Similar array feature-oriented measurement point cloud efficient processing method

Abstract

The invention belongs to the field of point cloud processing, and discloses a similar array feature-oriented measurement point cloud efficient processing method. The invention provides an efficient processing method for measuring point clouds oriented to similar array features, mainly aiming at the urgent demand that the utilization rate of similar features is insufficient in the process of processing the measuring point clouds with the similar array features in the existing point cloud simplification method, and the processing precision and efficiency of large-scale point cloud data need to be further improved, and provides the efficient processing method for the measuring point clouds oriented to the similar array features. According to the method, a feature region is identified through a clustering method; identifying an edge shape based on normal vector analysis; accurately determining the feature region based on the edge shape; and point cloud simplification is realized through voxel segmentation of the edge shape. According to the processing method, judgment can be carried out according to feature similar information, so that the processing speed is increased, data information with higher quality is obtained, and the processing method has a good application prospect in actual production and processing.

Description

Technical Field

[0001] The invention belongs to the field of point cloud processing and relates to an efficient processing method for measurement point clouds oriented to similar array features. Background Art

[0002] In recent years, modeling and analyzing workpieces through three-dimensional point clouds to guide workpiece processing has become a research hotspot in the industrial field. This technology overcomes the problems of random deformation and inaccurate positioning that are difficult to handle in traditional CNC machining. However, the original point cloud data is huge, the data is discrete, and there are a large number of redundant points. There are many disadvantages, which put forward higher requirements on the computer data transmission rate, storage capacity, computing power, and the accuracy of the calculation method. Point cloud simplification is an effective way to deal with the above data processing problems. Existing point cloud simplification methods usually reduce point cloud data by simplifying voxels, but this easily leads to the loss of detailed features. For point cloud data with complex features, such methods need to further subdivide voxels, resulting in long processing time and high requirements for operators. When processing point clouds with similar array features, it is impossible to extract feature similarity information for judgment to improve processing speed and obtain higher quality data information.

[0003] Xi'an Jiaotong University disclosed a data simplification method and system based on blade surface features in its invention patent "A data simplification method and system based on blade surface features", CN113469907A, which simplifies the single-view point cloud by retaining the blade detail features and using point cloud registration. However, this method requires manual selection when retaining features, and the simplification effect of point cloud simplification through point cloud registration method is limited. Shanghai University disclosed an adaptive point cloud simplification method based on point cloud feature partitioning in its invention patent "An adaptive point cloud simplification method based on point cloud feature partitioning", CN115294272A. This method partitions different areas by curvature, uses different simplification methods for each partition, and obtains the final simplified point cloud model. However, since this method needs to calculate the relevant information of all points, it takes a long time to process large-scale point cloud data, and when the point cloud data has many noise points, the local curvature is prone to mutation, resulting in partition errors, and too many detailed features are lost during the simplification process.

[0004] None of the above studies mentioned an efficient processing method for measured point clouds with similar array features. Summary of the invention

[0005] The present invention mainly aims at the fact that the existing point cloud simplification methods do not make sufficient use of similar features in the process of processing measurement point clouds with similar array features, and the urgent need to further improve the accuracy and efficiency of large-scale point cloud data processing. This paper proposes an efficient processing method for measurement point clouds with similar array features. The method identifies feature areas through clustering methods; identifies edge shapes based on normal vector analysis; accurately determines feature areas based on edge shapes; and simplifies point clouds through voxel segmentation of edge shapes.

[0006] The technical solution of the present invention:

[0007] An efficient processing method for measuring point clouds of similar array features is proposed. First, a line laser sensor is used to scan the workpiece to be processed to obtain point cloud data of its surface profile, and similar array features are identified by using a clustering method through drawings.

[0008] Secondly, the edge of each feature is identified by the PCA normal vector method, and the same edge features in different regions are fused to obtain the precise edge shape of similar feature arrays;

[0009] Then, edge feature points are further identified by combining shape level set with geometric flow;

[0010] Finally, based on the expanded edge shape and combined with multi-condition constraints, voxel segmentation is performed to achieve point cloud simplification while retaining feature points.

[0011] The specific steps are as follows:

[0012] Step 1: Clustering-based similar array feature recognition

[0013] The point cloud data of the surface profile of the workpiece to be processed is obtained by scanning with a line laser sensor. The point cloud data is initially filtered based on the height threshold, and the point cloud data is processed using a density-based clustering method in combination with the design size. The density of each point, that is, the number of points in the neighborhood, is calculated, and the points are sorted according to the density, and then classified into clusters C = {C1, C2..., C n}; The calculation formula is as follows:

[0014]

[0015] Among them, d(p i ,p j ) represents the point p in the point cloud data i and point p j distance; δ is the indicator function, when d(p,p i )<r, δ=1; otherwise, δ=0; N represents the total number of point clouds; the core points, i.e., the points with high density, are prioritized as the center points of the clusters, and the two clusters are gradually merged according to the center distance and average density of the two clusters. The specific formula is as follows:

[0016]

[0017] Among them, D avg (C i ,C j ) represents cluster C i and C j The average density, d(C i ,C j ) represents the distance between the centers of two clusters, D threshold and d threshold is the merge distance of density and distance; iterative merging is performed until the number of clusters is consistent with the number of designed similar array features n; the point cloud that does not belong to the cluster is divided into small voxels using the octree algorithm, and the points that belong to the cluster are processed in the next step within the cluster;

[0018] Step 2: Feature edge shape recognition and fusion

[0019] First, a plane is estimated using the neighboring points of any point q belonging to the cluster, and the normal vector of point q is estimated by minimizing an objective function:

[0020]

[0021] Among them, m is the number of points in the neighborhood of point q in cluster, q i is the coordinates of the normal point to be solved, q0 is the coordinates of the center point of the neighborhood, f is the optimization objective function, and the eigenvector corresponding to the minimum eigenvalue is the normal vector corresponding to the point;

[0022] After obtaining the corresponding normal vector, point q and its m neighboring points in the neighborhood are projected onto the tangent plane. With the projection point of point q as the starting point and the projection point of the neighboring points as the end point, m vectors can be defined. Select any vector as the starting vector, calculate the angle between two adjacent vectors, and obtain the set of angles A = {α1, α2, ..., α k}, calculate the maximum value α in the angle i , if the maximum value α i If it is greater than the threshold θ, point q is judged as a boundary point and retained as an edge point, otherwise it is deleted; the point cloud data is calculated point by point to complete edge extraction;

[0023] Since the workpiece point cloud may be limited by noise and structural occlusion during the scanning process, the edges of some similar feature arrays may be incomplete. In order to obtain data closer to the real edge features, the edges of the n different clusters obtained by clustering are extracted, smoothed and fused to obtain the comprehensive feature edge E of the similar feature array. complete ;

[0024] Step 3: Feature point recognition based on edge shape

[0025] Mark the comprehensive feature edge E complete The innermost edge E in , calculate the center of mass o of the inner circle edge in :

[0026]

[0027] Among them, n in is the innermost edge E in The number of points on (x i ,y i ,z i ) are the coordinates of each edge point, o in (x c ,y c ,z c ) is the inner edge E in The centroid of

[0028] Center of mass o in As the center, the angle θ f Perform circumferential segmentation on all edges, θ f The calculation formula is as follows:

[0029]

[0030] Among them, r is the average radius of the innermost circle edge, n r is the total number of points on the innermost edge, s i is the point on the edge of the innermost circle, d l is the length of the innermost circle edge;

[0031] After circumferential segmentation of all edges, the points in different sectors are divided by rotation angle and distance from the centroid o in The distance is used as a reference for sorting, to avoid the intersection of different edge curves, and the B-spline curve fitting is used to generate smooth edge curves at different radii;

[0032]

[0033] Among them, C(θ) is the edge curve function to be fitted, θ represents the angle parameter of each sector, and B l,k (θ) is the influence weight of the lth control point in the B-spline basis function, k represents the order of the spline, P i is the control point location;

[0034] The innermost edge of the B-spline fitting is used as the starting edge, and the edge of the next layer is used as the shape constraint item. The shape expansion is achieved by combining the shape level set with the geometric flow. The calculation formula is as follows:

[0035]

[0036] Among them, φ is the level set function of the current edge; F is the expansion speed function, represents the normal direction of the edge, α is the smoothing control parameter; κ is the curvature term; β is the shape control weight, φ target is the target level set function of the outer edge;

[0037] Record the continuous expansion process in the same sector, and use cosine similarity to calculate the change in the normal vector angle difference generated by adjacent expansion operations in the same sector, avoiding direct calculation of the angle value and simplifying the calculation process;

[0038]

[0039] in, is the normal vector between the current point and the previous expansion of the point position, is the normal vector between the current point and the next expansion of the point, θ s The change in the normal vector angle difference before and after the expansion;

[0040] Set a similarity threshold, and the points greater than the threshold are regarded as similar array feature edge feature points; the corresponding extended area of ​​this type of point is regarded as the feature area E;

[0041] Step 4: Simplify the feature points based on edge shape voxel segmentation

[0042] Due to the uncertainty of edge shape, when the edge shape changes suddenly, the feature information is easily lost. The density-weighted neighborhood collaborative curvature is used to determine whether it is a feature mutation:

[0043]

[0044] Among them, w j is the density weight of the jth voxel in the feature area E, N j is the number of points in voxel j, is the total number of feature area points; C t is the co-curvature of the t-th feature region, Curvature t is the curvature of the t-th feature region, Curvature j is the curvature of the jth voxel;

[0045] Based on voxel to centroid o in Dynamic subdivision of distance:

[0046]

[0047] Where M is the number of points in the neighborhood of the voxel, p k is a point in the neighborhood, D i For each point p iThe distance to the centroid C, λ is a dynamic partition parameter used to adjust the density of the partition; for D i >D th The points are further subdivided to ensure the integrity of the feature area;

[0048] Finally, the points in the non-feature area of ​​the cluster are segmented and simplified according to the octree voxel segmentation method, thus completing the rapid processing of the point cloud.

[0049] Beneficial effects of the invention: The invention proposes an array point cloud simplification method based on edge shape preservation, which recognizes edge shapes and simplifies point cloud data while maintaining edge shape invariance. At the same time, the traditional voxel shape is improved according to the edge shape, which effectively reduces the damage to the feature area and avoids the loss of detail features. When processing large-scale high-precision point clouds, the quality of data simplification is guaranteed while taking efficiency into account, which provides a strong guarantee for subsequent processing and is suitable for precision processing scenarios in intelligent manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a specific flow chart of the method of the present invention.

[0051] Figure 2 This is the physical model diagram of the workpiece to be processed.

[0052] Figure 3 It is a point cloud diagram of the processing parts involved in the present invention.

[0053] Figure 4 This is a point cloud image simplified by the method of the present invention.

[0054] Figure 5 This is the effect diagram of point cloud image feature recognition in the present invention. DETAILED DESCRIPTION

[0055] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0056] The embodiment of the present invention is a certain type of rocket engine injection disk with a diameter of 430 mm and a height of 300 mm. The injection disk has a plurality of identical injection hole features to be processed, such as Figure 5 As shown. Figure 1 The overall flow chart of the efficient processing method of the measurement point cloud for similar array features is as follows:

[0057] Step 1: Clustering-based similar array feature recognition

[0058] The point cloud data of the surface profile of the workpiece to be processed is obtained by scanning with a line laser sensor. Since the original profile point cloud has many noise points and data point redundancy, the data volume is extremely large. The collected point cloud data is about 80 million points. Data processing has extremely high requirements on the computer system, and the processing industrial control system cannot meet the conditions. The point cloud data is initially filtered based on the height threshold, and the point cloud data is processed using a density-based clustering method in combination with the workpiece drawing. According to the drawing, there are a total of 271 processing features. The density of each point (the number of points in the neighborhood) is calculated, and the points are sorted according to the density. The data is classified into clusters C = {C1, C2..., C 271}.

[0059] Step 2: Feature edge shape recognition and fusion

[0060] After calculating the normal vector of each point in the point cloud, the edges of different clusters are extracted and the influence of noise is reduced by smoothing. The edge fusion is performed to obtain the comprehensive feature edge E. complete :

[0061] Step 3: Feature point recognition based on edge shape

[0062] Mark the comprehensive feature edge E complete The inner edge E in , calculate the center of mass o of the inner circle edge in :

[0063]

[0064] Where n is the edge E in The number of points on (x i ,y i ,z i ) is the coordinate of each edge point, o in is the inner edge E in The calculated center position characteristic hole edge center o in The center of mass is (-0.494039, 0.2888). Repeat the operation for other holes to obtain the center of mass. in The center is divided into all edges in the circumferential direction. In this embodiment, the radius r of the iterative calculation is the hole radius 7.5, the number of characteristic hole edge points n at the center position is 3206, and the calculation adopts θ f It is 0.003.

[0065] After sorting the points of different edges in each sector, this embodiment uses cubic B-spline fitting to generate a smooth edge curve. The edge close to the centroid position of the B-spline fitting is used as the starting edge, and the next edge is used as the shape constraint item. The shape extension is achieved by combining the shape level set with the geometric flow. The calculation formula is as follows:

[0066]

[0067] Among them, φ is the level set function of the current edge, F is the expansion speed function with a constant value of 1.0 to control the basic speed of expansion, α is the smoothness control parameter with a value of 0.05 to adjust the influence of curvature and ensure the smoothness of the expanded edge, κ is the curvature term used to maintain the smoothness of the edge during the expansion process, β is the shape control weight with a value of 1.0 to adjust the influence of the shape constraint term, and θ target is the target level set function of the outer edge, and the zero level set corresponds to the position of the outer edge. (φ target -φ) is a shape constraint term, which represents the deviation between the current edge and the target outer edge. This term gradually guides the inner edge toward the outer edge.

[0068] The cosine similarity is used to calculate the angle difference between the normal vectors of the first extension and the second extension in the same sector to simplify the calculation process.

[0069]

[0070] The similarity threshold is set to 0.9, and points greater than the threshold are regarded as feature change points. The extended area is stored as feature area E.

[0071] Step 4: Simplify the feature points based on edge shape voxel segmentation

[0072] In this example, the resolution of the line laser is 0.025. Due to the uncertainty of the edge shape, when the edge shape changes suddenly, the feature information is easily lost. The neighborhood collaborative curvature based on density weighting is used to determine whether it is a feature mutation:

[0073]

[0074] Among them, w j is the density weight of the jth voxel in the feature area E, with a weight of 0.2, N j is the number of voxel j, N is the total number of feature area point clouds, in this example the number of feature point clouds is 3206, C t is the co-curvature of the t-th feature region, Curvature t is the curvature of the t-th feature region, Curvature j is the curvature of the jth voxel. Based on the voxel to the centroid o in Dynamic subdivision of distance:

[0075]

[0076] Where M is the number of points in the neighborhood of the voxel, p k is a point in the neighborhood, D i For each point pi The distance to the centroid C, λ is a dynamic partition parameter used to adjust the partition density, and is set to 0.2 in this embodiment. i >D th The point is further subdivided, in this embodiment D th =0.01 to ensure the integrity of the feature area. Finally, the points in the non-feature area of ​​the cluster are segmented and simplified according to the octree voxel segmentation method to complete the simplification.

[0077] The final simplified point cloud data is about 610,000, and the simplification rate is close to 99%. After testing, the feature error after simplification is 0.05mm, which meets the requirement of retaining feature accuracy.

[0078] The method described in the present invention can realize efficient and high-precision simplification of large-scale point cloud data, and completely retains the local processing features of parts. It is suitable for the digital processing of complex structural parts with multiple array processing features in the aerospace field, and can also be applied to related fields such as geometric feature measurement and three-dimensional surface reconstruction. The overall process is simple to operate, and the results are reliable and efficient, which effectively makes up for the shortcomings of existing methods.

[0079] The specific implementation cases described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation case of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An efficient processing method for measurement point clouds with similar array features, characterized in that: Here are the steps: Firstly, a line laser sensor is used to scan the workpiece to obtain the point cloud data of its surface profile, and similar array features are identified through the drawings using the clustering method; Secondly, the edge of each feature is identified by the PCA normal vector method, and the same edge features in different regions are fused to obtain the precise edge shape of similar feature arrays; Then, edge feature points are further identified by combining shape level set with geometric flow; Finally, based on the expanded edge shape and combined with multi-condition constraints, voxel segmentation is performed to achieve point cloud simplification while retaining feature points.

2. An efficient processing method for measurement point clouds with similar array features, characterized in that: The specific steps are as follows: Step 1: Clustering-based similar array feature recognition The point cloud data of the surface profile of the workpiece to be processed is obtained by scanning with a line laser sensor. The point cloud data is initially filtered based on the height threshold, and the point cloud data is processed using a density-based clustering method in combination with the design size. The density of each point, that is, the number of points in the neighborhood, is calculated, and the points are sorted according to the density, and then classified into clusters C = {C1, C2..., C n }; The calculation formula is as follows: Among them, d(p i ,p j ) represents the point p in the point cloud data i and point p j distance; δ is the indicator function, when d(p,p i )<r, δ=1; otherwise, δ=0; N represents the total number of point clouds; the core points, i.e., the points with high density, are prioritized as the center points of the clusters, and the two clusters are gradually merged according to the center distance and average density of the two clusters. The specific formula is as follows: Among them, D avg (C i ,C j ) represents cluster C i and C j The average density, d(C i ,C j ) represents the distance between the centers of two clusters, D threshold and d threshold is the merge distance of density and distance; iterative merging is performed until the number of clusters is consistent with the number of designed similar array features n; the point cloud that does not belong to the cluster is divided into small voxels using the octree algorithm, and the points that belong to the cluster are processed in the next step within the cluster; Step 2: Feature edge shape recognition and fusion First, a plane is estimated using the neighboring points of any point q belonging to the cluster, and the normal vector of point q is estimated by minimizing an objective function: Among them, m is the number of points in the neighborhood of point q in cluster, q i is the coordinates of the normal point to be solved, q0 is the coordinates of the center point of the neighborhood, f is the optimization objective function, and the eigenvector corresponding to the minimum eigenvalue is the normal vector corresponding to the point; After obtaining the corresponding normal vector, point q and its m neighboring points in the neighborhood are projected onto the tangent plane. The projection point of point q is used as the starting point and the projection point of the neighboring point is used as the end point to define m vectors. Any vector is selected as the starting vector, and the angle between two adjacent vectors is calculated to obtain the set of angles A = {α1, α2, ..., α k }, calculate the maximum value α in the angle i , if the maximum value α i If it is greater than the threshold θ, point q is judged as a boundary point and retained as an edge point, otherwise it is deleted; the point cloud data is calculated point by point to complete edge extraction; Since the point cloud data of the workpiece in the scanning process may be limited by noise and structural occlusion, the edges of some similar feature arrays may be incomplete. In order to obtain data closer to the real edge features, the edges of the n different clusters obtained by clustering are extracted, smoothed and fused to obtain the comprehensive feature edge E of the similar feature array. complete ; Step 3: Feature point recognition based on edge shape Mark the comprehensive feature edge E complete The innermost edge E in , calculate the center of mass o of the inner circle edge in : Among them, n in is the innermost edge E in The number of points on (x i ,y i ,z i ) are the coordinates of each edge point, o in (x c ,y c ,z c ) is the inner edge E in The centroid of Take the center of mass o in As the center, the angle θ f Perform circumferential segmentation on all edges, θ f The calculation formula is as follows: Among them, r is the average radius of the innermost circle edge, n r is the total number of points on the innermost edge, s i is the point on the edge of the innermost circle, d l is the length of the innermost circle edge; After circumferential segmentation of all edges, the points in different sectors are divided by rotation angle and distance from the centroid o in The distance is used as a reference for sorting, to avoid the intersection of different edge curves, and the B-spline curve fitting is used to generate smooth edge curves at different radii; Among them, C(θ) is the edge curve function to be fitted, θ represents the angle parameter of each sector, and B l,k (θ) is the influence weight of the lth control point in the B-spline basis function, k represents the order of the spline, P i is the control point location; The innermost edge of the B-spline fitting is used as the starting edge, and the edge of the next layer is used as the shape constraint item. The shape expansion is achieved by combining the shape level set with the geometric flow. The calculation formula is as follows: Among them, φ is the level set function of the current edge; F is the expansion speed function, represents the normal direction of the edge, α is the smoothing control parameter; κ is the curvature term; β is the shape control weight, φ target is the target level set function of the outer edge; Record the continuous expansion process in the same sector, and use cosine similarity to calculate the change in the normal vector angle difference generated by adjacent expansion operations in the same sector, avoiding direct calculation of the angle value and simplifying the calculation process; in, is the normal vector between the current point and the previous expansion of the point position, is the normal vector between the current point and the next expansion of the point, θ s The change in the normal vector angle difference before and after the expansion; Set a similarity threshold, and the points greater than the threshold are regarded as similar array feature edge feature points; the corresponding extended area of ​​this type of point is regarded as the feature area E; Step 4: Simplify the feature points based on edge shape voxel segmentation Due to the uncertainty of edge shape, when the edge shape changes suddenly, the feature information is easily lost. The density-weighted neighborhood collaborative curvature is used to determine whether it is a feature mutation: Among them, w j is the density weight of the jth voxel in the feature area E, N j It is a voxel j The number of points, N is the total number of feature area points; C t is the co-curvature of the t-th feature region, Curvature t is the curvature of the t-th feature region, Curvature j is the curvature of the jth voxel; Based on voxel to centroid o in Dynamic subdivision of distance: Where M is the number of points in the neighborhood of the voxel, p k is a point in the neighborhood, D i For each point p i The distance to the centroid C, λ is a dynamic partition parameter used to adjust the density of the partition; for D i >D th The points are further subdivided to ensure the integrity of the feature area; Finally, the points in the non-feature area of ​​the cluster are segmented and simplified according to the octree voxel segmentation method, thus completing the rapid processing of the point cloud.

Images