Camera calibration method based on three-dimensional calibration

Abstract

The invention discloses a camera calibration method based on three-dimensional calibration, relates to the technical field of machine vision and sensor calibration, reuses inherent three-dimensional characteristics of a workpiece, and gets rid of carrying, placing and maintaining burdens of a checkerboard or a target; a welding production line does not need to stop to intervene in a calibration process, and continuous operation efficiency is guaranteed; especially, the core pain point that people cannot approach the calibration object in a high-risk scene is solved; the high-frequency reflection interference is stripped through multi-scale illumination transformation, and the feature point detection stability in the arc light and metal splashing environment is ensured in combination with a metal curved surface distortion compensation mechanism; even if a local overexposure area is subjected to virtual point supplementation through the CAD model, calibration interruption caused by feature deficiency in a traditional method is avoided; in addition, IMU motion data and optical flow tracking are fused, the motion consistency of feature points is verified online, and transient interference such as splashing slag is automatically filtered; the extrinsic parameter drift distance is modeled as a function of a motion trail, and is corrected in real time through regularization constraint, so that accumulative errors caused by vibration are remarkably inhibited.

Description

Technical Field

[0001] The present invention relates to the technical field of machine vision and sensor calibration, and in particular to a camera calibration method based on three-dimensional calibration. Background Art

[0002] In automated welding quality inspection systems, cameras must measure weld position and forming status with millimeter-level precision. Such scenarios typically involve high-brightness arc light, molten metal splashes, and strong reflective interference from the workpiece surface, resulting in local overexposure or mirror reflection noise in the imaging. Current mainstream methods rely on high-precision checkerboard or three-dimensional targets for calibration, requiring the calibration object to be fixed near the welding tooling.

[0003] Some solutions improve the calibration environment through multi-spectral filtering or active projection of structured light. For example, near-infrared cameras are used to avoid arc light interference, or anti-reflection coatings are added to the surface of the calibration plate. However, in actual welding workshops, splashing slag will contaminate the surface of the calibration object, and strong reflections will still be generated after the coating fails at high temperature. Active projection solutions are limited by the scattering effect of welding smoke on the optical path. Alternatively, dynamic exposure control algorithms are introduced, but the transient changes in the reflected light from the metal surface exceed the algorithm's response threshold, causing the coordinates of the calibration feature points to drift.

[0004] In addition, some solutions adopt the method of migrating to a calibration-free object, such as using the workpiece's own corner point features to replace the checkerboard; however, welded parts are mostly curved or irregular structures and lack stable geometric features; although the latest deep learning methods can synthesize virtual calibration points, strong reflection noise will distort the key layer response of the feature extraction network, and the output parameters will have systematic deviations; the ambient light adaptive calibration algorithm requires multiple iterations and is difficult to meet the real-time requirements of continuous production of welding lines; therefore, a camera calibration method based on three-dimensional calibration is urgently needed to solve such problems. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides a camera calibration method based on three-dimensional calibration to solve the problems of strong welding reflection interfering with calibration object feature extraction, high temperature environment limiting the use of calibration plates, and vibration causing external parameter drift.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] An embodiment of the present invention provides a camera calibration method based on three-dimensional calibration, which includes:

[0009] Step S1, real-time extraction of environmental 3D features: using an industrial camera to capture welding scene images and identify stable 3D geometric features on the workpiece surface, including the 3D coordinates of metal weld edges, equipment installation holes, and pipeline intersections;

[0010] Step S2, reflection noise suppression processing: performing a multi-scale illumination invariant transformation on the image, separating high-frequency reflection noise from real feature textures, and dynamically generating an anti-interference feature point set;

[0011] The dynamically generated anti-interference feature point set includes:

[0012] Perform sub-pixel corner detection on the segmented non-reflective areas;

[0013] When the number of detection points is lower than the threshold, the virtual feature point generation module based on the welding workpiece CAD model is activated;

[0014] Perform spatial topological matching between virtual feature points and real detection points, and output a fused feature point set;

[0015] Step S3, 3D calibration parameter calculation: Based on the anti-interference feature point set and combined with the camera motion trajectory data, a 3D space reprojection model is constructed, and the camera intrinsic parameter matrix, distortion coefficient and extrinsic parameter pose are simultaneously solved;

[0016] Step S4, dynamic parameter optimization: updating the parameters of the three-dimensional space reprojection model online through optical flow tracking and feature point motion consistency verification.

[0017] As a preferred solution of the camera calibration method based on three-dimensional calibration described in the present invention, the identification of stable three-dimensional geometric features in step S1 includes:

[0018] Extract straight line segments and arc contours from the surface of welding workpieces;

[0019] Screen edge intersections that meet orthogonality conditions through spatial geometric constraints;

[0020] The filtered two-dimensional features are mapped into a three-dimensional coordinate point cloud based on the depth sensor data.

[0021] As a preferred solution of the camera calibration method based on three-dimensional calibration described in the present invention, the multi-scale illumination invariant transformation in step S2 includes:

[0022] Perform logarithmic domain Gaussian difference filtering on the input image;

[0023] Adaptively segment the reflection area according to the pixel gradient direction distribution;

[0024] Locate the coordinates of feature points in the segmented non-reflective area;

[0025] The adaptive segmentation of the reflection area includes:

[0026] Calculate the gradient direction histogram of the local area of ​​the image;

[0027] If the dispersion of the main gradient direction distribution exceeds the set range, it is determined to be a reflection noise area;

[0028] A morphological closing operation is performed on the noise region to connect the broken feature boundaries.

[0029] As a preferred solution of the camera calibration method based on three-dimensional calibration described in the present invention, in the process of logarithmic domain Gaussian difference filtering in step S2, the brightness span is first weakened in the logarithmic domain, then the geometric edges are highlighted by adaptive two-scale Gaussian difference, and finally the noise-resistant response map is generated by multi-scale weighted fusion. The specific steps include:

[0030] (1) Perform logarithmic domain compression:

[0031] L(x,y)=ln(I(x,y)+ε),

[0032] Where L(x,y) is the logarithmic domain pixel, I(x,y) is the normalized grayscale, and ε is a constant to prevent zero values ​​from appearing when taking the logarithm;

[0033] (2) Define the local brightness variance as:

[0034]

[0035] Among them, S w (x, y) is the local standard deviation of the window, W is the pixel set centered at (x, y), N win is the number of pixels in W, I(i,j) is the grayscale of the window pixel, is the average grayscale of the window;

[0036] (3) Define the adaptive fine scale as:

[0037] σ s (x,y)=σ min +(σ max -σ min )[1-S w (x,y)],

[0038] Among them, σ s (x,y) is the fine-scale standard deviation, σ min σ max are the upper and lower limits of the scale,

[0039] (4) Corresponding to the coarse scale σ l (x,y)=κσ s (x,y), where σ l (x,y) is the coarse-scale standard deviation, κ is the multiplication factor;

[0040] (5) The dual-scale convolution process is expressed as:

[0041]

[0042] Among them, L s (x,y) and L l (x, y) are the fine and coarse scale convolution results respectively, is the corresponding Gaussian kernel, represents a two-dimensional convolution operator;

[0043] (6) Perform differential response, the formula is: D(x,y)=L s (x,y)-L l (x,y), where D(x,y) is the dual-scale difference map;

[0044] (7) Perform global normalization, the formula is:

[0045]

[0046] Among them, D norm (x,y) is the standardized response, μ D is the global mean of D(x,y), σ D is the global standard deviation;

[0047] (8) Perform multi-scale construction. To obtain n-level fine scales, let σ m =σ min κ m-1 ,m=1,2,…,n, where, σ m is the mth level fine scale; repeat steps 5-7 to obtain D m,norm ;

[0048] (9) Perform weight normalization, the formula is:

[0049]

[0050] Among them, α m is the mth level weight, β is the dimensionless noise suppression tension coefficient, σ k is the kth level fine scale, n s is the number of scales, k is the weight sum index;

[0051] (10) Fusion output, expressed as:

[0052]

[0053] Among them, D fuse is the final multi-scale difference map, and m is the summation index.

[0054] As a preferred solution of the camera calibration method based on three-dimensional calibration described in the present invention, wherein: the camera motion trajectory data in step S3 is provided by the IMU sensor, including acceleration and angular velocity information; the three-dimensional space reprojection model is established by fusing the IMU pose initial value and the three-dimensional coordinates of the feature points;

[0055] The fused IMU pose initial value includes:

[0056] Perform sliding window mean filtering on IMU acceleration data to suppress welding vibration noise;

[0057] The filtered data is converted into camera pose changes through the pre-integration model;

[0058] The pose change is used as a regularization constraint for reprojection model optimization.

[0059] As a preferred solution of the camera calibration method based on three-dimensional calibration described in the present invention, the motion consistency check in step S4 includes:

[0060] Calculate the displacement vector of feature points in adjacent frames;

[0061] Eliminate transient interference points whose displacement direction deviates from the motion baseline;

[0062] Update the extrinsic rotation matrix based on the remaining static feature points.

[0063] As a preferred solution of the camera calibration method based on three-dimensional calibration described in the present invention, the segmentation of the reflection area adopts a dynamic threshold mechanism:

[0064] Adjust the segmentation threshold in real time according to the welding arc light intensity;

[0065] Enable the local template matching algorithm to complete feature points in the overexposed area.

[0066] As a preferred solution of the camera calibration method based on three-dimensional calibration described in the present invention, the distortion coefficient calculation in step S3 adopts a radial-tangential joint model, and its optimization objective function includes a nonlinear distortion compensation term caused by metal surface reflection.

[0067] As a preferred solution of the camera calibration method based on three-dimensional calibration described in the present invention, the distortion coefficient calculation process in step S3 includes: introducing a radial-tangential joint model and adding a metal surface reflection compensation term in the three-dimensional space reprojection model, and solving the distortion coefficient and intrinsic parameter simultaneously by minimizing the reprojection residual, specifically:

[0068] Perform pixel normalization:

[0069]

[0070] in, represents the normalized camera coordinates of the i-th feature point, K is the camera intrinsic parameter matrix, u i With v i are the horizontal and vertical coordinates of the pixel plane, respectively, and the number 1 is the homogeneous coordinate constant;

[0071] Calculate the radial distance using the formula:

[0072]

[0073] Among them, r i is the radial distance of point i, x i 、y i for The horizontal and vertical components;

[0074] Define the nonlinear distortion model:

[0075]

[0076] Among them, x d,i 、y d,i is the normalized coordinate after distortion, k1, k2, k3 are the first to third order radial distortion coefficients, q is the metal reflection compensation coefficient, dimensionless high-order radial term, p1, p2 are the tangential distortion coefficients;

[0077] Perform pixel reprojection:

[0078]

[0079] in, The pixel coordinates predicted by the model, f x ,f y is the focal length component, c x ,c y is the main point offset;

[0080] Construct the objective function:

[0081]

[0082] Where E(θ) is the optimization objective, θ=[k1,k2,k3,p1,p2,q] T is the coefficient vector to be estimated, T is the transpose symbol, N is the total number of feature points, u i is the observed pixel coordinate, λ is the regularization weight;

[0083] Finally, continue iterative update, the formula is:

[0084] θ t+1 =θ t -η(J TJ+λI) -1 J T e,

[0085] Among them, θ t+1 is the new round coefficient, θ t is the current coefficient, t is the iteration index, η is the step size, J is the Jacobian matrix of the objective function, I is the identity matrix, and e is the residual vector.

[0086] As a preferred solution of the camera calibration method based on three-dimensional calibration described in the present invention, it also includes: step S5, self-verification of the calibration result: back-projecting the calibration parameters to the original image, and triggering a parameter iterative optimization loop through the feature point reprojection residual.

[0087] The beneficial effects of the present invention are as follows: the present invention reuses the inherent three-dimensional features of the workpiece (weld edges, mounting holes), completely getting rid of the burden of carrying, placing and maintaining the checkerboard or target; the welding production line does not need to be shut down to intervene in the calibration process, ensuring continuous operation efficiency; especially solving the core pain point of manual access to the calibration object in high-risk scenarios (such as high-pressure pipeline welding); through multi-scale illumination transformation to remove high-frequency reflection interference, combined with the metal surface distortion compensation mechanism, to ensure the stability of feature point detection in arc light and metal splash environment. Even local overexposed areas can still be virtually supplemented by the CAD model, avoiding the calibration interruption caused by missing features of traditional methods. In addition, the IMU motion data and optical flow tracking are integrated to verify the consistency of feature point motion online, and automatically filter transient interference such as splashing slag. The external parameter drift is modeled as a function of the motion trajectory, and is corrected in real time through regularization constraints, which significantly suppresses the cumulative error caused by vibration.

[0088] The calibration results of the present invention trigger iterative optimization through reprojection residuals, forming an autonomous cycle of acquisition-calculation-verification; users do not need to manually verify the accuracy, and the system adaptively adjusts parameters until the millimeter-level accuracy threshold of welding detection is met, reducing the complexity of operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0090] Figure 1 Schematic diagram of the process of the camera calibration method based on three-dimensional calibration in Example 1. DETAILED DESCRIPTION

[0091] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0092] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0093] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0094] Example 1, with reference to Figure 1 This embodiment provides a camera calibration method based on three-dimensional calibration, including:

[0095] Step S1, real-time extraction of environmental 3D features: using an industrial camera to capture welding scene images and identify stable 3D geometric features on the workpiece surface. The 3D geometric features include the 3D coordinates of metal weld edges, equipment installation holes, and pipeline intersections.

[0096] The identification of stable three-dimensional geometric features in step S1 includes:

[0097] Extract straight line segments and arc contours from the surface of welding workpieces;

[0098] Screen edge intersections that meet orthogonality conditions through spatial geometric constraints;

[0099] Mapping the filtered two-dimensional features into a three-dimensional coordinate point cloud based on the depth sensor data;

[0100] Step S2, reflection noise suppression processing: perform multi-scale illumination invariant transformation on the image, separate high-frequency reflection noise from real feature texture, and dynamically generate an anti-interference feature point set;

[0101] Dynamically generated anti-interference feature point sets include:

[0102] Perform sub-pixel corner detection on the segmented non-reflective areas;

[0103] When the number of detection points is lower than the threshold, the virtual feature point generation module based on the welding workpiece CAD model is activated;

[0104] Perform spatial topological matching between virtual feature points and real detection points, and output a fused feature point set;

[0105] The multi-scale illumination invariant transformation in step S2 includes:

[0106] Perform logarithmic domain Gaussian difference filtering on the input image;

[0107] Adaptively segment the reflection area according to the pixel gradient direction distribution;

[0108] Locate the coordinates of feature points in the segmented non-reflective area;

[0109] Adaptive segmentation of reflection areas includes:

[0110] Calculate the gradient direction histogram of the local area of ​​the image;

[0111] If the dispersion of the main gradient direction distribution exceeds the set range, it is determined to be a reflection noise area;

[0112] Perform morphological closing operation on the noise area to connect the broken feature boundaries;

[0113] During the logarithmic domain Gaussian difference filtering in step S2, strong reflected light in the welding scene can easily mask the real texture. First, the brightness span is weakened in the logarithmic domain, then the geometric edges are highlighted by adaptive dual-scale Gaussian difference. Finally, the noise-resistant response map is generated through multi-scale weighted fusion. The specific steps include:

[0114] (1) Perform logarithmic domain compression:

[0115] L(x,y)=ln(I(x,t)+ε),

[0116] Where L(x,y) is the logarithmic domain pixel, I(x,y) is the normalized grayscale, and ε is a constant to prevent zero values ​​from appearing when taking the logarithm;

[0117] (2) Define the local brightness variance as:

[0118]

[0119] Among them, S w (x, y) is the local standard deviation of the window, W is the pixel set centered at (x, y), N win is the number of pixels in W, I(i,j) is the grayscale of the window pixel, is the average grayscale of the window;

[0120] (3) Define the adaptive fine scale as:

[0121] σ s (x,y)=σ min +(σ max -σ min )[1-S w (x,y)],

[0122] Among them, σs (x,y) is the fine-scale standard deviation, σ min σ max are the upper and lower limits of the scale,

[0123] (4) Corresponding to the coarse scale σ l (x,y)=κσ s (x,y), where σ l (x,y) is the coarse-scale standard deviation, κ is the multiplication factor;

[0124] (5) The dual-scale convolution process is expressed as:

[0125]

[0126] Among them, L s (x,y) and L l (x, y) are the fine and coarse scale convolution results respectively, is the corresponding Gaussian kernel, represents a two-dimensional convolution operator;

[0127] (6) Perform differential response, the formula is: D(x,y)=L s (x,y)-L l (x,y), where D(x,y) is the dual-scale difference map;

[0128] (7) Perform global normalization, the formula is:

[0129]

[0130] Among them, D norm (x,y) is the standardized response, μ D is the global mean of D(x,y), σ D is the global standard deviation;

[0131] (8) Perform multi-scale construction. To obtain n-level fine scales, let σ m =σ min κ m-1 ,m=1,2,…,n, where, σ m is the mth level fine scale; repeat steps 5-7 to obtain D m,norm ;

[0132] (9) Perform weight normalization, the formula is:

[0133]

[0134] Among them, α m is the mth level weight, β is the dimensionless noise suppression tension coefficient, σ k is the kth level fine scale, n sis the number of scales, k is the weight sum index;

[0135] (10) Fusion output, expressed as:

[0136]

[0137] Among them, D fuse is the final multi-scale difference map, m is the summation index;

[0138] Specifically, logarithmic domain compression weakens extremely bright areas, local variance measures texture richness and drives fine-scale adaptive magnification or contraction; two-scale convolution performs on-site comparison to remove low-frequency illumination and highlight edges; normalization eliminates exposure differences between scenes, and multi-scale ensembles expand the edge perception range; the tension coefficient suppresses the contribution of high-noise scales during weight normalization. The fusion result combines clear details with noise suppression, meeting the stability requirements of subsequent sub-pixel corner detection in weld scenes.

[0139] The reflection area segmentation adopts a dynamic threshold mechanism:

[0140] Adjust the segmentation threshold in real time according to the welding arc light intensity;

[0141] Enable local template matching algorithm to complete feature points in overexposed areas;

[0142] Step S3, 3D calibration parameter calculation: Based on the anti-interference feature point set and combined with the camera motion trajectory data, a 3D space reprojection model is constructed, and the camera intrinsic parameter matrix, distortion coefficient and extrinsic parameter pose are simultaneously solved;

[0143] In step S3, the camera motion trajectory data is provided by the IMU sensor, including acceleration and angular velocity information; the three-dimensional space reprojection model is established by fusing the IMU initial pose value and the three-dimensional coordinates of the feature points;

[0144] The initial value of the fused IMU pose includes:

[0145] Perform sliding window mean filtering on IMU acceleration data to suppress welding vibration noise;

[0146] The filtered data is converted into camera pose changes through the pre-integration model;

[0147] The pose change is used as a regularization constraint for reprojection model optimization;

[0148] In step S3, the distortion coefficient is calculated using a radial-tangential joint model, and its optimization objective function includes a nonlinear distortion compensation term caused by metal surface reflection;

[0149] The distortion coefficient calculation process in step S3 includes: introducing a radial-tangential joint model and adding a metal surface reflection compensation term into the 3D space reprojection model, and solving the distortion coefficient and internal parameter simultaneously by minimizing the reprojection residual, specifically:

[0150] Perform pixel normalization:

[0151]

[0152] in, represents the normalized camera coordinates of the i-th feature point, J is the camera intrinsic parameter matrix, u i With v i are the horizontal and vertical coordinates of the pixel plane, respectively, and the number 1 is the homogeneous coordinate constant;

[0153] Calculate the radial distance using the formula:

[0154]

[0155] Among them, r i is the radial distance of point i, x i 、y i for The horizontal and vertical components;

[0156] Define the nonlinear distortion model:

[0157]

[0158]

[0159] Among them, x d,i 、y d,i is the normalized coordinate after distortion, k1, k2, k3 are the first to third order radial distortion coefficients, q is the metal reflection compensation coefficient, dimensionless high-order radial term, p1, p2 are the tangential distortion coefficients;

[0160] Perform pixel reprojection:

[0161]

[0162] in, The pixel coordinates predicted by the model, f x ,f y is the focal length component, c x ,c y is the main point offset;

[0163] Construct the objective function:

[0164]

[0165] Where E(θ) is the optimization objective, θ=[k1,k2,k3,p1,p2,q] T is the coefficient vector to be estimated, T is the transpose symbol, N is the total number of feature points, u i is the observed pixel coordinate, λ is the regularization weight;

[0166] Finally, continue iterative update, the formula is:

[0167] θ t+1 =θ t -η(J T J+λI) -1 J T e,

[0168] Among them, θ t+1 is the new round coefficient, θ t is the current coefficient, t is the iteration index, η is the step size, J is the Jacobian matrix of the objective function, I is the identity matrix, and e is the residual vector;

[0169] Specifically, the model separates distortion and projection geometry in a normalized coordinate system, allowing radial and tangential terms to participate in the fitting simultaneously. The addition of an octal radial compensation term q captures high-order reflection distortions of metal surfaces, and the regularization term avoids polynomial divergence. The gradient damping update maintains stable parameter convergence even in highly reflective environments, reducing the reprojection residual by an average of 20%, providing a stable internal parameter foundation for subsequent extrinsic parameter solutions.

[0170] Step S4, dynamic parameter optimization: update the parameters of the 3D space reprojection model online through optical flow tracking and feature point motion consistency verification;

[0171] The motion consistency check in step S4 includes:

[0172] Calculate the displacement vector of feature points in adjacent frames;

[0173] Eliminate transient interference points whose displacement direction deviates from the motion baseline;

[0174] Update the extrinsic rotation matrix based on the remaining static feature points;

[0175] Step S5, self-verification of calibration results: back-project the calibration parameters to the original image, and trigger the parameter iterative optimization loop through the feature point reprojection residual.

[0176] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A camera calibration method based on three-dimensional calibration, characterized in that: The following steps are involved: Step S1, real-time extraction of environmental 3D features: using an industrial camera to capture welding scene images and identify stable 3D geometric features on the workpiece surface, including the 3D coordinates of metal weld edges, equipment installation holes, and pipeline intersections; Step S2, reflection noise suppression processing: performing a multi-scale illumination invariant transformation on the image, separating high-frequency reflection noise from real feature textures, and dynamically generating an anti-interference feature point set; The dynamically generated anti-interference feature point set includes: Perform sub-pixel corner detection on the segmented non-reflective areas; When the number of detection points is lower than the threshold, the virtual feature point generation module based on the welding workpiece CAD model is activated; Perform spatial topological matching between virtual feature points and real detection points, and output a fused feature point set; Step S3, 3D calibration parameter calculation: Based on the anti-interference feature point set and combined with the camera motion trajectory data, a 3D space reprojection model is constructed, and the camera intrinsic parameter matrix, distortion coefficient and extrinsic parameter pose are simultaneously solved; Step S4, dynamic parameter optimization: updating the parameters of the three-dimensional space reprojection model online through optical flow tracking and feature point motion consistency verification.

2. The camera calibration method based on three-dimensional calibration according to claim 1, characterized in that: The identification of stable three-dimensional geometric features in step S1 includes: Extract straight line segments and arc contours from the surface of welding workpieces; Screen edge intersections that meet orthogonality conditions through spatial geometric constraints; The filtered two-dimensional features are mapped into a three-dimensional coordinate point cloud based on the depth sensor data.

3. The camera calibration method based on three-dimensional calibration according to claim 1, wherein: The multi-scale illumination invariant transformation in step S2 includes: Perform logarithmic domain Gaussian difference filtering on the input image; Adaptively segment the reflection area according to the pixel gradient direction distribution; Locate the coordinates of feature points in the segmented non-reflective area; The adaptive segmentation of the reflection area includes: Calculate the gradient direction histogram of the local area of ​​the image; If the dispersion of the main gradient direction distribution exceeds the set range, it is determined to be a reflection noise area; A morphological closing operation is performed on the noise region to connect the broken feature boundaries.

4. The camera calibration method based on three-dimensional calibration according to claim 3, characterized in that: In the process of logarithmic domain Gaussian difference filtering in step S2, the brightness span is first weakened in the logarithmic domain, and then the geometric edges are highlighted by adaptive two-scale Gaussian difference. Finally, the noise-resistant response map is generated through multi-scale weighted fusion. The specific steps include: (1) Perform logarithmic domain compression: L(x,y)=ln(I(x,y)+ε), Where L(x,y) is the logarithmic domain pixel, I(x,y) is the normalized grayscale, and ε is a constant to prevent zero values ​​from appearing when taking the logarithm; (2) Define the local brightness variance as: Among them, S w (x, y) is the local standard deviation of the window, W is the pixel set centered at (x, y), N win is the number of pixels in W, I(i,j) is the grayscale of the window pixel, is the average grayscale of the window; (3) Define the adaptive fine scale as: s s (x,y)=σ min +(s max -s min )[1-S w (x,y)], Among them, σ s (x,y) is the fine-scale standard deviation, σ min σ max are the upper and lower limits of the scale, (4) Corresponding to the coarse scale σ l (x,y)=κσ s (x,y), where σ l (x,y) is the coarse-scale standard deviation, κ is the multiplication factor; (5) The dual-scale convolution process is expressed as: Among them, L s (x,y) and L l (x,y) are the fine and coarse scale convolution results respectively, is the corresponding Gaussian kernel, represents a two-dimensional convolution operator; (6) Perform differential response, the formula is: D(x,y)=L s (x,y)-L l (x,y), where D(x,y) is the dual-scale difference map; (7) Perform global normalization, the formula is: Among them, D norm (x,y) is the standardized response, μ D is the global mean of D(x,y), σ D is the global standard deviation; (8) Perform multi-scale construction. To obtain n-level fine scales, let σ m =σ min κ m-1 ,m=1,2,…,n, where, σ m is the mth level fine scale; repeat steps 5-7 to obtain D m,norm ; (9) Perform weight normalization, the formula is: Among them, α m is the mth level weight, β is the dimensionless noise suppression tension coefficient, σ k is the kth level fine scale, n s is the number of scales, k is the weight sum index; (10) Fusion output, expressed as: Among them, D fuse is the final multi-scale difference map, and m is the summation index.

5. The camera calibration method based on three-dimensional calibration according to claim 1, wherein: The camera motion trajectory data in step S3 is provided by the IMU sensor, including acceleration and angular velocity information; The three-dimensional space reprojection model is established by fusing the initial value of the IMU pose and the three-dimensional coordinates of the feature points; The fused IMU pose initial value includes: Perform sliding window mean filtering on IMU acceleration data to suppress welding vibration noise; The filtered data is converted into camera pose changes through the pre-integration model; The pose change is used as a regularization constraint for reprojection model optimization.

6. The camera calibration method based on three-dimensional calibration according to claim 1, characterized in that: The motion consistency check in step S4 includes: Calculate the displacement vector of feature points in adjacent frames; Eliminate transient interference points whose displacement direction deviates from the motion baseline; Update the extrinsic rotation matrix based on the remaining static feature points.

7. The camera calibration method based on three-dimensional calibration according to claim 3, characterized in that: The segmentation of the reflection area adopts a dynamic threshold mechanism: Adjust the segmentation threshold in real time according to the welding arc light intensity; Enable the local template matching algorithm to complete feature points in the overexposed area.

8. The camera calibration method based on three-dimensional calibration according to claim 1, wherein: The distortion coefficient calculation in step S3 adopts a radial-tangential joint model, and its optimization objective function includes a nonlinear distortion compensation term caused by metal surface reflection.

9. The camera calibration method based on three-dimensional calibration according to claim 8, characterized in that: The distortion coefficient calculation process in step S3 includes: introducing a radial-tangential joint model and adding a metal surface reflection compensation term in the three-dimensional space reprojection model, and solving the distortion coefficient and internal parameter simultaneously by minimizing the reprojection residual, specifically: Perform pixel normalization: in, represents the normalized camera coordinates of the i-th feature point, K is the camera intrinsic parameter matrix, u i With v i are the horizontal and vertical coordinates of the pixel plane, respectively, and the number 1 is the homogeneous coordinate constant; Calculate the radial distance using the formula: Among them, r i is the radial distance of point i, x i 、y i for The horizontal and vertical components; Define the nonlinear distortion model: Among them, x d,i 、y d,i is the normalized coordinate after distortion, k1, k2, k3 are the first to third order radial distortion coefficients, q is the metal reflection compensation coefficient, dimensionless high-order radial term, p1, p2 are the tangential distortion coefficients; Perform pixel reprojection: in, The pixel coordinates predicted by the model, f x ,f y is the focal length component, c x ,c y is the main point offset; Construct the objective function: Where E(θ) is the optimization objective, θ=[k1,k2,k3,p1,p2,q] T is the coefficient vector to be estimated, T is the transpose symbol, N is the total number of feature points, u i is the observed pixel coordinate, λ is the regularization weight; Finally, continue iterative update, the formula is: i t+1 =θ t -η(J T J+λI) -1 J T oh, Among them, θ t+1 is the new round coefficient, θ t is the current coefficient, t is the iteration index, η is the step size, J is the Jacobian matrix of the objective function, I is the identity matrix, and e is the residual vector.

10. The camera calibration method based on three-dimensional calibration according to claim 1, characterized in that: The method further includes: step S5, self-verification of the calibration result: back-projecting the calibration parameters to the original image, and triggering a parameter iterative optimization loop through the feature point reprojection residual.

Images