Computer vision-based method and test device for identifying vibration modal of large flexible structure

Abstract

The application provides a large flexible structure vibration modal identification method and test device based on computer vision, and belongs to the technical field of simulation testing. The method steps are as follows: camera modeling and calibration; image acquisition; image preprocessing; cooperative target profile extraction; cooperative target centroid calculation; cooperative target star coordinate solution; cooperative target star coordinate conversion; Kalman filtering processing; and spectrum analysis. Compared with the traditional vibration measurement method and device, the application has the characteristics of non-contact, high precision, low identification error, easy deployment and the like. Not only can the application perform multi-point measurement in a global range, but also can realize high-precision measurement of flexible accessory vibration displacement on the premise of reducing the complexity of on-orbit equipment.

Description

TECHNICAL FIELD

[0001] The application relates to a large flexible structure vibration modal identification method and test device based on computer vision and belongs to the technical field of simulation testing. BACKGROUND

[0002] With the development of space technology, the structure of space satellites is becoming more and more complex, and the structure gradually presents the characteristics of carrying light and large-size flexible accessories. The large flexible accessory structure has the characteristics of light weight, low damping and small vibration frequency, and is easily disturbed by torque output components such as flywheels and pushers to generate long-time and low-frequency vibrations, which not only affects the attitude pointing accuracy of the satellite, but also seriously affects the service life of the device. Therefore, how to suppress the flexible vibration of the large flexible accessory is very important, and the parameter measurement and identification method of the flexible mode is the key to the flexible suppression technology.

[0003] For the vibration measurement method of the large flexible accessory, according to the measurement means, it can be divided into contact measurement technology and non-contact measurement technology. The contact measurement technology uses the sensor installed on the accessory to obtain the vibration displacement information, but it will affect the mass distribution of the light flexible accessory and change its dynamic characteristics, so the measurement error is large, and it is not suitable for the vibration measurement of the light large flexible accessory. The non-contact measurement technology uses optical technology to indirectly measure the vibration displacement information of the accessory, which can not only measure multiple points in the global range, but also realize high-precision measurement of the vibration displacement of the flexible accessory under the premise of reducing the complexity of the on-orbit equipment. Many researches have been carried out in the field of non-contact measurement at home and abroad. Avitable et al. use the method of machine vision to measure the vibration of the flexible structure, and compare the measurement results with the accelerometer and laser vibration measuring instrument, which proves the feasibility of the visual measurement method. Qiu et al. use a visual sensor to measure the bending and torsional modal parameters of a flexible piezoelectric cantilever plate in real time, and provide a reliable feedback measurement method for the stable control of the cantilever plate. NASA uses camera measurement method to measure the vibration deformation information of the Hubble telescope solar sail under complex environment.

[0004] Due to the limitations of the on-orbit spacecraft space environment and the particularity of the flexible accessory, the non-contact measurement method based on a monocular camera is usually used to obtain the vibration parameters of the flexible accessory, therefore, according to the true on-orbit environment of the spacecraft, it is of great significance to research and build a ground demonstration simulation system. SUMMARY

[0005] The purpose of the present application is to solve the problems existing in the prior art, and to provide a large flexible structure vibration modal identification method and test device based on computer vision.

[0006] The purpose of the present application is realized by the following technical scheme:

[0007] A large flexible structure vibration modal identification method based on computer vision, comprising the following steps:

[0008] Step one: camera modeling and calibration

[0009] Analyze the imaging process of the camera, obtain the camera internal and external parameter matrix through camera calibration, and establish a pinhole imaging model of the monocular camera without considering the camera lens distortion;

[0010] Step two: image acquisition

[0011] Real-time high-speed acquisition of images containing cooperative targets on the flexible truss is realized through an industrial camera, and the images are transmitted to a measurement PC through an image acquisition card for subsequent image processing and coordinate calculation using the measurement PC;

[0012] Step three: image preprocessing

[0013] Image preprocessing operations include setting target areas, image filtering, color recognition binarization and morphological processing. Through the image preprocessing process, most of the interference noise information in the image can be filtered out, and the foreground image information containing the feature targets can be extracted from the image;

[0014] Step four: cooperative target contour extraction

[0015] The above-mentioned binarized image after image preprocessing only contains the foreground information of the measurement cooperative target, and the contour information of each cooperative target needs to be extracted and separated in order to calculate the centroid position of the target later;

[0016] Step five: cooperative target centroid calculation

[0017] For the above-mentioned cooperative target contour point information, a sub-pixel positioning method based on edge curve fitting is used to position the centroid of the feature target;

[0018] Step six: cooperative target star coordinate calculation

[0019] Through the above-mentioned image preprocessing algorithm and cooperative target centroid extraction algorithm, the center position information of multiple cooperative targets in the image is obtained, but since the camera in the system will rotate with the single-axis air floating platform, its imaging coordinate system is not fixed, and the modal vibration identification cannot be directly performed from the target position information in the image. Therefore, the image coordinates of the centroid need to be converted and unified to the static coordinate system, and then the vibration displacement curve is analyzed to identify the vibration modal parameters;

[0020] Step seven: cooperative target star coordinate conversion

[0021] Considering the case of rotation of the single-axis air-floating platform, a static coordinate system and a star coordinate system are established, and the static coordinate system is defined as the star coordinate system when the single-axis air-floating platform is static and the flexible truss is not stressed. When the single-axis air-floating platform rotates by an angle θ, the star coordinate system and the static coordinate system satisfy the rotation transformation relationship, and the star coordinate system at each time needs to be unified to the static coordinate system for identification of the modal vibration frequency.

[0022] Step eight: Kalman filtering processing

[0023] The vibration displacement curve in the static coordinate system contains interference noise signals, and the interference error mainly comes from the sensor noise of the camera and the inertial unit element, therefore, the Kalman filtering is adopted to eliminate the sensor interference noise, and the optimal estimation curve of the measurement value is obtained. The Kalman filtering is a linear filtering and prediction method, which utilizes the linear system state equation to perform optimal estimation on the system state through the system input and output observation data.

[0024] Step nine: spectrum analysis

[0025] The discrete Fourier transform (DFT) is adopted to perform spectrum analysis on the above displacement curve. The discrete Fourier transform is an effective method for frequency spectrum analysis of a finite-length discrete time domain sequence, which obtains the discrete sequence information in the frequency domain through transformation of the discrete points of finite sampling, obtains the frequency domain composition of the sampling signal, and the largest several principal component frequencies are the frequencies of each order of the vibration mode, so that the identification of the vibration modal parameters of the large flexible satellite is realized.

[0026] The test device of the large flexible structure vibration modal identification method based on computer vision comprises a measurement industrial computer, an industrial camera, a flexible accessory, a measurement cooperative target, a single-axis air-floating platform and an air-floating base;

[0027] The measurement industrial computer and the industrial camera are arranged on the single-axis air-floating platform, the measurement industrial computer is connected with the industrial camera, the measurement cooperative targets are connected through the flexible accessory, one side of the measurement cooperative targets is arranged on the single-axis air-floating platform, the other side of the measurement cooperative targets is arranged on the air-floating base, and the air-floating base supports the flexible accessory and the measurement cooperative targets.

[0028] The method has the advantages that:

[0029] This invention provides a computer vision-based method and experimental device for vibration mode identification of large flexible structures. Compared with traditional vibration measurement methods and devices, this invention features non-contact operation, high precision, low identification error, and easy deployment. Vibration measurement methods for large flexible attachments can be categorized into contact and non-contact measurement techniques. Contact measurement techniques use sensors mounted on the attachment to obtain its vibration displacement information, but this affects the mass distribution of lightweight flexible attachments and alters their dynamic characteristics, resulting in larger measurement errors and making it unsuitable for vibration measurement of lightweight large flexible attachments. Non-contact measurement techniques use optical technology to indirectly measure the vibration displacement information of the attachment. This not only enables multi-point measurement on a global scale but also achieves high-precision measurement of the vibration displacement of flexible attachments while reducing the complexity of on-orbit equipment. Attached Figure Description

[0030] Figure 1 This is a flowchart of the computer vision-based vibration mode identification method and experimental device for large flexible structures according to the present invention.

[0031] Figure 2 The stellar coordinate system is the basis for the computer vision-based vibration mode identification method and experimental device for large flexible structures in this invention.

[0032] Figure 3 This invention relates the transformation relationship between the computer vision-based vibration mode identification method for large flexible structures and the stellar coordinate system of the experimental device.

[0033] Figure 4 This is a schematic diagram of the experimental device for vibration mode identification of large flexible structures based on computer vision, according to the present invention.

[0034] In the attached diagram, 1 represents the measurement industrial control computer, 2 represents the industrial camera, 3 represents the flexible accessory, 4 represents the measurement cooperative target, 5 represents the single-axis air-bearing platform, and 6 represents the air-bearing base. Detailed Implementation

[0035] The present invention will be further described in detail below with reference to the accompanying drawings: This embodiment is implemented under the premise of the technical solution of the present invention, and detailed implementation methods are given, but the protection scope of the present invention is not limited to the following embodiments.

[0036] like Figure 1 As shown, the experimental device for the computer vision-based vibration mode identification method for large flexible structures involved in this embodiment includes: a measurement industrial control computer 1, an industrial camera 2, a flexible accessory 3, a measurement cooperative target 4, a single-axis air-bearing platform 5, and an air-bearing base 6.

[0037] The single-axis air-bearing platform 5 is equipped with a measuring industrial control computer 1 and an industrial camera 2. The measuring industrial control computer 1 is connected to the industrial camera 2. The measuring cooperative targets 4 are connected by a flexible attachment 3. One side of the measuring cooperative target 4 is set on the single-axis air-bearing platform 5, and the other side of the measuring cooperative target 4 is set on the air-bearing base 6. The air-bearing base 6 supports the flexible attachment 3 and the measuring cooperative target 4.

[0038] Example 1

[0039] Step 1: Camera Modeling and Calibration

[0040] The imaging process of the camera is analyzed, and the intrinsic and extrinsic parameter matrices of the camera are obtained through camera calibration. Without considering camera lens distortion, a pinhole imaging model of a monocular camera is established.

[0041] Based on the similar triangle theory in the pinhole imaging model, the correspondence between the image plane coordinate system and the camera coordinate system can be obtained:

[0042]

[0043] Where (x,y,1) T Let P be the homogeneous coordinates of the spatial object point P in the image plane, (X... c ,Y c Z c ,1) T Let f be the homogeneous coordinates of the spatial point P in the camera coordinate system, and f be the focal length of the camera lens.

[0044] Since the camera coordinate system and the world coordinate system satisfy the rigid body transformation law, the following correspondence exists:

[0045]

[0046] Among them, (X) w ,Y w Z w ,1) T Let P be the homogeneous coordinates of the spatial point P in the world coordinate system, and R and T be the rotation matrix and translation vector from the world coordinate system to the camera coordinate system, respectively.

[0047] Since the pixel coordinate system and the image plane coordinate system satisfy a proportional scaling relationship, and their zero points have a fixed offset, the following correspondence exists:

[0048]

[0049] Where (u,v,1) T Let be the homogeneous coordinates of the spatial object point P in the pixel coordinate system, dx and dy be the scaling factors in the horizontal and vertical directions respectively, and u0 and v0 be the offsets of the zero point in the horizontal and vertical directions respectively.

[0050] According to the above relations, a camera pinhole imaging model can be obtained as shown below:

[0051]

[0052] In the above formula, let:

[0053]

[0054]

[0055] Then:

[0056]

[0057] wherein M1 and M2 are internal parameter matrix and external parameter matrix of the camera respectively, which can be obtained by camera calibration using Zhang's calibration method, and are used to construct the pinhole imaging model of the monocular camera as prior conditions.

[0058] Step two: image acquisition

[0059] An image containing the cooperative targets on the flexible truss is acquired in real time and at high speed by an industrial camera, and the image is transmitted to a measurement PC through an image acquisition card for subsequent image processing and coordinate calculation by the measurement PC.

[0060] Step three: image preprocessing

[0061] The image preprocessing operation includes setting a target area, image filtering, color recognition binarization, and morphological processing. Through the image preprocessing process, most of the interference noise information in the image can be filtered out, and the foreground image information containing the characteristic targets can be extracted from the image.

[0062] Setting a target area: Since the flexible accessory and the cooperative targets only occupy a small part of the image area, most of the image is redundant environmental information, so it is unnecessary to process all the pixel points of the entire image one by one. Therefore, a suitable fixed target area is first set for the collected image, so that all the cooperative targets on the flexible accessory are stably located in this area in the image as the flexible accessory vibrates.

[0063] Image filtering: The purpose of image filtering is to filter out noise information in the image, reduce environmental light source interference in the image, and make the pixel change more smooth. In this method, Gaussian filtering is selected, the principle of which is to use a Gaussian template to convolve each pixel point, that is, to take the pixel values of the surrounding n fields and add the Gaussian weight, and the result is the new pixel value of the pixel point. The mathematical formula of Gaussian filtering is shown as follows:

[0064]

[0065]

[0066] Color recognition binarization: the purpose of color recognition binarization is to extract the foreground image information containing the feature target, and distinguish the foreground information from the background information in the image. The principle is that, first, the HSV space value range of the color of the feature target is obtained according to the HSV color space, that is, the HSV color space corresponding to the red color, and the range set is defined

[0067]

[0068] Where h, s, v respectively correspond to the hue, saturation and brightness of the pixel point, and H min and H max respectively represent the lower limit and upper limit of the hue value, S min and S max respectively represent the lower limit and upper limit of the saturation value, V min and V max respectively represent the lower limit and upper limit of the brightness value.

[0069] Compare the HSV three-channel value of each pixel in the image with the above range set , and judge whether the point pixel is a foreground point or a background point according to the comparison result, and the specific formula is as follows:

[0070]

[0071] Wherein, is the above defined range set, I Binary is the gray value of the binarized image, which can be 0 or 1, 0 is white and 1 is black, I hsv is the hsv three-channel value of the current pixel point.

[0072] Morphological processing: morphological processing mainly includes erosion and dilation operation. Through morphological processing, the isolated noise points in the above obtained binarized image can be eliminated, and the extraction of features is more obvious, which is convenient for the extraction of mark points. The principle of morphological processing is to use a specific kernel to do convolution operation with the image to get the new value of each pixel point.

[0073] The role of morphological erosion is to expand the black part of the image, and the specific calculation formula is as follows:

[0074]

[0075] The role of morphological dilation is to expand the white part of the image, and the specific calculation formula is as follows:

[0076]

[0077] Where (x, y) is a pixel point on the image, (x', y') is a pixel point on the kernel, dst(x, y) is the calculated value at pixel point (x, y), and src(x, y) is the original value at pixel point (x, y).

[0078] Step four: cooperative target profile extraction

[0079] The binary image obtained by the above image preprocessing only contains the foreground information of the measured cooperative target, and the profile information of each cooperative target needs to be extracted and separated in order to calculate the centroid position of the target subsequently.

[0080] The Canny edge detection algorithm is used to extract the edge information of each cooperative target, which mainly consists of the following parts: Gaussian filtering, gradient calculation, non-maximum suppression and edge connection.

[0081] The image is filtered using a two-dimensional Gaussian kernel as follows to remove noise interference in the image.

[0082]

[0083] Secondly, the amplitude and direction of the image gray value gradient are calculated, and the convolution operator used in the canny algorithm is shown in the following formula:

[0084]

[0085] The operator is used to convolve the image to obtain the gradient information of each pixel point. The point with a larger amplitude has a more intense gray change and is more likely to be an edge point in the image.

[0086] For all possible edge points in the image, the local edge points are obtained by non-maximum suppression. The main process is as follows: find the maximum value of each point in the gradient direction, which is the local extreme value, and mark the non-local extreme value points as 0. The image is cycled through each pixel point according to this rule.

[0087] Finally, the detected edge points are connected to form a closed contour. First, a high threshold is used to obtain an edge image, but the edge points may not be closed at this time, so a low threshold is further used to find the end points of the edge image until the entire edge contour is closed, and the closed contour information in the image is obtained.

[0088] Step five: cooperative target centroid calculation

[0089] For the above extracted cooperative target profile point information, a sub-pixel positioning method based on edge curve fitting is used to position the feature target centroid. Since the feature target is circular, the preset fitting curve L and the deviation function Q are as follows:

[0090] L: x 2 +y2 + ax + by + c = 0

[0091]

[0092] wherein a, b, c are undetermined parameters of the quadratic fitting curve L, and the fitting solution is a set of spatial vectors when the deviation function Q takes a minimum value;

[0093] According to the principle of least squares, the parameters of the fitting curve are optimized and solved, and Q(a, b, c) is taken with respect to a, b, and c, and is equal to 0, to form the following equation group:

[0094]

[0095] The minimum point is obtained by solving the above equation group, and the optimal estimation parameters of the fitting curve are finally obtained, that is, the sub-pixel coordinates (x0, y0) of the characteristic target centroid.

[0096]

[0097] Step six: cooperative target star coordinate solution

[0098] Through the above image preprocessing algorithm and cooperative target centroid extraction algorithm, the center position information of multiple cooperative targets in the image is obtained, but since the camera in the system can rotate with the single-axis air floating platform, the imaging coordinate system is not fixed, and the modal vibration cannot be directly identified from the target position information in the image. Therefore, the image coordinates of the centroid need to be converted and unified to the static coordinate system in advance, and then the frequency spectrum of the vibration displacement curve is analyzed to identify the vibration modal parameters.

[0099] The rigid connection end of the single-axis air floating platform and the flexible truss is taken as the coordinate origin, and the tangent and normal of the connection end are taken as the x-axis and y-axis to establish the star coordinate system, as shown in the definition diagram Figure 2 The star coordinate system rotates with the single-axis air floating platform, and because the industrial camera is fixed to the single-axis air floating platform through a support, the star coordinate system and the camera satisfy a rigid transformation relationship, so it is taken as the world coordinate system in camera calibration. Then, according to the monocular camera imaging model, the star coordinates of the cooperative target centroid can be calculated based on the extracted cooperative target centroid, as follows:

[0100] According to the pinhole imaging model of the monocular camera established above, the following formula is established:

[0101] P p = M1P n

[0102] wherein P p , P nrespectively, M1 is the camera intrinsic parameter matrix obtained by camera calibration, which is a reversible matrix, and its inverse matrix is

[0103]

[0104] wherein a x , a y is the scaling factor between the image coordinate system and the camera coordinate system, u0 and v0 are the offset vectors of the coordinate origin between the image coordinate system and the camera coordinate system, respectively;

[0105] The normalized coordinates of the feature target mass in the camera coordinate system P n can be obtained by back calculation:

[0106]

[0107] According to the definition of the normalized coordinates, the following formula is established:

[0108]

[0109] wherein P c is the camera coordinate system coordinate of the feature target mass, Z c is the depth estimation information, and the unknown quantity Z c needs to be solved in the above formula to calculate the coordinates in the camera coordinate system, and the specific method is as follows:

[0110] According to the pinhole imaging model of the monocular camera, the following formula is established:

[0111] P c =M2P w

[0112] wherein M2 is the camera extrinsic parameter matrix obtained by camera calibration, which is a reversible matrix, and its inverse matrix is P w is the coordinate of the cooperative target in the star coordinate system:

[0113]

[0114] Then, the left and right sides are multiplied by to obtain:

[0115]

[0116] Taking the third row, the following formula is obtained after expansion:

[0117] Z w =(r 31 X n +r 32 Yn +r 33 )Z c -(r 31 t1+r 32 t2+r 33 t3)

[0118] Since the artificial star coordinate system takes Z w = 0 as the truss motion plane, after substitution, Z c :

[0119]

[0120] After obtaining the unknown Z c , it can be brought into the formula to obtain the cooperative target mass center coordinates P c in the camera coordinate system, and then the camera coordinate is brought into the formula to obtain the feature target mass center coordinates P w in the world coordinate system, so that the two-dimensional image coordinates of the feature target are converted to obtain the star coordinates of each cooperative target in the star coordinate system.

[0121] Step seven: cooperative target star coordinate conversion

[0122] Considering the case of single-axis air floating platform rotation, the relationship between the static coordinate system O-XYZ and the star coordinate system O-X'Y'Z' is shown in Figure 3 , and the static coordinate system is defined as the star coordinate system when the single-axis air floating platform is static and the flexible truss is not under force. When the single-axis air floating platform rotates by an angle θ, the star coordinate system and the static coordinate system satisfy the rotation transformation relationship, and for modal vibration frequency identification, the star coordinates at each time need to be unified to the static coordinate system.

[0123] Since the single-axis air floating platform only makes one-dimensional rotational motion, according to the rigid body rotation theorem, the coordinates P' in the O-X'Y'Z' system and the coordinates P in the O-XYZ system have the following relationship:

[0124] P' = R Z P

[0125]

[0126] where R z is the rotation matrix between the coordinate systems:

[0127]

[0128] Therefore, for each cooperative target star coordinate P' calculated in each frame of image, the rotation angle information θ of the single-axis air floating platform at the current time can be obtained through the inertial element on the single-axis air floating platform, and the coordinates can be converted and unified to the static coordinate system through calculation, so as to facilitate subsequent modal vibration frequency identification work.

[0129] Step 8: Kalman Filtering

[0130] The vibration displacement curves in the static coordinate system described above contain interference noise signals. This interference error mainly originates from sensor noise in the camera and inertial navigation system components. Therefore, this method employs Kalman filtering to eliminate sensor interference noise and obtain the optimal estimation curve for the measured values. Kalman filtering is a linear filtering and prediction method that utilizes the state equations of a linear system to perform an optimal estimation of the system state using system input and output observation data.

[0131] The system's state equations and measurement equations are as follows:

[0132] x k =Fx k-1 +Bu k-1 +w k-1

[0133] z k =Hx k +v k

[0134] Where x k It is the state vector at time k; z k It is the measurement vector at time k; F is the one-step state transition matrix; u k B is the system input; B is the matrix that transforms the input vector into a state vector; H is the measurement model matrix; w k ~N(0,Q) k ) and v k ~N(0,R k These are uncorrelated process noise and measurement noise that follow a Gaussian distribution.

[0135] The recursive equations for the Kalman filter are given below:

[0136]

[0137] in, It is state x k-1 Filtered estimation, It is by The calculated pair of x k The next step of prediction. In this system, based on the state variable x from the previous time step... k-1 And real-time vibration displacement information z obtained from camera measurements k The target's position and velocity can be predicted in one step by calculating using the established system model.

[0138] Among them, K k P represents the filter gain matrix. k|k-1is the covariance of the last optimal estimate and the current prediction, P k and P k-1 represent the state vectors at k and k-1, respectively, H k represents the measurement model matrix, R k represents the noise variance, Q k represents the noise matrix;

[0139] Before the algorithm runs, appropriate filtering parameters need to be given, including initial position and velocity, noise characteristics of the environment, etc. Then the target acceleration information collected by the inertial unit on the single-axis air floating platform and obtained after data processing is taken as the input u k The vibration displacement measured from the camera is taken as the input z k Substitute the above information into the iteration equation, and the optimal estimation result of the vibration displacement at each step can be obtained.

[0140] Step nine: spectrum analysis

[0141] Discrete Fourier transform (DFT) is used to analyze the spectrum of the above displacement curve. Discrete Fourier transform is an effective method for frequency spectrum analysis of finite-length discrete time-domain sequences. It obtains discrete sequence information in the frequency domain by transforming discrete points of finite sampling, and obtains the frequency domain composition of the sampling signal, wherein the largest several principal component frequencies are the frequencies of each order of the vibration mode, so as to realize the identification of the vibration mode parameters of the large flexible satellite.

[0142] First, the vibration displacement discrete information obtained above is taken as a fixed sampling number N to form a discrete finite time sequence x(t), and then the sequence is converted into a discrete infinite sequence x(n), i.e. the sequence number n is used to replace the original time variable; next, the discrete infinite sequence is truncated to only take a part to form a discrete sequence:

[0143] x(n) n = 0, 1, 2, …, N-1.

[0144] Discrete Fourier transform is performed on the above discrete sequence:

[0145]

[0146] Wherein k is a multiple of a frequency interval, and finally X(k) is mapped to X(f):

[0147]

[0148] The frequency distribution of the discrete signal in the frequency domain f is obtained, and the identification of the vibration mode parameters of the flexible accessory of the large flexible satellite is realized.

[0149] The above merely describes preferred specific embodiments of the present application, which are based on different implementations of the overall concept of the present application, and the protection scope of the present application is not limited thereto. Any changes or replacements that are easily conceived by those skilled in the art within the technical scope disclosed by the present application shall be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A computer vision based method for vibration modal identification of large flexible structures, characterized in that, Comprising the following steps: Step one: camera modeling and calibration Analyzing the imaging process of the camera, obtaining the camera internal and external parameter matrices through camera calibration, and establishing a pinhole imaging model of the monocular camera without considering the camera lens distortion; According to the similar triangle theory in the pinhole imaging model, the corresponding relationship between the image plane coordinate system and the camera coordinate system can be obtained: where (x, y, 1) T are the homogeneous coordinates of the spatial object point P in the image plane, (X c , Y c , Z c , 1) T are the homogeneous coordinates of the spatial object point P in the camera coordinate system, and f is the focal length of the camera lens. Since the camera coordinate system and the world coordinate system satisfy the rigid body transformation law, the following correspondence relationship is obtained: wherein (X w ,Y w ,Z w ,1) T are the homogeneous coordinates of the spatial object point P in the world coordinate system, and R and T are the rotation matrix and the translation vector from the world coordinate system to the camera coordinate system, respectively. Since the pixel coordinate system and the image plane coordinate system satisfy the equal scaling relationship, and the zero point has a fixed offset, the following correspondence relationship is obtained: where (u, v, 1) T is the homogeneous coordinate of the spatial object point P in the pixel coordinate system, dx and dy are the scaling factors in the horizontal and vertical directions, respectively, and u0 and v0 are the offsets of the zero point in the horizontal and vertical directions, respectively. Combining the above relationships, the camera pinhole imaging model is as follows: In the above formula, let: Then: Wherein, M1 and M2 are the internal and external parameter matrices of the camera, respectively, which can be obtained by camera calibration using Zhang's calibration method, and are used as prior conditions to construct the pinhole imaging model of the monocular camera; Step two: image acquisition Real-time high-speed acquisition of images containing cooperative targets on the flexible truss by an industrial camera, and transmission of the images to a measurement PC through an image acquisition card for subsequent image processing and coordinate calculation by the measurement PC; Step three: image preprocessing Image preprocessing operations include setting a target area, image filtering, color recognition binarization, and morphological processing. Through the image preprocessing process, most of the interference noise information in the image can be filtered out, and the foreground image information containing the feature targets can be extracted from the image; Setting a target area: Since the flexible accessory and the cooperative target only occupy a small part of the image, most of the image is redundant environmental information, so it is unnecessary to process all the pixel points of the entire image. Therefore, a suitable fixed target area is first set for the collected image, so that all the cooperative targets on the flexible accessory are stable within this area in the image as the flexible accessory vibrates; Image filtering: The purpose of image filtering is to filter out noise information in the image, reduce environmental light source interference in the image, and make the pixel change more smooth. This method selects Gaussian filtering, which uses a Gaussian template to convolve each pixel point, i.e., taking the pixel values of the surrounding n fields and adding the Gaussian weight to obtain the new pixel value of the pixel point. The mathematical formula of Gaussian filtering is as follows: Color recognition binarization: the purpose of color recognition binarization is to extract foreground image information containing a feature target, and distinguish foreground information from background information in the image; the principle is that, first, the HSV space value range of the color of the feature target is obtained according to the HSV color space, that is, the HSV color space corresponding to red, and a range set is defined wherein h, s, v respectively correspond to hue, saturation and lightness of the pixel point, and H min and H max respectively represent the lower limit and upper limit of the hue value, S min and S max respectively represent the lower limit and upper limit of the saturation value, V min and V max respectively represent the lower limit and upper limit of the lightness value. The HSV three-channel values of each pixel in the image are compared with the above range set and the pixel is determined to be a foreground point or a background point according to the comparison result, and the specific formula is as follows: wherein, I is a range set defined above, Binary I is a gray value of the binarized image, which can be 0 or 1, 0 is white, and 1 is black, hsv I is an hsv three-channel value at the current pixel point; Morphological processing: Morphological processing mainly includes erosion and dilation operations. Through morphological processing, isolated noise points in the binarized image obtained above can be eliminated, and the extraction of features becomes more prominent, which is convenient for the extraction of marker points. The principle of morphological processing is to use a specific kernel to convolve the image to obtain the new value of each pixel point; The role of morphological erosion is to expand the black part of the image, and its specific calculation formula is as follows: The role of morphological dilation is to expand the white part of the image, and its specific calculation formula is as follows: Where (x, y) is a pixel point on the image, (x', y') is a pixel point on the kernel, dst(x, y) is the calculated value of pixel point (x, y), and src(x, y) is the original value of pixel point (x, y); Step four: cooperative target contour extraction The binarized image after image preprocessing only contains foreground information of the measured cooperative targets, and the contour information of each cooperative target needs to be extracted and separated in order to subsequently calculate the centroid position of the target; The Canny edge detection algorithm is used to extract the edge information of each cooperative target, and the Canny edge detection algorithm mainly consists of the following parts: Gaussian filtering, gradient calculation, non-maximum suppression and edge connection. The image is filtered using a two-dimensional Gaussian kernel as follows to remove noise interference in the image. Secondly, the amplitude and direction of the image gray value gradient are calculated, and the convolution operator used in the canny algorithm is shown in the following formula: The operator is used to convolve the image to obtain the gradient information of each pixel point, and the point with a larger amplitude value has a more intense gray change and is more likely to be an edge point in the image. For all possible edge points in the image, the local edge points are obtained through non-maximum suppression, and the main process is as follows: finding the maximum value of each point in the gradient direction, which is a local extreme value, marking the non-local extreme value points as 0, and repeating the above rule for each pixel point in the image. Finally, the detected edge points are connected into a closed contour. First, a high threshold value is used to obtain an edge image, but the edge points may not be closed at this time, so a low threshold value is further used to find the end points of the edge image until the entire edge contour is closed, and the closed contour information in the image is obtained. Step five: cooperative target centroid calculation For the cooperative target contour point information extracted above, an edge curve fitting based sub-pixel positioning method is used to position the feature target centroid. Since the feature target is circular, the preset fitting curve L and the deviation function Q are as follows: L: x 2 + y 2 + ax + by + c = 0 Where a, b, and c are the undetermined parameters of the quadratic fitting curve L, and the fitting solution is a set of spatial vectors when the deviation function Q takes the minimum value. According to the least squares method, the parameters of the fitting curve are optimized and solved, and Q(a, b, c) is taken with respect to a, b, and c and equal to 0 to form the following equation group: The minimum point is obtained by solving the above equation group, and the optimal estimation parameters of the fitting curve are obtained, i.e. the sub-pixel coordinates (x0, y0) of the feature target centroid. Step six: cooperative target star coordinate calculation Through the above image preprocessing algorithm and cooperative target centroid extraction algorithm, the center position information of multiple cooperative targets in the image is obtained. However, since the camera in the system can rotate with the single-axis air floating platform, the imaging coordinate system is not fixed, and the vibration mode cannot be directly identified from the target position information in the image. Therefore, the image coordinates of the centroid need to be converted and unified to the static coordinate system, and then the vibration displacement curve is analyzed to identify the vibration modal parameters. The rigid connection end of the single-axis air floating platform and the flexible truss is taken as the coordinate origin, and the tangent and normal of the connection end are taken as the x-axis and y-axis to establish a star coordinate system. The star coordinate system rotates with the single-axis air floating platform. Since the industrial camera is fixed to the single-axis air floating platform through a support, the star coordinate system and the camera satisfy a rigid transformation relationship, and thus the star coordinate system is taken as the world coordinate system in camera calibration. According to the monocular camera imaging model, the star coordinates of the cooperation target can be calculated according to the mass center of the cooperation target extracted above, and the specific process is as follows: According to the monocular camera pinhole imaging model, the following formula is established: P p = M1P n where P p , P n are the normalized coordinates of the feature target mass center in the image coordinate system and the camera coordinate system, respectively, M1 is the camera internal parameter matrix obtained by camera calibration, which is a reversible matrix, and its inverse matrix is where a x , a y are scaling factors between the image coordinate system and the camera coordinate system, and u0and v0are offset vectors of the coordinate origin between the image coordinate system and the camera coordinate system, respectively. Then the normalized coordinates of the feature target mass center in the camera coordinate system P can be obtained by back calculation n : According to the definition of the normalized coordinates, the following formula is established: where P c is the camera coordinate of the feature target mass center, Z c is the depth estimation information, and the unknown quantity Z c needs to be solved in the above formula to calculate the coordinates in the camera coordinate system, and the specific method is as follows: According to the monocular camera pinhole imaging model, the following formula is established: P c = M2P w Wherein, M2 is a camera exterior parameter matrix obtained by camera calibration, which is a reversible matrix, and its inverse matrix is P w is the coordinate of the cooperative target in the coordinate system of the star. Then the left and right sides of the same ride It can be obtained: Taking the third row, the following formula is obtained after expansion: Z w = (r 31 X n + r 32 Y n + r 33 )Z c - (r 31 t1+ r 32 t2+ r 33 t3) Since the artificial star coordinate system is defined with Z w = 0 as the truss movement plane, after substitution, Z c : Get the unknown quantity Z c The camera coordinate system can be brought into the formula to obtain the cooperative target centroid coordinates P in the camera coordinate system c The camera coordinate system can be brought into the formula to obtain the cooperative target centroid coordinates P in the camera coordinate system w Thus, the two-dimensional image coordinates of the feature target are converted to obtain the star coordinates, and the coordinates of each cooperative target in the star coordinate system are obtained; Step seven: star coordinate conversion of the cooperation target Considering the case when the single-axis air floating platform rotates, a static coordinate system O-XYZ and a star coordinate system O-X'Y'Z' are established. The static coordinate system is defined as the star coordinate system when the single-axis air floating platform is static and the flexible truss is not subjected to force. When the single-axis air floating platform rotates by an angle θ, the star coordinate system and the static coordinate system satisfy a rotation transformation relationship. In order to identify the modal vibration frequency, the star coordinates at each time are unified to the static coordinate system. Since the single-axis air floating platform only performs one-dimensional rotational motion, according to the rigid body rotation theorem, the coordinates P' in the O-X'Y'Z' system and the coordinates P in the O-XYZ system have the following relationship: P' = R Z P where R z is the rotation matrix between coordinate systems: Then, for the star coordinates P' of each cooperation target calculated in each frame of image, the rotation angle information θ of the single-axis air floating platform at the current time is obtained in real time through the inertial measurement unit on the single-axis air floating platform. The coordinates can be converted to the static coordinate system through calculation, so as to facilitate the subsequent modal vibration frequency identification work. Step eight: Kalman filter processing The vibration displacement curve in the static coordinate system contains interference noise signals, and the interference error mainly comes from the sensor noise of the camera and the inertial measurement unit. Therefore, the Kalman filter is used to eliminate the sensor interference noise to obtain the optimal estimation curve of the measured value. The Kalman filter is a linear filtering and prediction method, which uses the state equation of the linear system to perform optimal estimation on the system state through the input and output observation data of the system. The state equation and the measurement equation of the system are as follows: x k = Fx k-1 + Bu k-1 + w k-1 z k = Hx k + v k where x k is the state vector at time k; z k is the measurement vector at time k; F is a one-step state transition matrix; u k is the system input; B is a matrix that transforms the input vector to the state vector; H is a measurement model matrix; w k ~ N(0, Q k ) and v k ~ N(0, R k ) are mutually uncorrelated process noise and measurement noise that are Gaussian distributed. The recursive equation of the Kalman filter is given below: wherein, is the filtered estimate of the state x k-1 , is a one-step prediction of x computed from k ; in the present system, the one-step prediction of the target position and velocity is computed from the state variable x k-1 at the previous time instant and the real-time vibration displacement information z k from the camera measurement, through the established system model. where K k represents the filter gain matrix, P k|k-1 is the covariance of the last optimal estimate and the current prediction value, P k and P k-1 represent the state vectors at k and k-1, respectively, H k represents the measurement model matrix, R k represents the noise variance, Q k represents the noise matrix; The proper filtering parameters, including initial position and velocity, noise characteristics of the environment, etc. are given before the algorithm runs. Then the target acceleration information, which is collected by the inertial measurement unit on the single-axis air bearing platform and processed, is taken as the input u of each step k The vibration displacement measured by the camera is taken as the input z of each step k The optimal estimation of the vibration displacement of each step is obtained by substituting the above information into the iteration equation. Step nine: spectrum analysis Discrete Fourier transform (DFT) is used to perform spectrum analysis on the above displacement curve. Discrete Fourier transform is an effective method for frequency spectrum analysis of finite length discrete time domain sequences. It obtains discrete sequence information in the frequency domain by transforming the discrete points of finite sampling, and obtains the frequency domain composition of the sampling signal. The largest several principal component frequencies are the frequencies of each order of vibration mode, so as to realize the identification of the vibration modal parameters of the large flexible satellite. First, a fixed number N of vibration displacement discrete information is taken to form a discrete finite time sequence x(t), and then the sequence is converted into a discrete infinite sequence x(n), that is, the sequence number n is used instead of the original time variable. Next, the discrete infinite sequence is truncated to form a discrete sequence by taking only a part of it. x(n) n = 0, 1, 2,..., N-1. Discrete Fourier transform is made to the above discrete sequence: Where k is a multiple of frequency interval, and finally X(k) is mapped to X(f): The frequency distribution of the discrete signal in the frequency domain f is obtained, and the identification of the vibration modal parameters of the flexible attachment of the large flexible satellite is realized.

2. The test apparatus for computer vision based modal identification of large flexible structures according to claim 1, characterized in that, It comprises: A measurement industrial computer (1), an industrial camera (2), a flexible attachment (3), a measurement cooperative target (4), a single-axis air floating platform (5) and an air floating base (6); The single-axis air floating platform (5) is provided with the measurement industrial computer (1) and the industrial camera (2), the measurement industrial computer (1) is connected with the industrial camera (2), the measurement cooperative targets (4) are connected through the flexible attachment (3), one side of the measurement cooperative target (4) is arranged on the single-axis air floating platform (5), the other side of the measurement cooperative target (4) is arranged on the air floating base (6), and the air floating base (6) supports the flexible attachment (3) and the measurement cooperative target (4).

Images