Bending robot bending coordinate positioning method and system

Abstract

The application discloses a bending robot bending coordinate positioning method. The method comprises the following steps: obtaining three-dimensional coordinates of a deformed workpiece, obtaining denoising point cloud data after denoising, interpolation encryption and filtering processing, and obtaining uniform sampling data set through downsampling; extracting geometric features in the uniform sampling data set, fitting a geometric model and collecting feature parameters, and obtaining effective geometric constraint groups after spatial constraint consistency verification; performing point cloud registration on the effective geometric constraint groups and corresponding features of a theoretical model, and iteratively solving initial pose transformation parameters; constructing a least square target function through matching residuals according to the initial pose transformation parameters, and optimizing to obtain final pose transformation parameters; and performing coordinate transformation on the to-be-positioned point cloud to obtain a final coordinate positioning result. Through coordinate positioning, the application realizes accurate establishment of constraints reflecting the real geometric relationship of the workpiece from the plate point cloud, and realizes high-precision positioning of the deformed workpiece.

Description

Technical Field

[0001] This invention relates to the field of coordinate positioning technology, and in particular to a bending robot coordinate positioning method and system. Background Technology

[0002] Currently, industrial robots play a central role in modern manufacturing, especially in processes such as bending, cutting, welding, and assembly of sheet metal. Their operational precision directly determines the consistency of final product quality and production efficiency. When processing large workpieces like sheet metal, robots should rely on industrial vision technology to accurately identify the workpiece's position and orientation in space to establish the correct processing coordinate system.

[0003] In existing technologies, localization typically relies on a single vision sensor or a fixed template matching method based on an ideal CAD model. This approach assumes the workpiece is rigid and perfectly shaped. However, in actual production environments, sheet metal workpieces often experience varying degrees of deformation, warping, or local bending due to transportation compression, stacking gravity, or heat treatment during processing. This results in a significant deviation between the actual geometry of the workpiece surface and the theoretical design model. This geometric deviation makes traditional localization methods based on ideal models difficult to adapt, causing point cloud matching algorithms to get stuck in local optima or generate incorrect correspondences. This, in turn, leads to positional misalignment or posture inaccuracies between the robot's end effector and the actual sheet metal surface. This can result in minor issues like machining path deviations and bending angle errors, or more serious issues like tool collision damage or finished product scrap. Furthermore, sheet metal point cloud data itself is often accompanied by noise interference and data gaps. Without an effective geometric constraint mechanism, it is difficult to converge to globally consistent pose parameters under deformation interference.

[0004] In summary, existing technologies struggle to accurately establish constraints reflecting the true geometric relationships of workpieces from point clouds containing noise and deformation, making it impossible to achieve high-precision positioning of deformed workpieces. This results in poor positioning robustness and a high rate of scrap during processing. Summary of the Invention

[0005] This invention provides a bending robot bending coordinate positioning method and system to accurately establish constraints reflecting the true geometric relationship of the workpiece from the point cloud of sheet metal containing noise and deformation, thereby achieving high-precision positioning of deformed workpieces.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a bending coordinate positioning method for a bending robot, comprising: Obtain the three-dimensional coordinate set of the deformed workpiece, denoise the three-dimensional coordinate set and distinguish between deformed and undeformed regions, perform interpolation encryption processing on the deformed regions to obtain encrypted deformed regions, and merge the encrypted deformed regions and undeformed regions to obtain the original point cloud dataset. The original point cloud dataset is processed by statistical filtering to obtain a denoised point cloud dataset. For the aforementioned denoised point cloud dataset, the point density is reduced by voxel grid downsampling to obtain a uniformly sampled dataset; Geometric features are extracted from the uniformly sampled dataset. A geometric model is fitted based on the geometric features using the random sampling consensus method. The feature parameters of the geometric model are then summarized to obtain a set of feature parameters. Spatial constraint relationships are established based on the set of feature parameters, and the inter-constraint deviation between the feature parameters is calculated. If the inter-constraint deviation is less than a preset deviation threshold, the set of feature parameters is confirmed to satisfy consistency, and an effective geometric constraint group is obtained. Select the set of feature pairs that correspond to the pre-acquired theoretical model from the effective geometric constraint group, and obtain the initial pose transformation parameters by iteratively solving the rotation and translation matrix according to the set of feature pairs using the point cloud registration method. The point cloud to be located is transformed according to the initial pose transformation parameters, and the final pose transformation parameters are obtained by minimizing the geometric deviation objective function in combination with the theoretical model. The coordinate transformation of the point cloud to be located is then performed according to the final pose transformation parameters to obtain the final coordinate positioning result.

[0007] In one optional implementation, the step of obtaining a set of three-dimensional coordinates of the deformed workpiece, denoising the set of three-dimensional coordinates and distinguishing between deformed and undeformed regions, performing interpolation encryption on the deformed regions to obtain encrypted deformed regions, and merging the encrypted deformed regions and undeformed regions to obtain the original point cloud dataset includes: The laser echo signal acquired by the three-dimensional laser scanning equipment is collected, and the discrete spatial coordinates are analyzed based on the laser echo signal; The reflection intensity value is extracted from the laser echo signal, and the discrete spatial coordinates are associated with the reflection intensity value to obtain a three-dimensional coordinate set. Noise below a preset intensity threshold is removed from the three-dimensional coordinate set to obtain the effective surface coordinates. Calculate the rate of change of the normal vector of the effective surface coordinates, locate the deformation region based on the rate of change of the normal vector, and perform interpolation and densification processing to obtain the densified deformation region; The encrypted deformed region is merged with the undeformed region to obtain the original point cloud dataset.

[0008] In one optional implementation, processing the original point cloud dataset using statistical filtering to obtain a denoised point cloud dataset includes: A spatial neighborhood index structure is constructed based on the original point cloud dataset, and the nearest neighbor set of each point in the original point cloud dataset is retrieved through the spatial neighborhood index structure. The mean distance of the neighborhood is calculated based on the set of nearest points, and the statistical screening threshold is determined by combining the mean global distance and standard deviation of the original point cloud dataset. If the mean neighborhood distance exceeds the statistical filtering threshold, the corresponding query point is removed, and the remaining valid point data after removal is integrated to obtain a denoised point cloud dataset.

[0009] In one optional implementation, the step of reducing point density through voxel grid downsampling to obtain a uniformly sampled dataset for the denoised point cloud dataset includes: Calculate the three-dimensional coordinate extreme values ​​based on the denoised point cloud dataset, and determine the minimum three-dimensional spatial range surrounding the denoised point cloud dataset based on the three-dimensional coordinate extreme values. The three-dimensional space is divided according to the minimum three-dimensional spatial range to obtain multiple non-overlapping three-dimensional voxel units; Establish a spatial mapping index relationship between the three-dimensional voxel unit and each point in the denoised point cloud dataset, and calculate the geometric center coordinates of all points falling within the three-dimensional voxel unit based on the spatial mapping index relationship; By integrating the geometric center coordinates corresponding to all non-empty voxel units, a uniformly sampled dataset is obtained.

[0010] In one optional implementation, the step of extracting geometric features from the uniformly sampled dataset, fitting a geometric model based on the geometric features using the random sampling consensus method, and summarizing the feature parameters of the geometric model to obtain a feature parameter set, includes: A KD tree index structure is constructed based on the uniform sampling dataset. For each sampling point, the set of neighboring points in the K nearest neighbor domain of that sampling point is searched. The covariance matrix is ​​calculated based on the set of neighboring points, and eigenvalue decomposition is performed to obtain the normal vector and curvature value of each sampling point. Based on the curvature value, a curvature threshold is set, and the uniformly sampled dataset is divided into a planar feature point set and a surface feature point set according to the curvature threshold. The planar feature point set and the surface feature point set are then clustered to obtain multiple independent planar region clusters and surface region clusters. Based on each of the planar region clusters and the surface region clusters, the optimal planar fitting model and the optimal surface fitting model are constructed respectively using the random sampling consensus algorithm. Extract and integrate the planar feature parameters of all the optimal planar fitting models and the surface feature parameters of all the optimal surface fitting models to construct a feature parameter set.

[0011] In one optional implementation, the step of establishing spatial constraint relationships based on the set of feature parameters, calculating the inter-constraint deviations between the feature parameters, and confirming that the set of feature parameters satisfies consistency if the inter-constraint deviations are less than a preset deviation threshold, thereby obtaining a valid geometric constraint set, includes: Based on the set of feature parameters, calculate the Euclidean distance between geometric feature objects, determine the topological adjacency relationship based on the Euclidean distance, and extract the normal vector and center point coordinates; Calculate the angle constraint value and distance constraint value corresponding to the normal vector and the coordinates of the center point to obtain the initial constraint set; Obtain a preset standard geometric constraint template library, calculate the difference between the initial constraint set and the theoretical parameter values ​​in the standard geometric constraint template library, and obtain the constraint deviation value; If the deviation value between the constraints is less than the preset consistency judgment threshold, then the initial constraint set is confirmed to meet the consistency requirements, and a valid geometric constraint group is obtained.

[0012] In one optional implementation, the step of filtering the set of feature pairs corresponding to the pre-acquired theoretical model from the effective geometric constraint group, and obtaining the initial pose transformation parameters by iteratively solving the rotation and translation matrix using the point cloud registration method based on the set of feature pairs, includes: Obtain spatial constraint relationship data between the theoretical model and the effective geometric constraint group; retrieve the feature set of the theoretical model based on the spatial constraint relationship data; and construct a feature point pair set. The set of feature point pairs is processed using the singular value decomposition algorithm to obtain the covariance matrix. The covariance matrix is ​​then decomposed to obtain rotation matrix components and translation vector components. Construct a rigid body transformation matrix based on the rotation matrix components and the translation vector components, and calculate the residual value; If the residual value is greater than the preset residual threshold, the rigid body transformation matrix is ​​corrected using the iterative nearest point algorithm to obtain the initial pose transformation parameters.

[0013] In one optional implementation, the step of transforming the point cloud to be located according to the initial pose transformation parameters, and obtaining the final pose transformation parameters by minimizing the geometric deviation objective function in conjunction with the theoretical model, and then performing coordinate transformation on the point cloud to be located according to the final pose transformation parameters to obtain the final coordinate positioning result, includes: The coordinates of the point cloud to be located are transformed according to the initial pose transformation parameters, and the transformed point cloud data is matched with the theoretical model to obtain the initial residual set. Based on the initial residual set, a weighted least squares objective function is constructed and iteratively solved to obtain the updated pose parameters and the current residual value. If the current residual value is greater than the preset residual threshold, the weights are adjusted using the Huber kernel function and the weighted least squares objective function is re-solved until the current residual value meets the convergence condition, and the final pose transformation parameters are obtained. The coordinates of the point cloud to be located are transformed using the final pose transformation parameters, and the final coordinate positioning result is calculated.

[0014] In one optional implementation, the step of calculating the rate of change of the normal vector of the effective surface coordinates, locating the deformation region based on the rate of change of the normal vector, and performing interpolation and densification processing to obtain the densified deformation region includes: Calculate the rate of change of the angle between the normal vectors of adjacent sampling points of the effective surface coordinates. If the rate of change exceeds a preset angle threshold, then mark the region as a deformed region. An interpolation algorithm is used to reduce the point spacing within the deformation region from the original spacing to a preset encrypted spacing, thus obtaining an encrypted deformation region.

[0015] In a second aspect, the present invention provides a bending coordinate positioning system for a bending robot, comprising: The data acquisition module is used to acquire a set of three-dimensional coordinates of the deformed workpiece, denoise the set of three-dimensional coordinates and distinguish between deformed and undeformed regions, perform interpolation encryption processing on the deformed regions to obtain encrypted deformed regions, and merge the encrypted deformed regions and undeformed regions to obtain the original point cloud dataset. The data preprocessing module is used to process the original point cloud dataset using statistical filtering to obtain a denoised point cloud dataset. The sampling processing module is used to reduce the point density of the denoised point cloud dataset by voxel grid downsampling to obtain a uniformly sampled dataset. The feature extraction module is used to extract geometric features from the uniformly sampled dataset, fit a geometric model based on the geometric features using the random sampling consensus method, and summarize the feature parameters of the geometric model to obtain a feature parameter set. The constraint establishment and verification module is used to establish spatial constraint relationships based on the set of feature parameters, calculate the constraint deviation between the feature parameters, and if the constraint deviation is less than a preset deviation threshold, then the set of feature parameters is confirmed to meet the consistency requirement, and a valid geometric constraint group is obtained. The initial pose estimation module is used to filter the set of feature pairs that correspond to the pre-acquired theoretical model in the effective geometric constraint group, and to obtain the initial pose transformation parameters by iteratively solving the rotation and translation matrix according to the set of feature pairs through the point cloud registration method. The pose optimization and localization module is used to transform the point cloud to be located using the initial pose transformation parameters, and combined with the theoretical model, to obtain the final pose transformation parameters by minimizing the geometric deviation objective function. Based on the final pose transformation parameters, the point cloud to be located is subjected to coordinate transformation to obtain the final coordinate localization result.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention obtains the three-dimensional coordinate set of the deformed workpiece and performs interpolation and densification processing on the deformed area, and then combines statistical filtering to remove outlier noise. This process not only effectively preserves the key geometric details of the deformed area, but also eliminates environmental noise interference during the scanning process, thereby constructing a high-quality original point cloud dataset. This solves the problem of inaccurate feature extraction caused by sparse point clouds or noise in the prior art, and lays a solid data foundation for subsequent high-precision positioning.

[0017] (2) This invention extracts geometric features from a uniformly sampled dataset, establishes spatial constraint relationships for the feature parameter set based on these geometric features, and verifies geometric consistency by calculating the deviation between constraints. This method can screen out effective geometric constraint groups with stable topological relationships and no severe deformation damage from complex deformation data, and use these rigid features for initial registration, avoiding the algorithm from getting trapped in local optima due to local deformation, and significantly improving the reliability and accuracy of initial pose estimation.

[0018] (3) This invention optimizes the pose by constructing a least-squares objective function and introduces the Huber kernel function to dynamically adjust the weights when the residual is too large. This mechanism can adaptively reduce the influence of the deformed region on the overall positioning, thereby increasing the weight of the rigid region. This enables the accurate calculation of the final pose transformation parameters of the workpiece body even when there is non-rigid deformation, thus solving the problem of insufficient robot coordinate positioning accuracy under complex working conditions. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating an embodiment of a bending robot coordinate positioning method provided by the present invention; Figure 2 This is a flowchart illustrating an embodiment of obtaining initial pose transformation parameters provided by the present invention.

[0020] Figure 3 This is a schematic diagram of an embodiment of a bending robot bending coordinate positioning system provided by the present invention. Detailed Implementation The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 The first embodiment of the present invention provides a bending coordinate positioning method for a bending robot, comprising the following steps: Step S11: Obtain the three-dimensional coordinate set of the deformed workpiece, denoise the three-dimensional coordinate set and distinguish between deformed and undeformed regions, perform interpolation encryption processing on the deformed regions to obtain encrypted deformed regions, and merge the encrypted deformed regions and undeformed regions to obtain the original point cloud dataset. Step S12: Process the original point cloud dataset using statistical filtering to obtain a denoised point cloud dataset; Step S13: For the denoised point cloud dataset, the point density is reduced by voxel grid downsampling to obtain a uniformly sampled dataset; Step S14: Extract geometric features from the uniformly sampled dataset, fit a geometric model based on the geometric features using the random sampling consensus method, and summarize the feature parameters of the geometric model to obtain a feature parameter set; Step S15: Establish spatial constraint relationships based on the set of feature parameters, calculate the constraint deviation between the feature parameters, and if the constraint deviation is less than a preset deviation threshold, confirm that the set of feature parameters satisfies consistency and obtain an effective geometric constraint group. Step S16: Select the set of feature pairs in the effective geometric constraint group that correspond to the pre-acquired theoretical model. Based on the set of feature pairs, iteratively solve the rotation and translation matrix using the point cloud registration method to obtain the initial pose transformation parameters. Step S17: Transform the point cloud to be located using the initial pose transformation parameters, and combine the theoretical model to obtain the final pose transformation parameters by minimizing the geometric deviation objective function. Perform coordinate transformation on the point cloud to be located based on the final pose transformation parameters to obtain the final coordinate positioning result.

[0022] In step S11, the process involves obtaining a set of three-dimensional coordinates of the deformed workpiece, denoising the set of three-dimensional coordinates and distinguishing between deformed and undeformed regions, performing interpolation encryption on the deformed regions to obtain encrypted deformed regions, and merging the encrypted deformed regions and undeformed regions to obtain the original point cloud dataset, including: The laser echo signal acquired by the three-dimensional laser scanning equipment is collected, and the discrete spatial coordinates are analyzed based on the laser echo signal; The reflection intensity value is extracted from the laser echo signal, and the discrete spatial coordinates are associated with the reflection intensity value to obtain a three-dimensional coordinate set. Noise below a preset intensity threshold is removed from the three-dimensional coordinate set to obtain the effective surface coordinates. Calculate the rate of change of the normal vector of the effective surface coordinates, locate the deformation region based on the rate of change of the normal vector, and perform interpolation and densification processing to obtain the densified deformation region; The encrypted deformed region is merged with the undeformed region to obtain the original point cloud dataset.

[0023] First, laser echo signals acquired by a 3D laser scanning device are collected, and discrete spatial coordinates are derived from these signals. Specifically, a 3D laser scanning device, such as a line laser scanner or structured light camera, emits a laser beam onto the workpiece surface. The laser beam reflects back to the sensor after hitting the surface. An internal timer records the time difference between the laser pulse's emission and reception. Using triangulation principles and the positional deviation of the reflected light on the sensor, combined with the speed of light or geometric relationships, the spatial position information of each reflection point—that is, the discrete spatial coordinates—is calculated. Specifically, the 3D laser scanning device emits a laser beam onto the workpiece surface, and a timer records the round-trip time difference of the laser pulse. The distance is calculated as the speed of light multiplied by half the time difference. Combined with the deflection angle of the laser emission system, the spatial position information of each reflection point is calculated using the formula for converting spherical coordinates to rectangular coordinates.

[0024] Subsequently, the discrete spatial coordinates are correlated with the reflection intensity values ​​to obtain a three-dimensional coordinate set. Low-intensity noise is then removed from this set to obtain the effective surface coordinates. Specifically, the laser scanning device records the intensity value of the returned signal while recording the coordinates, binding each spatial coordinate to its corresponding intensity value to form a four-dimensional data vector. Next, an intensity filtering threshold is set. This threshold is obtained through pre-calibration testing of the reflectivity characteristics of the workpiece material, such as stainless steel or aluminum alloy, under standard lighting conditions. Typically, the 95th percentile of the background noise intensity distribution is selected as the intensity filtering threshold. All data points are iterated through, and points with intensity values ​​below the intensity filtering threshold are considered as dust scattering or edge artifact noise in the air and are removed. The coordinates of the remaining high-intensity points are the effective surface coordinates.

[0025] Next, the rate of change of the normal vector of the effective surface coordinates is calculated. Based on the rate of change of the normal vector, the deformation region is located and interpolated for refinement, resulting in a refined deformation region, including: Calculate the rate of change of the angle between the normal vectors of adjacent sampling points of the effective surface coordinates. If the rate of change exceeds a preset angle threshold, then mark the region as a deformed region. An interpolation algorithm is used to reduce the point spacing within the deformation region from the original spacing to a preset encrypted spacing, thus obtaining an encrypted deformation region.

[0026] First, the rate of change of the angle between the normal vectors of adjacent sampling points on the effective surface is calculated. If the rate of change exceeds a preset angle threshold, the region is marked as a deformed region. Specifically, for each effective coordinate point, all points within its 3mm radius are searched. Principal component analysis is used to calculate the covariance matrix of this neighborhood. Eigenvalue decomposition is performed on the covariance matrix, and the eigenvector corresponding to the smallest eigenvalue is the normal vector of that point. Then, the dot product of the normal vector of this point and the normal vectors of its adjacent points is calculated, and the angle is obtained using the inverse cosine function. If the rate of change of the angle between the normal vectors of adjacent points exceeds a preset angle threshold, the region is marked as a deformed region. The preset angle threshold is determined by scanning a standard undeformed sheet material, statistically analyzing the natural fluctuation range of the normal vector caused by its surface micro-irregularities, and taking 1.5 times the maximum value of this fluctuation range as the judgment threshold.

[0027] It should be noted that the search radius of 3cm mentioned above is adaptively determined based on the average point spacing of the effective surface coordinates. Specifically, the Euclidean distance between all adjacent sampling points is first calculated and the average value is taken as the average point spacing. Then, the search radius is set to n times the average point spacing, where n is a proportionality coefficient, and the typical value space is [3,10], to ensure that the neighborhood contains a sufficient number of points for stable calculation.

[0028] Subsequently, an interpolation algorithm is used to reduce the point spacing within the deformed region from the original spacing to a preset encrypted spacing, resulting in an encrypted deformed region. Specifically, for the point cloud marked as the deformed region, a cubic polynomial function is constructed between every two adjacent original data points within the deformed region using cubic spline interpolation. The polynomial coefficients are then solved using boundary conditions, ensuring that the curve not only passes through all known sample points but also maintains continuity in both the first derivative (slope) and the second derivative (curvature) at the connection points. New virtual points are generated between the original points using cubic spline interpolation, reducing the point spacing within the deformed region from the original spacing to the preset encrypted spacing, thereby obtaining a high-density encrypted deformed region. The preset encrypted spacing is set based on the sampling theorem of geometric signals and the minimum geometric feature size of the deformed region. First, the minimum radius of curvature of the deformed region is analyzed. The encrypted spacing is typically set to one-tenth of the minimum radius of curvature. For example, if the bending radius is 2 mm, the encrypted spacing is set to 0.2 mm.

[0029] Finally, the encrypted deformed regions are merged with the undeformed regions to obtain the original point cloud dataset. Specifically, the original sparse point cloud that was not marked as a deformed region is retained, and it is combined with the high-density deformed region point cloud that has undergone interpolation and encryption processing in the same coordinate system to obtain the original point cloud dataset.

[0030] In step S12, processing the original point cloud dataset using statistical filtering to obtain a denoised point cloud dataset includes: A spatial neighborhood index structure is constructed based on the original point cloud dataset, and the nearest neighbor set of each point in the original point cloud dataset is retrieved through the spatial neighborhood index structure. The mean distance of the neighborhood is calculated based on the set of nearest points, and the statistical screening threshold is determined by combining the mean global distance and standard deviation of the original point cloud dataset. If the mean neighborhood distance exceeds the statistical filtering threshold, the corresponding query point is removed, and the remaining valid point data after removal is integrated to obtain a denoised point cloud dataset.

[0031] First, a spatial neighborhood index structure is constructed based on the original point cloud dataset. This index structure is then used to retrieve the nearest neighbor set of each point in the original point cloud dataset. Specifically, to improve the retrieval efficiency of massive point cloud data, a KD-tree algorithm is used to construct the spatial index, recursively partitioning the three-dimensional space until each leaf node contains fewer than 150 points. For each query point in the dataset, this index structure is used to quickly search for its 50 nearest neighbors in Euclidean space. These 50 points constitute the nearest neighbor set of the query point. The number m of the nearest neighbor set is adaptively set based on the point cloud density, typically within the range [20, 100]. One feasible approach is to use m = max(20, min(100, ρπR). 2 ), where ρ is the estimated local surface density of the point cloud and R is the neighborhood radius.

[0032] Subsequently, the mean neighborhood distance is calculated based on the nearest neighbor set, and a statistical filtering threshold is determined by combining the global mean and standard deviation of the original point cloud dataset. Specifically, for each point, the Euclidean distance to all points in the nearest neighbor set is calculated, and the arithmetic mean of these distances is taken as the mean neighborhood distance of that point. After traversing all points, the distribution of the mean neighborhood distance of the entire point cloud dataset is statistically obtained, and the global mean and standard deviation of this distribution are calculated. The statistical filtering threshold is the global mean plus three times the standard deviation.

[0033] Finally, if the mean neighborhood distance exceeds the statistical filtering threshold, the corresponding query point is removed, and the remaining valid point data after removal are integrated to obtain a denoised point cloud dataset. Specifically, the mean neighborhood distance of each point is compared with the statistical filtering threshold. If the mean neighborhood distance of a point is greater than the statistical filtering threshold, it means that the distance between the point and its neighboring points is significantly greater than the average level, belonging to outlier noise points far from the main body, such as dust points suspended in the air or measurement flypoints, and the point is removed from the dataset. All remaining points that have not been removed are reorganized to obtain the denoised point cloud dataset.

[0034] For example, the input raw point cloud dataset contains 800,000 points. After constructing a KD-tree, the system sets the number of nearest neighbors to be searched, k=30. The calculated global average distance is 0.15mm, and the standard deviation is 0.05mm, so the preset distance threshold is 0.15 + 3 × 0.05 = 0.3mm. During processing, it is found that the average neighborhood distance of a certain point is 0.45mm, which is greater than 0.3mm. The system determines this point as an outlier and removes it. After traversal processing, approximately 15,000 noisy points are removed, resulting in a final output denoised point cloud dataset containing 785,000 points.

[0035] In step S13, the step of reducing the point density by voxel grid downsampling to obtain a uniformly sampled dataset for the denoised point cloud dataset includes: Calculate the three-dimensional coordinate extreme values ​​based on the denoised point cloud dataset, and determine the minimum three-dimensional spatial range surrounding the denoised point cloud dataset based on the three-dimensional coordinate extreme values. The three-dimensional space is divided according to the minimum three-dimensional spatial range to obtain multiple non-overlapping three-dimensional voxel units; Establish a spatial mapping index relationship between the three-dimensional voxel unit and each point in the denoised point cloud dataset, and calculate the geometric center coordinates of all points falling within the three-dimensional voxel unit based on the spatial mapping index relationship; By integrating the geometric center coordinates corresponding to all non-empty voxel units, a uniformly sampled dataset is obtained.

[0036] First, the extreme values ​​of the three-dimensional coordinates are calculated based on the denoised point cloud dataset. These extreme values ​​are then used to determine the minimum three-dimensional spatial range that encloses the denoised point cloud dataset. Specifically, the x, y, and z coordinates of all points in the denoised point cloud dataset are traversed, and the maximum and minimum values ​​along the x, y, and z axes are found respectively. These six extreme values ​​define a minimum bounding box parallel to the coordinate axes; the spatial range of this bounding box is the minimum three-dimensional spatial range that can completely contain the current point cloud.

[0037] Subsequently, the three-dimensional space is divided according to the minimum three-dimensional spatial range, resulting in multiple non-overlapping three-dimensional voxel units. Specifically, the size parameters of the voxel grid are set, which are determined based on the minimum feature size of the workpiece, typically set to 1 / 10 to 1 / 5 of the minimum feature size to ensure that key geometric features are not lost after sampling. Based on the length, width, height of the bounding box and the voxel size, the number of grids required in the x, y, and z directions is calculated, thereby cutting the entire minimum three-dimensional spatial range into a series of small cubic units, i.e., three-dimensional voxel units.

[0038] Next, a spatial mapping index relationship is established between the three-dimensional voxel unit and each point in the denoised point cloud dataset, and the geometric center coordinates of all points falling within the three-dimensional voxel unit are calculated based on the spatial mapping index relationship. Specifically, for any point in the point cloud, its voxel index is calculated using integer division. A hash table is used to store the mapping relationship between voxel indices and the point set list, and all points are assigned to the corresponding voxels. For each non-empty voxel unit containing point cloud data, the geometric center coordinates of all points within it are calculated. The method for calculating the geometric center coordinates is to sum the x, y, and z coordinates of all points within the voxel and then divide by the total number of points within that voxel.

[0039] Finally, the geometric center coordinates corresponding to all non-empty voxel units are integrated to obtain a uniformly sampled dataset. Specifically, the calculated geometric center coordinates are used to replace all the original points within the voxel, and the centroids of all non-empty voxels are collected to obtain a uniformly sampled dataset.

[0040] For example, the bounding box range of the denoised point cloud is: X[-500,500], Y[-300,300], Z[0,50] (unit: mm). The voxel size is set to L=5mm. The system divides the space into 200×120×10 grids. In the voxel cell with index (10,5,2), 150 points originally fell into the grid, and the geometric center coordinates of these 150 points were calculated to be (52.3,27.8,12.1). The system retains only this one centroid. After full-field processing, the original 785,000 points are reduced to approximately 45,000 uniformly distributed sampling points.

[0041] In step S14, geometric features are extracted from the uniformly sampled dataset, a geometric model is fitted based on the geometric features using the random sampling consensus method, and the feature parameters of the geometric model are summarized to obtain a feature parameter set, including: A KD tree index structure is constructed based on the uniform sampling dataset. For each sampling point, the set of neighboring points in the K nearest neighbor domain of that sampling point is searched. The covariance matrix is ​​calculated based on the set of neighboring points, and eigenvalue decomposition is performed to obtain the normal vector and curvature value of each sampling point. Based on the curvature value, a curvature threshold is set, and the uniformly sampled dataset is divided into a planar feature point set and a surface feature point set according to the curvature threshold. The planar feature point set and the surface feature point set are then clustered to obtain multiple independent planar region clusters and surface region clusters. Based on each of the planar region clusters and the surface region clusters, the optimal planar fitting model and the optimal surface fitting model are constructed respectively using the random sampling consensus algorithm. Extract and integrate the planar feature parameters of all the optimal planar fitting models and the surface feature parameters of all the optimal surface fitting models to construct a feature parameter set.

[0042] First, a KD-tree index structure is constructed based on the uniformly sampled dataset. For each sampled point, a set of neighboring points within its K nearest neighbor region is searched. The covariance matrix is ​​calculated based on this set of neighboring points, and eigenvalue decomposition is performed to obtain the normal vector and curvature value of each sampled point. Specifically, a KD-tree is constructed again for the uniformly sampled point cloud, and 50 nearest neighbors are searched for each point. The covariance matrix of these 50 points is calculated, and eigenvalue decomposition is performed on the covariance matrix to obtain three non-negative eigenvalues ​​and their corresponding eigenvectors. The eigenvector corresponding to the smallest eigenvalue is the estimated normal vector of that point. The curvature value is calculated by the proportion of the smallest eigenvalue in the sum of the eigenvalues; that is, the curvature value is equal to the ratio of the smallest eigenvalue to the sum of the eigenvalues.

[0043] Subsequently, a curvature threshold is set based on the curvature values, and the uniformly sampled dataset is divided into a planar feature point set and a surface feature point set. The planar feature point set and the surface feature point set are then clustered to obtain multiple independent planar region clusters and surface region clusters. Specifically, the preset curvature threshold is determined by statistically analyzing the curvature noise level of a standard planar point cloud. High-precision point cloud data of a standard planar calibration board (the theoretical curvature of which is 0) is collected using the same 3D scanning device. Then, the curvature values ​​of all points in the standard planar point cloud are calculated using the aforementioned eigenvalue decomposition method. Due to sensor noise and algorithm errors, these curvature values ​​are not actually zero, but rather exhibit a certain probability distribution, typically a skewed distribution close to zero. The mean and standard deviation of this distribution are calculated, and the mean plus three times the standard deviation is taken as the preset curvature threshold. If the curvature value of a point is less than the preset curvature threshold, it is determined to be a planar point; otherwise, it is determined to be a surface point. Using the Euclidean distance growing clustering algorithm, spatially continuous planar points and surface points are classified separately to form several independent point cloud clusters. Specifically, a label array is established to record the processing status of all points, initially set to unprocessed. An unprocessed point is selected from a certain type of point set (such as a planar point set) as a seed point, added to the current cluster, and pushed into a queue. All neighboring points of the seed point within a voxel sampling size are searched. If neighboring points belong to the same geometric category (i.e., both are planar or surface points) and are unprocessed, they are added to the current cluster and pushed into a queue. The above process is recursively executed until the queue is empty, completing the extraction of an independent point cloud cluster. The above steps are repeated until all points have been processed, thereby segmenting the discrete point cloud into multiple independent geometric region clusters.

[0044] Next, based on each planar region cluster and the curved surface region cluster, the optimal planar fitting model and the optimal curved surface fitting model are constructed respectively using the random sampling consensus algorithm. Specifically, for a planar region cluster, the random sampling consensus algorithm iteratively executes the following steps: randomly selects 3 points to calculate the plane equation parameters; calculates the distance from all points in the cluster to the plane, and counts the number of interior points whose distance is less than a distance threshold; repeats the above process multiple times, and selects the plane equation with the most interior points as the optimal model. For the curved surface region cluster, points are randomly selected to fit the quadratic surface equation, and the optimal curved surface fitting model is also determined by maximizing the number of interior points. In one feasible approach, for common geometric features of bent workpieces, a cylindrical surface model is preferentially used for surface fitting. Specifically, randomly selecting 3 non-collinear points can uniquely determine the initial parameters (axial direction and radius) of a cylindrical surface. Calculate the Euclidean distance from all points within the cluster to the assumed cylindrical surface, and count the number of interior points whose distances are less than a set threshold. If the interior point rate of the cylindrical surface fitting for a certain surface region cluster is consistently lower than an acceptable level (e.g., 60%), then attempt to fit other quadratic surface models (e.g., conical surfaces). Through multiple random samplings, select the model with the highest interior point rate as the optimal cylindrical surface fitting model.

[0045] Finally, the planar feature parameters of all the optimal planar fitting models and the surface feature parameters of all the optimal surface fitting models are integrated to construct a feature parameter set. Specifically, the planar feature parameters include the normal vector and the distance to the origin. By calculating the geometric centroid of all points in the cluster, a covariance matrix about the points is constructed and eigenvalue decomposition is performed. The eigenvector corresponding to the smallest eigenvalue is selected as the normal vector. The absolute value of the projection length of the geometric centroid in the direction of the normal vector is calculated as the distance to the origin. The surface feature parameters include the surface center coordinates, principal axis direction, and radius of curvature. By constructing a nonlinear objective function of the sum of squared distances from points to spatial straight lines, the Gauss-Newton method is used to iteratively solve the problem. The spatial straight line that minimizes the distance error is obtained as the axis, and the direction vector of this axis is the principal axis direction. The point on the axis closest to the centroid of the points in the cluster is determined as the surface center coordinates. The arithmetic mean of the Euclidean distances from all points in the cluster to this axis is calculated to obtain the radius of curvature. The planar feature parameters and the surface feature parameters are summarized to obtain the feature parameter set.

[0046] For example, in the processed sampling point set, the system calculates a curvature threshold of 0.01 by pre-scanning the curvature noise distribution of a standard flat plate. The average curvature value of a certain region is calculated to be 0.003, which is less than the threshold of 0.01, and it is marked as a planar point. Another region has an average curvature value of 0.15 and is marked as a curved point. Through clustering, three large planar clusters and two cylindrical curved surface clusters are separated. For one of the planar clusters, the random sampling consensus algorithm fits the plane equation as 0.02x + 0.99y - 0.01z + 120 = 0, achieving an interior point ratio of 98%. For the curved surface cluster at the bend, cylindrical surface parameters with a radius of 15mm are fitted. The final feature parameter set contains the mathematical descriptions of these five geometric primitives.

[0047] In step S15, establishing spatial constraint relationships based on the set of feature parameters, calculating the constraint deviations between the feature parameters, and confirming that the set of feature parameters satisfies consistency if the constraint deviations are less than a preset deviation threshold, yields a valid geometric constraint group, including: Based on the set of feature parameters, calculate the Euclidean distance between geometric feature objects, determine the topological adjacency relationship based on the Euclidean distance, and extract the normal vector and center point coordinates; Calculate the angle constraint value and distance constraint value corresponding to the normal vector and the coordinates of the center point to obtain the initial constraint set; Obtain a preset standard geometric constraint template library, calculate the difference between the initial constraint set and the theoretical parameter values ​​in the standard geometric constraint template library, and obtain the constraint deviation value; If the deviation value between the constraints is less than the preset consistency judgment threshold, then the initial constraint set is confirmed to meet the consistency requirements, and a valid geometric constraint group is obtained.

[0048] First, based on the set of feature parameters, the Euclidean distance between geometric feature objects is calculated to determine the topological adjacency relationship and extract the normal vectors and center point coordinates. Specifically, the Euclidean distance between each pair of geometric primitives (planes or surfaces) in the feature parameters is calculated using their centroids or geometric centers. If the minimum distance between two geometric features is less than a threshold, they are considered to be physically adjacent and have a topological connection. The normal vectors and center point coordinates of these adjacent features are then extracted. The threshold is set to 5 times the average point spacing of the uniformly sampled dataset. For example, if the average point spacing is 1 mm, the threshold is set to 5 mm.

[0049] Subsequently, a spatial constraint relationship model is constructed based on the topological adjacency relationship. The angle constraint value and distance constraint value corresponding to the normal vector and center point coordinates are calculated to obtain an initial constraint set. Specifically, for each pair of adjacent features, the angle between their normal vectors is calculated as the angle constraint value; the length of the line connecting their center points or the distance between planes is calculated as the distance constraint value. These geometric values ​​of all adjacent pairs are combined to form an initial constraint set describing the overall topological shape of the workpiece.

[0050] Next, a preset standard geometric constraint template library is obtained, and the difference between the initial constraint set and the theoretical parameter values ​​in the standard geometric constraint template library is calculated to obtain the constraint deviation value. Specifically, the standard geometric constraint template library is generated based on the CAD design model of the workpiece and contains the theoretical standard angles and standard distances between each geometric surface. The system compares the actual constraint values ​​extracted from the point cloud with the corresponding theoretical values ​​in the template library, and calculates the absolute value of the difference between the standard angles and the standard distances, which is the constraint deviation value.

[0051] Finally, if the deviation values ​​between constraints are less than a preset consistency judgment threshold, the initial constraint set is confirmed to meet the consistency requirements, and a valid geometric constraint group is obtained. Specifically, the preset consistency judgment threshold is divided into an angle deviation threshold and a distance deviation threshold. These thresholds are set according to the manufacturing tolerance level of the workpiece. For the angle deviation threshold, it is set as the theoretically permissible elastic deformation angle plus three times the angle measurement uncertainty of the scanning equipment. For example, if the process-permissible elastic rebound is within 1 degree and the scanner error is 0.1 degree, the angle threshold can be set to 1.3 degrees. For the distance deviation threshold, it is set as the theoretical geometric tolerance plus three times the linear measurement uncertainty of the scanning equipment. For example, if the workpiece feature spacing tolerance is ±0.5 mm and the scanner linear error is 0.1 mm, the distance threshold can be set to 0.8 mm. If the deviation values ​​between constraints, including the distance deviation value and the angle deviation value, are all within the allowable range, it indicates that the extracted features are structurally consistent with the theoretical model, and no serious structural damage or misidentification has occurred. At this time, these verified feature pairs are confirmed as a valid geometric constraint group. If the deviation exceeds the allowable range, it indicates that the region may have experienced severe non-rigid deformation, plastic distortion, or feature extraction errors. In this case, the system marks the feature pair as invalid or deformed and discards it, retaining only feature pairs that meet the consistency requirements.

[0052] like Figure 2 As shown, in step S16, the step of filtering the feature pair set corresponding to the pre-acquired theoretical model in the effective geometric constraint group, and obtaining the initial pose transformation parameters by iteratively solving the rotation and translation matrix using the point cloud registration method based on the feature pair set, includes: Step S161: Obtain spatial constraint relationship data between the theoretical model and the effective geometric constraint group; retrieve the feature set of the theoretical model based on the spatial constraint relationship data; and construct a feature point pair set. Step S162: The set of feature point pairs is processed using the singular value decomposition algorithm to obtain the covariance matrix. The covariance matrix is ​​then decomposed to obtain rotation matrix components and translation vector components. Step S163: Construct a rigid body transformation matrix based on the rotation matrix components and the translation vector components, and calculate the residual value; Step S164: If the residual value is greater than the preset residual threshold, the rigid body transformation matrix is ​​corrected using the iterative nearest point algorithm to obtain the initial pose transformation parameters.

[0053] First, spatial constraint relationship data of the effective geometric constraint group is obtained. Based on the spatial constraint relationship data, the feature set of the theoretical model is retrieved, and a feature point pair set is constructed. Specifically, using the geometric features in the effective geometric constraint group, such as the center point of a plane and the midpoint of the axis of a curved surface, the corresponding feature points are found in the point cloud of the theoretical CAD model. These one-to-one corresponding points are combined into a matching point pair set to obtain the feature point pair set.

[0054] It should be noted that the feature set of the theoretical CAD model needs to be extracted offline and parameterized from the CAD design file of the workpiece in advance. Specifically, this involves extracting all geometric patches such as planes and cylinders in the CAD model, calculating their respective parameterized representations (such as plane equations, cylinder axis equations and radii), and calculating the theoretical geometric center coordinates of each patch.

[0055] Subsequently, the feature point pair set is processed using a singular value decomposition algorithm to obtain a covariance matrix. This covariance matrix is ​​then decomposed to obtain rotation matrix components and translation vector components. Specifically, the centroids of the two point cloud sets are first calculated, and all point coordinates are subtracted from the centroid coordinates to achieve centering. The cross-covariance matrix of the centered point pairs is then calculated. Singular value decomposition is performed on the covariance matrix to obtain a left singular vector matrix and a right singular vector matrix. The rotation matrix is ​​then obtained by multiplying the left singular matrix and the transpose of the right singular matrix, and the translation vector is obtained by subtracting the rotated measured point cloud centroid from the theoretical point cloud centroid.

[0056] Next, a rigid body transformation matrix is ​​constructed based on the rotation matrix components and the translation vector components, and the residual value is calculated. Specifically, a 4x4 homogeneous transformation matrix is ​​constructed, with the following structure: the upper left 3x3 submatrix is ​​the rotation matrix, the upper right 3x1 submatrix is ​​the translation vector, the lower left 1x3 submatrix is ​​the zero vector [0,0,0], and the lower right element is 1. This matrix is ​​used to transform the measured feature points to the theoretical coordinate system, that is, to expand the feature points to homogeneous coordinates (x,y,z,1). T Then, multiply the homogeneous transformation matrix by left to calculate the average Euclidean distance between the transformed point and the corresponding theoretical point, which is used as the residual value.

[0057] Finally, if the residual value is greater than a preset residual threshold, the rigid body transformation matrix is ​​corrected using the iterative nearest point algorithm to obtain the initial pose transformation parameters. Specifically, the preset residual threshold is set to 2-3 times the nominal accuracy of the scanning device. For example, if the accuracy of the scanning device is 0.05mm, the residual threshold can be set to 1-1.5mm to determine whether the coarse registration accuracy is sufficient. If not, the ICP algorithm is started: based on the current pose, the nearest point in the theoretical model is found for each point in the point cloud to be located. The rotation and translation matrices are recalculated based on these new correspondences. The process of finding corresponding points, calculating transformations, and applying transformations is repeated until the residual converges or the maximum number of iterations is reached. The final transformation matrix is ​​the initial pose transformation parameter.

[0058] In step S17, the transformation of the point cloud to be located based on the initial pose transformation parameters, combined with the theoretical model, and the solution to obtain the final pose transformation parameters by minimizing the geometric deviation objective function, followed by coordinate transformation of the point cloud to be located based on the final pose transformation parameters to obtain the final coordinate positioning result, includes: The coordinate transformation of the point cloud to be located is performed according to the initial pose transformation parameters, and the transformed point cloud data is matched with the theoretical model to obtain the initial residual set. Based on the initial residual set, a weighted least squares objective function is constructed and iteratively solved to obtain the updated pose parameters and the current residual value. If the current residual value is greater than the preset residual threshold, the weights are adjusted using the Huber kernel function and the weighted least squares objective function is re-solved until the current residual value meets the convergence condition, and the final pose transformation parameters are obtained. The coordinates of the point cloud to be located are transformed using the final pose transformation parameters, and the final coordinate positioning result is calculated.

[0059] First, the point cloud to be located is transformed according to the initial pose transformation parameters. The transformed point cloud data is then matched with the theoretical model to obtain an initial residual set. Specifically, the initial pose matrix is ​​applied to transform all measured point clouds to the vicinity of the theoretical coordinate system. For each transformed measured point, a KD-tree is used to search for the nearest neighbor in the theoretical model point cloud, and the distance vector between the two is calculated. The distance values ​​of all points constitute the initial residual set.

[0060] Subsequently, a weighted least squares objective function is constructed based on the initial residual set and solved iteratively to obtain the updated pose parameters and the current residual value. Specifically, the objective function is constructed as the sum of the squares of the weighted Euclidean distances between all corresponding point pairs. The nonlinear least squares problem is linearized and iteratively solved using the Gauss-Newton method, including performing a first-order Taylor expansion of the nonlinear residual function at the current pose; and constructing the normal equations. ,in This is a diagonal matrix composed of weights, where the diagonal elements are weight parameters, and the initial value of the weight parameters is 1. For the current residual vector, Given the Jacobian matrix of the residuals relative to the pose parameters, solve this system of linear equations to obtain the pose parameter increments that minimize the objective function. The current pose parameters are updated using the displacement parameters, and the updated current residual value is calculated.

[0061] Subsequently, if the current residual value is greater than a preset residual threshold, the weights are adjusted using the Huber kernel function, and the weighted least squares objective function is re-solved until the current residual value meets the convergence condition, thus obtaining the final pose transformation parameters. Specifically, to resist the influence of non-rigid deformation regions in deformed workpieces on overall rigid positioning, the Huber kernel function is introduced as a robust estimator to dynamically adjust the weights. It should be noted that the determination process of the threshold parameter of the Huber kernel function is an adaptive parameter learning process, which specifically includes: collecting a large amount of representative point cloud data of deformed sheet metal workpieces as a sample set; for each sample, first performing standard rigid registration and calculating the residual distribution histogram; analyzing the residual distribution, where the residuals of the rigid part usually conform to a Gaussian distribution, while the residuals of the deformed part exhibit a long-tailed distribution; and calculating the median absolute deviation of the residual distribution using statistical methods, and setting the Huber parameter to 1.3 times the median deviation. During the optimization process, in each iteration, the weight of each point is dynamically calculated based on the current residual value and the set Huber parameters. The weighted objective function is then re-solved until the pose change between the two iterations is less than the convergence threshold or the residual meets the requirements. Finally, the pose transformation parameters are output.

[0062] Finally, the coordinate transformation of the point cloud to be located is performed using the final pose transformation parameters to calculate the final coordinate positioning result. Specifically, the optimized and converged 4x4 transformation matrix is ​​applied to each point of the original point cloud to be located, and the new coordinates obtained are the precise position of the workpiece in the robot coordinate system, thus obtaining the final positioning result.

[0063] For example, after initial registration, the overall residual mean is 0.8mm, but due to local warping of the sheet metal, the residual in some areas reaches 5mm. When constructing the objective function, the system sets the Huber threshold to 1.5mm based on pre-statistical parameters. For points with a residual of 0.5mm, the weight is set to 1; for points with a residual of 5mm, the weight is automatically reduced to 1.5 / 5 = 0.3. After 8 iterations of the LM algorithm, the objective function converges. At this point, although the physical deviation in the warped area still exists, the matching residual of the rigid area (undeformed part) that occupies the main body has been reduced to 0.04mm, ensuring the accuracy of the robot's grasping and positioning reference. The final output coordinate transformation matrix accurately aligns the workpiece coordinate system to the robot base coordinate system.

[0064] In summary, this invention discloses a bending coordinate positioning method for a bending robot, comprising: acquiring a set of three-dimensional coordinates of a deformable workpiece; denoising the set of three-dimensional coordinates and distinguishing between deformed and undeformed regions; performing interpolation densification on the deformed regions to obtain densified deformed regions; merging the densified deformed regions and undeformed regions to obtain an original point cloud dataset; processing the original point cloud dataset using a statistical filtering method to obtain a denoised point cloud dataset; reducing the point density of the denoised point cloud dataset by voxel mesh downsampling to obtain a uniformly sampled dataset; extracting geometric features from the uniformly sampled dataset; fitting a geometric model based on the geometric features using a random sampling consensus method; and summarizing the feature parameters of the geometric model to obtain feature parameters. The invention employs a set of feature parameters; establishes spatial constraints based on the set of feature parameters, calculates the inter-constraint deviations between the feature parameters, and if the inter-constraint deviations are less than a preset deviation threshold, confirms that the set of feature parameters satisfies consistency, thus obtaining an effective geometric constraint group; filters the feature pair set in the effective geometric constraint group that corresponds to the pre-acquired theoretical model, and iteratively solves the rotation and translation matrix using the point cloud registration method based on the feature pair set to obtain initial pose transformation parameters; transforms the point cloud to be located based on the initial pose transformation parameters, and combines it with the theoretical model to obtain the final pose transformation parameters by minimizing the geometric deviation objective function; and performs coordinate transformation on the point cloud to be located based on the final pose transformation parameters to obtain the final coordinate positioning result. This invention achieves high-precision positioning of deformed workpieces by accurately establishing constraints reflecting the true geometric relationships of the workpiece from a point cloud containing noise and deformation of a sheet metal.

[0065] Reference Figure 3 The second embodiment of the present invention provides a bending coordinate positioning system for a bending robot, comprising: The data preprocessing module is used to process the original point cloud dataset using statistical filtering to obtain a denoised point cloud dataset. The sampling processing module is used to reduce the point density of the denoised point cloud dataset by voxel grid downsampling to obtain a uniformly sampled dataset. The feature extraction module is used to extract geometric features from the uniformly sampled dataset, fit a geometric model based on the geometric features using the random sampling consensus method, and summarize the feature parameters of the geometric model to obtain a feature parameter set. The constraint establishment and verification module is used to establish spatial constraint relationships based on the set of feature parameters, calculate the constraint deviation between the feature parameters, and if the constraint deviation is less than a preset deviation threshold, then the set of feature parameters is confirmed to meet the consistency requirement, and a valid geometric constraint group is obtained. The initial pose estimation module is used to filter the set of feature pairs that correspond to the pre-acquired theoretical model in the effective geometric constraint group, and to obtain the initial pose transformation parameters by iteratively solving the rotation and translation matrix according to the set of feature pairs through the point cloud registration method. The pose optimization and localization module is used to transform the point cloud to be located using the initial pose transformation parameters, and combined with the theoretical model, to obtain the final pose transformation parameters by minimizing the geometric deviation objective function. Based on the final pose transformation parameters, the point cloud to be located is subjected to coordinate transformation to obtain the final coordinate localization result.

[0066] It should be noted that the bending robot bending coordinate positioning system provided in this embodiment of the invention is used to execute all the process steps of the bending robot bending coordinate positioning method in the above embodiment. The working principle and beneficial effects of the two are one-to-one, so they will not be described again.

[0067] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0068] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A bending coordinate positioning method for a bending robot, characterized in that, include: Obtain the three-dimensional coordinate set of the deformed workpiece, denoise the three-dimensional coordinate set and distinguish between deformed and undeformed regions, perform interpolation encryption processing on the deformed regions to obtain encrypted deformed regions, and merge the encrypted deformed regions and undeformed regions to obtain the original point cloud dataset. The original point cloud dataset is processed by statistical filtering to obtain a denoised point cloud dataset. For the aforementioned denoised point cloud dataset, the point density is reduced by voxel grid downsampling to obtain a uniformly sampled dataset; Geometric features are extracted from the uniformly sampled dataset. A geometric model is fitted based on the geometric features using the random sampling consensus method. The feature parameters of the geometric model are then summarized to obtain a set of feature parameters. Spatial constraint relationships are established based on the set of feature parameters, and the inter-constraint deviation between the feature parameters is calculated. If the inter-constraint deviation is less than a preset deviation threshold, the set of feature parameters is confirmed to satisfy consistency, and an effective geometric constraint group is obtained. Select the set of feature pairs that correspond to the pre-acquired theoretical model from the effective geometric constraint group, and obtain the initial pose transformation parameters by iteratively solving the rotation and translation matrix according to the set of feature pairs using the point cloud registration method. The point cloud to be located is transformed according to the initial pose transformation parameters, and the final pose transformation parameters are obtained by minimizing the geometric deviation objective function in combination with the theoretical model. The coordinate transformation of the point cloud to be located is then performed according to the final pose transformation parameters to obtain the final coordinate positioning result.

2. The bending coordinate positioning method for a bending robot according to claim 1, characterized in that, The process involves obtaining a set of three-dimensional coordinates for the deformed workpiece, denoising the set of coordinates and distinguishing between deformed and undeformed regions, performing interpolation and densification processing on the deformed regions to obtain densified deformed regions, and merging the densified deformed regions and undeformed regions to obtain the original point cloud dataset, including: The laser echo signal acquired by the three-dimensional laser scanning equipment is collected, and the discrete spatial coordinates are analyzed from the laser echo signal. The reflection intensity value is extracted from the laser echo signal, and the discrete spatial coordinates are associated with the reflection intensity value to obtain a three-dimensional coordinate set. Noise below a preset intensity threshold is removed from the three-dimensional coordinate set to obtain the effective surface coordinates. Calculate the rate of change of the normal vector of the effective surface coordinates, locate the deformation region based on the rate of change of the normal vector, and perform interpolation and densification processing to obtain the densified deformation region; The encrypted deformed region is merged with the undeformed region to obtain the original point cloud dataset.

3. The bending coordinate positioning method for a bending robot according to claim 1, characterized in that, The process of processing the original point cloud dataset using statistical filtering to obtain a denoised point cloud dataset includes: A spatial neighborhood index structure is constructed based on the original point cloud dataset, and the nearest neighbor set of each point in the original point cloud dataset is retrieved through the spatial neighborhood index structure. The mean distance of the neighborhood is calculated based on the set of nearest points, and the statistical screening threshold is determined by combining the mean global distance and standard deviation of the original point cloud dataset. If the mean neighborhood distance exceeds the statistical filtering threshold, the corresponding query point is removed, and the remaining valid point data after removal is integrated to obtain a denoised point cloud dataset.

4. The bending coordinate positioning method for a bending robot according to claim 1, characterized in that, The process of reducing point density through voxel grid downsampling to obtain a uniformly sampled dataset for the denoised point cloud dataset includes: Calculate the three-dimensional coordinate extreme values ​​based on the denoised point cloud dataset, and determine the minimum three-dimensional spatial range surrounding the denoised point cloud dataset based on the three-dimensional coordinate extreme values. The three-dimensional space is divided according to the minimum three-dimensional spatial range to obtain multiple non-overlapping three-dimensional voxel units; Establish a spatial mapping index relationship between the three-dimensional voxel unit and each point in the denoised point cloud dataset, and calculate the geometric center coordinates of all points falling within the three-dimensional voxel unit based on the spatial mapping index relationship; By integrating the geometric center coordinates corresponding to all non-empty voxel units, a uniformly sampled dataset is obtained.

5. The bending coordinate positioning method for a bending robot according to claim 1, characterized in that, The process involves extracting geometric features from the uniformly sampled dataset, fitting a geometric model based on these features using the random sampling consensus method, and summarizing the feature parameters of the geometric model to obtain a feature parameter set, including: A KD tree index structure is constructed based on the uniform sampling dataset. For each sampling point, a set of neighboring points in the K nearest neighbor region is searched. The covariance matrix is ​​calculated based on the set of neighboring points, and eigenvalue decomposition is performed to obtain the normal vector and curvature value of each sampling point. Based on the curvature value, a curvature threshold is set, and the uniformly sampled dataset is divided into a planar feature point set and a surface feature point set according to the curvature threshold. The planar feature point set and the surface feature point set are then clustered to obtain multiple independent planar region clusters and surface region clusters. Based on each of the planar region clusters and the surface region clusters, the optimal planar fitting model and the optimal surface fitting model are constructed respectively using the random sampling consensus algorithm. Extract and integrate the planar feature parameters of all the optimal planar fitting models and the surface feature parameters of all the optimal surface fitting models to construct a feature parameter set.

6. The bending coordinate positioning method for a bending robot according to claim 1, characterized in that, The step of establishing spatial constraint relationships based on the set of feature parameters, calculating the inter-constraint deviation between the feature parameters, and confirming that the set of feature parameters satisfies consistency if the inter-constraint deviation is less than a preset deviation threshold, yields a valid geometric constraint set, including: Based on the set of feature parameters, calculate the Euclidean distance between geometric feature objects, determine the topological adjacency relationship based on the Euclidean distance, and extract the normal vector and center point coordinates; Calculate the angle constraint value and distance constraint value corresponding to the normal vector and the coordinates of the center point to obtain the initial constraint set; Obtain a preset standard geometric constraint template library, calculate the difference between the initial constraint set and the theoretical parameter values ​​in the standard geometric constraint template library, and obtain the constraint deviation value; If the deviation value between the constraints is less than the preset consistency judgment threshold, then the initial constraint set is confirmed to meet the consistency requirements, and a valid geometric constraint group is obtained.

7. The bending coordinate positioning method for a bending robot according to claim 1, characterized in that, The process involves selecting a set of feature pairs from the effective geometric constraint group that correspond to the pre-acquired theoretical model, and then iteratively solving the rotation and translation matrix using the point cloud registration method based on the set of feature pairs to obtain the initial pose transformation parameters, including: Obtain spatial constraint relationship data between the theoretical model and the effective geometric constraint group; retrieve the feature set of the theoretical model based on the spatial constraint relationship data; and construct a feature point pair set. The set of feature point pairs is processed using the singular value decomposition algorithm to obtain the covariance matrix. The covariance matrix is ​​then decomposed to obtain rotation matrix components and translation vector components. Construct a rigid body transformation matrix based on the rotation matrix components and the translation vector components, and calculate the residual value; If the residual value is greater than the preset residual threshold, the rigid body transformation matrix is ​​corrected using the iterative nearest point algorithm to obtain the initial pose transformation parameters.

8. The bending coordinate positioning method for a bending robot according to claim 1, characterized in that, The process of transforming the point cloud to be located based on the initial pose transformation parameters, and obtaining the final pose transformation parameters by minimizing the geometric deviation objective function in conjunction with the theoretical model, and then performing coordinate transformation on the point cloud to be located based on the final pose transformation parameters to obtain the final coordinate positioning result, includes: The coordinates of the point cloud to be located are transformed according to the initial pose transformation parameters, and the transformed point cloud data is matched with the theoretical model to obtain the initial residual set. Based on the initial residual set, a weighted least squares objective function is constructed and iteratively solved to obtain the updated pose parameters and the current residual value. If the current residual value is greater than the preset residual threshold, the weights are adjusted using the Huber kernel function and the weighted least squares objective function is re-solved until the current residual value meets the convergence condition, and the final pose transformation parameters are obtained. The coordinates of the point cloud to be located are transformed using the final pose transformation parameters, and the final coordinate positioning result is calculated.

9. The bending coordinate positioning method for a bending robot according to claim 2, characterized in that, The calculation of the rate of change of the normal vector of the effective surface coordinates, the location of the deformation region based on the rate of change of the normal vector, and the interpolation and densification processing to obtain the densified deformation region include: Calculate the rate of change of the angle between the normal vectors of adjacent sampling points of the effective surface coordinates. If the rate of change exceeds a preset angle threshold, the corresponding region is marked as a deformation region. An interpolation algorithm is used to reduce the point spacing within the deformation region from the original spacing to a preset encrypted spacing, thus obtaining an encrypted deformation region.

10. A bending coordinate positioning system for a bending robot, characterized in that, include: The data acquisition module is used to acquire a set of three-dimensional coordinates of the deformed workpiece, denoise the set of three-dimensional coordinates and distinguish between deformed and undeformed regions, perform interpolation encryption processing on the deformed regions to obtain encrypted deformed regions, and merge the encrypted deformed regions and undeformed regions to obtain the original point cloud dataset. The data preprocessing module is used to process the original point cloud dataset using statistical filtering to obtain a denoised point cloud dataset. The sampling processing module is used to reduce the point density of the denoised point cloud dataset by voxel grid downsampling to obtain a uniformly sampled dataset. The feature extraction module is used to extract geometric features from the uniformly sampled dataset, fit a geometric model based on the geometric features using the random sampling consensus method, and summarize the feature parameters of the geometric model to obtain a feature parameter set. The constraint establishment and verification module is used to establish spatial constraint relationships based on the set of feature parameters, calculate the constraint deviation between the feature parameters, and if the constraint deviation is less than a preset deviation threshold, then the set of feature parameters is confirmed to meet the consistency requirement, and a valid geometric constraint group is obtained. The initial pose estimation module is used to filter the set of feature pairs that correspond to the pre-acquired theoretical model in the effective geometric constraint group, and to obtain the initial pose transformation parameters by iteratively solving the rotation and translation matrix according to the set of feature pairs through the point cloud registration method. The pose optimization and localization module is used to transform the point cloud to be located using the initial pose transformation parameters, and combined with the theoretical model, to obtain the final pose transformation parameters by minimizing the geometric deviation objective function. Based on the final pose transformation parameters, the point cloud to be located is subjected to coordinate transformation to obtain the final coordinate localization result.

Images