A high-precision edge three-dimensional reconstruction method for groove cutting

Abstract

This invention discloses a high-precision 3D edge reconstruction method for bevel cutting. The method first acquires the original point cloud and grayscale image of the workpiece using a binocular vision system; then, it uses the filtered clean point cloud to fit the plane equation of the workpiece surface as a geometric constraint; next, it performs adaptive feature enhancement on the grayscale image to extract high-precision sub-pixel 2D edges; finally, it combines the plane equation and the 2D edges, calculates the 3D coordinates of the edges using a ray intersection model, and generates the robot cutting trajectory. This invention effectively solves the problems of point cloud loss and distortion caused by edge burrs, defects, and reflections, achieving sub-millimeter-level edge reconstruction and significantly improving the consistency of bevel cutting edges and trajectory accuracy without manual intervention.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and industrial visual inspection, and more specifically, to a method for high-precision three-dimensional reconstruction of beveled edges. Background Technology

[0002] In heavy industries such as shipbuilding, marine engineering, and large steel bridge construction, metal sheets must be beveled before welding to ensure weld penetration and joint strength. Traditional beveling methods often use mechanical planers or semi-automatic flame cutting machines, which suffer from low efficiency, poor precision, and insufficient flexibility. With the development of intelligent manufacturing, laser or plasma beveling workstations based on industrial robots are gradually becoming mainstream. In these workstations, high-precision vision guidance is the core of automated cutting. Typically, a structured light 3DS scanner is used to acquire point cloud data of the workpiece surface, and algorithms are used to extract the bevel edge trajectory to guide the robot in cutting. The accuracy of the reconstructed bevel edge directly determines the uniformity of the root face and the accuracy of the bevel angle after cutting, thus affecting the subsequent welding quality.

[0003] The technical solutions of this invention and existing technologies involve the following well-known technologies: MEMS structured light projection: This involves projecting a specifically coded stripe pattern onto the surface of an object using a micro-mirror of a micro-electromechanical system (MEMS), acquiring deformed stripe images using a binocular camera, and calculating the three-dimensional depth information of the object's surface using triangulation principles. Point cloud: This refers to a massive set of points representing the spatial distribution and surface characteristics of a target within the same spatial reference frame. Hand-eye calibration: This is a technique for determining the relative positional relationship between the coordinate system of a vision sensor and the coordinate system of the robot's end effector flange or base. Commonly used solution methods include equation-based solutions based on rotation matrices.

[0004] A search revealed several existing technical solutions for visual inspection or edge repair of bevel cutting. For example, Chinese invention patent application CN113305849A discloses a smart bevel cutting system and method based on composite vision. This application utilizes a composite vision system composed of 3D vision, 2D vision, and structured light sensors to perform position correction and trajectory planning for parts. Although this solution introduces multiple visual methods to improve the positioning success rate, when dealing with point cloud gaps caused by physical defects at workpiece edges (such as oxide scale and burrs), it mainly relies on matching and fitting the overall model, lacking a high-precision sub-pixel-level repair mechanism for local edges. Chinese invention patent application CN116542914A discloses a weld extraction and fitting method based on 3D point clouds. This method preprocesses the point cloud through point cloud stitching, statistical filtering, and uniform sampling, and uses cubic B-spline curves to perform segmented fitting of the weld trajectory. This scheme is suitable for situations where the weld seam trajectory is relatively complete. However, when irregular gaps appear at the workpiece edge due to the heat of cutting or collision, or when high light reflection causes a large area of ​​point cloud to be lost, simple B-spline fitting is prone to producing trajectory smoothing over or local oscillation, making it difficult to guarantee the geometric true value of the edge.

[0005] Although the aforementioned existing technologies have solved the positioning and trajectory generation problems of automated cutting to some extent, the following major drawbacks still exist in the complex industrial beveling scenarios:

[0006] Insufficient anti-interference capability and poor edge point cloud quality: In actual working conditions, the edges of the sheet metal to be cut often have burrs due to the rolling process, or local highlights or black holes due to bumps during hoisting or rusting. Existing vision algorithms are mostly based on filtering the overall point cloud, which makes it difficult to effectively separate the real point cloud of the workpiece edge from the burr noise, resulting in "false edges" or severe noise fluctuations in the extracted edge lines.

[0007] Low accuracy in repairing missing areas leads to inconsistent bevel edges: When point clouds are missing in edge areas due to reflection or occlusion, existing commercial algorithms typically use linear interpolation or polynomial fitting to fill the gaps. This method ignores the geometric constraints of the workpiece surface, and the generated virtual edges often deviate from the actual physical edges, resulting in inconsistent widths of the bevel edges (blunt edges) after cutting. In severe cases, secondary manual grinding is required, increasing manufacturing costs.

[0008] Large cumulative error in hand-eye calibration: Traditional hand-eye calibration methods based on matrix decomposition often separate the calculation of rotation and translation, which easily introduces cumulative calculation errors and makes it difficult to meet the stringent requirements of bevel cutting for trajectory accuracy to reach the sub-millimeter level (such as within ±0.5mm). Summary of the Invention

[0009] To address the chain-like defects in existing beveling techniques, such as missing or distorted point clouds due to burrs, corrosion, or defects on the workpiece edges, leading to large beveling trajectory deviations, inconsistent edge widths, and poor cutting accuracy, this invention proposes a high-precision 3D edge reconstruction method for beveling. This method integrates feature enhancement, edge topology repair, and iterative fitting based on physical constraints. By combining adaptive feature enhancement in the image domain with sub-pixel edge extraction technology, and applying physical geometric constraints using the workpiece surface plane equations for 2D-3D joint reconstruction, it effectively solves the problems of low reconstruction accuracy and trajectory jitter caused by incomplete edge data under complex working conditions. This achieves sub-millimeter level high-precision automated beveling without manual intervention. The high-precision 3D edge reconstruction method for beveling described in this invention includes the following steps:

[0010] S1: Data acquisition and preprocessing. The workpiece is scanned using a calibrated binocular vision system to obtain the original point cloud and the distortion-free grayscale image of the workpiece. The original point cloud is a three-dimensional point set reconstructed from the absolute phase map calculated based on Gray code and phase shift code.

[0011] S2: Workpiece surface plane fitting. The original point cloud from step S1 is filtered to remove outliers and background points, resulting in a clean point cloud containing only workpiece surface data. Plane fitting is performed on the clean point cloud using the least squares method to obtain the workpiece plane equation in the camera coordinate system. The calculation formula is as follows:

[0012]

[0013] in, All are planar parameters, and satisfy... ;

[0014] S3: Subpixel-level two-dimensional edge extraction. Adaptive feature enhancement processing is applied to the distortion-reduced grayscale image from step S1 to obtain the enhanced image. The calculation formula for adaptive feature enhancement processing is:

[0015]

[0016] in, This is the original, distorted grayscale image. To enhance the image, To enhance the coefficient, It is a constant. It is a local Gaussian weighted mean. The standard deviation is a local Gaussian weighted average.

[0017] In the enhanced image, sub-pixel extremum locations are calculated along the gradient direction to extract the sub-pixel edge point set of the workpiece edge in the image pixel coordinate system. ;

[0018] S4: 3D reconstruction of edge points, from the sub-pixel edge point set in step S3 Each point in Convert to camera normalized planar coordinates Combined with the workpiece plane equation in step S2, the coordinates of the edge three-dimensional points in the camera coordinate system are calculated using the ray-plane intersection model. The formula for calculating the coordinates of three-dimensional points on the edge is:

[0019]

[0020] S5: Cutting trajectory generation, based on the pre-calibrated hand-eye transformation matrix, converts the edge 3D point coordinates from step S4. The final bevel cutting trajectory point sequence is generated by converting to the tool coordinate system of the industrial robot's end-effector cutting gun.

[0021] As a preferred embodiment of the present invention, step S1 involves scanning the workpiece using a calibrated binocular vision system, specifically including the following steps:

[0022] A hybrid encoding strategy is adopted to control the projection component of the binocular vision system to sequentially project Gray code pattern sequence, Gray code complement pattern sequence, and binary line shift auxiliary code pattern sequence onto the workpiece surface; the image sequence acquired by the binocular vision system is decoded to calculate a highly robust absolute phase map; based on the principle of binocular vision stereo matching, the three-dimensional coordinates of the workpiece surface are calculated using the absolute phase map to generate the original point cloud.

[0023] As a preferred embodiment of the present invention, step S2 involves filtering the original point cloud from step S1 to remove outliers and background points, specifically including the following sub-steps:

[0024] S2-1: Use the Canny edge detection algorithm to perform edge detection on the absolute phase map obtained in step S1 and extract two-dimensional boundary points;

[0025] S2-2: Calculate the unit vector of the gradient direction at the two-dimensional boundary points. and the direction of the outer normal The calculation formula is:

[0026]

[0027] in, This represents the gradient of the current point along the x and y directions of the image.

[0028] S2-3: Offset each two-dimensional boundary point along the outward normal direction. 1 pixel, to obtain the new boundary point The calculation formula is:

[0029]

[0030] S2-4: Based on the new boundary point set after offset, the two-dimensional region is defined as the points projected outside the two-dimensional region in the original point cloud and removed, while the points inside the region are retained as the point cloud to be processed.

[0031] As a preferred embodiment of the present invention, in step S3, adaptive feature enhancement processing is performed on the distortion-free grayscale image from step S1. The specific calculation process is as follows: an adaptive enhancement model is constructed, and local statistics are used to dynamically amplify image details. The calculation formula is:

[0032]

[0033] in, This indicates the enhanced output image at the pixel level. grayscale value at that location This indicates the original distorted grayscale image at the pixel level. grayscale value at that location Indicates the enhancement coefficient. Indicates a self-reinforcing term; self-reinforcing term The calculation formula is:

[0034]

[0035] in, Represents pixels Local mean at that location Represents pixels Local standard deviation at , It is a very small constant used to prevent the denominator from being zero;

[0036] Local mean and local standard deviation All calculations are based on Gaussian weighting functions, defining pixel points. The neighborhood window is The calculation formulas are as follows:

[0037]

[0038]

[0039] in, For neighborhood windows Pixel coordinates within, The weighting function is Gaussian.

[0040] Gaussian weighting function The calculation formula is:

[0041]

[0042] in, These represent the neighboring pixels and the center pixel respectively. shaft and Distance along the axis, The standard deviation parameter represents the Gaussian function.

[0043] As a preferred embodiment of the present invention, step S3, which calculates the sub-pixel extreme value position along the gradient direction and extracts the sub-pixel edge point set of the workpiece edge in the image pixel coordinate system, specifically includes the following steps:

[0044] Computational image enhancement gradient magnitude of each pixel and gradient direction angle The specific calculation formula is as follows:

[0045]

[0046] in, The images are respectively in direction and The first gradient of the direction;

[0047] Determine the current edge point gradient direction unit vector The specific calculation formula is as follows:

[0048]

[0049] unit vector along the gradient direction At the current edge point Gradient magnitudes were obtained by sampling from both sides and the center. The specific calculation formulas are as follows:

[0050]

[0051] Constructing a quadratic function Fit the gradient magnitudes of the three sampling points, where, Calculate the offset of the extreme value position of the quadratic function. The specific calculation formula is as follows:

[0052]

[0053] The coordinates of the current edge point are corrected based on the extreme value position offset to obtain the sub-pixel edge point coordinates. The specific calculation formula is as follows:

[0054]

[0055] in, This represents the offset of the extreme value position. For the current edge point, These are the coordinates of the subpixel edge point.

[0056] As a preferred embodiment of the present invention, after extracting the sub-pixel edge point set, the false edge points are further removed. This removal operation specifically includes the following steps:

[0057] Calculate any two edge points Directional difference between The calculation formula is:

[0058]

[0059] in, Points The gradient direction angle, This indicates that the angle is normalized to Functions over an interval;

[0060] Statistical analysis of each edge point Within its radius Within the neighborhood of a given point, the number of neighboring points that satisfy the direction consistency condition is considered if this number is less than a preset threshold. Then determine the edge point. False edge points are identified and removed; the formula for calculating the judgment criteria is as follows:

[0061]

[0062] in, Point With point The Euclidean distance between them The angle threshold for determining directional consistency. This represents a counting function.

[0063] As a preferred embodiment of the present invention, after removing false edge points, the method further includes the steps of edge chain growth and gap filling of the sub-pixel edge point set. The specific calculation process is as follows:

[0064] Based on connection weight Connect discrete sub-pixel edge points into an edge chain, where any two edge points... Connection weights between The calculation formula is:

[0065]

[0066] in, These are the distance weighting coefficient and the direction weighting coefficient, respectively. Point With point The Euclidean distance between them The maximum distance threshold allowed for connections. For edge points The difference in gradient direction between them;

[0067] For the generated broken edge chain, calculate the endpoints of any two edge chains. Connection cost between If the cost is less than a preset threshold Then, connect these two endpoints and fill the gap; connection cost The calculation formula is:

[0068]

[0069] in, These are the weight coefficients for the distance term, angle term, and gradient confidence term, respectively. The normalized distance constant, As endpoints The gradient direction difference at that point The cost is the gradient confidence level.

[0070] Gradient confidence cost The calculation formula is:

[0071]

[0072] in, The number of sampling points on the line connecting the two endpoints. For sampling index, The gradient magnitude of the sampling point location in the enhanced image. This indicates a floor operation to obtain the pixel coordinates. This represents the maximum gradient magnitude across the entire graph.

[0073] As a preferred embodiment of the present invention, after edge chain growth and gap filling, the method further includes a step of regularizing and smoothing the edge chains. The specific calculation process is as follows:

[0074] A one-dimensional Gaussian weighted smoothing algorithm is used to perform convolution processing on the point sequence on the edge chain to remove edge jitter; for any point on the edge chain Its smoothed coordinates The calculation formula is:

[0075]

[0076] in, To smooth the window radius, For edge chains The coordinates of the points in the neighborhood of the center;

[0077] Weighting function The Gaussian function is a one-dimensional function, and its calculation formula is:

[0078]

[0079] in, To control the standard deviation parameter of smoothness, For neighborhood points relative to the center point The index offset.

[0080] As a preferred embodiment of the present invention, the hand-eye transformation matrix in step S5 is pre-calculated using a hand-eye calibration method based on dual quaternions. The specific calculation process is as follows:

[0081] Controlling an industrial robot to drive a binocular vision system to photograph a calibration board in different poses, acquiring multiple sets of robot end-effector poses. and camera observation pose ; Position the robot's end effector and camera observation pose Convert each to dual quaternion form and Construct a dual quaternion equation that satisfies the hand-eye relationship; the calculation formula is as follows:

[0082]

[0083] in, The hand-eye transformation dual quaternion to be solved is... This represents the operation of dual quaternion multiplication;

[0084] The dual quaternion is obtained by solving the dual quaternion equation. The real and dual parts are then separated to obtain the rotation matrix. Translation vector To form a hand-eye transformation matrix .

[0085] Compared with the relevant prior art, the beneficial effects of the present invention are:

[0086] This invention significantly improves edge reconstruction accuracy and solves the trajectory jitter problem caused by missing point clouds. Existing technologies often directly interpolate and fit the missing point cloud regions, which is prone to errors. This invention employs a joint reconstruction strategy of "2D sub-pixel edges + 3D planar constraints," utilizing the fitted plane equation of the workpiece surface as a geometric constraint, combined with high-precision sub-pixel edge rays extracted from the image domain for solving. This method effectively avoids fitting deviations caused by physical missing or sparse edge point clouds, controlling edge reconstruction errors within a limited range. It is significantly superior to traditional interpolation algorithms, ensuring the smoothness and accuracy of the bevel trajectory.

[0087] The algorithm's robustness and anti-interference capabilities under harsh working conditions are enhanced. Addressing common interference factors on workpiece edges in industrial environments, such as burrs, rust, highlights, or defects, this invention introduces an adaptive feature enhancement algorithm based on local statistics in the image processing stage to dynamically enhance edge contrast. Simultaneously, combined with a filtering strategy based on Canny edge detection and outward normal direction offset, it effectively eliminates false edges and background noise. This allows the system to adapt to various steel surface conditions of different thicknesses and strengths without requiring manual patching or grinding.

[0088] This invention improves the overall positioning and cutting accuracy of the system. In the system calibration stage, it employs a hand-eye calibration method based on dual quaternions. Compared to traditional methods that separate rotation and translation calculations, this method unifies the calculation of rotation and translation, reduces accumulated errors, and exhibits higher numerical stability. Combined with a high-precision vision algorithm, the final overall cutting accuracy of the workpiece bevel (including edge cut error) is maintained at a high level. Within this range, the consistency of the distance along the edge is excellent.

[0089] This invention achieves fully automated operation, significantly reducing production costs and improving efficiency. It can automatically generate high-precision robot cutting trajectories, and the algorithm has good real-time performance, meeting [the requirements of]... The beveling cutting requirement is met. Due to the high cutting quality and consistent edge width, the secondary manual grinding step required in traditional processes is eliminated, significantly reducing manufacturing costs and improving the first-pass yield rate. Attached Figure Description

[0090] Figure 1 A flowchart of a high-precision three-dimensional reconstruction method for bevel cutting provided by the present invention;

[0091] Figure 2 The diagrams provided in this embodiment of the invention are experimental process diagrams, wherein 2-A is the original image, 2-B is the phase image, 2-C is the original point cloud image, and 2-D is the edge cutting line image after three-dimensional reconstruction;

[0092] Figure 3 The images provided in this embodiment of the invention are as follows: 3-A is an actual on-site environment image, 3-B is an original workpiece image, and 3-C is a workpiece image after being cut using the applied algorithm. Detailed Implementation

[0093] The solutions provided by the present invention will be further described below with reference to the accompanying drawings. However, the present invention can be implemented in many different ways and should not be construed as limited to the embodiments shown; rather, these embodiments provide those skilled in the art with implementation methods that meet applicable legal requirements.

[0094] Example 1: This example provides a high-precision 3D edge reconstruction method for beveling in complex industrial environments. This method relies on an intelligent robot workstation system integrating visual perception and execution capabilities, aiming to solve the automated cutting challenges caused by poor workpiece surface quality in shipbuilding and steel structure manufacturing. Please refer to the appendix of the instruction manual. Figure 1 To be continued Figure 3 The physical carrier in this embodiment is as follows: Figure 3 As shown, it mainly includes a high-precision six-axis industrial robot, a binocular structured light vision sensing system installed at the robot's end effector, and a dedicated flame or plasma cutting torch. In actual operation scenarios, such as... Figure 3 As shown in Figure -A, the industrial robot, acting as a motion actuator, has a vision sensor and a cutting gun rigidly connected to its end flange. These two components are precisely designed with an optical-mechanical connection to maintain a fixed relative position, enabling them to adapt to factory environments filled with metal dust and electromagnetic interference. The workpiece to be processed is placed on the grid-like cutting platform below, as shown... Figure 3 As shown in -B, these workpieces are usually thick steel plates that have undergone preliminary cutting. Their surfaces are covered with dark brown oxide scale and rust. Some areas even have oil stains, white marker pen marks, and edge chips caused by bumps during hoisting. This "dirty, messy, and poor" surface condition is an obstacle that traditional visual algorithms find difficult to overcome.

[0095] To achieve high-precision bevel cutting, the system must first execute a strict initialization process, namely... Figure 1 Step S1 involves system calibration. In this stage, the operator controls the robot to move the vision sensor to capture images of the high-precision calibration board in different poses. The system uses Zhang's calibration method to calculate the intrinsic parameters and distortion coefficients of the left and right cameras, eliminating the influence of lens distortion on measurement accuracy. Subsequently, a high-precision hand-eye calibration algorithm based on dual quaternions is used to establish a rigid transformation relationship between the vision sensor coordinate system and the robot base coordinate system. This process uniformly solves for the rotation and translation components, effectively suppressing the accumulation of calibration errors and ensuring that the edge coordinates "seen" by the vision system can be accurately executed by the robot's "hand," laying a spatial reference for subsequent sub-millimeter-level cutting.

[0096] After calibration is complete, the system enters an automatic operation cycle, first executing... Figure 1 Step S2 involves acquiring the original point cloud. When the robot moves above the workpiece, the MEMS projection component projects a series of specially designed coded patterns onto the workpiece surface at high speed, including Gray code sequences, Gray code complement sequences, and binary line shifting auxiliary code sequences. This hybrid encoding strategy effectively suppresses high-gloss reflections from the metal surface, improving the robustness of decoding. A binocular camera simultaneously acquires these stripe images, such as... Figure 2As shown in -A, due to the corrosion and light absorption characteristics of the workpiece surface, the overall brightness of the original left camera image is low. The grid texture of the cutting platform in the background is mixed with the workpiece edge, and the edge appears as a blurry dark area due to the reflection from the chamfer. The processing unit inside the system then decodes these image sequences to generate images such as... Figure 2 -B represents the absolute phase map. Compared to the original grayscale image, the workpiece outline in the phase map is clearer, and the surface rust texture is transformed into continuous phase information, providing a reliable basis for depth calculation. Based on this phase map and the principle of binocular stereo matching, the system further generates, as shown in Figure 1. Figure 2 -C represents the original 3D point cloud. Figure 2 As can be seen in -C, although the point cloud roughly restores the sheet-like structure of the workpiece, it is limited by physical defects at the edges (such as burrs and occlusions). The point cloud edges exhibit severe jaggedness, discretization, and even large gaps, and are accompanied by outlier noise caused by the background grid. If the cutting trajectory is directly generated based on this point cloud, it will inevitably cause the cutting gun to vibrate violently, which will not meet the process requirements of welding bevel.

[0097] To address the issue of poor quality in the original point cloud, this embodiment then proceeds to... Figure 1 The core step S3 is the 3D reconstruction of the edge point cloud. The system first filters and cleans the original point cloud. Using edge gradient information from the phase map, it calculates the outward normal direction of the edge and offsets it outward by a certain number of pixels to construct a mask region. Background noise projected outside this region is removed, resulting in a clean point cloud containing only the workpiece itself. Subsequently, given the unreliability of the workpiece edge data, the system employs a "roundabout" strategy, using the least squares method to fit the point cloud of the flat area inside the workpiece, solving for a high-precision plane equation representing the ideal surface of the workpiece. This plane serves as the spatial geometric constraint for subsequent edge reconstruction. Next, the algorithm focuses on the 2D image domain, performing adaptive feature enhancement processing on the distortion-free grayscale image. This processing uses a local Gaussian weighted mean and standard deviation to construct a self-enhancing term, dynamically stretching the contrast of the edge region, revealing edge details that were previously hidden in darkness. Based on this, the system performs sub-pixel-level extreme value search along the gradient direction, using a quadratic curve fitting algorithm to improve the edge localization accuracy from the pixel level to the sub-pixel level, extracting continuous and detailed 2D edge point sequences.

[0098] In order to cope with Figure 3-B. For physical defects on the workpiece edges (such as chipped material or dents left by gas cutting), the system further performs edge topology repair and gap filling algorithms in the image domain. By calculating the Euclidean distance between the endpoints of the broken edges, directional consistency, and gradient confidence cost, the system can intelligently identify broken segments belonging to the same edge and use smooth curves to "fill in" the gaps, restoring the original geometry of the workpiece, thus correcting the physical defects of the raw material at the software level. Finally, the system performs joint 2D and 3D reconstruction, projecting the repaired high-precision sub-pixel 2D edge points along the camera's projection ray onto the previously fitted 3D workpiece plane. By solving for the intersection of the ray and the plane, the final 3D coordinates of the edge are calculated. Figure 2 As shown in -D, the reconstructed edge lines obtained after this series of processing exhibit a perfect L-shaped structure, with smooth lines, no noise, and no breaks, in line with... Figure 2 The incomplete original point cloud in -C forms a stark contrast, perfectly reproducing the true outline of the workpiece.

[0099] Finally, the system executes. Figure 1 Step S4 in the process generates the cutting trajectory. Based on the reconstructed high-precision edge data and preset beveling process parameters (such as beveling angle and blunt edge height), the system automatically plans the robot's six-axis motion trajectory. Figure 3 In the on-site implementation shown in -C, the robot-guided cutting gun completed the cutting operation stably and smoothly. Observation of the cut workpiece reveals a bright metallic luster on the cut surface, indicating a good match between cutting speed and energy. Crucially, the width of the blunt edge below the cut remained consistently high throughout the entire L-shaped path, without any "wide or narrow edge" or "missed cuts" caused by edge recognition errors. The rounded corners were smooth and rounded, without any sharp edges. Actual measurements showed that the edge reconstruction error was controlled within ±0.15mm, and the overall cutting accuracy met the stringent standard of ±0.5mm. The entire process required no manual intervention for grinding or patching, truly achieving automated and efficient "scan and cut" operation, verifying the superior performance of the method in this embodiment under harsh working conditions.

[0100] Example 2: Building upon Example 1, this example further demonstrates the adaptability and robustness of the invention in dealing with extreme edge interference and workpieces of different materials through a specific complex working condition process. In actual shipbuilding sites, in addition to common rusted carbon steel plates, it is often necessary to process stainless steel plates with strong surface reflection or thick plates with severe oxide scale. The edges of these workpieces are often abnormally rough due to the slag (fouling) from the previous plasma cutting, forming a large number of pseudo-edge interferences. This example will focus on describing how the algorithm achieves accurate edge extraction in such scenarios through logical judgment and parameter optimization.

[0101] During the data acquisition phase, for workpieces with strong surface reflection or extremely high blackness, the system automatically adjusts the exposure time and stripe contrast of the MEMS projection component, employing multi-exposure fusion technology to acquire high dynamic range original images, ensuring no overexposure in highly reflective areas and clear texture in dark areas. After generating the original point cloud, traditional statistical filtering often struggles to distinguish between real edges and burrs closely adhering to the edges when faced with "burr" point clouds caused by slag at the edges. This embodiment introduces a pseudo-edge removal mechanism based on direction consistency in step S3. In the sub-pixel edge extraction stage, the system not only calculates the gradient magnitude of the edge point but also rigorously verifies the gradient direction consistency within the neighborhood of that point. Since slag and burrs are usually chaotic, their gradient directions change drastically in a local range, while real workpiece edges (even curves) have directional continuity at the microscale. Utilizing this characteristic, the algorithm sets a direction consistency threshold, identifying chaotic edge points in the neighborhood as "pseudo-edges" and removing them, thus scraping away the noise attached to the edges like a "razor," retaining only the geometric contour of the workpiece body.

[0102] Furthermore, addressing the common edge "jump" problem in thick plate cutting—that is, due to the large thickness of the plate and poor side perpendicularity, the vision system easily misjudges the bottom edge of the plate as the top surface edge—this embodiment utilizes the high-precision top surface plane equation fitted in step S2 as a strong geometric constraint. During 2D to 3D reconstruction, any solution point deviating from this plane by more than a certain threshold (e.g., 0.5mm) will be forcibly corrected or discarded. This means that even if the image algorithm mismatches in a certain local area and captures the texture of the plate side, it will be filtered out by the physical constraint mechanism because the point cannot fall on the "workpiece top surface" reference plane in space. This dual insurance mechanism of "image features + physical constraints" greatly improves the algorithm's anti-interference ability.

[0103] In terms of implementation results, this embodiment demonstrates the process of repairing and cutting a steel plate with continuous serrated edges and an oil-stained surface. During image processing, the adaptive enhancement algorithm successfully suppressed localized reflective spots caused by the oil stains. In the topology connection stage, the algorithm identified the collinearity of the edges on both sides of the serrated edges and smoothly connected the breakpoints using Bézier curve interpolation, generating a straight virtual cutting path. The final cutting result shows that the robot smoothly removed the original serrated edges, resulting in a smooth and clean bevel surface with uniform edge dimensions. Compared to the control group using the traditional interpolation method, this embodiment not only avoids the frequent acceleration and deceleration vibrations caused by the cutting gun chasing the serrated edges but also saves approximately 30 minutes of subsequent manual welding and grinding time, improving the single-piece processing efficiency by more than 40%. This result fully demonstrates that the method proposed in this embodiment is not only a visual measurement algorithm but also an intelligent manufacturing solution capable of tolerating upstream process defects and possessing process correction capabilities, greatly improving the flexibility and intelligence level of the automated cutting workstation.

[0105] The above embodiments merely illustrate implementation methods of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A high-precision edge three-dimensional reconstruction method for groove cutting, characterized in that: The method includes the following steps: S1: Data acquisition and preprocessing. The workpiece is scanned using a calibrated binocular vision system to obtain the original point cloud and the distortion-free grayscale image of the workpiece. The original point cloud is a three-dimensional point set reconstructed from the absolute phase map calculated based on Gray code and phase shift code. S2: Workpiece surface plane fitting. The original point cloud from step S1 is filtered to remove outliers and background points, resulting in a clean point cloud containing only workpiece surface data. Plane fitting is performed on the clean point cloud using the least squares method to obtain the workpiece plane equation in the camera coordinate system. The calculation formula is as follows: wherein are plane parameters and satisfy ; S3: Subpixel-level two-dimensional edge extraction. Adaptive feature enhancement processing is applied to the distortion-reduced grayscale image from step S1 to obtain the enhanced image. The calculation formula for adaptive feature enhancement processing is: wherein, is the original undistorted grayscale image, is the enhanced image, is the enhancement coefficient, is a constant, is the local Gaussian weighted mean, is the local Gaussian weighted standard deviation; On the enhanced image, the sub-pixel extreme position is calculated along the gradient direction, and a set of sub-pixel edge points of the workpiece edge in the image pixel coordinate system is extracted ; S4: three-dimensional reconstruction of edge points, converting each point in the sub-pixel edge point set in step S3 into a camera normalized plane coordinate , and combining the workpiece plane equation in step S2, calculating the edge three-dimensional point coordinates in the camera coordinate system through a ray and plane intersection model ; the calculation formula of the edge three-dimensional point coordinates is:​​ S5: Cutting trajectory generation, based on the pre-calibrated hand-eye transformation matrix, converts the edge 3D point coordinates from step S4. The final bevel cutting trajectory point sequence is generated by converting to the tool coordinate system of the industrial robot's end-effector cutting gun.

2. The method for high-precision three-dimensional reconstruction of bevel cutting edges according to claim 1, characterized in that: Step S1 involves scanning the workpiece using a calibrated binocular vision system, specifically including the following steps: A hybrid coding strategy is adopted to control the projection component of the binocular vision system to sequentially project Gray code pattern sequence, Gray code complement code pattern sequence and binary line shift auxiliary code pattern sequence onto the workpiece surface. The image sequence acquired by the binocular vision system is decoded to calculate a robust absolute phase map; Based on the principle of stereo matching in binocular vision, the three-dimensional coordinates of the workpiece surface are calculated using an absolute phase map to generate the original point cloud.

3. The method for high-precision three-dimensional reconstruction of bevel cutting edges according to claim 1, characterized in that: Step S2 involves filtering the original point cloud from step S1 to remove outliers and background points. This includes the following sub-steps: S2-1: Use the Canny edge detection algorithm to perform edge detection on the absolute phase map obtained in step S1 and extract two-dimensional boundary points; S2-2: Calculate the unit vector of the gradient direction at the two-dimensional boundary points. and the direction of the outer normal The calculation formula is: in, This represents the gradient of the current point along the x and y directions of the image. S2-3: Offset each two-dimensional boundary point along the outward normal direction. 1 pixel, to obtain the new boundary point The calculation formula is: S2-4: Based on the new boundary point set after offset, the two-dimensional region is defined as the points projected outside the two-dimensional region in the original point cloud and removed, while the points inside the region are retained as the point cloud to be processed.

4. The method for high-precision three-dimensional reconstruction of bevel cutting edges according to claim 1, characterized in that: In step S3, adaptive feature enhancement processing is performed on the distortion-free grayscale image from step S1. The specific calculation process is as follows: an adaptive enhancement model is constructed, and local statistics are used to dynamically amplify image details. The calculation formula is: in, This indicates the enhanced output image at the pixel level. grayscale value at that location This indicates the original distorted grayscale image at the pixel level. grayscale value at that location Indicates the enhancement coefficient. Indicates a self-reinforcing term; self-reinforcing term The calculation formula is: in, Represents pixels Local mean at that location Represents pixels Local standard deviation at , It is a very small constant used to prevent the denominator from being zero; Local mean and local standard deviation All calculations are based on Gaussian weighting functions, defining pixel points. The neighborhood window is The calculation formulas are as follows: in, For neighborhood windows Pixel coordinates within, The weighting function is Gaussian. Gaussian weighting function The calculation formula is: in, These represent the neighboring pixels and the center pixel respectively. shaft and Distance along the axis, The standard deviation parameter represents the Gaussian function.

5. The method for high-precision three-dimensional reconstruction of bevel cutting edges according to claim 1, characterized in that: Step S3 involves calculating the sub-pixel extremum positions along the gradient direction and extracting the sub-pixel edge point set of the workpiece edge in the image pixel coordinate system. This specifically includes the following steps: Computational image enhancement gradient magnitude of each pixel and gradient direction angle The specific calculation formula is as follows: in, The images are respectively in direction and The first gradient of the direction; Determine the current edge point gradient direction unit vector The specific calculation formula is as follows: unit vector along the gradient direction At the current edge point Gradient magnitudes were obtained by sampling from both sides and the center. The specific calculation formulas are as follows: Constructing a quadratic function Fit the gradient magnitudes of the three sampling points, where, Calculate the offset of the extreme value position of the quadratic function. The specific calculation formula is as follows: The coordinates of the current edge point are corrected based on the extreme value position offset to obtain the sub-pixel edge point coordinates. The specific calculation formula is as follows: in, This represents the offset of the extreme value position. For the current edge point, These are the coordinates of the subpixel edge point.

6. The method for high-precision three-dimensional reconstruction of bevel cutting edges according to claim 5, characterized in that: After extracting the sub-pixel edge point set, the false edge points are further removed. This removal operation specifically includes the following steps: Calculate any two edge points Directional difference between The calculation formula is: in, Points The gradient direction angle, This indicates that the angle is normalized to Functions over an interval; Statistical analysis of each edge point Within its radius Within the neighborhood of a given point, the number of neighboring points that satisfy the direction consistency condition is considered if this number is less than a preset threshold. Then determine the edge point. False edge points are identified and removed; the formula for calculating the judgment criteria is as follows: in, Point With point The Euclidean distance between them The angle threshold for determining directional consistency. This represents a counting function.

7. The method for high-precision three-dimensional reconstruction of bevel cutting edges according to claim 6, characterized in that: After removing false edge points, the process also includes edge chain growth and gap filling of the sub-pixel edge point set. The specific calculation process is as follows: Based on connection weight Connect discrete sub-pixel edge points into an edge chain, where any two edge points... Connection weights between The calculation formula is: in, These are the distance weighting coefficient and the direction weighting coefficient, respectively. Point With point The Euclidean distance between them The maximum distance threshold allowed for connections. For edge points The difference in gradient direction between them; For the generated broken edge chain, calculate the endpoints of any two edge chains. Connection cost between If the cost is less than a preset threshold Then, connect these two endpoints and fill the gap; connection cost The calculation formula is: in, These are the weight coefficients for the distance term, angle term, and gradient confidence term, respectively. The normalized distance constant, As endpoints The gradient direction difference at that point The cost is the gradient confidence level. Gradient confidence cost The calculation formula is: in, The number of sampling points on the line connecting the two endpoints. For sampling index, The gradient magnitude of the sampling point location in the enhanced image. This indicates a floor operation to obtain the pixel coordinates. This represents the maximum gradient magnitude across the entire graph.

8. The method for high-precision three-dimensional reconstruction of bevel cutting edges according to claim 7, characterized in that: After edge chain growth and gap filling, the process also includes a step of regularizing and smoothing the edge chains. The specific calculation process is as follows: A one-dimensional Gaussian weighted smoothing algorithm is used to perform convolution processing on the point sequence on the edge chain to remove edge jitter; for any point on the edge chain Its smoothed coordinates The calculation formula is: in, To smooth the window radius, For edge chains The coordinates of the points in the neighborhood of the center; Weighting function The Gaussian function is a one-dimensional function, and its calculation formula is: in, To control the standard deviation parameter of smoothness, For neighborhood points relative to the center point The index offset.

9. The method for high-precision three-dimensional reconstruction of bevel cutting edges according to claim 1, characterized in that: In step S5, the hand-eye transformation matrix is ​​pre-calculated using a hand-eye calibration method based on dual quaternions. The specific calculation process is as follows: Controlling an industrial robot to drive a binocular vision system to photograph a calibration board in different poses, acquiring multiple sets of robot end-effector poses. and camera observation pose ; Position the robot's end effector and camera observation pose Convert each to dual quaternion form and Construct a dual quaternion equation that satisfies the hand-eye relationship; the calculation formula is as follows: in, The hand-eye transformation dual quaternion to be solved is... This represents the operation of dual quaternion multiplication; The dual quaternion is obtained by solving the dual quaternion equation. The real and dual parts are then separated to obtain the rotation matrix. Translation vector To form a hand-eye transformation matrix .

Images