A material cutting method and system based on multi-camera cooperative positioning

Abstract

The application provides a material cutting method and system based on multi-camera cooperative positioning, and relates to the technical field of program-controlled mechanical hand, which comprises the following steps: unifying a profile camera and a 3D camera into the same coordinate system; performing full-line scanning on a workpiece by using each profile camera on a cross beam to obtain point cloud data in the respective coordinate systems; performing fusion processing on the point cloud data obtained by each profile camera; generating a cutting path according to the global point cloud data; performing smoothing processing on the cutting path; performing global path posture calibration on the cutting path based on the overall observation result of the profile camera on the workpiece; performing local high-precision scanning by using a 3D camera at the end of each robot to perform local path fine compensation on the cutting path; performing weighted fusion on the global path posture calibration result and the local path fine compensation result; taking the fused real-time target path point as the robot end motion control input to drive two industrial cutting robots to perform a cutting task.

Description

Technical Field

[0001] This invention relates to the field of program-controlled robotic arms, and in particular to a material cutting method and system based on multi-camera collaborative positioning. Background Technology

[0002] In the cutting and machining of complex curved surface workpieces such as large wind turbine shrouds, traditional technical solutions generally face problems such as insufficient accuracy, poor system rigidity, and accumulation of path planning and positioning errors. Currently, mainstream methods mostly employ single robots or non-rigid structures for scanning and cutting operations, which limits the overall stability of the system and makes it difficult to meet the requirements of high-precision synchronous operations. Especially when processing large-sized (such as a 30m x 5m water tank platform) and complex-shaped workpieces, the lack of effective cooperative motion mechanisms and global-local fusion positioning strategies leads to frequent cutting path deviations and jitter, affecting the final machining quality.

[0003] In existing technologies, most cutting systems rely on fixed trajectories programmed offline, failing to adequately consider the impact of actual workpiece placement deviations and dynamic deformations on path execution. For example, when using laser cutting, if real-time 3D point cloud scanning of the workpiece is not performed and a pose compensation model in a unified coordinate system is not established, significant deviations may occur between the tool tip position and the theoretical path. Furthermore, current vision positioning systems typically only possess single-scale measurement capabilities, unable to simultaneously meet the needs of large-area contour scanning and precise positioning in small areas, thus limiting the system's adaptability and flexibility.

[0004] In path planning, existing methods often neglect curvature constraints and geometric continuity requirements, making the generated cutting trajectories prone to distortion or unevenness on complex surfaces. Especially in dual-robot collaborative operation scenarios, without the introduction of smoothing algorithms, it will lead to difficulties in path connection between the two robots in the intersection area, and may even cause vibration or impact, further reducing processing efficiency and surface quality.

[0005] In summary, current technologies still have significant shortcomings in multi-robot collaborative localization, cross-scale visual fusion modeling, and path generation. There is an urgent need for a material cutting method that can achieve high precision and high stability to meet the high-efficiency processing requirements of large and complex workpieces. Summary of the Invention

[0006] In view of the shortcomings of the prior art, the purpose of this invention is to provide a material cutting method and system based on multi-camera cooperative localization, which can solve the technical problems of current technology in multi-robot cooperative localization, cross-scale visual fusion modeling and path generation.

[0007] A first aspect of this invention provides a material cutting method based on multi-camera cooperative positioning, comprising:

[0008] S1: Using the top platform of the water tank as the reference datum of the coordinate system, unify the contour camera and the 3D camera into the same coordinate system;

[0009] S2: Use the contour cameras on the crossbeam to scan the workpiece along its entire length and obtain point cloud data in their respective coordinate systems;

[0010] S3: The point cloud data acquired by each of the contour cameras are fused to obtain global point cloud data;

[0011] S4: Generate a cutting path based on the global point cloud data;

[0012] S5: Smooth the cutting path;

[0013] S6: Divide the smoothed cutting path into two segments, left and right, according to the center line of the water tank, and assign them to the two industrial cutting robots respectively.

[0014] S7: Based on the overall observation results of the workpiece by the contour camera, perform global path attitude calibration on the cutting path;

[0015] S8: Use the 3D camera at the end of each robot to perform local high-precision scanning, adjust the robot end motion trajectory in real time, and perform local path fine compensation on the cutting path.

[0016] S9: Weighted fusion of global path attitude calibration results and local path fine compensation results is performed to form a fused real-time target path;

[0017] S10: The fused real-time target path points are used as the motion control input for the robot end effector to drive the two industrial cutting robots to perform the cutting tasks of the corresponding path segments.

[0018] A second aspect of the present invention provides a material cutting system based on multi-camera cooperative positioning, comprising: a water tank base, a support platform, a guide rail, an industrial cutting robot, a crossbeam, a vision positioning module, a waste recycling module, a processor, and a memory;

[0019] The support platform is located above the water tank base. The support platform is used to support the workpiece to be cut, and the water tank base is used to absorb the wastewater generated during the cutting process and settle some of the dust.

[0020] Guide rails are laid on both sides of the water tank, and an industrial cutting robot is installed on each guide rail. The industrial cutting robot moves independently or synchronously on the guide rail.

[0021] A detachable crossbeam is provided between the two industrial cutting robots;

[0022] The visual positioning module includes a contour camera and a 3D camera. Multiple contour cameras are spaced apart on the crossbeam, and the 3D camera is mounted on the end effector of the industrial cutting robot.

[0023] The contour camera is used to quickly acquire the overall three-dimensional point cloud data of the workpiece;

[0024] The 3D camera is used to perform a fine scan of the area to be cut in order to obtain local geometric information;

[0025] The waste recycling module is used to collect wastewater and debris in real time during the cutting process to keep the working area clean;

[0026] The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the material cutting method based on multi-camera cooperative positioning described above.

[0027] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0028] In this embodiment of the invention, macroscopic contour scanning is combined with microscopic precise positioning for the first time to construct a multi-scale, multi-source visual fusion positioning system. Furthermore, multi-source point cloud fusion technology is innovatively applied to generate the cutting path in path planning, and the cutting path is smoothed, effectively solving the problems of path distortion and uneven junctions caused by neglecting geometric characteristics in traditional methods. This method not only improves cutting accuracy and efficiency but also possesses good environmental adaptability and process compatibility, making it suitable for automated processing scenarios of various large and complex curved surface materials. Attached Figure Description

[0029] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0030] Figure 1 This is a schematic flowchart of a material cutting method based on multi-camera collaborative positioning provided by an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram of a material cutting system based on multi-camera cooperative positioning provided in an embodiment of the present invention.

[0032] Explanation of reference numerals in the attached drawings: 1-Water tank base; 2-Bearing platform; 3-Guide rail; 4-Crossing beam; 5-Contour camera; 6-3D camera. Detailed Implementation

[0033] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions 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, not all embodiments. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0034] The material cutting system based on multi-camera cooperative positioning provided by the present invention will be described in detail below with reference to the accompanying drawings, through specific embodiments and application scenarios.

[0035] Reference manual attached Figure 1 The diagram shows a flowchart of a material cutting method based on multi-camera cooperative positioning provided by an embodiment of the present invention.

[0036] This invention provides a material cutting method based on multi-camera cooperative positioning, which may include the following steps:

[0037] S1: Using the top platform of the water tank as the reference datum of the coordinate system, the contour camera and the 3D camera are unified into the same coordinate system.

[0038] It should be noted that by using the top platform of the water tank as the reference benchmark of the world coordinate system and unifying the contour camera and the 3D camera into the same coordinate system, spatial consistency of multi-source visual data can be achieved, avoiding the accumulation of coordinate errors caused by differences in installation position and posture between different cameras.

[0039] S2: Use the contour cameras on the crossbeam to scan the entire workpiece and obtain point cloud data in their respective coordinate systems.

[0040] S3: The point cloud data acquired by each contour camera is fused to obtain global point cloud data.

[0041] In one possible implementation, S3 specifically includes sub-steps S301 to S304:

[0042] S301: Using the extrinsic parameter matrix, the point cloud data acquired by each contour camera is transformed into the world coordinate system.

[0043] S302: Principal component analysis algorithm is used to perform coarse registration processing on the point cloud data acquired by each contour camera.

[0044] Principal Component Analysis (PCA) is used to perform coarse registration of the point cloud data acquired by each contour camera. This is done by analyzing the overall distribution characteristics of the point cloud to quickly estimate its principal orientation.

[0045] Specifically, by centering the point cloud and calculating its covariance matrix, the principal eigenvectors of the point cloud can be extracted to represent its main extension direction in space. Aligning the principal eigenvectors of point clouds from different cameras, and constructing rotation and translation vectors, preliminary registration of the point clouds at the macroscopic level can be achieved, ensuring consistency in the principal directions across different viewpoints. This coarse registration process significantly reduces initial pose differences between point clouds, providing better initial conditions for subsequent fine registration algorithms, thereby improving the efficiency and stability of overall point cloud fusion. This is already a well-established technique and will not be elaborated upon further in this invention.

[0046] S303: The iterative nearest point algorithm is used to perform fine registration on the point cloud data after coarse registration, and to align the point cloud data in the overlapping areas.

[0047] Among them, the Iterative Closest Point (ICP) algorithm is a commonly used point cloud fine registration method. Its core idea is to iteratively solve the optimal rigid body transformation between two sets of point clouds so that they are gradually aligned in space.

[0048] Traditional iterative nearest-neighbor algorithms rely solely on Euclidean distance for point registration, lacking the ability to discriminate point cloud geometric features. This leads to failures or insufficient accuracy in several scenarios: First, when the point cloud has significant initial pose deviations or originates from complex surfaces, some points in the source cloud may perform "nearest neighbor matching" with incorrect locations in the target cloud, accumulating erroneous point pairs and causing the obtained rotation and translation matrices to deviate from the true transformation. Second, when the point cloud contains noise, occlusion, missing regions, or uneven point density, distance-based matching results in a large number of unreliable points being included in the solution process, amplifying registration errors, causing iterative oscillations, or even non-convergence. Third, traditional ICP (Integrated Point Coding) cannot identify different geometric structures such as planes, boundaries, and curved surfaces. Especially in large industrial components (such as long curved surfaces and corner areas of wind turbine fairings), where local geometric changes in the point cloud are significant, Euclidean distance cannot represent the true correspondence between points. Therefore, this invention introduces curvature feature similarity to improve the traditional iterative nearest-neighbor algorithm.

[0049] Optionally, sub-step S303 specifically includes S3031 to S3037:

[0050] S3031: For the first point cloud data P and the second point cloud data Q after coarse registration, calculate the curvature feature vector of each data point.

[0051] Optionally, the curvature feature vector includes: a first principal curvature, a second principal curvature, a Gaussian curvature, and a mean curvature.

[0052] The first principal curvature describes the degree of curvature of the surface in the steepest direction at a given point; that is, the curvature of the curve with the largest curvature value among all normal sections at that point. The second principal curvature describes the degree of curvature of the surface in the gentlest direction at a given point; that is, the curvature of the curve with the smallest curvature value among all normal sections. Gaussian curvature is equal to the product of the first and second principal curvatures; it reflects the inherent curvature property of the surface and is an important quantity describing the geometric properties of the surface. Mean curvature is defined as the arithmetic mean of the two principal curvatures; it describes the overall curvature of the surface in a localized manner and is an important quantity for measuring the external curvature behavior of the surface.

[0053] Specifically, an accurate geodesic algorithm is used to calculate the discrete exponential mapping around the data point. A set of sampling points is extracted from the geodesic disk centered on the current data point, based on a pre-defined two-dimensional sampling template. An MLS energy function is constructed based on this sampling point set. The average curvature and Gaussian curvature of the points are then calculated using the MLS energy function.

[0054] Specifically, by solving the energy function and its partial derivatives with respect to position variables, the principal curvature directions and bending characteristics of the local surface are obtained. Furthermore, the average curvature and Gaussian curvature of a point can be directly calculated using the second derivative of the surface and the local shape operator obtained through MLS.

[0055] Then, based on the relationship between the first principal curvature K1, the second principal curvature K2, the Gaussian curvature K, and the mean curvature H, the two principal curvatures are solved:

[0056]

[0057] Where K1 represents the first principal curvature, K2 represents the second principal curvature, K represents the Gaussian curvature, and H represents the mean curvature.

[0058] S3032: For each data point in the first point cloud data, search for the nearest point in the second point cloud data by Euclidean distance.

[0059] S3033: Calculate the curvature feature similarity between each data point in the first point cloud data and its corresponding nearest point.

[0060] Optionally, curvature feature similarity is obtained by calculating the cosine similarity of the curvature feature vectors of the two data points.

[0061] S3034: Determine if the curvature feature similarity between data point pairs is greater than the preset similarity. If yes, proceed to the next step. Otherwise, discard the current data point pair and search again.

[0062] Those skilled in the art can set the preset similarity level according to actual needs, and this invention does not limit this.

[0063] In this embodiment of the invention, curvature feature similarity is introduced to measure the consistency of two points in their local geometric structure. A curvature feature vector is constructed by integrating features such as principal curvature, mean curvature, and Gaussian curvature. Feature similarity is used as a filtering condition; only when the curvature feature similarity is higher than a preset threshold is the point pair considered a valid corresponding point for subsequent rotation matrix and pose calculation. This effectively eliminates incorrect corresponding points with inconsistent geometric properties, avoids mismatches caused by relying solely on distance, makes the registration iteration process more stable and reliable, and significantly improves the accuracy and robustness of point cloud fusion in complex curved surface scenes.

[0064] S3035: Uses the traditional iterative nearest-point algorithm to solve for the rotation matrix and translation vector.

[0065] Specifically, in the traditional iterative nearest neighbor algorithm, the rotation matrix and translation vector are solved by minimizing the sum of squared Euclidean distances between corresponding points. After finding the nearest neighbor in the target point cloud for each point in the source point cloud, two sets of corresponding points are constructed, their centroids are calculated, and the decentralized coordinates of the point pair are obtained after eliminating the translation effect using the centroids. Subsequently, the optimal rotation matrix that maximizes the overlap of the point pair is solved by constructing the covariance matrix and using the quaternion method or singular value decomposition (SVD). After the rotation matrix is ​​determined, the optimal translation vector is obtained by the difference between the centroid of the source point cloud and the centroid of the source point cloud after rotation in the target point cloud. Through the above steps, a rigid body transformation that minimizes the overall registration error can be obtained, allowing the source point cloud to gradually approach the target point cloud in space, achieving accurate alignment of the point clouds. This is a very mature existing technology, and will not be elaborated further in this invention.

[0066] S3036: Update the point cloud based on the rotation matrix and translation vector.

[0067] S3037: Repeat S3031 to S3036 until the convergence condition is met.

[0068] In this embodiment of the invention, by introducing curvature feature vectors and curvature feature similarity judgment during the fine registration process of point clouds, and combining it with the iterative solution process of traditional ICP, the accuracy and robustness of point cloud registration under complex workpiece surface conditions can be significantly improved. Specifically, by using curvature features to filter corresponding point pairs, erroneous matches based solely on Euclidean distance can be effectively eliminated. This ensures that the registration process not only depends on spatial distance but also on the similarity of the local geometry of the two points, thereby preventing the point cloud from getting stuck in local optima or diverging under conditions of noise, occlusion, uneven density, or large initial attitude deviation. Subsequently, the traditional ICP solution for rotation and translation is performed on the filtered high-quality corresponding point set, which can improve iterative efficiency and obtain more reliable optimal rigid body transformations. By continuously iterating and updating the point cloud and checking the convergence conditions, the entire fine registration step can achieve high-precision alignment while ensuring computational efficiency, providing a more stable and accurate geometric basis for the final multi-camera point cloud fusion and path generation.

[0069] S304: The point cloud data acquired by each contour camera is fused to obtain global point cloud data.

[0070] Optionally, sub-step S304 specifically includes S3041 to S3044:

[0071] S3041: Calculate the registration fit of point cloud data in overlapping regions.

[0072] Optionally, the registration fit is calculated as follows:

[0073]

[0074] Where τ represents the registration fit, and D m This represents the median of the Euclidean distances between all valid point pairs. After point cloud registration is complete, many valid point pairs will be obtained. Calculate the Euclidean distances between these valid point pairs, sort them in ascending order, and the median number is the median of the Euclidean distances.

[0075] S3042: When the registration fit of the point cloud data in the overlapping area is greater than the first preset fit, voxel filtering downsampling is directly performed on all the finely registered point clouds to generate global point cloud data.

[0076] It should be noted that when the fit is high, global point clouds can be generated directly by downsampling, ensuring efficiency.

[0077] S3043: When the registration fit of the point cloud data in the overlapping region is between the second preset fit and the first preset fit, statistical filtering is used to remove outliers, and then downsampling is performed to generate global point cloud data.

[0078] It should be noted that when the fit is in the medium range, statistical filtering is used to remove outliers before downsampling, which can effectively avoid noise spread caused by local misregistration.

[0079] S3044: When the registration fit of the point cloud data in the overlapping area is less than the second preset fit, discard the point cloud data with the low registration fit, fuse the other point cloud data, and issue an alarm. This guides the user to reacquire the point cloud data.

[0080] It should be noted that when the fit is below the threshold, timely removal of data from abnormal perspectives and issuance of an alarm can prevent mismatched point clouds from damaging the overall modeling and ensure the system's security and data quality in abnormal environments.

[0081] The second preset fit is less than the first preset fit.

[0082] Those skilled in the art can set the magnitude of the first preset fit degree and the second preset fit degree according to the actual situation, and the present invention does not limit them.

[0083] Furthermore, by dynamically selecting different point cloud fusion and filtering strategies based on the registration fit of the overlapping point clouds, the stability of multi-view point cloud fusion and the reliability of the final global point cloud model can be significantly improved. Through this hierarchical processing mechanism, the system can enhance its robustness to noise, occlusion, illumination changes, and equipment malfunctions while ensuring high accuracy, thereby obtaining more reliable and stable global point cloud data.

[0084] In this embodiment of the invention, by employing extrinsic parameter matrix transformation, PCA coarse registration, ICP fine registration, and final point cloud fusion processing, the spatial consistency and alignment accuracy of point clouds from multi-contour cameras can be significantly improved.

[0085] S4: Generate cutting paths based on global point cloud data.

[0086] In one possible implementation, S4 specifically includes sub-steps S401 and S402:

[0087] S401: Perform surface fitting and reconstruction on noisy global point cloud data, reconstructing the point cloud data while adhering to the basic geometric shape of the global point cloud data.

[0088] Specifically, the objective function for point cloud reconstruction is set to fit the basic geometry of the original point set and suppress excessive clustering between output points:

[0089]

[0090] Where Q represents the output point set, argmin represents the expression that minimizes the value, E1 represents the approximation term used to make the output point set fit the basic geometry of the original point set, and P... J Represents the original point set, E2 represents the regularization term used to suppress excessive clustering between output points, and q i p represents the i-th data point in the output point set, I represents the index set of the output point set, and p j Let J represent the j-th data point in the original point set, where J represents the index set of the original point set. θ represents the distance between the i-th data point in the output set and the j-th data point in the original set, θ represents the weight kernel function, and γ represents the weight coefficient of the regularization term, used to adjust the ratio between approximation and uniform distribution. This represents the index value of other data points in the output point set that are different from the i-th data point. This indicates the number of data points in the output point set that are different from the i-th data point. Other data points, This indicates that the i-th data point in the output point set is related to the i-th data point. The distance between each other data point, η represents the repulsion function used to penalize data points that are too close to each other, and σ is the distance between each other. i This represents the local directionality of the i-th data point determined by principal component analysis.

[0091] It should be noted that by constructing a point cloud reconstruction objective function that simultaneously includes approximation and regularization terms, the structural uniformity and stability of the reconstructed point set can be significantly improved while ensuring geometric accuracy. The approximation term, by introducing a weighted distance between the output points and the original points, ensures that the output point set effectively conforms to the basic geometry of the original point cloud, thereby restoring the true contour of the workpiece surface. The regularization term, through a repulsive function, applies a repulsive force to output points that are too close to each other, and adjusts the repulsion strength using a local directionality index, preventing collapse and clustering in the reconstructed point cloud, thus achieving a point set representation with uniform spatial distribution and a clear topological structure. The combined effect of these two terms forms a dynamic balance mechanism of "attraction-repulsion," enabling the point cloud reconstruction process to eliminate noise interference, enhance local continuity, and maintain the integrity and accuracy of global geometric features. This provides a high-quality, structurally stable 3D data foundation for subsequent cutting path extraction, sorting, and smoothing processing.

[0092] Alternatively, a rapidly decaying Gaussian kernel function can be used to construct the weight kernel function:

[0093]

[0094] Where θ represents the weight kernel function, r represents the distance, e represents the natural constant, and h represents the support radius.

[0095] It should be noted that the influence weight of points in the optimization process is adjusted according to the distance *r* between them. When two points are close, the function value is close to 1, indicating a significant contribution. However, as the distance increases, the function value rapidly decays with the exponential term, approaching 0, making distant points almost insignificant in the calculation. The finite support radius *h* is used to control the effective range of the kernel function. By introducing this kernel function, the point cloud reconstruction algorithm can perform geometric approximation and distribution adjustment within the local neighborhood, thereby enhancing the stability and locality of point cloud processing and avoiding interference from distant noise points in the output results.

[0096] Alternatively, an inverse proportional function can be used to construct the repulsion function:

[0097]

[0098] Where η represents the repulsive force function and r represents the distance.

[0099] It should be noted that the repulsion function describes the repulsive force between points that varies with distance; its value is inversely proportional to the distance *r* between the two points. When two points are close to each other, a strong repulsive force is generated, effectively preventing excessive clustering of output points. As the distance increases, the repulsive force gradually decreases, making the repulsive effect between distant points negligible. By introducing it into the regularization term, it ensures that the output point set exhibits a stable spacing and shape in its geometry, thus better meeting the continuity of toolpath generation and machining requirements.

[0100] Optionally, principal component analysis determines the local directionality of data points as follows: for any point q in the output point set Q... i First, collect all other points within its local neighborhood. The neighborhood points are then weighted according to a distance-weighted kernel function to construct a local weighted difference vector. Subsequently, all weighted difference vectors are summed to form a 3×3 weighted covariance matrix. This matrix is ​​then subjected to eigenvalue decomposition to obtain three eigenvalues ​​sorted by size. The direction corresponding to the largest eigenvalue λ2 reflects the main trend of change in the neighborhood point cloud, and a local directionality index is defined accordingly:

[0101]

[0102] Where λ0, λ1, and λ2 represent the three eigenvalues ​​of the weighted covariance matrix.

[0103] The following describes how to reconstruct point cloud data based on the objective function described above:

[0104] First, let the gradient of the objective function of the output point set Q be 0. Then, each data point in the output point set Q satisfies the following:

[0105]

[0106]

[0107] Where, α ij This represents the weighted pulling coefficient between the i-th data point in the output point set and the j-th data point in the original point set. This indicates that the i-th data point in the output point set is related to the i-th data point. The weighted repulsion coefficient between the other data points, ∂ represents the partial derivative operation.

[0108] It should be noted that, let the objective function be applied to q i The gradient is 0, which is essentially the tension equals the repulsion, used to ensure that the output point position reaches equilibrium.

[0109] make

[0110] The derivation yields:

[0111]

[0112] This can then be transformed into a fixed-point iteration form:

[0113] .

[0114] It should be noted that a practically executable iterative update formula is constructed so that the point cloud can gradually approach the optimal solution.

[0115] After that, , here d j This can be understood as the density compensation weight of the j-th data point in the original point set. When the point cloud distribution in the region where the j-th point is located is relatively dense, the sum of its neighborhood response values ​​is larger, and the corresponding d... j The smaller the value, the weaker its influence on the objective function. When the region containing the j-th point is sparse, the sum of its neighborhood response values ​​is smaller, and the corresponding d... j The value is relatively large, thus enhancing the contribution at that point. Substituting d... j The final iterative formula is obtained, and the iterative result can be expressed as:

[0116]

[0117] in, This represents the result of the (k+1)th iteration for the i-th data point. This represents the weighted tension coefficient between the i-th data point in the output set and the j-th data point in the original set during the k-th iteration. This indicates that the i-th data point in the output point set during the k-th iteration is related to the i-th data point in the output point set during the k-th iteration. The weighted repulsion coefficient between the other data points.

[0118] The above iterative formula is used for iteration. When the iteration offset is less than the offset threshold or the number of iterations reaches the upper limit, the iteration is terminated, the reconstructed point set is output, and the point cloud data is reconstructed.

[0119] Optionally, the iteration offset is specifically:

[0120]

[0121] Where, ε k This represents the iteration offset at the k-th iteration. This indicates the size of the output point set.

[0122] It should be noted that by solving the objective function of "tension approximation term + repulsion regularization term" using fixed-point iteration, the output point set is made to approximate the real surface while maintaining a uniform distribution. Combined with directional constraints and density compensation mechanisms, a high-quality reconstructed point cloud with geometric accuracy, uniform structure, and good noise suppression is finally obtained.

[0123] In this embodiment of the invention, by solving the gradient of the objective function for point cloud reconstruction to zero and constructing a fixed-point iterative form, the complex optimization problem can be transformed into an easy-to-implement and stable convergent iterative update process, so that the output point set achieves a dynamic balance between approximating the original geometry and maintaining the uniformity of distribution.

[0124] S402: Sort the reconstructed point cloud data and generate cutting paths.

[0125] Specifically, the point cloud data sorting problem can be viewed as finding the minimum spanning tree on a weighted undirected graph. Each point in the point cloud data can be treated as a vertex in the graph, and an extended graph can be constructed using the squared Euclidean distance between points as the edge weights, resulting in geometrically adjacent points having smaller weights. Then, by solving the minimum spanning tree (MST) of this weighted graph, a tree structure that covers all points and has the smallest total weight can be obtained, thus establishing reasonable connections between points globally. Since the minimum spanning tree is inherently non-cyclic, to generate a single-chain path suitable for tool cutting, the tree structure can be further pruned, removing short branches and degenerating it into a continuous linked list structure. This sorting technique is existing technology and will not be elaborated upon here.

[0126] In this embodiment of the invention, by constructing an optimization objective that includes attraction and repulsion terms, the output point set can both conform to the true geometry of the original point cloud and avoid excessive aggregation between points, thereby generating a reconstructed point set with uniform structure and good noise suppression, providing a stable data foundation for subsequent path generation. Simultaneously, a local directionality and density compensation mechanism is added to ensure that the reconstructed point set maintains correct topological and distribution characteristics even in complex curved surface regions. Then, a graph-based sorting method is employed, transforming the sorting problem into a minimum spanning tree solution combined with pruning, which effectively ensures the continuity and global consistency of the path, avoids the occurrence of loops and jump points, and makes the final generated cutting path smoother, more reasonable, and in line with the dynamic requirements of industrial robot execution. Through the above processing flow, high-quality cutting paths can be stably obtained under noisy data and complex geometric scenes, significantly improving material cutting accuracy and overall system reliability.

[0127] S5: Smooth the cutting path.

[0128] In one possible implementation, S5 specifically involves smoothing the cutting path with the goal of improving curvature smoothness, point spacing uniformity, and offset retention.

[0129] Specifically, a discrete-point smoothing method can be introduced, modeling the smoothing problem as a quadratic programming problem. The coordinates of each discrete point in the path are used as variables to be optimized. By constructing an objective function and linear constraints, the path smoothing process is transformed into an optimization problem with a clear mathematical form. A cost function for smoothing can be set:

[0130]

[0131] Where C represents the cost function, c1 represents the curvature smoothing term, ρ1 represents the weight coefficient of the curvature smoothing term, c2 represents the point spacing uniformity term, ρ2 represents the weight coefficient of the point spacing uniformity term, c3 represents the offset preservation term, and ρ3 represents the weight coefficient of the offset preservation term.

[0132] Those skilled in the art can set the weight coefficients of the curvature smoothing term, the point spacing uniformity term, and the offset preservation term according to the actual situation; this invention does not impose any limitations.

[0133]

[0134] Among them, b u b represents the u-th path point on the smoothed cutting path. u-1 b represents the (u-1)th path point on the smoothed cutting path. u+1 This represents the (u+1)th path point on the cut path after smoothing.

[0135] Furthermore, the curvature smoothing term penalizes sharp turns in local paths, making the path exhibit a more natural and continuous geometric trend overall, avoiding jitter or trajectory instability caused by sudden changes in robot posture during movement. The point spacing uniformity term constrains the distance between adjacent path points, making the path distribution more uniform, which helps improve the consistency of tool cutting speed and cutting quality, and reduces machining errors caused by excessively dense or sparse local areas. The offset preservation term ensures that the smoothed path is optimized without deviating from the original geometric features, making the path smooth without compromising the accuracy requirements of the workpiece contour.

[0136] In this embodiment of the invention, by constructing a comprehensive cost function that includes a curvature smoothing term, a point spacing uniformity term, and an offset preservation term, the continuity and executability of the path can be significantly improved while ensuring the geometric accuracy of the cutting path.

[0137] The specific boundary conditions for smoothing include:

[0138]

[0139] in, This represents the minimum permissible offset in the x-direction. Indicates the maximum permissible offset in the x-direction, x u This represents the x-coordinate of the u-th path point. This represents the x-coordinate of the u-th path point on the cutting path. This represents the minimum permissible offset in the y-direction. This represents the maximum allowed offset in the y-direction of the u-th path point. u This represents the y-coordinate of the u-th path point. This represents the z-coordinate of the u-th path point on the cutting path. This represents the minimum permissible offset in the z-direction. This represents the maximum permissible offset in the z-direction. u This represents the z-coordinate of the u-th path point. μ represents the z-coordinate of the u-th path point on the cutting path. u This represents the curvature at the position of the u-th path point. This represents the upper limit of the curvature.

[0140] In this embodiment of the invention, by introducing spatial offset constraints and curvature upper limit constraints for path points during the smoothing optimization process, the smoothed cutting path can be effectively balanced between shape preservation, safety and executability.

[0141] Based on the aforementioned cost function and constraints, a quadratic programming solver can be used to obtain a smooth path. The solution process of the quadratic programming solver is existing technology and will not be described in detail here.

[0142] Furthermore, based on the aforementioned cost function and constraints, optimization algorithms such as genetic algorithms and particle swarm optimization can be used to search for and determine a smooth path. The specific optimization methods of these algorithms are existing technologies and will not be elaborated upon here.

[0143] S6: Divide the smoothed cutting path into two segments, left and right, according to the center line of the water tank, and assign them to two industrial cutting robots respectively.

[0144] S7: Based on the overall observation results of the workpiece by the contour camera, perform global path attitude calibration on the cutting path.

[0145] It should be noted that global path orientation calibration can align the robot's cutting path with the workpiece's actual orientation on a macroscopic scale.

[0146] In one possible implementation, S7 specifically involves: based on the overall observation results of the contour camera on the workpiece, comparing the expected pose of the cutting path in the world coordinate system with the actual placement posture of the workpiece, correcting the overall position and orientation of the cutting path through rigid body transformation, and performing global path pose calibration on the cutting path.

[0147] In this embodiment of the invention, global path pose calibration of the cutting path is performed based on the overall observation results of the contour camera, ensuring that the generated cutting path remains consistent with the actual placement posture of the workpiece on a macroscopic scale. Since the actual position and angle of the workpiece on the support platform often deviate, directly using the theoretically generated path will result in overall misalignment of the cutting trajectory. By comparing the expected pose of the path with the actual posture of the workpiece, and using rigid body transformation to correct the overall position and orientation of the path for consistency, workpiece placement errors can be effectively eliminated. This ensures that the tool trajectory accurately covers the target area when the robot performs cutting, thereby improving cutting accuracy, avoiding miscuts or missed cuts, and enhancing the robustness and processing reliability of the system.

[0148] S8: Utilizes a 3D camera at the end of each robot to perform local high-precision scanning, adjusts the robot's end-effector trajectory in real time, and performs fine compensation for the local path of the cutting path.

[0149] It should be noted that fine compensation of the local path allows the tool to accurately conform to the actual workpiece surface before cutting.

[0150] In one possible implementation, S8 specifically involves: using the 3D camera at the end of each robot to perform a local high-precision scan; extracting the real cutting feature points near the current tool based on the local point cloud obtained from the scan; calculating the position error of the real cutting feature points in the robot end coordinate system; using the position error as a local compensation amount; and using Kalman filtering to adjust the robot end motion trajectory in real time to perform local path fine compensation on the cutting path.

[0151] In this embodiment of the invention, by utilizing a 3D camera at the robot's end effector to perform real-time high-precision scanning of the area near the cutting tool, and extracting the positional deviation of the actual cutting feature points based on the scanned point cloud, and then combining this deviation with Kalman filtering for robust estimation and dynamic compensation, the robot can correct its end effector trajectory in real time before cutting, ensuring that the cutting tool precisely conforms to the workpiece surface. Since the local morphology of the workpiece often exhibits minute warping, unevenness, deformation, or clamping errors, relying solely on the global path cannot guarantee the cutting accuracy of the cutting tool in local areas. Introducing fine compensation for the local path effectively eliminates these local geometric deviations, achieving trajectory following with millimeter-level or even higher precision, avoiding tool suspension, overcutting, or off-center cutting, thereby significantly improving cutting quality, kerf consistency, and processing stability, while also enhancing the system's robustness to environmental disturbances.

[0152] S9: Weighted fusion of global path attitude calibration results and local path fine compensation results to form fused real-time target path points.

[0153] Optionally, the specific method for fusing real-time target path points is as follows:

[0154]

[0155] Among them, f v This shows the v-th real-time target path point after fusion. ω represents the global path attitude calibration result at point v. G This represents the weighting coefficients for global path attitude calibration. ω represents the local path fine-compensation result at point v. L This represents the weighting coefficients for fine-grained compensation of local paths.

[0156] Those skilled in the art can set the weight coefficients for global path attitude calibration and local path fine compensation according to actual conditions; this invention does not impose any limitations.

[0157] In this embodiment of the invention, by weighted fusion of the global path attitude calibration results and the local path fine compensation results, a dynamic balance between macroscopic alignment and local accuracy of the cutting path can be achieved. Global path attitude calibration ensures the overall orientation of the robot cutting trajectory is correct and consistent with the workpiece placement posture. Meanwhile, local path fine compensation corrects minor geometric deviations near the tool in real time, improving local fitting accuracy. By setting the weight coefficients for both global and local paths and calculating the fused real-time target path points proportionally, error amplification caused by relying solely on a single path can be avoided. This ensures that the final generated robot trajectory possesses both global stability and responsiveness to local changes, thereby significantly improving the continuity, accuracy, and robustness of the cutting process.

[0158] S10: The fused real-time target path points are used as the motion control input for the robot's end effector, driving the two industrial cutting robots to perform the cutting tasks of the corresponding path segments.

[0159] Reference manual attached Figure 2 The diagram shows a structural schematic of a material cutting system based on multi-camera cooperative positioning provided by an embodiment of the present invention.

[0160] This invention provides a material cutting system 20 based on multi-camera collaborative positioning, comprising: a water tank base 1, a support platform 2, a guide rail 3, an industrial cutting robot 4, a crossbeam, a vision positioning module, a waste recycling module, a processor, and a memory.

[0161] The support platform 2 is set above the water tank base 1. The support platform 2 is used to support the workpiece to be cut, and the water tank base 1 is used to absorb the wastewater generated during the cutting process and settle some of the dust.

[0162] Guide rails 3 are laid on both sides of the water tank base 1. An industrial cutting robot 4 is installed on each guide rail 3. The industrial cutting robot 4 moves independently or synchronously on the guide rail 3.

[0163] A detachable crossbeam is installed between the two industrial cutting robots 4.

[0164] The visual positioning module includes a contour camera 5 and a 3D camera 6. Multiple contour cameras 5 are spaced apart on the crossbeam, and the 3D camera 6 is mounted on the end effector of the industrial cutting robot 4.

[0165] The contour camera 5 is used to quickly acquire overall 3D point cloud data of the workpiece.

[0166] The 3D camera 6 is used to perform fine scanning of the area to be cut in order to obtain local geometric information.

[0167] The waste recycling module is used to collect wastewater and debris in real time during the cutting process, keeping the work area clean.

[0168] The memory stores programs or instructions that can run on the processor. When the program or instructions are executed by the processor, they implement the steps of the material cutting method based on multi-camera cooperative positioning described above and achieve the same technical effect. To avoid repetition, the present invention will not repeat the above.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. A material cutting method based on multi-camera cooperative positioning, characterized in that, include: S1: Using the top platform of the water tank as the reference datum of the coordinate system, the contour camera and the 3D camera are unified into the same coordinate system; S2: Use the contour cameras on the crossbeam to scan the workpiece along its entire length and obtain point cloud data in their respective coordinate systems; S3: The point cloud data acquired by each of the contour cameras are fused to obtain global point cloud data; S4: Generate a cutting path based on the global point cloud data; S5: Smooth the cutting path; S6: Divide the smoothed cutting path into two segments, left and right, according to the center line of the water tank, and assign them to two industrial cutting robots respectively. S7: Based on the overall observation results of the contour camera on the workpiece, perform global path attitude calibration on the cutting path; S8: Use the 3D camera at the end of each robot to perform local high-precision scanning, adjust the robot end motion trajectory in real time, and perform local path fine compensation on the cutting path. S9: Weighted fusion of global path attitude calibration results and local path fine compensation results to form fused real-time target path points; S10: Use the fused real-time target path points as the motion control input for the robot end effector to drive the two industrial cutting robots to perform the cutting tasks of the corresponding path segments respectively; Specifically, S3 includes: S301: Using the extrinsic parameter matrix, the point cloud data acquired by each of the contour cameras is transformed into the world coordinate system; S302: The principal component analysis algorithm is used to perform coarse registration processing on the point cloud data acquired by each of the contour cameras; S303: The iterative nearest point algorithm is used to perform fine registration on the point cloud data after coarse registration, and to align the point cloud data in the overlapping areas. S304: The point cloud data acquired by each of the contour cameras is fused to obtain the global point cloud data; Specifically, S303 includes: S3031: For the first and second point cloud data after coarse registration, calculate the curvature feature vector of each data point; S3032: For each data point in the first point cloud data, search for the nearest point in the second point cloud data by Euclidean distance; S3033: Calculate the curvature feature similarity between each data point in the first point cloud data and its corresponding nearest point; S3034: Determine whether the curvature feature similarity between data point pairs is greater than the preset similarity; if yes, proceed to the next step; otherwise, discard the current data point pair and search again; S3035: Uses the traditional iterative nearest point algorithm to solve for the rotation matrix and translation vector; S3036: Update the point cloud according to the rotation matrix and the translation vector; S3037: Repeat S3031 to S3036 until the convergence condition is met.

2. The multi-camera co-location based material cutting method of claim 1, wherein, The curvature feature vector includes: a first principal curvature, a second principal curvature, a Gaussian curvature, and a mean curvature; The curvature feature similarity is obtained by calculating the cosine similarity of the curvature feature vectors of two data points.

3. The multi-camera co-location based material cutting method of claim 1, wherein, Specifically, S304 includes: S3041: Calculate the registration fit of point cloud data in the overlapping region; S3042: When the registration fit of the point cloud data in the overlapping area is greater than the first preset fit, voxel filtering downsampling is directly performed on all the finely registered point clouds to generate the global point cloud data. S3043: When the registration fit of the point cloud data in the overlapping region is between the second preset fit and the first preset fit, statistical filtering is used to remove outliers, and then downsampling is performed to generate the global point cloud data; S3044: When the registration fit of the point cloud data in the overlapping area is less than the second preset fit, discard the point cloud data with the low registration fit, fuse the other point cloud data, and issue an alarm. Wherein, the second preset fit degree is less than the first preset fit degree.

4. The multi-camera co-location based material cutting method of claim 1, wherein, S4 specifically includes: S401: Perform surface fitting and reconstruction on the noisy global point cloud data, and reconstruct the point cloud data while conforming to the basic geometric shape of the global point cloud data; S402: Sort the reconstructed point cloud data to generate the cutting path.

5. The multi-camera co-location based material cutting method of claim 1, wherein, Specifically, S5 is: The cutting path is smoothed to improve curvature smoothness, point spacing uniformity, and offset retention.

6. The multi-camera co-location based material cutting method of claim 1, wherein, Specifically, S7 is: Based on the overall observation results of the workpiece by the contour camera, the expected pose of the cutting path in the world coordinate system is compared with the actual placement posture of the workpiece. The overall position and orientation of the cutting path are corrected by rigid body transformation, and the global path pose calibration of the cutting path is performed.

7. The multi-camera co-location based material cutting method of claim 1, wherein, Specifically, S8 is: A high-precision local scan is performed using a 3D camera at the end of each robot. Based on the local point cloud obtained from the scan, the real cutting feature points near the current cutting tool are extracted, and the position error of the real cutting feature points in the robot end coordinate system is calculated. The position error is used as a local compensation amount, and Kalman filtering is used to adjust the robot end motion trajectory in real time to perform fine local path compensation on the cutting path.

8. A multi-camera cooperative positioning based material cutting system, comprising: include: Water tank base, support platform, guide rail, industrial cutting robot, crossbeam, vision positioning module, waste recycling module, processor and memory; The support platform is located above the water tank base. The support platform is used to support the workpiece to be cut, and the water tank base is used to absorb the wastewater generated during the cutting process and settle some of the dust. The water tank base is provided with guide rails on both sides, and the industrial cutting robot is installed on each guide rail. The industrial cutting robot moves independently or synchronously on the guide rail. A detachable crossbeam is provided between the two industrial cutting robots; The visual positioning module includes a contour camera and a 3D camera. Multiple contour cameras are spaced apart on the crossbeam, and the 3D camera is mounted on the end effector of the industrial cutting robot. The contour camera is used to quickly acquire the overall three-dimensional point cloud data of the workpiece; The 3D camera is used to perform a fine scan of the area to be cut in order to obtain local geometric information; The waste recycling module is used to collect wastewater and debris in real time during the cutting process to keep the working area clean; The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the material cutting method based on multi-camera cooperative positioning as described in any one of claims 1 to 7.

Images