Machine vision-based high-precision measurement method for shield machine cutter head component size

Abstract

The application discloses a shield cutter component size high-precision measurement method based on machine vision and belongs to the technical field of industrial vision measurement. The method comprises the following steps: coating a random speckle coating containing fluorescent particles on the surface of the cutter component; synchronously collecting speckle fluorescence images under the excitation of an ultraviolet light source through four industrial cameras from four orthogonal visual angles; performing distortion correction according to the camera internal parameter matrix and the distortion coefficient; calculating the matching relationship of the same points of adjacent cameras based on a zero-mean normalized cross-correlation function to obtain an initial disparity field; generating a sub-pixel level disparity map through median filtering and sub-pixel optimization of a quadratic surface fitting; converting three-dimensional point cloud data by using external parameters; extracting a set of points with the maximum gradient change of the outer edge of the cutter ring as edge point cloud; and performing least square space circle fitting to obtain cutter ring outer diameter and roundness deviation measurement values. The application realizes non-contact high-precision size measurement and effectively overcomes the influence of contact measurement on the damage of the cutter disc surface and complex light interference.

Description

Technical Field

[0001] This invention relates to the field of industrial vision measurement technology, specifically a high-precision measurement method for the dimensions of tunnel boring machine cutterhead components based on machine vision. Background Technology

[0002] The dimensional accuracy of the cutterhead components of a tunnel boring machine (TBM) directly affects the assembly quality and tunneling efficiency. The outer diameter and roundness deviation of the cutterhead ring are key control indicators. For such large rotating components, contact measuring tools such as calipers and micrometers can only measure local points and cannot reflect overall contour deviations. While coordinate measuring machines (CMMs) offer high accuracy, their measurement efficiency is low, they are sensitive to on-site temperature and humidity, and establishing measurement benchmarks for large, complex-shaped components is a cumbersome process.

[0003] Non-contact measurement methods based on machine vision reconstruct 3D topography and perform dimensional analysis by acquiring surface images of components, offering advantages such as high efficiency and full-scene coverage. Existing solutions often employ methods like spraying black and white paint or projecting speckle patterns to construct random textures on the component surface, and then use digital image correlation to match corresponding points to obtain 3D information. However, tunnel boring machine cutterhead components are often inspected in workshop environments with abundant stray light of uncertain direction and strong reflected light from metal surfaces. Ordinary speckle patterns, under white light illumination, result in insufficient image contrast, easily leading to incorrect matching or decreased accuracy during the matching process, directly affecting the realism of point cloud reconstruction and the reliability of subsequent dimensional calculations. Furthermore, due to machining textures, microburrs, and surface oxide layers, the outer edge of the cutter ring does not have a smooth geometric boundary represented in the 3D point cloud. Conventional edge extraction algorithms easily misinclude these local irregular points in the outer edge set, causing significant deviations in subsequent circle fitting results and making it difficult to consistently obtain dimensional evaluation data that meets industrial standards.

[0004] To address the above issues, it is necessary to eliminate the disturbance of ambient light to speckle image acquisition at the source, improve the signal-to-noise ratio and contrast of speckle images, and establish a measurement path that accurately extracts the point cloud of the cutterhead outer edge and robustly fits the spatial circle, thereby achieving high-precision and reliable measurement of the dimensions of the tunnel boring machine cutterhead components under field conditions. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a high-precision measurement method for the dimensions of shield machine cutterhead components based on machine vision, so as to overcome the defects of existing technology, such as the degradation of speckle image quality caused by ambient light interference, unstable extraction of point cloud on the outer edge of the cutter ring, and insufficient accuracy of spatial circle fitting.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a high-precision measurement method for the dimensions of tunnel boring machine cutterhead components based on machine vision, the method comprising: A random speckle coating containing fluorescent particles is formed on the surface of the cutter head component. Preferably, fluorescent particles and epoxy resin are mixed at a mass ratio of 1:10 and uniformly sprayed onto the surface of the cutter head component using a pneumatic spray gun. The spraying pressure is controlled between 0.3 MPa and 0.5 MPa. The sprayed cutter head component is then placed in a 60°C constant temperature oven for curing for 4 hours, forming a fluorescent speckle coating with a thickness of 0.1 mm to 0.3 mm. This coating can generate fluorescence signals with wavelengths of 500 nm to 550 nm under ultraviolet light irradiation, thereby effectively eliminating ambient light interference and obtaining speckle images with a high signal-to-noise ratio.

[0007] Four industrial cameras are used to simultaneously acquire speckle fluorescence images of the cutter head component under ultraviolet light excitation from four orthogonal perspectives. In one technical solution of the present invention, the optical axes of the four industrial cameras are at a 90-degree angle, and the cameras are 1m to 2m away from the surface of the cutter head component. An external trigger signal generator is used to simultaneously trigger the exposure of the four cameras at a frequency of 20Hz, and the exposure time is set to 10ms to 20ms. A bandpass filter with a center wavelength of 525nm and a full width at half maximum (FWHM) of 20nm is installed in front of each camera lens to allow only the fluorescence signal to pass through and block the ultraviolet excitation light, ensuring the clarity and consistency of the multi-view images.

[0008] The speckle fluorescence image is distorted by performing distortion correction on the speckle image based on the pre-calibrated intrinsic parameter matrix and distortion coefficient of each camera, thereby obtaining the corrected speckle image.

[0009] The initial disparity field is obtained by calculating the matching relationship of corresponding points of adjacent cameras in the corrected speckle image based on the zero-mean normalized cross-correlation function. Preferably, one camera image is used as the reference image and the other as the target image. Sub-regions of size 32×32 pixels are divided in both the reference and target images. For the center pixel of each sub-region in the reference image, the zero-mean normalized cross-correlation function value is calculated by searching along the epipolar direction in the target image with an integer pixel step size. The pixel coordinate difference at the maximum function value is used as the initial disparity value of the center pixel. This process is repeated for all sub-region center pixels to form an initial disparity field covering the entire image. This process can stably establish a dense correspondence between adjacent camera images.

[0010] The initial disparity field is optimized by median filtering and quadratic surface fitting to generate a sub-pixel level disparity map. In a preferred embodiment, a 3×3 window median filter is first used to remove isolated noise points from the initial disparity field, resulting in a smooth disparity field. Then, a least-squares quadratic surface equation is constructed using the disparity values ​​of each pixel and its nine neighboring pixels in the smooth disparity field. The coordinates of the extreme points of this quadratic surface equation are solved, and the resulting coordinate offsets are used as sub-pixel displacement increments and superimposed on the corresponding integer-pixel disparity values. This improves the disparity accuracy to the sub-pixel level, effectively breaking through the limitations of integer-pixel resolution and providing reliable data for high-precision 3D reconstruction.

[0011] The subpixel-level disparity map is converted into 3D point cloud data of the cutter head component surface using pre-calibrated extrinsic parameters. Specifically, the rotation matrix and translation vector of the four cameras are obtained based on the Zhang Zhengyou calibration method and unified into a world coordinate system with the center of the cutter head component as the origin. According to the principle of binocular vision triangulation, the 3D coordinates of each corresponding point in the world coordinate system are calculated by combining the rotation matrix, translation vector, and disparity value in the subpixel-level disparity map. The 3D coordinates calculated by the four cameras are transformed and fused, and outliers with a distance exceeding the preset tolerance range of 0.5mm are removed to obtain a complete and accurate 3D point cloud of the cutter head component surface.

[0012] The set of points with the maximum gradient change on the outer edge of the cutter ring of the cutter head component is extracted from the three-dimensional point cloud data and used as the edge point cloud. In a preferred embodiment of the present invention, the three-dimensional point cloud data is projected onto a plane perpendicular to the axis of the cutter ring to generate a two-dimensional projection point map; the normal angle between each point and its 30 nearest neighbors is calculated, and candidate points with a normal angle greater than 60 degrees are selected. The normal angle is obtained through principal component analysis. Principal component analysis is performed on each point and its neighbors, and the eigenvector corresponding to the smallest eigenvalue is taken as the normal vector of that point. The angle between this normal vector and the normal vectors of other points in the neighborhood is then calculated, and the average value is taken as the normal angle of that point. The gray-level gradient magnitude is calculated for the candidate points, and the points with the gradient magnitude in the top 20% are selected as the initial edge points; distance-based Euclidean clustering is performed on the initial edge points to remove isolated outliers and retain the cluster with the highest density as the edge point cloud of the cutter ring outer edge. This process can accurately extract the outer edge contour of the cutter ring from the complex point cloud and avoid interference from pseudo-edges such as welding spatter and surface roughness.

[0013] The edge point cloud is fitted with a least-squares spatial circle to obtain the measured values ​​of the cutter ring's outer diameter and roundness deviation. Preferably, three non-collinear points are randomly selected from the edge point cloud to determine the center and radius of the initial spatial circle, establishing a spatial circle parameter model. The sum of squared residuals between the distances from all points in the edge point cloud to the center of the initial spatial circle and the radius is calculated. This sum of squared residuals is minimized using the Levenberg-Marquardt iterative algorithm to obtain the optimized coordinates of the spatial circle's center and radius. Then, the absolute values ​​of the differences between the distances from each point in the edge point cloud to the optimized circle's center and the optimized radius are calculated. The maximum value of these differences is taken as the roundness deviation measurement value, and twice the optimized radius is taken as the cutter ring's outer diameter measurement value. Through the above nonlinear optimization process, the measurement accuracy and repeatability of the outer diameter and roundness deviation of large-size cutter head components are significantly improved.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: A random speckle coating is formed by mixing fluorescent particles with epoxy resin in a specific ratio and spraying the mixture onto the surface of the cutterhead component using a pneumatic spray gun. Under ultraviolet light excitation, the fluorescent particles emit fluorescence signals in a specific wavelength band. Simultaneously, a bandpass filter with a center wavelength matching the fluorescence signal is installed in front of the industrial camera lens, allowing only the fluorescence signal to enter the sensor. This method utilizes the band selectivity of fluorescence emission and filter reception to completely isolate stray white light and reflected light from metal surfaces in the workshop environment. This results in speckle images carrying only high-intensity fluorescent pixels, significantly enhancing the grayscale difference between the speckle background and fundamentally improving the image signal-to-noise ratio and contrast compared to black-and-white speckle under conventional white light illumination. The high-contrast fluorescent speckle image makes the extreme regions of the zero-mean normalized cross-correlation function sharper, greatly reducing the probability of mismatches during the initial search of corresponding points. Even with some perspective distortion in the image from various viewpoints of large-sized components, a dense and reliable initial parallax field can still be obtained, providing an accurate input basis for subsequent sub-pixel-level parallax optimization and 3D point cloud reconstruction.

[0015] The initial disparity field is denoised using median filtering to eliminate isolated erroneous disparity values ​​caused by residual local specular reflections or micro-defects in the coating. Then, a local quadratic surface equation is constructed using the disparity values ​​of each pixel and its nine neighboring pixels, and extreme coordinates are calculated. The sub-pixel displacement increments corresponding to these extreme values ​​are superimposed on the integer-pixel disparity to generate a sub-pixel level disparity map. This process does not directly rely on image interpolation but utilizes the continuity constraint of the disparity field in local space to improve matching accuracy to the sub-pixel level, effectively suppressing the interference of high-frequency image noise on sub-pixel estimation. In the point cloud processing stage, when extracting the outer edge of the tool ring from the 3D point cloud, a dual constraint of normal angle gradient and grayscale gradient is introduced. Candidate edge points with abrupt changes in normal and prominent gradient magnitudes are selected. Isolated outliers are then removed using distance-based Euclidean clustering, retaining the highest-density cluster as the outer edge point cloud of the tool ring. This processing strategy can accurately distinguish between pseudo-edges formed by processing textures or micro-burrs and the true geometric edges of the tool ring, resulting in a purer point set for circle fitting. The Levenberg-Marquardt iterative algorithm is used to fit the spatial circle of the above edge point cloud. Nonlinear optimization is performed with the goal of minimizing the sum of squared geometric distance residuals. The center and radius of the spatial circle obtained can reflect the macroscopic design profile of the outer edge of the tool ring to the greatest extent. The roundness deviation and outer diameter values ​​calculated in this way are not easily disturbed by the irregular morphology of the local surface, and the repeatability and accuracy of the measurement results are significantly guaranteed. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of a high-precision measurement method for the dimensions of tunnel boring machine cutterhead components based on machine vision. Figure 2 This is a flowchart of the fluorescent speckle coating fabrication and testing method; Figure 3 This is a graph showing the measurement and uniformity analysis of ultraviolet light source irradiance; Figure 4 This is a sub-region similarity search result graph based on the zero-mean normalized cross-correlation function. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 This invention provides a high-precision measurement method for the dimensions of tunnel boring machine cutterhead components based on machine vision, comprising: coating the surface of the cutterhead component with a random speckle coating containing fluorescent particles; synchronously acquiring speckle fluorescence images of the cutterhead component under ultraviolet light excitation from four orthogonal perspectives using four industrial cameras; performing distortion correction on the speckle fluorescence images according to the pre-calibrated intrinsic parameter matrix and distortion coefficient of each camera to obtain a corrected speckle image; calculating the matching relationship of corresponding points of adjacent cameras in the corrected speckle image based on the zero-mean normalized cross-correlation function to obtain an initial disparity field; performing sub-pixel optimization of the initial disparity field by median filtering and quadratic surface fitting to generate a sub-pixel level disparity map; converting the sub-pixel level disparity map into three-dimensional point cloud data of the cutterhead component surface using pre-calibrated extrinsic parameters; extracting the set of points with the maximum gradient change at the outer edge of the cutter ring of the cutterhead component from the three-dimensional point cloud data as an edge point cloud; and performing least-squares spatial circle fitting on the edge point cloud to obtain the measured values ​​of the outer diameter and roundness deviation of the cutter ring.

[0020] Example 1 In specific implementation, please refer to Figure 2The fluorescent particles and epoxy resin are mixed at a mass ratio of 1:10 using a pneumatic spray gun and then uniformly sprayed onto the surface of the cutter head component. The fluorescent particles are rare-earth phosphors that emit fluorescence with wavelengths from 500 nm to 550 nm under ultraviolet light excitation, and the average particle size is controlled within the range of 10 to 50 micrometers. The epoxy resin is a low-viscosity two-component epoxy resin system, comprising component A (epoxy resin main agent) and component B (amine curing agent). During mixing, the fluorescent particles and epoxy resin component A are premixed at a mass ratio of 1:10 and stirred for 15 minutes at 200 rpm using a mechanical stirrer to ensure uniform dispersion of the fluorescent particles in component A. Then, an equal mass of epoxy resin component B (curing agent) is added, and stirring continues for 5 minutes to obtain the mixed spray material of fluorescent particles and epoxy resin. The mixed spray material is then injected into the storage tank of the pneumatic spray gun, whose nozzle diameter is selected from 0.8 mm to 1.2 mm. Adjust the pressure valve of the compressed air source to control the spraying pressure between 0.3 MPa and 0.5 MPa. During spraying, maintain a vertical distance of 200 mm to 300 mm between the nozzle of the pneumatic spray gun and the surface of the cutter head component. Move the spray gun in a serpentine motion along the surface of the cutter head component at a speed of 50 mm to 100 mm per second, with an overlap rate of 30% to 50% between adjacent spraying paths, thereby obtaining a wet coating of uniform thickness. Maintain the spraying environment temperature between 15°C and 30°C, and the relative humidity below 60%.

[0021] After spraying, the sprayed cutter head component is placed in a 60°C constant temperature oven for four hours to cure. The oven uses a forced-air circulation system, and the temperature fluctuation is controlled within ±2°C. During the curing process, the epoxy resin undergoes a cross-linking reaction, gradually changing from a liquid to a solid state, and firmly encapsulating the fluorescent particles within the coating. After curing, the cutter head component is allowed to cool naturally to room temperature, and the coating thickness is measured at ten randomly selected locations on the surface of the cutter head component using a coating thickness gauge. If the measured average thickness is within the range of 0.1 mm to 0.3 mm, and the standard deviation of the thickness is less than 0.05 mm, the fluorescent speckle coating is considered complete. If the average thickness is less than 0.1 mm, the spraying and curing steps are repeated until the required thickness is achieved. If the local thickness exceeds 0.3 mm, it is lightly sanded with fine sandpaper for correction. In the formed fluorescent speckle coating, the fluorescent particles are randomly distributed in the epoxy resin matrix, forming a high-contrast random speckle pattern.

[0022] During the detection phase, ultraviolet (UV) light irradiation excites the fluorescent speckle coating to generate fluorescence signals with wavelengths from 500 nm to 550 nm, eliminating ambient light interference. The UV light source is an LED array with a wavelength of 365 nm, and the UV irradiance is 10 milliwatts per square centimeter at a distance of 100 mm from the light source. The UV light source is symmetrically arranged around the blade assembly, ensuring a uniform distribution of UV irradiance received by the fluorescent speckle coating surface, with a difference in irradiance not exceeding 10%. After absorbing UV light energy, the fluorescent particles in the speckle coating emit visible fluorescence with a center wavelength of approximately 525 nm. A bandpass filter installed in front of the industrial camera lens allows only light beams with wavelengths from 500 nm to 550 nm to pass through, effectively suppressing interference from UV excitation light and ambient visible light. This results in a high signal-to-noise ratio in the acquired speckle fluorescence image, with a grayscale contrast between the speckle particles and the background exceeding 0.6.

[0023] Example 2 In practice, four industrial cameras are mounted around the cutter head component. The optical axis of each industrial camera forms a 90-degree angle with the optical axis of the adjacent industrial camera. The cutter head component is placed at the center of the inspection platform. Using the center of the cutter head component as a reference, an orthogonal coordinate system is established in the horizontal plane, with the four industrial cameras arranged along the four coordinate axes of this orthogonal coordinate system. The industrial cameras are monochrome CCD industrial cameras with a resolution of 2448 x 2048 pixels, and a pixel size of 3.45 micrometers x 3.45 micrometers. The focal length of each industrial camera lens is selected based on the working distance and the size of the cutter head component, ensuring that the cutter head component is completely imaged within the field of view of the industrial camera. The distance from the front face of each industrial camera to the surface of the cutter head component is one to two meters. The optical platform base of each industrial camera is equipped with a three-axis fine-tuning translation stage and a pitch and yaw adjustment mechanism. By adjusting the translation stage and the adjustment mechanism, the optical axis of each industrial camera is precisely passed through the geometric center of the cutter head component, and the angle between the optical axes of adjacent industrial cameras is 90 degrees ±0.1 degrees.

[0024] An external trigger signal generator was used to simultaneously trigger exposure of four industrial cameras at a frequency of 20 Hz. The external trigger signal generator was a programmable pulse signal generator with an output impedance of 50 ohms and an output level of 5V TTL. The square wave signal output by the pulse signal generator had a period of 50 milliseconds, a frequency of 20 Hz, and a duty cycle of 20%. All four industrial cameras were set to external trigger mode and connected in parallel to the output port of the pulse signal generator via coaxial cables, ensuring that the edge time deviation of the trigger signal received by the four industrial cameras was less than one microsecond. At the rising edge of each trigger pulse, the four industrial cameras simultaneously began exposure. The exposure time for each industrial camera was set to 10 to 20 milliseconds. The exposure time was set based on the fluorescence signal intensity and the full-well capacity of the industrial camera, ensuring that the maximum grayscale value of the acquired speckle fluorescence image was between 70% and 90% of the dynamic range of the industrial camera, avoiding pixel saturation and fully utilizing the quantization level of the analog-to-digital conversion. At a trigger frequency of 20 Hz, 20 sets of four-view speckle fluorescence images were acquired per second, with each set containing four speckle fluorescence images simultaneously acquired from four orthogonal viewpoints.

[0025] A bandpass filter corresponding to the wavelength band is installed in front of the lens of each industrial camera. The bandpass filter is an interference type filter with a center wavelength of 525 nm and a full width at half maximum (FWHM) of 20 nm. The bandpass filter is installed in a filter holder with a threaded interface at the front of the industrial camera lens. The filter holder is perpendicular to the lens optical axis, with a tilt deviation not exceeding 0.1 degrees. The fluorescent particles emitted under 365 nm ultraviolet light excitation have a peak fluorescence wavelength of 525 nm and a FWHM of approximately 30 nm. The bandpass filter's transmission band is from 515 nm to 535 nm, covering the main energy wavelength range of the fluorescence signal. The transmittance of the bandpass filter to 365 nm ultraviolet excitation light is less than 0.05%, allowing the fluorescence signal to pass through while blocking ultraviolet excitation light from entering the industrial camera. Meanwhile, visible light illumination is turned off in the detection area where the cutter head component is located, and only ultraviolet light source is retained. After the ambient stray light is suppressed by the bandpass filter, the photons that reach the industrial camera image sensor are mainly contributed by the fluorescence signal, which improves the contrast and repeatability of speckle fluorescence images.

[0026] See Figure 3 In the graph, the horizontal axis represents the measurement location number, ranging from 0 to 500, and the vertical axis represents the ultraviolet irradiance, in milliwatts per square centimeter. The green dots represent the actual measured ultraviolet irradiance values ​​at different measurement locations, the red solid line represents the average ultraviolet irradiance at each measurement location, approximately 10 milliwatts per square centimeter horizontally, and the gray dashed lines represent the ±10% deviation limits of the average irradiance, corresponding to 9 milliwatts per square centimeter and 11 milliwatts per square centimeter, respectively.

[0027] As shown in the figure, the ultraviolet irradiance measurements are relatively uniformly distributed across the entire measurement location, with the vast majority of measurements falling within a deviation range of 9 to 11 milliwatts per square centimeter, without any significant abnormal fluctuations or deviations. The red average irradiance curve is approximately flat, indicating that the ultraviolet light source provides good irradiance uniformity within the detection area, and the irradiance difference does not exceed the design requirement of ±10%.

[0028] Example 3 In the specific implementation, four speckle fluorescence images acquired by four industrial cameras and corrected for distortion are paired in pairs to cover the entire surface of the cutterhead component. Two industrial cameras with adjacent viewpoints are selected as binocular stereo vision units. The corrected speckle image corresponding to one of the industrial cameras is used as the reference image, and the corrected speckle image corresponding to the other industrial camera is used as the target image. Sub-regions of size 32 by 32 pixels are divided in both the reference and target images. The overlap between adjacent sub-regions in the column and row directions is set to 16 pixels, so that the center pixels of the sub-regions form a grid distribution on the reference image, with a grid spacing of 16 pixels.

[0029] For the center pixel of each sub-region in the reference image, the epipolar line corresponding to the center pixel of the reference image sub-region is determined in the target image. Based on the pre-calibrated fundamental matrix of the four industrial cameras and the homogeneous coordinates of the center pixels of the reference image sub-regions, the epipolar line equation in the target image is calculated, expressed in linear parameter form. A search is performed along the epipolar line direction in the target image with an integer pixel step size. The search range is set to ±128 pixels along the epipolar line direction, and the search step size is one pixel. At each search position, the pixel corresponding to the search position is taken as the center pixel of the target image sub-region, and a 32x32 pixel target image sub-region with the same size as the reference image sub-region is extracted.

[0030] Calculate the zero-mean normalized cross-correlation function between the reference image sub-region and the target image sub-region. Traverse all integer-pixel search positions along the epipolar direction and record the zero-mean normalized cross-correlation value corresponding to each search position.

[0031] The column offset corresponding to the search position where the zero-mean normalized cross-correlation values ​​reach their maximum value. and row direction offset The initial disparity value is used as the center pixel of the reference image sub-region. The horizontal component of the initial disparity value is... The vertical component is The process iterates through all center pixels of sub-regions in the reference image, calculates the initial disparity value for each center pixel, and arranges all initial disparity values ​​according to the pixel coordinates of the reference image to form an initial disparity field covering the entire reference image. Each coordinate position in the initial disparity field records a two-dimensional disparity vector, and the resolution of the initial disparity field is consistent with the grid density of the center pixels of the sub-regions in the reference image.

[0032] After obtaining the initial disparity field, a 3x3 window median filter is used to remove isolated noise points. Using each integer pixel coordinate in the initial disparity field as the center of the filtering window, nine initial disparity values ​​are selected at three rows and three columns surrounding the center. The horizontal and vertical components of these nine initial disparity values ​​are then sorted according to their numerical values, and the median of the sorted values ​​is taken as the disparity value for the smoothed disparity field at the center of the filtering window. After performing a 3x3 window median filter on all integer pixel coordinates in the initial disparity field, a smoothed disparity field is obtained.

[0033] Based on the smoothed disparity field, sub-pixel optimization based on quadratic surface fitting is performed. For each integer pixel coordinate position in the smoothed disparity field, a least-squares quadratic surface equation is constructed using the nine disparity values ​​of that integer pixel coordinate position and its eight neighboring integer pixel coordinate positions. The quadratic surface equation is a bivariate quadratic polynomial of disparity with respect to image coordinates, containing constant terms, first-order terms, and second-order terms. The column coordinates and row coordinates of the nine integer pixel coordinate positions, along with the corresponding disparity values ​​in the smoothed disparity field, constitute nine data points. Substituting these into the bivariate quadratic polynomial forms an overdetermined system of equations. The least-squares method is used to solve the overdetermined system of equations, yielding a closed-form solution for the coefficients of the bivariate quadratic polynomial.

[0034] After obtaining the coefficients of the bivariate quadratic polynomial, partial derivatives are calculated in both the column and row directions. These partial derivatives are then set to zero, and the system of two linear equations is solved to obtain the coordinates of the extreme points of the quadratic surface. The difference between the column coordinates corresponding to the extreme points and the integer pixel column coordinates is used as the subpixel horizontal displacement increment, and the difference between the row coordinates corresponding to the extreme points and the integer pixel row coordinates is used as the subpixel vertical displacement increment. The subpixel horizontal displacement increments are then superimposed on the corresponding integer pixel horizontal disparity values ​​in the smoothed disparity field, and the subpixel vertical displacement increments are superimposed on the corresponding integer pixel vertical disparity values ​​in the smoothed disparity field, yielding the subpixel-level disparity value at that integer pixel coordinate position. After traversing all integer pixel coordinate positions in the smoothed disparity field, all subpixel-level disparity values ​​are arranged according to image coordinates to generate a subpixel-level disparity map. Each pixel position in the subpixel-level disparity map records a subpixel-precision two-dimensional disparity vector.

[0035] See Figure 4The figure shows the performance curves of multi-sub-region disparity matching based on the zero-mean normalized cross-correlation function in Example 3. The horizontal axis represents the search offset in pixels, ranging from -128 pixels to 128 pixels, covering the search range set along the epipolar direction in Example 3; the vertical axis represents the zero-mean normalized cross-correlation coefficient, ranging from 0 to 1, indicating the degree of matching between the reference image sub-region and the target image sub-region at the corresponding search position.

[0036] The figure contains five matching curves for different sub-regions (sub-region 1 to sub-region 5), distinguished by different line types such as solid lines, dashed lines, and dotted lines. Each curve exhibits a distinct peak-shaped structure, with the peak values ​​concentrated near the zero offset position. This indicates that the zero-mean normalized cross-correlation coefficient of the matching reaches its maximum value at this location, demonstrating high-precision matching of corresponding points.

[0037] Within the search offset range of approximately -20 pixels to 20 pixels, the zero-mean normalized cross-correlation coefficient of each sub-region rapidly rises to a high value above 0.9, reaching a peak near offset 0, with the peak value reaching approximately 0.95. This indicates that the matching effect is optimal within this range, and the starting point for sub-pixel disparity calculation is accurate and reliable. To the left and right of the peak, the cross-correlation coefficient curve shows a symmetrical decreasing trend. After the absolute value of the offset exceeds 60 pixels, the cross-correlation coefficient rapidly drops to near zero, indicating that the image regions within this range do not match.

[0038] The peak amplitudes and shapes of the curves in each sub-region are basically consistent, and the peak positions are all concentrated near zero, indicating that the sub-region matching algorithm based on the zero-mean normalized cross-correlation function has good stability and consistency, effectively supporting subsequent median filtering and smoothing of the disparity field construction. The subtle fluctuations in the curves reflect the noise and texture differences in the actual speckle fluorescence image, but the overall trend is clear, which is beneficial for achieving accurate disparity estimation.

[0039] Example 4 In the specific implementation, the rotation matrix and translation vector of four industrial cameras are obtained based on Zhang Zhengyou's calibration method. Images of the checkerboard calibration board in different poses are acquired. Each square on the checkerboard calibration board has a physical size of 5 mm by 5 mm, and the checkerboard pattern contains 12 columns by 9 rows, totaling 108 corner points. The four industrial cameras simultaneously acquire images of the calibration board in 15 different spatial poses, obtaining 15 sets of four-view calibration image pairs. For each set of images, the sub-pixel image coordinates of the checkerboard corner points are extracted. Based on the homography relationship between the world coordinates and image coordinates of the checkerboard corner points, the intrinsic parameter matrix and distortion coefficients of each industrial camera are solved. After obtaining the intrinsic parameters, keeping them constant, the rotation matrix and translation vector between every two adjacent industrial cameras are solved using the calibration images acquired by the four industrial cameras at the same time.

[0040] The extrinsic parameters of the four industrial cameras were unified to a world coordinate system with the center of the cutterhead component as the origin. Three non-collinear reference markers were set on the cutterhead component inspection platform. The three-dimensional coordinates of the three reference markers in the world coordinate system were precisely determined by a laser tracker, with a measurement uncertainty better than 0.03 mm. The four industrial cameras simultaneously acquired images containing the reference markers. The rotation matrix and translation vector of each industrial camera's coordinate system relative to the world coordinate system were calculated using the resection method, thus unifying the coordinate system transformation of the four industrial cameras to the same world coordinate system.

[0041] Based on the principle of binocular vision triangulation, and combining the rotation matrix, translation vector, and disparity values ​​in the sub-pixel disparity map, the 3D coordinates of each corresponding point in the world coordinate system are calculated. For a binocular stereo vision unit composed of two adjacent industrial cameras, the rotation matrix of the left industrial camera is set as follows: The translation vector of the left industrial camera is The rotation matrix of the right industrial camera is The translation vector of the right industrial camera is For pixel coordinate positions in sub-pixel disparity maps Corresponding spatial point spatial point 3D coordinates in the left industrial camera coordinate system Calculate using the following formula: in, For spatial points Depth coordinates in the left industrial camera coordinate system, in millimeters; The baseline length between the optical centers of the left and right industrial cameras is extracted from the rotation matrix and translation vector, and is in millimeters. The equivalent focal length of the left industrial camera is given by the intrinsic parameter matrix of the left industrial camera. Quantity and The average value of the components is expressed in pixels; Pixel coordinates in the subpixel disparity map The horizontal component of the disparity vector, in pixels. Spatial point. Horizontal coordinates in the left industrial camera coordinate system Through formula Calculation, where The coordinates of the principal points in the intrinsic parameter matrix of the left industrial camera; spatial points Vertical coordinates in the left industrial camera coordinate system Through formula Calculation, where The coordinates of the principal points in the intrinsic parameter matrix of the left industrial camera are given. The spatial points are... 3D coordinates in the left industrial camera coordinate system The pose transformation relationship in the world coordinate system is used to transform it into three-dimensional coordinates in the world coordinate system.

[0042] Four sets of binocular stereo vision units are processed sequentially, and the 3D coordinates calculated by the four industrial cameras are transformed and fused. After initially registering the four sets of 3D point clouds to a unified world coordinate system using an iterative nearest-neighbor algorithm, the distances between corresponding points in the overlapping region are statistically analyzed. A spherical region with a radius of 0.5 mm is set as the distance tolerance range. For any point in the overlapping region, its nearest corresponding point in other point clouds is searched. If the Euclidean distance between the two points exceeds 0.5 mm, both points are marked as outliers and removed. After removing outliers, the four sets of 3D point clouds are merged into a complete 3D point cloud dataset. Each point in the 3D point cloud dataset contains coordinates in the world coordinate system. Coordinate values coordinate values ​​and Coordinate values.

[0043] The 3D point cloud data is projected onto a plane perpendicular to the cutter ring axis to generate a 2D projected point map. A subset of the 3D point cloud data for the cutter ring region of the cutter head component is extracted from the 3D point cloud data. Principal component analysis is used to calculate the three principal directions of the cutter ring point cloud subset, and the eigenvector corresponding to the smallest eigenvalue among the three principal directions is taken as the axial vector of the cutter ring. A projection plane is constructed, with its normal vector parallel to the axial vector of the cutter ring. Each 3D point in the subset of the 3D point cloud data for the cutter ring region is orthographically projected onto the projection plane along the axial vector direction, obtaining the 2D projected coordinates of each point on the projection plane. The set of all 2D projected coordinates constitutes the 2D projected point map.

[0044] For each point in the 2D projected point map, select the thirty nearest neighboring points in space and perform principal component analysis (PCA) on the local point set formed by the point and its thirty neighbors. The covariance matrix of PCA is a 3x3 real symmetric matrix. Solve for the eigenvalues ​​and corresponding eigenvectors of the covariance matrix. Take the eigenvector corresponding to the smallest eigenvalue as the normal vector of the point. Calculate the angle between the normal vector of the point and the normal vectors of each of the thirty neighboring points. Sum the thirty angle values ​​and divide by thirty to obtain the average normal angle of the point, in degrees. In the 2D projected point map, select points with a normal angle greater than 60 degrees as a candidate point set.

[0045] For each candidate point in the candidate point set, the grayscale gradient magnitude is calculated. During the initial acquisition of the 3D point cloud data, each 3D point records the grayscale value of the corresponding pixel in the industrial camera image. Within the local neighborhood of a candidate point, the first-order differences of the grayscale values ​​of each point in the neighborhood are taken along the horizontal and vertical directions. The square root of the sum of the squares of the horizontal and vertical first-order differences is used to obtain the grayscale gradient magnitude of the candidate point. All candidate points are sorted from largest to smallest grayscale gradient magnitude, and the candidate points with the highest grayscale gradient magnitudes (top 20%) are selected as the initial edge points.

[0046] Perform distance-based Euclidean clustering on the initially selected edge points. Set the distance threshold for Euclidean clustering to twice the average distance between the initially selected edge points on the projection plane. Randomly select a point from the initial set of edge points as a seed point and add it to a cluster set. Search for the neighborhood of the initially selected edge points with the seed point as the center and the distance threshold as the radius. Add the found neighborhood points to the same cluster set and use them as new seed points to continue expanding the search until no new neighborhood points are found, at which point the growth of the cluster set ends. Iterate through the points in the initial set of edge points that have not yet been clustered and repeat the above clustering process to generate multiple cluster sets. Count the number of points contained in each cluster set, and remove the cluster sets with fewer than 5% of the total number of initially selected edge points. Keep the cluster set with the largest number of points as the outer edge point cloud of the blade circle.

[0047] Example 5 In practical implementation, three non-collinear points are randomly selected from the edge point cloud to determine the center and radius of the initial spatial circle, thus establishing a spatial circle parameter model. The edge point cloud is a set of points represented by three-dimensional coordinates in the world coordinate system, and the total number of points in the edge point cloud is denoted as [missing information]. The indices of the three randomly selected points in the edge point cloud are respectively... , and The corresponding world coordinates are as follows: , and Through vectors with vector To determine if the three points are collinear, check if the cross product modulus is greater than zero. If the cross product modulus is greater than zero, the three points are not collinear, and proceed to the next step of calculation; if the cross product modulus is equal to zero, randomly select three points again until the three points are not collinear.

[0048] The method for determining a spatial circle using three non-collinear points is as follows: Calculate the normal vectors of the planes containing the three points; solve the equations that the three points are equidistant from the center of the spatial circle; and obtain the coordinates of the center of the spatial circle in the world coordinate system. This coordinate is denoted as the initial center. Calculate the Euclidean distances from the three points to the initial center of the circle, and take the average of the three distances as the initial radius. The initial parameter vector for constructing the spatial circular parametric model using the initial center and initial radius is denoted as . .

[0049] Calculate the sum of squared residuals between the distances from all points in the edge point cloud to the center of the initial spatial circle and the radius of the initial spatial circle. Use the sum of squared geometric distances from all edge points to the spatial circle as the objective function. Define the objective function. for: in, Represents the parameter vector to be optimized; This represents the total number of points in the edge point cloud. This is the index number of a point in the edge point cloud, with a value ranging from... arrive Integers; Indicating the first edge point cloud The three-dimensional coordinates of a point in the world coordinate system; Denotes the center of the spatial circle to be solved. Axis coordinate components; Denotes the center of the spatial circle to be solved. Axis coordinate components; Denotes the center of the spatial circle to be solved. Axis coordinate components; This represents the radius of the spatial circle to be solved. In each iteration, the parameter vector... The value is updated, and the objective function is... The value gradually decreases.

[0050] Minimize the objective function using the Levonberg-Marquardt iterative algorithm. This yields the optimized coordinates of the center and radius of the spatial circle. The parameters of the Levenberg-Marquardt iterative algorithm are set as follows: The parameter vector to be optimized... The initial value is set to Damping factor The initial value is set to 0.001. The reason for setting it to 0.001 is that the scale of the edge point cloud distribution area is on the order of one meter. This damping factor value can keep the parameter increment within a reasonable range in the initial iteration step, avoiding iteration divergence caused by excessive difference between the initial parameters and the optimal parameters. The iteration termination condition is that the L2 norm of the parameter vector change is less than 0.001. Or the relative decrease in the objective function value is less than Or the number of iterations reaches the maximum of two hundred.

[0051] In each Levenberg-Marquardt iteration, based on the current parameter vector Calculate the residual vector, the th residual vector Each component is Calculate the Jacobian matrix. The dimension of the Jacobian matrix is OK Column, number Row corresponding residual components For four parameters The partial derivatives. Construct the incremental equation. ,in It is a fourth-order identity matrix. For parameter increment vectors, Let be the residual vector. Solving the incremental equation yields... Update parameters as follows Calculate the objective function value under the new parameters. ,like Then reduce the damping factor ,make Accept this iteration; otherwise, increase the damping factor. ,make If the current iteration is rejected, the incremental equation is solved again. This process is repeated until the iteration termination condition is met, and the optimized parameter vector is output. ,in The optimized coordinates of the center of the spatial circle. The radius of the optimized spatial circle.

[0052] After obtaining the optimized coordinates of the center and radius of the spatial circle, the absolute value of the difference between the distance from each point in the edge point cloud to the optimized center and the optimized radius is calculated, and the maximum value of these differences is taken as the roundness deviation measurement value. For the edge point cloud with index... The point, which is the center of the circle after optimization. The distance is The absolute value of the difference between the distance and the radius corresponding to this point is .all The absolute values ​​of the differences constitute a set. Roundness deviation measurement value Take the maximum value in the set, that is Measurement of the outer diameter of the cutter ring Take twice the optimized radius, that is .

[0053] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A high-precision measurement method for the dimensions of tunnel boring machine cutterhead components based on machine vision, characterized in that, include: A random speckle coating containing fluorescent particles is applied to the surface of the cutter head component; The speckle fluorescence images of the cutter head component under ultraviolet light source excitation were simultaneously acquired from four orthogonal perspectives using four industrial cameras. The speckle fluorescence image is distorted according to the pre-calibrated intrinsic parameter matrix and distortion coefficient of each camera to obtain the corrected speckle image; The initial disparity field is obtained by calculating the matching relationship of corresponding points of adjacent cameras in the corrected speckle image based on the zero-mean normalized cross-correlation function. The initial disparity field is optimized by median filtering and quadratic surface fitting to generate a sub-pixel level disparity map. The subpixel-level disparity map is converted into three-dimensional point cloud data of the cutter head component surface using pre-calibrated extrinsic parameters; Extract the set of points with the maximum gradient change at the outer edge of the cutter ring of the cutter head component from the three-dimensional point cloud data, and use it as the edge point cloud; The edge point cloud is fitted with a least-squares spatial circle to obtain the measured values ​​of the outer diameter and roundness deviation of the blade ring.

2. The high-precision measurement method for the dimensions of tunnel boring machine cutterhead components based on machine vision as described in claim 1, characterized in that, The process of coating the surface of the cutter head component with a random speckle coating containing fluorescent particles includes: Fluorescent particles and epoxy resin are mixed at a mass ratio of 1:10 using a pneumatic spray gun and then evenly sprayed onto the surface of the cutter head component. The spraying pressure is controlled between 0.3 MPa and 0.5 MPa. The sprayed cutter head component was placed in a 60-degree Celsius constant temperature oven for four hours to cure, forming a fluorescent speckle coating with a thickness of 0.1 mm to 0.3 mm. The fluorescent speckle coating is excited by ultraviolet light to generate a fluorescent signal with a wavelength of 500 to 550 nanometers, thereby eliminating ambient light interference.

3. The high-precision measurement method for the dimensions of tunnel boring machine cutterhead components based on machine vision as described in claim 1, characterized in that, The process of simultaneously acquiring speckle fluorescence images of the cutter head component under ultraviolet light excitation from four orthogonal perspectives using four industrial cameras includes: Four industrial cameras are installed around the cutter head component with their optical axes at a 90-degree angle. The distance between the cameras and the surface of the cutter head component is one to two meters. An external trigger signal generator is used to simultaneously trigger the exposure of four cameras at a frequency of 20 Hz, with the exposure time set to 10 to 20 milliseconds. A bandpass filter with a corresponding wavelength is installed in front of each camera lens. The center wavelength of the bandpass filter is 525 nanometers and the full width at half maximum (FWHM) is 20 nanometers, so as to allow fluorescence signals to pass through and block ultraviolet excitation light.

4. The high-precision measurement method for the dimensions of tunnel boring machine cutterhead components based on machine vision as described in claim 1, characterized in that, The calculation of the corresponding point matching relationship between adjacent cameras in the corrected speckle image based on the zero-mean normalized cross-correlation function to obtain the initial disparity field includes: Using one camera image as the reference image and the other camera image as the target image, sub-regions of size 32 by 32 pixels are divided in both the reference image and the target image. For the center pixel of each sub-region in the reference image, the zero-mean normalized cross-correlation function value is calculated by searching along the epipolar direction in the target image with an integer pixel step size. The pixel coordinate difference corresponding to the search position where the zero-mean normalized cross-correlation function value reaches its maximum value is used as the initial disparity value of the center pixel. Traverse the center pixels of all sub-regions to form an initial disparity field covering the entire image.

5. The high-precision measurement method for the dimensions of tunnel boring machine cutterhead components based on machine vision as described in claim 1, characterized in that, The sub-pixel optimization of the initial disparity field by median filtering and quadratic surface fitting to generate a sub-pixel level disparity map includes: A 3x3 window median filter is used to remove isolated noise points in the initial disparity field to obtain a smooth disparity field; A least-squares quadratic surface equation is constructed using the disparity values ​​of each pixel and its nine neighboring pixels in the smooth disparity field. The quadratic surface equation is a bivariate quadratic polynomial of disparity with respect to image coordinates. Solve for the coordinates of the extreme points of the quadratic surface equation, and use the coordinate offset corresponding to the extreme point as the sub-pixel displacement increment; The subpixel displacement increment is superimposed on the corresponding integer pixel disparity value of the smooth disparity field to obtain a subpixel-level disparity map.

6. The high-precision measurement method for the dimensions of tunnel boring machine cutterhead components based on machine vision as described in claim 1, characterized in that, The process of converting the subpixel-level disparity map into three-dimensional point cloud data of the cutterhead component surface using pre-calibrated extrinsic parameters includes: The rotation matrix and translation vector of the four cameras were obtained based on Zhang Zhengyou's calibration method and unified to the world coordinate system with the center of the cutter head component as the origin. Based on the principle of binocular vision triangulation, the three-dimensional coordinates of each corresponding point in the world coordinate system are calculated by combining the rotation matrix, translation vector and disparity value in the sub-pixel disparity map. The three-dimensional coordinates calculated by the four cameras are transformed and fused, and outliers that are more than 0.5 mm away from the preset tolerance are removed to obtain the final three-dimensional point cloud data of the cutter head component surface.

7. The high-precision measurement method for the dimensions of tunnel boring machine cutterhead components based on machine vision as described in claim 1, characterized in that, Extracting the set of points with the maximum gradient change at the outer edge of the cutter ring of the cutter head component from the three-dimensional point cloud data as the edge point cloud includes: The three-dimensional point cloud data is projected onto a plane perpendicular to the axis of the cutter ring to generate a two-dimensional projection point map; In the two-dimensional projection point map, calculate the normal angle between each point and its nearest thirty neighboring points, and filter out candidate points with a normal angle greater than sixty degrees. The grayscale gradient magnitude of the candidate points is calculated, and the points with gradient magnitudes in the top 20% are selected as the initial edge points. The initially selected edge points are subjected to distance-based Euclidean clustering to remove isolated outliers and retain the cluster with the highest density as the outer edge point cloud of the blade ring.

8. The high-precision measurement method for the dimensions of tunnel boring machine cutterhead components based on machine vision as described in claim 1, characterized in that, The step of performing least-squares spatial circle fitting on the edge point cloud to obtain the measured values ​​of the outer diameter and roundness deviation of the blade ring includes: Three non-collinear points are randomly selected from the edge point cloud to determine the center and radius of the initial spatial circle, and a spatial circle parameter model is established. Calculate the sum of squared residuals between the distances from all points in the edge point cloud to the center of the initial spatial circle and the radius of the initial spatial circle. Minimize the sum of squared residuals using the Levenberg-Marquardt iterative algorithm to obtain the optimized coordinates of the center and radius of the spatial circle. Calculate the absolute value of the difference between the distance from each point in the edge point cloud to the optimized circle center and the optimized radius. Take the maximum value of these differences as the roundness deviation measurement value, and take twice the optimized radius as the outer diameter measurement value of the tool ring.

9. The high-precision measurement method for the dimensions of tunnel boring machine cutterhead components based on machine vision as described in claim 7, characterized in that, The steps for calculating the normal angle are as follows: perform principal component analysis on each point and its 30 nearest neighbors, take the eigenvector corresponding to the smallest eigenvalue as the normal vector of the point, then calculate the angle between the normal vector and the normal vectors of other points in the neighborhood, and take the average value as the normal angle of the point.

10. The high-precision measurement method for the dimensions of tunnel boring machine cutterhead components based on machine vision as described in claim 8, characterized in that, The least-squares spatial circle fitting uses the Levenberg-Marquardt algorithm for nonlinear optimization, with the objective function being the sum of the squares of the geometric distances from all edge points to the spatial circle.

Images