A second master control module is used to calculate the normal vector divergence feature and the tangent plane deviation feature of the seed point, and perform three-dimensional point cloud feature fusion according to the features to form a point cloud fusion feature;

[0123] A third master control module is used to determine whether the seed point belongs to the weld region by using a dynamic threshold updating method according to the workpiece point cloud fusion feature. If yes, the K-neighborhood of the weld feature point is included in the seed set as the seed point set for subsequent determination of whether it belongs to the weld region. If not, it is determined whether the seed point set is empty. If not empty, the seed point is updated and the updated seed point is input to the second master control module. If empty, it is determined that the complete weld region segmentation is completed;

[0124] A fourth master control module is used to extract the weld point cloud and contour information after segmentation when the third master control module determines that the complete weld region segmentation is completed, and complete the weld region segmentation of the three-dimensional point cloud.

[0125] Those skilled in the art can easily understand that the above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for weld point cloud region segmentation based on multi-feature dynamic threshold analysis, characterized in that, The method comprises the following steps: Step one, obtaining three-dimensional point cloud data containing a weld area, selecting an arbitrary point in the weld as an initial seed point for region growing; Step two, calculating the seed point normal vector divergence feature and the tangent plane deviation feature, and performing three-dimensional point cloud feature fusion based on the same to form a point cloud fusion feature; The calculation of the seed point normal vector divergence feature comprises: (211) calculating the K-neighborhood of each point in the three-dimensional point cloud data and calculating the L1 median point thereof; (212) estimating the normal vector of each point in the three-dimensional point cloud data based on the L1 median point by using a PCA method; (213) calculating a local normal vector divergence feature factor based on the normal vector; The calculation of the seed point tangent plane deviation feature comprises: (221) slicing the feature point along the scanning line direction, calculating the projection of each point on the tangent plane, and forming a two-dimensional coordinate; (222) performing curve fitting on the two-dimensional coordinate by using a sliding average filter; (223) constructing a deviation feature factor, calculating a tangent plane deviation feature, and the deviation feature factor comprises: wherein d i is the deviation of the original point of the two-dimensional coordinate from its corresponding point on the parent curve; The three-dimensional point cloud feature fusion comprises: combining the normal vector divergence feature and the tangent plane deviation feature information by weighting to form a new weld feature factor: θ i = ω1α i + ω2β i where ω1 and ω2 are weight values, and α i is a local normal vector divergence feature factor; Step three, judging whether the seed point belongs to the weld area by using a dynamic threshold updating method based on the workpiece point cloud fusion feature, if yes, including the K-neighborhood of the weld feature point in the seed set as a seed point set for subsequent judgment of whether it belongs to the weld area, if not, judging whether the seed point set is empty, if not, updating the seed point and returning to step two, if empty, judging that the complete weld area segmentation has been completed, and entering step four; Step four, extracting the weld point cloud and contour information after the segmentation is completed, and completing the three-dimensional point cloud weld area segmentation.

2. The method of claim 1, wherein, In step one, a line laser scanner is used to scan the weld area to obtain a long and narrow point cloud containing the weld area, and the long and narrow point cloud is taken as the three-dimensional point cloud data.

3. The method of claim 1, wherein, In step (211), let the three-dimensional point cloud data point set P k ={p1,p2,…,p j ,…,p k} where p k ∈R 3 is the k different points near p j , β i is the weight of p j , for the point q ∈ R j in the neighborhood of p 3 , C(q) is the weighted sum of the distance of q and p1, p2,…, p k , the equation of C(q) includes: where d i (q) = ||q - p i || denotes the Euclidean distance of q and p i , Find a point q e R 3 Equation to minimize C(q): M = M(p1,..., p k ) = argmin{C(q); q e R k ) = argmin{C(q); q e R 3} The solution of M is the L1 median point in the K-neighborhood.

4. The method of claim 3, wherein, Step (212) comprises: Configuration C 3×3 covariance matrix of: Wherein, K is the number of neighborhood points; The eigenvalue of the covariance matrix is calculated, and the eigenvector corresponding to the minimum eigenvalue is taken as the normal vector.

5. The method of claim 4, wherein, Step (213) comprises: The local normal vector mean value is defined as: A local normal vector divergence feature factor calculation model is constructed according to the local normal vector mean value: where n ij is the unit normal vector of each point in the K neighborhood, K is the number of points in the neighborhood, and a i is the local normal vector divergence feature factor.

6. The method of claim 5, wherein the method further comprises: Step (221) comprises: Let the feature point p A = [x A , y A , z A ] T The distance between the point p j and the tangent plane C is: d j =|[p A -p j ] T ·n C | Where, n C Let be the normal vector of the tangent plane, when the distance d j Less than threshold d thre At that time, point p is considered j Located on the tangent plane; A tangent plane coordinate system is constructed with the tangent plane normal vector, the scanning line direction vector and the direction vector perpendicular to the two vectors, a homogeneous transformation matrix T of the tangent plane coordinate system and the space coordinate system is calculated, and the three-dimensional point cloud coordinates are converted into two-dimensional coordinates in the tangent plane coordinate system.

7. The method of claim 6, wherein the method further comprises: In step (222), the formula of the sliding average filter comprises: where N represents the length of the filter, p M-i denotes the value within the filter range; The deviation d of the original point of the two-dimensional coordinate from its corresponding point on the parent curve is calculated by the following formula i : wherein [x pi ,y pi ] represents the corresponding coordinate point after smoothing, and [x i ,y i ] represents the initial coordinate before smoothing.

8. The method of claim 1-7, wherein, Step three comprises the following steps: (311) collecting and preparing data samples, combining part of the obtained prior weld feature information to form a weld area feature set; (312) training a Gaussian process model, inputting the feature information into the Gaussian process model, setting a suitable kernel function k(x, x'), modeling the correlation of the input data, and obtaining the mean function μ(x) and the covariance function σ(x) of the output by training the Gaussian process model; (313)Predict the probability distribution of new points, for a new feature point x new , predict its mean μ(x new ) and variance σ(x new ) using a Gaussian process model, compute P(X>Threshold) using the cumulative probability function of normal distribution F(z); (314) calculating a threshold value of cumulative probability reaching 99%; (315) judging whether the seed point belongs to the weld feature point according to the threshold value, if yes, the point is included in the prior data sample, which is used for subsequent threshold updating, if not, the threshold value is unchanged, judging whether the seed point set is empty, if not, the seed point is updated, and returning to step two, if empty, judging that the complete weld region segmentation is completed, and entering step four.

9. A multi-feature dynamic threshold analysis based weld point cloud region segmentation system, comprising: Comprise: A first main control module for acquiring three-dimensional point cloud data containing a weld region, and selecting an arbitrary point in the weld as an initial seed point for region growing; A second main control module for calculating the normal vector divergence feature and the tangent plane deviation feature of the seed point, and performing three-dimensional point cloud feature fusion according to the two features to form a point cloud fusion feature; The calculation of the normal vector divergence feature of the seed point comprises: (211) calculating the K-neighbors of each point in the three-dimensional point cloud data, and calculating the L1 median point thereof; (212) estimating the normal vector of each point in the three-dimensional point cloud data based on the L1 median point by using the PCA method; (213) calculating the local normal vector divergence feature factor of the feature point based on the normal vector; The calculation of the tangent plane deviation feature of the seed point comprises: (221) slicing the feature point along the scanning line direction, and calculating the projection of each point on the tangent plane to form a two-dimensional coordinate; (222) performing curve fitting on the two-dimensional coordinate by using a sliding average filter; (223) constructing a deviation feature factor to calculate the tangent plane deviation feature, wherein the deviation feature factor comprises: wherein d i is the deviation of the origin of the two-dimensional coordinates from its corresponding point on the parent curve; The three-dimensional point cloud feature fusion comprises: combining the normal vector divergence feature and the tangent plane deviation feature information by weighting to form a new weld feature factor: θ i = ω1α i + ω2β i where ω1 and ω2 are weight values, and α i is a local normal vector divergence feature factor; A third main control module for judging whether the seed point belongs to the weld region by using a dynamic threshold updating method according to the workpiece point cloud fusion feature, if yes, the K-neighborhood of the weld feature point is included in the seed set as the seed point set for subsequent judgment of whether it belongs to the weld region, if not, judging whether the seed point set is empty, if not, updating the seed point, and inputting the updated seed point into the second main control module, if empty, judging that the complete weld region segmentation is completed; A fourth main control module for extracting the weld point cloud and contour information after the segmentation is completed after the third main control module judges that the complete weld region segmentation is completed, and completing the three-dimensional point cloud weld region segmentation.

Images