Multi-type polishing area detection algorithm for casting autonomous polishing

Abstract

The invention relates to the technical field of visual identification and detection of polishing robots, in particular to a multi-type polishing area detection algorithm for casting autonomous polishing. Comprising the following steps of S1, process type classification, S2, defect type classification, S3, bulge type defect detection, S4, crator defect detection, S5, rough surface defect detection and S6, point-by-point grinding process detection. According to a grinding target area detection algorithm provided by the invention, a position sensing basis is provided for grinding track planning by classifying grinding areas; different detection algorithm libraries are formed and deployed in a system of the grinding robot, the adaptability to detection of the to-be-ground area can be enhanced, and the working capacity of the grinding robot is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of polishing robot visual recognition, and in particular to a multi-type polishing area detection algorithm for autonomous polishing of castings. Background Art

[0002] The steel industry is an important part of the manufacturing industry and an important basic industry of the national economy. Steel castings are widely used in various industries, such as shipyards, car factories, construction, etc., which brings with it a huge demand for steel casting grinding. At present, grinding still relies heavily on manual work. Due to the poor working environment, uneven skills, high labor intensity and health threats of workers, the emergence of grinding robots can reduce the labor intensity of workers and improve production efficiency, which is in line with the current development trend of intelligent manufacturing and equipment renewal.

[0003] Due to the complexity of the polishing and grinding process, there is a lack of mature companies and products in terms of technology, products, and market applications for this type of robot. The current casting grinding robot technology still has the following problems: (1) The grinding process is too complicated and the process is changeable. Different casting shapes and grinding conditions require the grinding robot to have autonomous adaptability and learn from the workers' experience to perform grinding operations; (2) Ideally, only industrial robots with high repeatability accuracy are used to complete the grinding of a single product. Since there is no need to change the program, its grinding efficiency is the highest; however, most grinding manufacturers frequently switch between different casting products for grinding, which limits the production efficiency of the grinding robot and increases the requirements for the robot's flexible production capacity; (3) Due to product deformation and raw material differences, the consistency of incoming materials is poor. If offline programming or teaching programming is used, the grinding tool will operate according to a predetermined path, which will cause some castings to be under-polished or over-polished, resulting in poor operational flexibility.

[0004] At present, robot grinding has low efficiency and poor autonomous adaptability. To verify the maturity of the casting grinding robot technology route, too many comprehensive factors need to be considered, such as the rigidity and stability of the force on the end of the industrial robot, the perception ability of the robot system combined with machine vision, trajectory accuracy, and the high requirements for the adaptability of grinding process parameters.

[0005] Polishing robots with visual perception capabilities have the potential to replace some manual work and are expected to be applied on a large scale; autonomous polishing robots usually use depth cameras to collect data, obtain spatial information of the area to be polished, obtain point cloud data, and then analyze the area to be polished to guide the robot to perform polishing operations; since there are many types of areas to be polished on the surface of bamboo swords with different shapes, how to quickly and accurately detect different types of areas to be polished is one of the key technologies of autonomous polishing robots. Summary of the invention

[0006] The object of the present invention is to provide a multi-type grinding area detection algorithm for autonomous grinding of castings, which can quickly and accurately detect and identify various types of areas to be ground on the surface of castings.

[0007] Based on the above purpose, the present invention adopts the following technical solution:

[0008] A multi-type grinding area detection algorithm for autonomous grinding of castings includes the following steps:

[0009] S1. Process type classification: According to the grinding process type, the surface defects of castings are classified into full surface grinding process and point-by-point grinding process;

[0010] S2. Classification of defect types: Classify the surface defects of the casting according to the grinding process type in step S1. The full-surface grinding process includes convex defects formed by welding after the casting defect, weld scar defects formed by welding, and surface defects to be rusted or rough; the point-by-point grinding process includes point scars formed by reasons such as molten steel dripping;

[0011] S3, convex defect detection: detecting the convex defects classified in step S2 based on the height difference;

[0012] S4, weld scar defect detection: detecting weld scar defects formed by welding classified in step S2 based on the change of normal vector;

[0013] S5, rough surface defect detection: detecting the surface defects to be removed or roughness classified in step S2 based on the flatness method;

[0014] S6. Point-by-point grinding process detection: Based on the Laplace skeleton extraction algorithm, the point-by-point grinding process type defects classified in step S1 are detected.

[0015] Preferably, the specific process of the convex defect detection in step S3 includes:

[0016] S3.1, grid splitting: according to the height difference between the raised edge of the raised grinding target and the casting surface, the point cloud of the area is gridded and the point cloud belonging to a certain grid is split;

[0017] S3.2, height detection: by traversing all grids in parallel, calculating the height difference in each grid, and marking the grids that exceed the height difference threshold;

[0018] S3.3, grid clustering: by merging all the marked grids, the convex edges are detected, and then clustering is performed to distinguish different convexities.

[0019] Preferably, the specific process of grid splitting in step S3.1 includes:

[0020] For the polishing area point cloud, project it onto the XOY plane and divide it into grids according to the projection boundary and the side length of the grid. The formula is as follows:

[0021]

[0022] in, Represents the maximum and minimum boundaries, L g is the grid side length, X num , Y num It is the number of grids in the X-axis and Y-axis directions.

[0023] Preferably, the specific process of step S4 welding scar defect detection includes:

[0024] S4.1. Normal calculation: Calculate the point cloud normal of the casting surface to be tested;

[0025] S4.2, curvature change detection: Since the normal line has mutation in the scar type area, the edge of the weld scar can be detected by detecting the change characteristics of the normal line, that is, the change of the surface curvature of the point in the point cloud;

[0026] S4.3, weld scar extraction: Based on the statistical results obtained in step S4.2, select a point set whose curvature is greater than a set threshold and whose corresponding point number is less than the set threshold as the weld scar target detection result.

[0027] Preferably, the specific process of normal calculation in step S4.1 includes:

[0028] The vector perpendicular to the tangent plane at a certain point is defined as the normal at that point. Therefore, solving the normal problem of a point on the surface is equivalent to estimating the normal of a tangent plane of the surface. After transformation, it can be transformed into a plane fitting estimation problem; using the least squares method for estimation, the formula is as follows:

[0029]

[0030] in, is the normal vector at the point and is the center of the plane, k represents P i The number of neighboring points;

[0031] The above formula indicates that: for point P in the local point cloud i , i = 1, 2, ..., k, find a plane that passes through the center of the plane The normal vector satisfy In the normal vector The sum of the projections on is minimal;

[0032] Normal vector The solution can be simplified to the problem of solving the eigenvalue and eigenvector of principal component analysis. The formula is as follows:

[0033]

[0034] where M is a symmetric semi-positive definite covariance matrix, λ j is the jth eigenvalue of the covariance matrix, is the eigenvalue λ j The corresponding eigenvector. i Normal vector of a point is the eigenvector corresponding to the minimum eigenvalue.

[0035] Preferably, the specific process of curvature change detection in step S4.2 includes:

[0036] The curvature of a point is defined as:

[0037]

[0038] in, Indicates P i The curvature of the point, λ 0 , 1 , 2 YesP i The eigenvalue of the covariance matrix M corresponding to the point, and λ 0 ≤λ 1 ≤λ 2 ; Divide the point cloud into different intervals by counting the curvature of the points in the point cloud.

[0039] Preferably, the specific process of rough surface defect detection in step S5 includes:

[0040] S5.1, Iterative surface fitting: Perform quadratic surface fitting on the casting surface, use iterative fitting method to solve the quadratic surface parameters, and obtain the fitting point cloud;

[0041] S5.2. Rough part extraction: By comparing the collected point cloud with the fitted point cloud, the area above the comparison threshold is extracted as the rough part to be polished.

[0042] Preferably, the specific process of iterative surface fitting in step S5.1 includes:

[0043] The fitting equation is as follows:

[0044]

[0045] Among them, P i =(X i ,Y i ,Z i), i = 1, 2, ..., n represents the point coordinates of the polishing area point cloud, n is the number of polishing area points, Z (X i ,Y i ) is the Z-axis fitting coordinate of the i-th point; the quadratic surface fitting parameter is S p =[a,b,c,d,e,f];

[0046] The iterative fitting method is used to solve the quadratic surface parameters and obtain the fitting point cloud: First, convert equation (6) into the least squares form and use the singular value decomposition (SVD) method to solve S p ; Then, set the iteration error threshold ε = 2mm and set S p Substitute into formula (1) to calculate Z(X i ,Y i ), filter out |Z(X i ,Y i )-Z i |≥ε; Finally, recalculate S for the filtered point cloud p , until every point satisfies |Z(X i ,Y i )-Z i |<ε, then the corresponding S p As the final quadratic surface fitting parameters.

[0047] Preferably, the specific process of step S6 of point-by-point polishing process detection includes:

[0048] S6.1. Parameter initialization: define pixel points and distribution according to Zhang-Suen algorithm;

[0049] S6.2. Delaunay triangulation of point cloud: Generate triangular geometry for a given point cloud. k The set of points in the neighborhood P = {P 1 ,P 2 ,…,P k}, get the triangulation set And it needs to satisfy: 1) the triangle endpoints corresponding to the triangulation set constitute the point set P; 2) the triangulation set is the convex hull of P; 3) the edges of any two triangulations cannot intersect.

[0050] S6.3, Laplace matrix construction: For the triangles in step S6.2 Delaunay triangulation, establish a graph topology relationship, use the vertices of the triangles as the nodes of the graph, the edges of the triangles as the edges of the graph, and use the Laplace cotangent weights as the weights of the graph to establish the Laplace matrix L k; The rows and columns of the matrix represent the vertices of the triangulation, and the elements of the matrix represent the cotangent weights of the edges between the vertices of the triangulation. The cotangent weights are expressed as:

[0051]

[0052] S6.4, iterative skeleton shrinkage: The Laplace skeleton shrinks the framework to reduce the volume of the three-dimensional point cloud to 0, but keeps the overall shape of the point cloud unchanged; the Laplace skeleton extraction algorithm considers two mutually balanced factors, one is to eliminate the detailed information of the original point cloud model to reduce the volume, and the other factor is to ensure that the shrunk shape can represent the topological structure of the original point cloud; as the shrinkage process proceeds, the point cloud will be gradually refined, making the calculation of the cutting plane more difficult and the accuracy reduced. Therefore, in the iterative process, the neighborhood information calculated in the first iteration and the edge connection relationship of the Delaunay triangulation are retained, and the subsequent iteration process only needs to update the Laplace weight;

[0053] S6.5, point-by-point key point extraction: Through the iterative shrinkage process of step S6.4, point-by-point key points are obtained as polishing key points.

[0054] The beneficial effects of the present invention include:

[0055] The casting grinding area detection algorithm proposed in the present invention classifies the grinding areas according to the process type, and designs different detection algorithms according to the different characteristics of defects of different categories, so that it can accurately identify grinding targets such as protrusions, rough surfaces, and weld scars; based on the point cloud image of the casting surface, it can accurately identify the type, area position and shape of defects that need to be polished, effectively locate the grinding target, and guide the autonomous grinding robot to perform grinding operations.

[0056] The present invention proposes an algorithm for detecting raised polished areas based on height differences. The raised edges can be identified by rasterizing the point cloud of the area, calculating the height difference in each grid, and marking the grids that exceed the threshold. The height difference detection method can accurately and stably detect raised defects on the surface of castings, and can accelerate the processing speed through parallel calculation of grids, thereby effectively improving the detection efficiency.

[0057] The present invention proposes an algorithm for detecting the weld scar type grinding area based on the change of normal vector. Based on the surface normal vector attribute of the point cloud, the surface curvature change of the points in the point cloud of the casting surface is detected, and the points with large change in point cloud curvature are counted, so as to obtain the contour of the weld scar type grinding target.

[0058] The present invention proposes an algorithm for detecting rough surface type grinding areas based on flatness. Since the areas to be polished of castings in industrial scenarios are commonly relatively homogeneous curved surfaces and planes, the information of the areas to be polished can be obtained by quadratic surface fitting; the quadratic surface parameters are solved by an iterative fitting method to obtain the surface and plane fitting results that are close to the surface shape of the casting; the collected point cloud is compared with the fitted point cloud, and the area above the comparison threshold can be extracted and regarded as the rough part to be polished.

[0059] The present invention proposes an improved Laplace skeleton extraction algorithm for grinding a gap (scar) type grinding area whose significant feature is a slightly convex gap (scar); through the improved Laplace contraction framework, the key point detection accuracy in the point-by-point grinding trajectory is effectively improved, so that the point-by-point grinding and polishing trajectory can be determined more accurately, thereby improving the working accuracy of the grinding robot.

[0060] The present invention classifies the polishing areas, forms different detection algorithm libraries for different types of defects, and deploys them in the system of the polishing robot, thereby enhancing the adaptability of detection of the area to be polished and effectively improving the operating ability of the polishing robot. There is no need to frequently adjust program parameters when used in scenarios with frequent switching of product types or poor consistency of incoming materials, thereby effectively improving the work efficiency of the polishing robot in a complex working environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a flow chart of casting grinding area detection of the present invention;

[0062] Figure 2 Schematic diagram of casting surface defect types;

[0063] Figure 3 The result of the height difference detection algorithm after rasterizing the point cloud image in the detection of the raised polishing area;

[0064] Figure 4 for Figure 3 The result after raster clustering and edge expansion in the image;

[0065] Figure 5 Grinding target detection results for weld scars;

[0066] Figure 6 Grinding the surface of curved and flat castings before quadratic surface fitting;

[0067] Figure 7 Grinding the surface of curved and flat castings after quadratic surface fitting;

[0068] Figure 8 Comparison diagram of the roughness of the polished surfaces of curved and flat castings;

[0069] Fig. 9 Definition and distribution of pixels for Zhang-Suen algorithm;

[0070] Fig.10 Schematic diagram of point-by-point polishing key points extracted by the algorithm. DETAILED DESCRIPTION

[0071] The following is a further explanation of the present invention in conjunction with specific embodiments. Figure 1 As shown, this embodiment is a multi-type grinding area detection algorithm for autonomous grinding of castings, which mainly includes the following steps:

[0072] S1. Process type classification: According to the type of grinding process, casting surface defects are classified into full-surface grinding process and point-by-point grinding process. Full-surface grinding process refers to enclosing the detected grinding area with a bounding box and grinding and polishing it all smooth. Point-by-point grinding process mainly targets the shape of weld scars and is a point-by-point grinding trajectory of weld scars planned according to the target shape.

[0073] S2. Classification of defect types: In the gradual production, there are mainly three types of workpieces that need to be polished, namely, the bulge formed by welding after the casting defect, the surface to be derusted or rough, the scar formed by molten steel dripping or the weld formed by welding (hereinafter collectively referred to as weld scar), such as Figure 2 As shown; the surface defects of the casting are classified according to the grinding process type in step S1. The full-surface grinding process is used for convex defects formed by welding after the casting defect, weld scars formed by welding, and surface defects that need to be derusted or rough; the point-by-point grinding process is used for weld scar-shaped defects caused by reasons such as molten steel dripping.

[0074] S3, raised defect detection: The raised defects classified in step S2 are detected based on the height difference. Since there is a height difference between the raised edge of the raised grinding target and the casting surface, the raised edge can be detected by detecting the height difference, thereby determining the position and area of ​​the raised grinding target.

[0075] like Figure 3 As shown, firstly, the point cloud of the area is rasterized, and the point cloud belonging to a certain grid is segmented; for the point cloud of the polishing area, it is projected onto the XOY plane, and the grid is divided according to the projection boundary and the side length of the grid. The formula is as follows:

[0076]

[0077] in, Represents the maximum and minimum boundaries, L g is the grid side length, X num , Y num It is the number of grids in the X-axis and Y-axis directions.

[0078] Then, by traversing all grids in parallel, the height difference is calculated in each grid, and the grids with height difference exceeding the threshold are marked as green; then, by merging all marked grids, the raised edges can be detected; clustering is then performed to distinguish different raised areas, such as Figure 4 As shown, after clustering the green grids and expanding the edges, the raised grinding targets can be located.

[0079] S4. Weld scar defect detection: The point cloud surface normal vector is an important attribute of the point cloud geometric features, which describes the orientation of the surface where the point is located. The vector perpendicular to the tangent plane at a certain point is defined as the normal of the point. Therefore, solving the normal problem of a point on the surface is equivalent to estimating the normal problem of a tangent plane of the surface, and the point cloud normal of the weld scar type workpiece is calculated accordingly.

[0080] Since the normal line in the weld scar type area has mutation, the edge of the scar can be detected based on the change characteristics of the normal line, that is, the change of the surface curvature of the point in the point cloud is detected; the vector perpendicular to the tangent plane at a certain point is defined as the normal line at that point, so solving the normal line problem of a point on the surface is equivalent to estimating the normal line of a tangent plane of the surface, which can be transformed into a plane fitting estimation problem after transformation; using the least squares method for estimation, the formula is as follows:

[0081]

[0082] in, is the normal vector at the point and is the center of the plane, k represents P i The number of neighboring points;

[0083] The above formula indicates that: for point P in the local point cloud i , i = 1, 2, ..., k, find a plane that passes through the center of the plane The normal vector satisfy In the normal vector The sum of the projections on is minimal;

[0084] Normal vector The solution can be simplified to the problem of solving the eigenvalue and eigenvector of principal component analysis. The formula is as follows:

[0085]

[0086] where M is a symmetric semi-positive definite covariance matrix, λ j is the jth eigenvalue of the covariance matrix, is the eigenvalue λ j The corresponding eigenvector. iNormal vector of a point is the eigenvector corresponding to the minimum eigenvalue.

[0087] Since the normal line has mutation in the scar type area, the edge of the weld scar can be detected by detecting the change characteristics of the normal line, that is, the change of the surface curvature of the point in the point cloud; the curvature of the point is defined as:

[0088]

[0089] in, Indicates P i The curvature of the point, λ 0 , 1 , 2 YesP i The eigenvalue of the covariance matrix M corresponding to the point, and λ 0 ≤λ 1 ≤λ 2 ; By counting the curvature of the points in the point cloud, the point cloud image can be divided into different intervals; based on the result of curvature change detection, a threshold is set, and the point set with a curvature greater than the set threshold and a corresponding point number less than the set threshold is selected as the result of weld scar target detection in the full surface grinding process, such as Figure 5 shown.

[0090] S5. Rough surface defect detection: Since the areas to be polished in castings in industrial scenarios are usually relatively homogeneous curved surfaces and planes, in order to obtain information about the areas to be polished, a quadratic surface fitting method can be used. In this embodiment, an iterative fitting method is used to solve the quadratic surface parameters, and a quadratic surface fitting is performed on the casting surface. The fitting equation is as follows:

[0091]

[0092] Among them, P i =(X i ,Y i ,Z i ), i = 1, 2, ..., n represents the point coordinates of the polishing area point cloud, n is the number of points in the polishing area, and the Z-axis fitting coordinate of the i-th point is Z(X i ,Y i ), the quadratic surface fitting parameter is S p =[a,b,c,d,e,f].

[0093] The iterative fitting method is used to solve the quadratic surface parameters to obtain the fitting point cloud. First, the equation (6) is converted into the least squares form, and the SVD method is used to solve S p ; Then, set the iteration error threshold ε = 2mm and set S p Substitute into formula (1) to calculate Z(X i ,Yi ), filter out |Z(X i ,Y i )-Z i |≥ε; Finally, recalculate S for the filtered point cloud p , until every point satisfies |Z(X i ,Y i )-Z i |<ε, then the corresponding S p As the final quadratic surface fitting parameters.

[0094] The rough surface of curved and flat castings before fitting is as follows Figure 6 As shown in the figure, after the quadratic surface fitting, the surface and plane fitting results close to the casting surface shape are as follows Figure 7 As shown in the figure, the casting surface information fitted by iterative method is of great significance. The surface of the casting after grinding should be consistent with the iterative fitting surface, indicating that the grinding is completed.

[0095] Compare the collected point cloud with the fitted point cloud, and then extract the area above the comparison threshold as the rough area to be polished, such as Figure 8 As shown; Figure 8 The green area in the middle is the fitted casting surface, and the red area is the rough part to be polished. In actual production, for the convenience of polishing, the rough surface to be polished is usually fully polished. Figure 8 The entire green area is polished.

[0096] S6. Point-by-point grinding process inspection: For the grinding area whose notable feature is a slightly convex gap (scar), the grinding trajectory should be consistent with the shape of the weld scar, which is called a point-by-point polishing grinding trajectory. For this type of grinding scenario, the core is to fit the key points of the weld scar shape.

[0097] In this embodiment, the key point detection problem is converted into the problem of extracting the weld scar skeleton. The weld scar is extracted from the 2D image to refine the pixel center line, and the pixel center line is fused into the 3D geometric space through the correspondence relationship to form a 3D grinding and polishing trajectory: First, the pixel points and distribution are defined according to the Zhang-Suen algorithm, such as Fig. 9 As shown, p 1 ,p 2 ,…,p 9 Represents the pixel point of the image, p 1 is the center pixel, and the remaining pixels are p 1 The eight-neighborhood pixel point is the center.

[0098] Then the point cloud image is Delaunay triangulated: the triangle geometry is generated for the given point cloud. k The set of points in the neighborhood P = {P1 ,P 2 ,…,P k}, get the triangulation set And the following three points need to be met:

[0099] (1) The triangle endpoints corresponding to the triangulation set constitute the point set P;

[0100] (2) The triangulation set is the convex hull of P;

[0101] (3) The edges of any two triangulations cannot intersect.

[0102] Then, for the triangles in the Delaunay triangulation, a graph topology relationship is established, with the vertices of the triangles as the nodes of the graph, the edges of the triangles as the edges of the graph, and the weights of the graph using the Laplace cotangent weights to establish the Laplace matrix L k ; The rows and columns of the matrix represent the vertices of the triangulation, and the elements of the matrix represent the cotangent weights of the edges between the vertices of the triangulation. The cotangent weights are expressed as:

[0103]

[0104] Finally, the improved Laplace shrinkage framework is used to improve the accuracy of key point detection for point-by-point polishing: the volume of the three-dimensional point cloud is reduced to 0 through the Laplace skeleton shrinkage framework, but the overall shape of the point cloud remains unchanged; the Laplace skeleton extraction algorithm considers two mutually balanced factors, one is to eliminate the detailed information of the model of the original point cloud to reduce the volume, and the other factor is to ensure that the shrunk shape can represent the topological structure of the original point cloud; as the shrinkage process proceeds, the point cloud will be gradually refined, making the calculation of the cutting plane more difficult and the accuracy reduced. Therefore, in the iterative process, the neighborhood information calculated in the first iteration and the edge connection relationship of the Delaunay triangulation are retained, and the subsequent iteration process only needs to update the Laplace weight.

[0105] According to the results of the iterative shrinkage process of the improved Laplace shrinkage framework, point-by-point key points are obtained and used as polishing key points. The detection results are as follows Fig.10 shown.

[0106] The above description is only a further explanation of the present invention in combination with specific embodiments. All descriptions do not limit the scope of protection of the present invention. Any changes or replacement schemes that can be easily thought of by any technician in this field within the technical scope disclosed by the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A multi-type grinding area detection algorithm for autonomous grinding of castings, characterized in that: The following steps are involved: S1. Process type classification: According to the grinding process type, the surface defects of castings are classified into full surface grinding process and point-by-point grinding process; S2. Classification of defect types: Classify the surface defects of the casting according to the grinding process type in step S1. The full-surface grinding process includes convex defects formed by welding after the casting defect, weld scar defects formed by welding, and surface defects to be rusted or rough; the point-by-point grinding process includes point scars formed by reasons such as molten steel dripping; S3, convex defect detection: detecting the convex defects classified in step S2 based on the height difference; S4, weld scar defect detection: detecting weld scar defects formed by welding classified in step S2 based on the change of normal vector; S5, rough surface defect detection: detecting the surface defects to be removed or roughness classified in step S2 based on the flatness method; S6. Point-by-point grinding process detection: Based on the Laplace skeleton extraction algorithm, the point-by-point grinding process type defects classified in step S1 are detected.

2. The multi-type grinding area detection algorithm for autonomous grinding of castings according to claim 1 is characterized in that: The specific process of the convex defect detection in step S3 includes: S3.1, grid splitting: according to the height difference between the raised edge of the raised grinding target and the casting surface, the point cloud of the area is gridded and the point cloud belonging to a certain grid is split; S3.2, height detection: by traversing all grids in parallel, calculating the height difference in each grid, and marking the grids that exceed the height difference threshold; S3.3, grid clustering: by merging all the marked grids, the convex edges are detected, and then clustering is performed to distinguish different convexities.

3. The multi-type grinding area detection algorithm for autonomous grinding of castings according to claim 2 is characterized in that: The specific process of grid splitting in step S3.1 includes: For the polishing area point cloud, project it onto the XOY plane and divide it into grids according to the projection boundary and the side length of the grid. The formula is as follows: in, Represents the maximum and minimum boundaries, L g is the grid side length, X num , Y num It is the number of grids in the X-axis and Y-axis directions.

4. The multi-type grinding area detection algorithm for autonomous grinding of castings according to claim 1 is characterized in that: The specific process of the welding scar defect detection in step S4 includes: S4.

1. Normal calculation: Calculate the point cloud normal of the casting surface to be tested; S4.2, Curvature change detection: Since the normal line has mutation in the scar type area, the edge of the weld scar can be detected by detecting the change characteristics of the normal line, that is, the change of the surface curvature of the point in the point cloud; S4.3, weld scar extraction: Based on the statistical results obtained in step S4.2, select a point set whose curvature is greater than a set threshold and whose corresponding point number is less than the set threshold as the weld scar target detection result.

5. The multi-type grinding area detection algorithm for autonomous grinding of castings according to claim 4 is characterized in that: The specific process of normal calculation in step S4.1 includes: The vector perpendicular to the tangent plane at a certain point is defined as the normal at that point. Therefore, solving the normal problem of a point on the surface is equivalent to estimating the normal of a tangent plane of the surface. After transformation, it can be transformed into a plane fitting estimation problem; using the least squares method for estimation, the formula is as follows: in, is the normal vector at the point and is the center of the plane, k represents P i The number of neighboring points; The above formula indicates that: for point P in the local point cloud i , i = 1, 2, ..., k, find a plane that passes through the center of the plane The normal vector satisfy In the normal vector The sum of the projections on is minimal; Normal vector The solution can be simplified to the problem of solving the eigenvalue and eigenvector of principal component analysis. The formula is as follows: where M is a symmetric semi-positive definite covariance matrix, λ j is the jth eigenvalue of the covariance matrix, is the eigenvalue λ j The corresponding eigenvector. i Normal vector of a point is the eigenvector corresponding to the minimum eigenvalue.

6. The multi-type grinding area detection algorithm for autonomous grinding of castings according to claim 4 is characterized in that: The specific process of curvature change detection in step S4.2 includes: The curvature of a point is defined as: in, Indicates P i The curvature of the point, λ0, λ1, λ2 are P i The eigenvalue of the covariance matrix M corresponding to the point, and λ0≤λ1≤λ2; different intervals are divided by counting the curvature of the points in the point cloud.

7. The multi-type grinding area detection algorithm for autonomous grinding of castings according to claim 1 is characterized in that: The specific process of the rough surface defect detection in step S5 includes: S5.1, Iterative surface fitting: Perform quadratic surface fitting on the casting surface, use iterative fitting method to solve the quadratic surface parameters, and obtain the fitting point cloud; S5.

2. Rough part extraction: By comparing the collected point cloud with the fitted point cloud, the area above the comparison threshold is extracted as the rough part to be polished.

8. The multi-type grinding area detection algorithm for autonomous grinding of castings according to claim 7 is characterized in that: The specific process of iterative surface fitting in step S5.1 includes: The fitting equation is as follows: Among them, P i =(X i , Y i , Z i ), i = 1, 2, ..., n represents the point coordinates of the polishing area point cloud, n is the number of polishing area points, Z(X i , Y i ) is the Z-axis fitting coordinate of the i-th point; the quadratic surface fitting parameter is S p = [a, b, c, d, e, f]; The iterative fitting method is used to solve the quadratic surface parameters and obtain the fitting point cloud: First, convert equation (6) into the least squares form and use the singular value decomposition (SVD) method to solve S p ; Then, set the iteration error threshold ε = 2mm and set S p Substitute into formula (1) to calculate Z(X i , Y i ), filter out |Z(X i , Y i )-Z i |≥ε; Finally, recalculate S for the filtered point cloud p , until every point satisfies |Z(X i , Y i )-Z i |<ε, then the corresponding S p As the final quadratic surface fitting parameters.

9. The multi-type grinding area detection algorithm for autonomous grinding of castings according to claim 1, characterized in that: The specific process of the point-by-point polishing process detection in step S6 includes: S6.

1. Parameter initialization: define pixel points and distribution according to Zhang-Suen algorithm; S6.2, Point Cloud Delaunay Triangulation: Generate triangular geometry for a given point cloud, and for a set of points in the neighborhood, obtain a set of triangulations; S6.3, Laplace matrix construction: For the triangles in step S6.2 Delaunay triangulation, establish a graph topology relationship, use the vertices of the triangles as the nodes of the graph, the edges of the triangles as the edges of the graph, and use the Laplace cotangent weights as the weights of the graph to establish the Laplace matrix L k ; S6.4, iterative skeleton contraction: the Laplace skeleton is used to contract the frame so that the volume of the 3D point cloud is reduced to 0, but the overall shape of the point cloud remains unchanged; S6.5, point-by-point key point extraction: Through the iterative shrinkage process of step S6.4, point-by-point key points are obtained as polishing key points.

Images