A low-invasive structure prior method for two-camera measurement of width

Abstract

This invention discloses a low-intrusion structural prior method for measuring width using dual cameras. This method requires only single-camera calibration of both cameras and quickly achieves scale alignment using a reference scale in the overlapping area. Based on adaptive weighting of edge fitting residuals, a shared normal is constructed by fusing weak prior normals in the transport direction, and width error compensation is achieved through parallelization correction. This invention eliminates the need for complex binocular stereo calibration, requires only 2 minutes for type-change calibration, has minimal computational load in the SPR module, and can reduce the standard deviation of width measurement by 20-40% without altering existing detection procedures, with virtually no additional systematic bias.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology, and more specifically to a low-intrusion structural prior method for measuring width using dual cameras. Background Technology

[0002] In industrial settings, online measurement of the width of transparent or semi-transparent workpieces such as glass and sheet metal has long faced the challenge of balancing high-speed production with micrometer-level measurement accuracy. In recent years, machine vision-based non-contact measurement solutions have gradually replaced traditional contact measurement methods, becoming the mainstream technology. However, different imaging schemes still have significant limitations in practical applications: While line scan cameras offer advantages such as continuous imaging and high resolution, their measurements are highly dependent on the uniform and stable movement of the workpiece. In on-site production, conditions such as longitudinal jitter, momentary slippage, and emergency stops during glass transport can all cause the image to be stretched or compressed in the direction of motion, thus introducing systematic errors in the width measurement direction. On the other hand, area scan cameras, limited by a fixed field of view, cannot simultaneously achieve overall coverage of a wide workpiece and high-precision imaging of local details, and are prone to decreased measurement accuracy due to insufficient edge imaging resolution. Therefore, in industrial settings, two cameras are often used to observe the left and right edges of the glass respectively, and the width is measured by calculating the edge spacing to balance the measurement range and accuracy requirements.

[0003] However, existing binocular vision measurement solutions still have many shortcomings: First, traditional binocular stereo calibration processes are complex. In addition to simultaneously calibrating the intrinsic parameters of the left and right cameras, additional baseline correction and disparity calculations are required to obtain stereo correction parameters. The calibration cycle is long and requires high levels of skill from the operating environment and personnel. At the same time, the baseline distance is easily affected by mechanical vibration and temperature deformation of the on-site support, and according to the principle of depth measurement, the baseline error will be further amplified with depth Z, directly reducing the stability and accuracy of width measurement. Second, existing solutions cannot simultaneously improve measurement speed and accuracy: the computational complexity of binocular stereo matching algorithms is high, making it difficult to adapt to high-speed production cycles; while simplifying the matching process will lead to a decrease in noise robustness, failing to meet the requirements of high-precision measurement. Third, if only single-target calibration is performed on two cameras, the lack of a unified scale benchmark provided by stereo calibration makes the measurement results susceptible to factors such as camera installation errors and uneven lighting, thus limiting the application scenarios. On the other hand, the monocular vision glass width measurement scheme is limited by the inherent defects in scale estimation and edge positioning. Under complex working conditions such as reflection from transparent materials, weak textures, and edge defects, the error amplification effect is particularly significant, making it difficult to meet the stable application requirements in industrial settings.

[0004] In view of this, the present invention is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a low-intrusion structural prior method for measuring width using dual cameras, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention specifically adopts the following technical solution: A low-intrusive structural prior method for measuring width using dual cameras includes the following steps: S1. Camera Deployment: Place two industrial cameras at the same height, with their main axes parallel to each other and perpendicular to the target measurement plane; S2. Single-camera calibration and plane mapping: Perform single-camera calibration and distortion correction on the two cameras respectively, solve the homography matrix from the pixel to the measurement plane, and establish the mapping from the left and right monocular coordinate systems to the same plane scale; determine the relative distance between the principal axes of the two cameras based on the precise reference objects in the overlapping area of ​​the images. S3. Edge Extraction and Reference Width Calculation: Perform sub-pixel edge extraction and line fitting on the measured target, solve for the normal distance between the two edge lines, and obtain the reference width by combining the relative distance of the principal axes; S4. Weak normal fusion and parallelization width calculation: Weak prior fusion of the normals of the two fitted edges is performed, and the reprojection bias under the shared normal direction is calculated to obtain the parallelization width. S5. Uncertainty Analysis and Linear Shrinkage Output: Uncertainty analysis is performed on the reference width and parallelized width, and the final measured width is obtained through linear shrinkage. S6. Online monitoring and anomaly alarm: Record the residual from the edge point to the fitted line. When the residual abnormally increases, trigger an environmental alarm to achieve online quality detection.

[0007] Preferably, the specific method of step S1 includes: Image streams from industrial cameras are used as algorithm inputs. During installation, two cameras are placed at the same height, with their main axes parallel to each other on both sides of the production line, and the shooting direction is directly facing the glass object being measured. An industrial camera with a focal length of f=16mm, a frame rate of 4.5fps, and a resolution of 5120×5120 is used. The camera posture is adjusted with the help of a level to make the main axis parallel to both sides of the production line and perpendicular to the glass surface.

[0008] Preferably, the specific method of step S2 includes: Two fixed cameras are used, and a checkerboard calibration board is placed at 20 different positions and poses to capture images. The intrinsic parameters of the two cameras are calibrated, and the homography matrix H from the pixels to the target plane is calculated, satisfying the following relationship:

[0009] H is solved in the distortion domain and applied to the distorted pixels during runtime to ensure coordinate domain consistency and establish a mapping relationship between the left and right monocular coordinate systems and the same plane scale. The reference block is placed at the same height as the glass, and the relative distance between the two camera principal axes is obtained by using the reference block in the overlapping area of ​​the image.

[0010] Preferably, the specific method of step S3 includes: Within a narrow-band ROI near the outer edge of the image, gradient detection is performed every few rows, and a sub-pixel edge point set P is obtained by three-point quadratic interpolation near the gradient peak. L With P R After projecting the edge points onto the millimeter domain, unit-form edge lines are obtained by fitting them respectively:

[0011] Unit direction parallel to both sides of the transport line As a reference, based on the principal axis distance between the two cameras -c Z n Z Calculate the reprojection offset distance W Z The following relation is satisfied:

[0012] Where n L n R These are the unit normals of the two edge lines, respectively; the reference width W. base The sum of the normal distance between the two edge lines and the offset distance of the principal axis reprojection satisfies the following relationship: .

[0013] Preferably, the specific method of step S4 includes: Based on the unit normal of the two edge lines and their respective fitted orthogonal residual standard deviation σ L σ R Construct weights Introduce a weak directional prior n0 along the transportation direction and assign it a minimal weight λ. Obtain the shared normal through weighted fusion, satisfying the following relation:

[0014] For two target edge lines in a scene with strong reflection, weak texture, and edge defects, let their true normal be n. ∗ The closest points of the two edge lines under the true normal direction are respectively The camera principal axis direction vector is n Z Its closest point parameter to the origin in the true direction is c. Z The normal deviation fitted by each of the two edge lines is δ. nLδ nR Then the base line width W base satisfy:

[0015] in, At the point closest to the origin in the true method; In shared legal direction Next, calculate the offset of the two edge lines. Offset relative to the distance between the two camera axes They respectively satisfy the following relation:

[0016] The corresponding distance between the two camera axes is The following bias is:

[0017] Thus, a parallelized width can be determined:

[0018]

[0019] Preferably, the specific method of step S5 includes: For the reference width W base and parallelization width W spr Perform error analysis and denote δ nL δ nR For the first-order perturbation of the normal error, δ CL δ CR For offset error, the reference width W base The error term is:

[0020] Parallelization width W spr The error term is:

[0021] in For the normalized first-order perturbation; when n L ≈n R At the same level of pixel noise, the parallelization width W spr The variance is less than the reference width W base The variance; the final width W is calculated through linear shrinkage. final and its variance V ar (W final ).

[0022] Preferably, the final width W is calculated through linear shrinkage. final and its variance Var (W final The relationship is as follows:

[0023] Where α∈[0.1,0.2], in order to reduce the measurement standard deviation while keeping the mean stable.

[0024] Preferably, the specific method of step S6 includes: Calculate the orthogonal distance from each edge point to its respective fitted line, and record the residual σ. L σ R The variance of the final width; when the residual term increases abnormally, the software and hardware interaction between the algorithm and the alarm is realized through the programmable development board and Arduino program to trigger the environmental alarm and complete the online quality inspection of the glass; at the same time, the image is saved for tracing and identifying the specific environmental conditions.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The two cameras of this invention only require single-target calibration, eliminating the need for complex dual-target calibration, epipolar correction, and parallax calculation processes, significantly reducing the difficulty of on-site calibration and maintenance, and shortening downtime. This solution achieves scale alignment by using a reference ruler in the overlapping area, allowing on-site calibration updates to be completed within 2 minutes during model changes. The process is simple and efficient, ensuring that the production line operating speed remains constant while meeting accuracy requirements. It is suitable for online real-time detection scenarios and can be extended to various geometric dimension detection tasks involving bilateral parallel measurements. 2. This invention utilizes edge fitting residuals to set weights, automatically suppressing interference from local defects such as reflections and edge chipping. It introduces extremely weak directional priors with very low intrusion and combines this with linear shrinkage to output the final width, effectively reducing measurement errors caused by inconsistencies in normals. The SPR module is located in the output layer, involving only vector weighting and one reprojection operation, with negligible computational load. It requires no modification to the existing system's intrinsic parameter calibration, homography solution, edge extraction, and line fitting processes, making it highly practical for engineering applications. Its control parameter α supports online adjustment; when α∈[0.1,0.2], it can reduce the standard deviation of width measurement by 20-40% with almost no additional system bias. 3. This invention is based on adaptive weighting of edge fitting residuals, which can effectively suppress interference caused by strong reflection, edge chipping, and local defects; through weak normal fusion and parallelization correction, it eliminates systematic errors caused by small-angle deviation of double edge normals, thus improving measurement robustness. Attached Figure Description

[0026] Figure 1 This is a diagram showing the camera installation location in actual production of this invention; Figure 2 Capture glass subpixel edge point sets using two cameras; Figure 3 This is a simplified diagram showing the width of the glass. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The materials and instruments used in the following embodiments are all commercially available.

[0028] A low-intrusive structural prior method for measuring width using dual cameras, comprising steps S1 to S6.

[0029] S1. Camera Deployment: Place two industrial cameras at the same height, with their main axes parallel to each other and perpendicular to the target measurement plane.

[0030] like Figure 1 As shown, the specific method of step S1 includes: Image streams from industrial cameras are used as algorithm inputs. During installation, two cameras are placed at the same height, with their main axes parallel to each other on both sides of the production line, and the shooting direction is directly facing the glass object being measured. An industrial camera with a focal length of f=16mm, a frame rate of 4.5fps, and a resolution of 5120×5120 is used. The camera posture is adjusted with the help of a level to make the main axis parallel to both sides of the production line and perpendicular to the glass surface.

[0031] S2. Single-camera calibration and plane mapping: Perform single-camera calibration and distortion correction on the two cameras respectively, solve the homography matrix from the pixel to the measurement plane, and establish the mapping from the left and right monocular coordinate systems to the same plane scale; determine the relative distance between the principal axes of the two cameras based on the precise reference objects in the overlapping area of ​​the images.

[0032] The specific methods for step S2 include: Two fixed cameras are used, and a checkerboard calibration board is placed at 20 different positions and poses to capture images. The intrinsic parameters of the two cameras are calibrated, and the homography matrix H from the pixels to the target plane is calculated, satisfying the following relationship:

[0033] H is solved in the distortion domain and applied to the distorted pixels during runtime to ensure coordinate domain consistency and establish a mapping relationship between the left and right monocular coordinate systems and the same plane scale. The reference block is placed at the same height as the glass, and the relative distance between the two camera principal axes is obtained by using the reference block in the overlapping area of ​​the image.

[0034] S3. Edge Extraction and Reference Width Calculation: Sub-pixel edge extraction and line fitting are performed on the measurement target to solve the normal distance between the two edge lines. The reference width is obtained by combining the relative distance of the principal axes.

[0035] like Figure 2 As shown, the specific method of step S3 includes: Within a narrow-band ROI near the outer edge of the image, gradient detection is performed every few rows, and a sub-pixel edge point set P is obtained by three-point quadratic interpolation near the gradient peak. L With P R After projecting the edge points onto the millimeter domain, unit-form edge lines are obtained by fitting them respectively:

[0036] Unit direction parallel to both sides of the transport line As a reference, based on the principal axis distance between the two cameras -c Z n Z Calculate the reprojection offset distance W Z The following relation is satisfied:

[0037] Where n L n R These are the unit normals of the two edge lines, respectively; for example... Figure 3 As shown, the reference width W base The sum of the normal distance between the two edge lines and the offset distance of the principal axis reprojection satisfies the following relationship: .

[0038] S4. Weak normal fusion and parallelization width calculation: Weak prior fusion is performed on the normals of the two fitted edges, and the reprojection bias under the shared normal is calculated to obtain the parallelization width.

[0039] The specific methods for step S4 include: Based on the unit normal of the two edge lines and their respective fitted orthogonal residual standard deviation σ L σ R Construct weights Introduce a weak directional prior n0 along the transportation direction and assign it a minimal weight λ. Obtain the shared normal through weighted fusion, satisfying the following relation:

[0040] For two target edge lines in a scene with strong reflection, weak texture, and edge defects, let their true normal be n. ∗ The closest points of the two edge lines under the true normal direction are respectively The camera principal axis direction vector is nZ Its closest point parameter to the origin in the true direction is c. Z The normal deviation fitted by each of the two edge lines is δ. nL δ nR Then the base line width W base satisfy:

[0041] in, At the point closest to the origin in the true method; In shared legal direction Next, calculate the offset of the two edge lines. Offset relative to the distance between the two camera axes They respectively satisfy the following relation:

[0042] The corresponding distance between the two camera axes is The following bias is:

[0043] Thus, a parallelized width can be determined:

[0044]

[0045] S5. Uncertainty Analysis and Linear Shrinkage Output: Uncertainty analysis is performed on the reference width and parallelized width, and the final measurement width is obtained through linear shrinkage.

[0046] The specific methods for step S5 include: Considering that introducing a parallelization width can significantly reduce the error caused by normal inconsistency, while introducing too many weak priors can also introduce certain systematic biases and affect the measurement results, we conduct an uncertainty analysis on the baseline width and parallelization width. For the reference width W... base and parallelization width W spr Perform error analysis and denote δ nL δ nR For the first-order perturbation of the normal error, δ CL δ CR For offset error, the reference width W base The error term is:

[0047] Parallelization width W spr The error term is:

[0048] in For the normalized first-order perturbation; when n L ≈n R At the same level of pixel noise, the parallelization width W spr The variance is less than the reference width W base The variance is relatively small, and the weak parallelization method can greatly reduce the variance, but it will also introduce a small system offset. Finally, the final width W is calculated through linear contraction. final and its variance V ar (W final The relationship is as follows:

[0049] Where α∈[0.1,0.2], in order to reduce the measurement standard deviation while keeping the mean stable.

[0050] S6. Online monitoring and anomaly alarm: Record the residual from the edge point to the fitted line. When the residual abnormally increases, trigger an environmental alarm to achieve online quality detection.

[0051] The specific methods for step S6 include: Calculate the orthogonal distance from each edge point to its respective fitted line, and record the residual σ. L σ R The variance of the final width; when the residual term increases abnormally, the software and hardware interaction between the algorithm and the alarm is realized through the programmable development board and Arduino program to trigger the environmental alarm and complete the online quality inspection of the glass; at the same time, the image is saved for tracing and identifying the specific environmental conditions.

[0052] In summary, the low-intrusion structural prior method for dual-camera width measurement provided by this invention, on the one hand, establishes a simple and unified dimensional benchmark for bilateral width measurement in photovoltaic glass production lines, effectively improving the reliability of single-target calibration and planar mapping, and reducing the system's dependence on complex stereo calibration processes; on the other hand, by introducing a weak normal fusion and parallelization correction mechanism, it provides stable and low-latency output results for width measurement under harsh conditions such as high-speed movement, strong reflection, and edge chipping, thereby enhancing the robustness and engineering applicability of the photovoltaic glass online quality inspection system.

Claims

1. A low-intrusive structural prior method for measuring width using dual cameras, characterized in that, Includes the following steps: S1. Camera Deployment: Place two industrial cameras at the same height, with their main axes parallel to each other and perpendicular to the target measurement plane; S2. Single-camera calibration and plane mapping: Perform single-camera calibration and distortion correction on the two cameras respectively, solve the homography matrix from the pixel to the measurement plane, and establish the mapping from the left and right monocular coordinate systems to the same plane scale; determine the relative distance between the principal axes of the two cameras based on the precise reference objects in the overlapping area of ​​the images. S3. Edge Extraction and Reference Width Calculation: Perform sub-pixel edge extraction and line fitting on the measured target, solve for the normal distance between the two edge lines, and obtain the reference width by combining the relative distance of the principal axes; S4. Weak normal fusion and parallelization width calculation: Weak prior fusion of the normals of the two fitted edges is performed, and the reprojection bias under the shared normal direction is calculated to obtain the parallelization width. S5. Uncertainty Analysis and Linear Shrinkage Output: Uncertainty analysis is performed on the reference width and parallelized width, and the final measured width is obtained through linear shrinkage. S6. Online monitoring and anomaly alarm: Record the residual from the edge point to the fitted line. When the residual abnormally increases, trigger an environmental alarm to achieve online quality detection.

2. The low-intrusive structural prior method for measuring width using dual cameras according to claim 1, characterized in that, The specific method of step S1 includes: Image streams from industrial cameras are used as algorithm inputs. During installation, two cameras are placed at the same height, with their main axes parallel to each other on both sides of the production line, and the shooting direction is directly facing the glass object being measured. An industrial camera with a focal length of f=16mm, a frame rate of 4.5fps, and a resolution of 5120×5120 is used. The camera posture is adjusted with the help of a level to make the main axis parallel to both sides of the production line and perpendicular to the glass surface.

3. The low-intrusive structural prior method for measuring width using dual cameras according to claim 1, characterized in that, The specific method of step S2 includes: Two fixed cameras are used, and a checkerboard calibration board is placed at 20 different positions and poses to capture images. The intrinsic parameters of the two cameras are calibrated, and the homography matrix H from the pixels to the target plane is calculated, satisfying the following relationship: ； H is solved in the distortion domain and applied to the distorted pixels during runtime to ensure coordinate domain consistency and establish a mapping relationship between the left and right monocular coordinate systems and the same plane scale. The reference block is placed at the same height as the glass, and the relative distance between the two camera principal axes is obtained by using the reference block in the overlapping area of ​​the image.

4. The low-intrusive structural prior method for measuring width using dual cameras according to claim 1, characterized in that, The specific method of step S3 includes: Within a narrow-band ROI near the outer edge of the image, gradient detection is performed every few rows, and a sub-pixel edge point set P is obtained by three-point quadratic interpolation near the gradient peak. L With P R After projecting the edge points onto the millimeter domain, unit-form edge lines are obtained by fitting them respectively: ； Unit direction parallel to both sides of the transport line As a reference, based on the principal axis distance between the two cameras -c Z n Z Calculate the reprojection offset distance W Z The following relation is satisfied: ； Where n L n R These are the unit normals of the two edge lines, respectively; the reference width W. base The sum of the normal distance between the two edge lines and the offset distance of the principal axis reprojection satisfies the following relationship: 。 5. The low-intrusive structural prior method for measuring width using dual cameras according to claim 1, characterized in that, The specific method of step S4 includes: Based on the unit normal of the two edge lines and their respective fitted orthogonal residual standard deviation σ L σ R Construct weights Introduce a weak directional prior n0 along the transportation direction and assign it a minimal weight λ. Obtain the shared normal through weighted fusion, satisfying the following relation: ； For two target edge lines in a scene with strong reflection, weak texture, and edge defects, let their true normal be n. ∗ The closest points of the two edge lines under the true normal direction are respectively The camera principal axis direction vector is n Z Its closest point parameter to the origin in the true direction is c. Z The normal deviation fitted by each of the two edge lines is δ. nL δ nR Then the base line width W base satisfy: ； in, At the point closest to the origin in the true method; In shared legal direction Next, calculate the offset of the two edge lines. Offset relative to the distance between the two camera axes They respectively satisfy the following relation: ； The corresponding distance between the two camera axes is The following bias is: ； Thus, a parallelized width can be determined: ； 。 6. The low-intrusive structural prior method for measuring width using dual cameras according to claim 1, characterized in that, The specific method of step S5 includes: For the reference width W base and parallelization width W spr Perform error analysis and denote δ nL δ nR For the first-order perturbation of the normal error, δ CL δ CR For offset error, the reference width W base The error term is: ； Parallelization width W spr The error term is: ； in For the normalized first-order perturbation; when n L ≈n R At the same level of pixel noise, the parallelization width W spr The variance is less than the reference width W base The variance; the final width W is calculated through linear shrinkage. final and its variance V ar (W final ).

7. The low-intrusive structural prior method for measuring width using dual cameras according to claim 6, characterized in that, The final width W is calculated using linear shrinkage. final and its variance V ar (W final The relationship is as follows: ； Where α∈[0.1,0.2], in order to reduce the measurement standard deviation while keeping the mean stable.

8. The low-intrusive structural prior method for measuring width using dual cameras according to claim 1, characterized in that, The specific method of step S6 includes: Calculate the orthogonal distance from each edge point to its respective fitted line, and record the residual σ. L σ R The variance of the final width; when the residual term increases abnormally, the software and hardware interaction between the algorithm and the alarm is realized through the programmable development board and Arduino program to trigger the environmental alarm and complete the online quality inspection of the glass; at the same time, the image is saved for tracing and identifying the specific environmental conditions.

Images