WO2017113535A1

WO2017113535A1 - Method and apparatus for geometric calibration of camera

Info

Publication number: WO2017113535A1
Application number: PCT/CN2016/078908
Authority: WO
Inventors: 赵博
Original assignee: 完美幻境(北京)科技有限公司
Priority date: 2015-12-30
Filing date: 2016-04-08
Publication date: 2017-07-06
Also published as: CN105678740B; CN105678740A

Abstract

Provided in the embodiments of the present invention are a method and apparatus for the geometric calibration of a camera, the method comprising: determining a camera parameter starting value, and acquiring a first cosine distance of a point pair calculated on the basis of the starting value, the point pair being the coordinates of two pixel points corresponding to the same spatial coordinate position in two adjacent pictures; determining whether the first cosine distance matches a preset value, and if not, then acquiring first iteration parameters, and using the first iteration parameters to adjust the starting value to acquire a first optimised value; acquiring a second cosine distance of the point pair calculated on the basis of the optimised value; determining whether the second cosine distance matches the preset value, and if so, then determining that the first optimised value is the camera parameter calibration value. The present solution greatly reduces the amount of calculation involved in the processing, and is beneficial for the implementation of real-time online processing.

Description

Camera geometric calibration processing method and device

Technical field

The present invention relates to the field of image processing, and in particular, to a camera geometric calibration processing method and apparatus.

Background technique

Panoramic shooting usually refers to a method of shooting 360 degrees and 180 degrees vertically with a certain point as the center, and stitching a plurality of pictures taken into a single panoramic picture and a picture stitching method. In general, panoramic shooting can include at least a panoramic image and a panoramic video.

Usually, when stitching a plurality of original pictures taken into one panoramic picture, mapping and stitching are involved. The mapping can be understood as projecting the pixel points on the original picture to the corresponding positions of the panoramic picture, and the splicing can be understood as a fusion transition of the overlapping areas of the two adjacent original pictures.

In order to determine the relationship between the three-dimensional geometric position of a point on the surface of the space object and its corresponding point in the original picture, the camera parameters can be obtained by means of camera geometric calibration, so that the camera point can be subsequently used for pixel point projection. Typically, camera parameters may include external parameters of the camera (eg, roll, yaw, pitch, Tx, Ty, Tz) and internal parameters of the camera (eg, u, v, w, alpha, beta, gamma). Where (Tx, Ty, Tz) represents the translation vector, (roll, yaw, pitch) represents the rotation matrix, which represents the z-axis rotation angle γ around the camera coordinate system, the rotation angle around the y-axis is β, and the rotation angle around the x-axis Let α; (u, v, w) denote the deflection distortion, and (α, β, γ) denote the fisheye lens imaging model parameters.

At present, most of them use the Levenberg-Marquardt algorithm (abbreviated as L-M algorithm) to perform iterative calculation to realize camera geometric calibration. In this method, each iterative process needs to obtain a second-order partial derivative for each parameter to be estimated, and obtain a Hessian matrix and a Jacobian matrix. When adopting N cameras for panoramic shooting, the calibration process involves 12*N parameters to be estimated, which is computationally intensive and requires a lot of calculation time. It is often used for offline calibration of the server, and real-time online processing is not yet available.

In addition, based on the L-M algorithm for geometric calibration, if the determinant of the two matrices is zero in the iterative process, that is, the transformation matrix is singular, then the local optimal value of the camera parameters can be considered. Generally, the local optimum value is affected by the initial value of the camera parameters. Different initial values may cause the variation matrix singularity to occur in different iterations, thus obtaining different local optimal values. Therefore, the method selects the initial value. There are also higher requirements.

Summary of the invention

The embodiment of the invention provides a camera geometric calibration processing method and device, which can greatly reduce the calculation amount involved in the processing process, and is beneficial to real-time online processing.

A camera geometric calibration processing method, the method comprising:

Determining a camera parameter initial value, and obtaining a first cosine distance of the pair of points calculated based on the initial value, the point pair being coordinates of two pixel points corresponding to the same position of the spatial coordinate in two adjacent pictures;

Determining whether the first cosine distance is consistent with a preset value, if not, acquiring a first iteration parameter, and adjusting the initial value by using the first iteration parameter to obtain a first optimization value;

Obtaining a second cosine distance of the pair of points calculated based on the first optimized value;

Determining whether the second cosine distance matches the preset value, and if so, determining the first optimized value as a camera parameter calibration value.

Preferably, if the second cosine distance does not match the preset value, the method further includes:

Obtaining a second iteration parameter, and adjusting the first optimization value by using the second iteration parameter to obtain a second optimization value;

Obtaining a third cosine distance of the pair of points calculated based on the second optimized value;

Determining whether the third cosine distance matches the preset value, and if so, determining the second optimized value as a camera parameter calibration value.

Preferably, the determining the initial value of the camera parameter comprises:

Determining an a priori value of the camera parameter as the initial value; or

A random disturbance is added based on the a priori value of the camera to obtain the initial value.

Preferably, the way to obtain the iterative parameters is:

Machine learning is performed on the preset samples to obtain the iterative parameters.

Preferably, the machine learning is performed on the preset sample, and the manner of obtaining the first iteration parameter is:

Where (M ₀ , N ₀ ) represents the first iteration parameter and C ^j represents the camera identity number.

Indicates the camera parameter calibration value of the jth camera,

Indicates the initial value of the camera parameter of the jth camera,

Represents the first cosine distance of the pair of points calculated based on (M ₀ , N ₀ ).

Preferably, the adjusting the initial value by using the first iteration parameter to obtain a first optimized value

The way is:

A camera geometric calibration processing device, the device comprising:

a cosine distance calculation unit, configured to determine an initial value of the camera parameter, and obtain a first cosine distance of the pair of points calculated based on the initial value, where the pair of points is two of the two adjacent pictures corresponding to the same position of the space coordinate The coordinates of the pixel;

An optimization value adjustment unit, configured to determine whether the first cosine distance is consistent with a preset value, and if not, acquiring a first iteration parameter, and adjusting the initial value by using the first iteration parameter to obtain a first optimization value;

The cosine distance calculation unit is further configured to obtain a second cosine distance of the pair of points calculated based on the first optimization value;

The calibration value determining unit is configured to determine whether the second cosine distance matches the preset value, and if yes, determine the first optimization value as a camera parameter calibration value.

Preferably, the device further comprises:

The optimization value adjustment unit is further configured to: when the second cosine distance does not match the preset value, acquire a second iteration parameter, and adjust the first optimization value by using the second iteration parameter to obtain Second optimized value;

The cosine distance calculation unit is further configured to obtain a third cosine distance of the pair of points calculated based on the second optimized value;

The calibration value determining unit is further configured to determine whether the third cosine distance matches the preset value, and if yes, determine the second optimization value as a camera parameter calibration value.

Preferably, the device further comprises:

An iterative parameter obtaining unit, configured to perform machine learning on the preset sample, to obtain the first iteration parameter:

Indicates the camera parameter calibration value of the jth camera,

Indicates the initial value of the camera parameter of the jth camera,

Preferably, the optimization value adjustment unit is specifically configured to be used according to

Calculate the first optimized value

Compared with the prior art, the scheme of the present invention uses the Taylor polynomial to fit the mapping relationship in the image mosaic process, and converts the optimization problem of the image model parameters into a nonlinear convex quadratic optimization problem of the error between the polynomial and the mapping function. Specifically, the pixel point projection may be performed based on the current parameter of the camera to obtain a cosine distance between the pair of projected points, and then it is determined according to the cosine distance whether camera parameter optimization is needed, and if not, the current parameter may be used as a camera parameter calibration value. If necessary, machine learning can be used to obtain iterative parameters to optimize the current parameters, and the optimized parameters are used for iterative calculation until the camera parameter calibration value is obtained. Such a scheme can greatly reduce the amount of calculation involved in the processing process and contribute to real-time online processing.

DRAWINGS

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly described below. It is obvious that the drawings in the following description are only some embodiments of the present invention. Other drawings may also be obtained from those of ordinary skill in the art in view of the drawings.

1 is a flow chart of a method for processing geometric calibration of a camera of the present invention;

2 is a diagram showing an effect of image stitching implemented according to the solution of the present invention;

FIG. 3 is a diagram showing an effect of image stitching implemented according to the prior art solution; FIG.

4 is a schematic structural view of a camera geometric calibration processing apparatus of the present invention.

detailed description

The technical solutions in the embodiments of the present invention are clearly and completely described in the following with reference to the accompanying drawings in the embodiments of the present invention. It is obvious that the described embodiments are only a part of the embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts are The scope of the invention is protected.

Before introducing the specific scheme of the present invention, the design idea of the present invention will be briefly introduced.

When performing panoramic shooting, there may be a partial overlapping area in the original picture taken by two adjacent cameras, and the overlapping area is merged to obtain a panoramic picture. Corresponding to this, we can understand that for the overlapping area, there are the following scenes: two adjacent original pictures A and B, one pixel point a exists in picture A, one pixel point b exists in picture B, and both a and b Corresponds to the same position in the space coordinates. Corresponding to this, the pixel points a and b can be referred to as a point pair.

During the splicing process, the pixel points a and b on the original picture can be mapped into the Equirectangular projection image respectively based on the current camera parameters, and a and b corresponding to the point pairs a' and b' on the Equirectangular image are obtained. In general, if a' and b' can be overlapped, the best stitching effect can be achieved. As an example, the cosine distance between the vectors a' and b' in the three-dimensional space can be calculated. The cosine distance can reflect the difference between a' and b', and then it can be known whether the current camera parameters are suitable for Panorama picture stitching.

That is to say, the cosine distance can be used to judge whether the current camera parameters are suitable, and if appropriate, the panoramic picture stitching can be performed; if not, the camera parameters can be optimized to determine the camera parameters that can be used for the panorama picture stitching. Calibration value. In the solution of the present invention, determining whether the camera parameter is appropriate can be understood as whether the cosine distance matches the preset value, that is, whether the cosine distance falls within a range allowed by the preset value. For example, the preset value may be 0.1. Generally, the smaller the preset value is, that is, the closer the cosine distance is to 0, the better the effect of the panoramic picture stitched based on the optimized calibration value is. The specific value of the preset value is not limited and may be determined by the actual application.

In summary, we can convert the panoramic camera geometry calibration process to find a gradient direction step size Δx to solve the following optimization problems:

among them,

a ⁱ and b ⁱ represent the i-th point pair of the adjacent original picture coincidence region;

h(*) denotes a function formed by camera parameters to realize coordinate transformation from a point on the original picture to a point on the equidistant spherical image based on camera parameters, eg, a ^{i is} converted to a ^i' , b ^{i is} converted to b ^i′ ; h(*) can be embodied in various forms according to practical applications, and the present invention does not specifically limit this, as long as coordinate conversion can be performed based on camera parameters;

C(*,*) represents the cosine distance of two points on the equidistant spherical image, such as the cosine distance of point pairs a ^i' and b ^{i '} .

The scheme of the present invention will be exemplified below.

Referring to FIG. 1, a flowchart of a camera geometric calibration processing method according to an embodiment of the present invention is shown, which may include the following steps:

S101. Determine a camera parameter initial value, and obtain a first cosine distance of the pair of points calculated based on the initial value, where the pair of points is a coordinate of two pixel points corresponding to the same position of the spatial coordinate in two adjacent pictures.

The invention determines whether the camera parameter optimization is needed by judging whether the cosine distance of the point pair falls within the allowable range of the preset value, so the initial value of a camera parameter may be selected first, and the point pair is performed based on the initial value. Mapping, and then extracting the first cosine distance of the pair of points corresponding to the initial value.

For example, the initial value in the solution of the present invention may be set in combination with actual operational experience; or, considering that the prior value of the camera parameter is generally a superior value, the solution in the solution of the present invention may also be set according to the prior value. The initial value, for example, may determine the a priori value of the camera parameter as an initial value; or, the random disturbance may be added to the camera's prior value to obtain an initial value. For example, the random disturbance value may be set according to the actual operation experience, or the random disturbance value may be set to ± (a priori value / 100), which may not be specifically limited in the embodiment of the present invention.

For example, C ^j represents the identity number of the jth camera at the time of panorama shooting.

Indicates the initial value of the camera parameter of the jth camera, a ^{j, i} and b ^{j, i} represents the i-th point pair on the original picture taken by the jth camera, which can be

Substituting the formula f(x), the first cosine distance d _{1 of the} pair of points calculated based on the initial value is obtained.

S102. Determine whether the first cosine distance is consistent with a preset value. If not, obtain a first iteration parameter, and adjust the initial value by using the first iteration parameter to obtain a first optimization value.

After obtaining d ₁ , a judgment as to whether or not to optimize the camera parameters can be performed, that is, whether or not d ₁ matches the preset value. Still taking the preset value as 0.1 as an example, the first cosine distance matching the preset value can be understood as d ₁ ≤ 0.1, otherwise the first cosine distance is considered to be inconsistent with the preset value.

If it is determined that the camera parameter is required to be optimized, the first iteration parameter may be obtained from the predetermined set of iterative parameters, and the initial value in S101 is adjusted by using the first iteration parameter to obtain the first optimized value of the camera parameter.

For example, camera parameters can be optimized by the following formula:

Where (M _k-1 , N _k-1 ) represents the kth iteration parameter,

Indicates the camera parameters that the jth camera needs to be optimized,

Indicates the camera parameters optimized for the jth camera.

In this step, the first optimized value

(M ₀ , N ₀ ) represents the first iteration parameter,

Indicates the initial value of the camera parameter of the jth camera,

Represents the first cosine distance d _{1 of the} pair of points calculated based on (M ₀ , N ₀ ).

In the solution of the present invention, an iterative parameter set can be obtained by performing machine learning on a preset sample. Generally, a set of iterative parameters is required for each iteration. For the manner in which the present invention obtains iterative parameters, reference is made to the following description, which will not be described in detail herein.

S103. Obtain a second cosine distance of the pair of points calculated based on the first optimized value.

S104. Determine whether the second cosine distance matches the preset value. If yes, determine the first optimization value as a camera parameter calibration value.

Similar to S101, obtaining the first optimized value

After that, you will be

Substituting the formula f(x), the second cosine distance d _{2 of the} pair of points corresponding to the first optimized value is obtained, and then the determination of whether to optimize the camera parameter is performed again by using d ₂ .

Specifically, if the judgment result indicates that d ₂ matches the preset value, the iterative process may be ended, and the first optimized value is obtained.

It is determined as the parameter calibration value of the jth camera, and the pixel stitching based on the calibration value can optimize the stitching effect of the panoramic picture.

In summary, the present invention uses the Taylor polynomial to fit the mapping relationship in the image stitching process, and transforms the image model parameter optimization problem into a nonlinear convex quadratic optimization problem between the polynomial and the mapping function. The camera geometric calibration based on the solution of the present invention does not need to obtain the second-order partial derivative according to the existing scheme, which can greatly reduce the calculation amount involved in the processing, and is helpful for real-time online processing. In addition, the invention is based on a machine learning algorithm to fit the gradient descent direction in the optimization problem, and also simplifies the process flow, which helps accelerate the convergence speed. Improve the convergence efficiency of optimization problems. In addition, the camera parameter calibration value determined based on the solution of the present invention is not easy to fall into local optimum, so that the optimization precision of the solution of the present invention is higher.

Optionally, if the judgment result of S104 indicates that d ₂ does not match the preset value, it indicates that the camera parameter calibration value is further obtained through a further iterative process. Specifically, the second iteration parameter may be obtained, and the first optimization value is adjusted by using the second iteration parameter to obtain a second optimization value; and the third point of the point pair calculated based on the second optimization value is obtained. a cosine distance; determining whether the third cosine distance matches the preset value, and if so, determining the second optimized value as a camera parameter calibration value. In this example, the manner of obtaining the second optimized value, the manner of calculating the third cosine distance by using the second optimized value, the manner of determining whether the third cosine distance matches the preset value, the manner of performing subsequent processing according to the determination result, and the like For example, reference may be made to the descriptions made at S102 to S104 above, and details are not described herein again.

The following briefly introduces the manner in which the present invention obtains an iterative parameter set.

In the first method, a set of iterative parameters can be obtained by means of machine learning.

Specifically, through the idea of supervised learning in machine learning, training samples are used to learn the set of iterative parameters {M ₀ , M ₁ , . . . , M _k-1 , M _k } and {N ₀ , which are required for each iteration. N ₁ ,...,N _k-1 ,N _k }.

The parameters involved in the training sample can be reflected as follows:

a series of camera encodings for panoramic photography {C ^j };

Camera parameter calibration value

The camera parameter calibration value representing the jth camera is a 12*n-dimensional vector;

A series of point pairs {a ^j,i ,b ^j,i }, a ^j,i and b ^j,i represent the i-th point pair on the original picture taken by the jth camera;

Camera parameter initial value

Indicates the initial value of the camera parameter of the jth camera.

Combining the above sample parameters, the first iteration parameter (M ₀ , N ₀ ) can be obtained by solving the following linear optimization problem:

After obtaining (M ₀ , N ₀ ), it can be based on

Calculate the first optimized value

At the same time, it can also

Substituting the formula f(x), the calculation is obtained

The second iterative parameter (M ₁ , N ₁ ) in the solution of the invention is obtained according to the following formula:

By continuously calculating with reference to the above method, the set of iterative parameters required for the inventive scheme can be determined. For example, when the above learning process finds that k=5, better results are obtained, and after the end of the learning, the following iterative parameter sets {M ₀ , M ₁ , M ₂ , M ₃ , M ₄ , M ₅ } are obtained. And {N ₀ , N ₁ , N ₂ , N ₃ , N ₄ , N ₅ }. The solution of the present invention does not specifically limit the value of k and the value of the iterative parameter, and may be determined by practical applications.

In the second method, an iterative parameter set can be obtained by combining machine learning and parameter verification.

Specifically, the learning may be performed according to the process shown in the first method, and the learned iterative parameter set is referred to as a test iteration parameter set. Then using test samples, such as a series of point pairs {a ⁱ , b ⁱ } of a test camera, using the inventive scheme, starting from the initial value of the camera parameters, using the above test iterative parameter set for camera parameter optimization, and k times after the iteration, obtaining points {a ^{^i,} b ^i} on the image projected spherical surface equidistant ^{^{{a i ', b i'}} }, and thus obtain error parameters of the test camera. For example, the error parameter can be an average pixel error and a highest pixel error.

Generally, if the error parameter matches the preset threshold, it is considered that the test iteration parameter set obtained in the learning phase is available, and is determined as the iterative parameter set in the solution of the present invention; if the error parameter does not match the preset threshold, the test phase is considered to be obtained. The iterative parameter set is not available, corresponding to this, the initial value of the camera parameter in the learning phase can be adjusted.

The machine learning is performed based on the adjusted initial value of the camera parameter until the error parameter obtained in the parameter verification phase matches the preset threshold. As an example, adjust the camera parameter initial value during the learning phase.

It is possible to add a random disturbance based on the initial value.

In order to better verify the beneficial effects of the inventive solution, the experimental comparison results shown in the table below are also provided. It should be noted that the test environment involved in the verification is Intel i5 processor (3.3 GHz), Win7 Pro system.

	平均像素误差Average pixel error	最高像素误差Highest pixel error	迭代次数Number of iterations	运算时间(s)Operation time (s)
本发明方案Solution of the invention	1.851.85	5.205.20	55	33
现有技术current technology	3.253.25	8.758.75	2020	1616

Table 1 9 million pixel (2048 × 1536 × 3) three-head panoramic camera calibration results

In addition to the above-mentioned experimental comparison results, the solution of the present invention is also superior to the prior art in terms of the effect of the picture stitching. For details, refer to the effect display diagram of the picture stitching implemented according to the solution of the present invention shown in FIG. The effect display diagram of the picture stitching realized by the prior art scheme, especially the circled area in the figure.

Corresponding to the method described above, the embodiment of the present invention further provides a camera geometric calibration processing device. Referring to FIG. 4, the device may include:

a cosine distance calculation unit 201, configured to determine a camera parameter initial value, and obtain a first cosine distance of the pair of points calculated based on the initial value, where the pair of points is two of the two adjacent pictures corresponding to the same position of the space coordinate The coordinates of the pixels;

The optimization value adjustment unit 202 is configured to determine whether the first cosine distance matches the preset value, and if not, acquire the first iteration parameter, and adjust the initial value by using the first iteration parameter to obtain the first Optimization value

The cosine distance calculation unit 201 is further configured to obtain a second cosine distance of the pair of points calculated based on the first optimization value;

The calibration value determining unit 203 is configured to determine whether the second cosine distance matches the preset value, and if yes, determine the first optimization value as a camera parameter calibration value.

Optionally, the device further includes:

Indicates the camera parameter calibration value of the jth camera,

Indicates the initial value of the camera parameter of the jth camera,

Optionally, the optimization value adjustment unit is specifically configured to be used according to

Calculate the first optimized value

It should be noted that each embodiment in the specification is described in a progressive manner, and each embodiment focuses on differences from other embodiments, and the same similar parts between the embodiments are referred to each other. can. For the device type embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiment.

Finally, it is also to be understood that the term "comprises", "comprising" or any other variants thereof is intended to encompass a non-exclusive inclusion, such that a process, method, article, or device comprising a plurality of elements includes Those elements, but also other elements not explicitly listed, or elements that are inherent to such a process, method, item or equipment. An element that is defined by the phrase "comprising a ..." does not exclude the presence of additional equivalent elements in the process, method, item, or device that comprises the element.

The foregoing provides a detailed description of the solution provided by the present invention. The principles and embodiments of the present invention are described herein by using specific examples. The description of the above embodiments is only for helping to understand the method and the core idea of the present invention; The present invention is not limited by the scope of the present invention, and the details of the present invention are not limited by the scope of the present invention.

Claims

A camera geometric calibration processing method, characterized in that the method comprises:

Determining a camera parameter initial value, and obtaining a first cosine distance of the pair of points calculated based on the initial value, the point pair being coordinates of two pixel points corresponding to the same position of the spatial coordinate in two adjacent pictures;

Determining whether the first cosine distance is consistent with a preset value, if not, acquiring a first iteration parameter, and adjusting the initial value by using the first iteration parameter to obtain a first optimization value;

Obtaining a second cosine distance of the pair of points calculated based on the first optimized value;

Determining whether the second cosine distance matches the preset value, and if so, determining the first optimized value as a camera parameter calibration value.
The method according to claim 1, wherein if the second cosine distance does not match the preset value, the method further comprises:

Obtaining a second iteration parameter, and adjusting the first optimization value by using the second iteration parameter to obtain a second optimization value;

Obtaining a third cosine distance of the pair of points calculated based on the second optimized value;

Determining whether the third cosine distance matches the preset value, and if so, determining the second optimized value as a camera parameter calibration value.
The method of claim 1 wherein said determining camera parameter initial values comprises:

Determining an a priori value of the camera parameter as the initial value; or

A random disturbance is added based on the a priori value of the camera to obtain the initial value.
The method according to any one of claims 1 to 3, characterized in that the manner of obtaining the iterative parameters is:

Machine learning is performed on the preset samples to obtain the iterative parameters.
The method according to claim 4, wherein the machine learning is performed on the preset sample, and the first iteration parameter is obtained by:

Where (M 0 , N 0 ) represents the first iteration parameter and C j represents the camera identity number.
Indicates the camera parameter calibration value of the jth camera,
Indicates the initial value of the camera parameter of the jth camera,
The first cosine distance of the pair of points calculated based on (M 0 , N 0 ) is expressed.
The method according to claim 5, wherein said adjusting said initial value by said first iteration parameter to obtain a first optimized value
The way is:
A camera geometric calibration processing device, characterized in that the device comprises:

a cosine distance calculation unit, configured to determine an initial value of the camera parameter, and obtain a first cosine distance of the pair of points calculated based on the initial value, where the pair of points is two of the two adjacent pictures corresponding to the same position of the space coordinate The coordinates of the pixel;

An optimization value adjustment unit, configured to determine whether the first cosine distance is consistent with a preset value, and if not, acquiring a first iteration parameter, and adjusting the initial value by using the first iteration parameter to obtain a first optimization value;

The cosine distance calculation unit is further configured to obtain a second cosine distance of the pair of points calculated based on the first optimization value;

The calibration value determining unit is configured to determine whether the second cosine distance matches the preset value, and if yes, determine the first optimization value as a camera parameter calibration value.
The device according to claim 7, wherein the device further comprises:

The optimization value adjustment unit is further configured to: when the second cosine distance does not match the preset value, acquire a second iteration parameter, and adjust the first optimization value by using the second iteration parameter to obtain Second optimized value;

The cosine distance calculation unit is further configured to obtain a third cosine distance of the pair of points calculated based on the second optimized value;

The calibration value determining unit is further configured to determine whether the third cosine distance matches the preset value, and if yes, determine the second optimization value as a camera parameter calibration value.
The device according to claim 7, wherein the device further comprises:

An iterative parameter obtaining unit, configured to perform machine learning on the preset sample, to obtain the first iteration parameter:

Where (M 0 , N 0 ) represents the first iteration parameter and C j represents the camera identity number.
Indicates the camera parameter calibration value of the jth camera,
Indicates the initial value of the camera parameter of the jth camera,
Represents the first cosine distance of the pair of points calculated based on (M 0 , N 0 ).
The device of claim 9 wherein:

The optimization value adjustment unit is specifically configured to be used according to
Calculate the first optimized value