Far-field target pose vision measurement method based on center standardization

Abstract

The invention discloses a far-field target pose vision measurement method based on center standardization, and the method comprises the steps: adjusting the pose of a far-field target to be measured, and extracting a far-field target feature point; the method comprises the steps that a far field target pose initial value is solved, a target function is established by minimizing image feature point back projection errors, initial nonlinear optimization is conducted through a Levenberg-Marquardt algorithm to obtain the pose initial value, and parameter pre-optimization is achieved; and after center standardization processing is carried out on the pose parameter Jacobian matrix, nonlinear optimization is carried out by adopting a Levenberg-Marquardt algorithm again, and an optimal solution of the far-field target pose is obtained. The method has the characteristics of reducing the sensitivity of pose parameters to image noise, reducing external parameter coupling, avoiding falling into local minimum and the like, and has important significance in improving far-field target high-precision pose vision measurement in the fields of aerospace and the like.

Description

Technical Field

[0001] The invention relates to the technical field of optical measurement, and in particular to a far-field target posture visual measurement method based on center standardization. Background Art

[0002] As an important means to detect and perceive targets at long distances, far-field target posture visual measurement has the characteristics of long distance, large space, and high precision. It is widely used in the fields of space-based early warning, earth observation, and ground-based optical detection, and is of great significance to improving space / ground-based far-field target tracking, monitoring and early warning. For example, space-based satellite monitoring and early warning of space vehicles, attitude monitoring of cooperative and non-cooperative spacecraft in space, and real-time tracking and monitoring of the posture of aircraft carriers at sea. However, since the object distance in the far-field target visual measurement system is much larger than the focal length of the camera, the target is far away, the speed is fast, and it is accompanied by image noise and other factors, it is extremely sensitive to image noise when solving the target posture. In addition, when solving the conventional PnP (Perspective-n-Point) posture directly, the external parameters are coupled with each other, resulting in large errors in the solved target posture parameters, which makes it difficult to meet the needs of precise guidance, strikes and other scenarios.

[0003] The PnP-based target pose visual measurement process usually includes two steps: pose initial value acquisition and nonlinear optimization. Among them, the acquisition of pose initial value is the basis of pose measurement and optimization process, and largely determines the convergence speed and convergence quality of the optimization algorithm. Many scholars at home and abroad have conducted in-depth research on the efficient and accurate acquisition of target pose, and proposed a variety of initial pose acquisition and optimization methods. Roberto Opromolla et al. proposed a pose solution architecture combining PnP with the Levenberg-Marquardt algorithm for geostationary orbit space robot capture operation. Henrik Rehbinder et al. improved the accuracy and robustness of pose estimation by integrating computer vision and inertial navigation technology. Sumant Sharma et al. used Monte Carlo simulation to test the performance of various PnP solution methods in initial pose estimation, including PosIt, Coplanar PosIt, EPnP and Newton–Raphson method. Michele Maestrini et al. used the EPnP method to obtain the initial pose in the pose observation of spacecraft and resident space objects, and then used BB (Bounding Box) information to detect outliers, and used RoI (Region of Interest)-based estimation to correct the translation vector to achieve pose optimization. Jin Liu et al. proposed an anchor point prediction (APP) algorithm in remote sensing target pose observation to predict the position coordinates of the four corner points of the target object, and finally calculated the three-dimensional pose through homography matrix decomposition. Nicolas Bourdis et al. proposed a visual servoing method based on visual information feedback to adjust the camera control strategy in real-time monitoring of aerial targets, so that the system can gradually obtain and optimize the precise pose. Dianqi Sun et al. used a time-of-flight (TOF) camera to estimate the relative pose of on-orbit space targets. Ravi Kumar et al. used the unscented Kalman filter (PP-UKF) and the integrated unscented Kalman filter (PP-IUKF) to preprocess the data to reduce the noise variance in the sensor measurement, so as to improve the tracking performance of the remote target in the high noise environment. Kelsey et al. matched the model edge image with the contour projected from the internal wireframe model based on the reweighted least squares iteration method to achieve accurate pose estimation during the autonomous rendezvous and docking of the aircraft. Laurent Jospin et al., in the UAV pose observation, took into account the projected light intensity factor of the target point on the image during image acquisition, as well as the influence of motion blur and non-ideal lens system, and adopted the method of extracting the target's photometric model according to the projection intensity and iteratively minimizing the photometric error to obtain a more accurate optimization effect.Andrew Lee et al. used the iterative closest point (ICP) algorithm to estimate the pose of an aerial refueling system using the 3D point cloud generated by a stereo vision system. Kenneth Alberto Funes Mora et al. added constraints to the experimental observation angles to increase the accuracy of optimization in their study of gaze estimation of far-field three-dimensional objects. Jie Li et al. used the unscented Kalman filter (UKF) combined with the Clohessy-Wiltshire (CW) equation to measure the pose and other motion parameters of an on-orbit spacecraft target. Yinlong Liu et al. used the branch-and-bound (BnB) algorithm to find the global optimal rotation and translation pose parameters.

[0004] When measuring far-field targets, the above pose estimation methods will have serious impacts on their accuracy and robustness due to tiny image errors or sensor system errors, resulting in reduced accuracy of the final measurement results. Secondly, the nonlinear optimization of pose parameters usually requires fast convergence speed, high solution accuracy, and low coupling. However, the existing optimization algorithms are limited by simple models and computing resources in the process of optimizing far-field target pose parameters, making it difficult to achieve fast and high-precision solutions. Summary of the invention

[0005] In order to solve the above technical problems, the main purpose of the present invention is to provide a far-field target pose visual measurement method based on center normalization, focusing on using a monocular telephoto camera to achieve high-precision pose measurement of far-field targets and subsequent efficient optimization. This method has the characteristics of low sensitivity to image noise, weak coupling of external parameters and avoiding falling into local minima, which can effectively improve the accuracy of far-field target pose visual measurement.

[0006] In order to achieve the above object, the present invention adopts the following technical scheme:

[0007] A far-field target posture visual measurement method based on center standardization includes the following steps:

[0008] Step 11, design a far-field target posture visual measurement system based on a monocular telephoto camera, fix the monocular telephoto camera on a dual-axis turntable, use the far-field target as the far-field target to be measured, adjust the posture of the far-field target to be measured, and use power supply, triggering, data acquisition and processing devices to achieve high-precision posture observation of the distant target;

[0009] Step 12: Using a telephoto optical camera to image the far-field target and accurately extract the far-field target feature points;

[0010] Step 13, obtaining the initial values ​​of the pose parameters of the far-field target by a non-iterative RPnP method;

[0011] Step 14: Establish an objective function by minimizing the back-projection error of the target image feature points, perform nonlinear pre-optimization through the Levenberg-Marquardt algorithm, and obtain the initial values ​​of the pose parameters for subsequent optimization;

[0012] Step 15: Perform central normalization on the Jacobian matrix of the pose parameters, and use the Levenberg-Marquardt algorithm for nonlinear optimization again to obtain the optimal solution of the pose parameters;

[0013] Step 16: Adjust the far-field target pose, collect multiple sets of image feature point images in sequence, and finally substitute them into the above steps to finally solve and obtain multiple sets of pose parameters.

[0014] The beneficial effects of the present invention compared with the prior art are:

[0015] The present invention uses a close-range short-focus camera and a long-focus optical camera to synchronously measure the pose of a far-field target, and multiple experiments are conducted to verify the effectiveness and accuracy of the present invention. The present invention has the characteristics of reducing the sensitivity of pose parameters to image noise, reducing external parameter coupling, and avoiding falling into local minima, which is of great significance for improving high-precision pose visual measurement of far-field targets in fields such as aerospace. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of a far-field target posture visual measurement method based on center standardization of the present invention;

[0017] Figure 2 This is a schematic diagram of the far-field target pose visual measurement principle of the monocular camera of the present invention; in the figure, Represents the camera coordinate system; Represents the world coordinate system; represents the far-field object coordinate system; Indicates the far field distance in km; Indicates that the far field distance is much greater than the camera focal length; Represents the characteristic points of far-field objects; Represents the projection points of feature points of far-field objects.

[0018] Figure 3 This is the principle diagram for the verification of the accuracy test of the far-field target posture visual measurement; in the figure, Represents the telephoto camera coordinate system; Represents the short-focus camera coordinate system; represents the far-field target coordinate system; The pose parameters representing the pose transformation of the target; Represents the target's pose parameters relative to the telephoto camera; Represents the pose parameters of the target relative to the short-focus camera.

[0019] Figure 4 The RMS error of the pose parameters is obtained for the simulation experiment.

[0020] Figure 5 is the reprojection RMS error of the far-field target in the simulation experiment. DETAILED DESCRIPTION

[0021] The basic idea of ​​the present invention is: using a telephoto optical camera to image it and accurately extract the feature points of the far-field target; using the RPnP method to solve the initial value of the far-field target posture, to establish the objective function by minimizing the back-projection error of the image feature points, and to obtain the initial value of the posture through the Levenberg-Marquardt algorithm for preliminary nonlinear optimization to achieve parameter pre-optimization; performing center normalization on the Jacobian matrix of the posture parameter, and again using the Levenberg-Marquardt algorithm for nonlinear optimization to obtain the optimal solution for the far-field target posture. Using a close-range short-focus camera and a telephoto optical camera to synchronously measure the posture of the far-field target, multiple experiments are conducted to verify the effectiveness and accuracy of the present invention.

[0022] The present invention is further described in detail below by taking the device and method for rapid and high-precision visual measurement of far-field target posture as an example.

[0023] like Figure 1 As shown, a far-field target posture visual measurement method based on center standardization of the present invention mainly includes the following steps:

[0024] Step 11: In a ground laboratory environment, build a camera far-field target pose visual measurement device based on telephoto optics.

[0025] The monocular telephoto optical camera used in the experiment was fixed, and the far-field target posture to be measured was adjusted at the same time to build a far-field target posture visual measurement device and system suitable for a telephoto optical camera.

[0026] Step 12: Accurately extract image feature points, including modeling and extraction.

[0027] A high-precision image feature point extraction method based on the optimal scale is used to achieve high-precision extraction of feature points under different scale conditions, including:

[0028] set up is the image grayscale distribution function, any pixel of the image The Hessian matrix is:

[0029] (1)

[0030] in, is the pixel coordinate, , , is the image grayscale distribution function The second-order partial derivative after Gaussian convolution is obtained by the image grayscale distribution function It is obtained by convolving with the second-order derivative convolution kernel of the Gaussian kernel. Represents the transpose of the matrix. Determine the optimal value of the second-order derivative convolution kernel of the Gaussian kernel according to the size of the feature point image to ensure the accuracy of feature point image positioning.

[0031] The present invention adopts coordinate normalization to obtain the normalization operator :

[0032] (2)

[0033] in, Normalization operator The value of the intermediate parameter , is the spot radius, in pixels. According to the spatial distribution of the normalized operator, it can be concluded that the spot center condition is the normalized operator is greater than 0 and is a local maximum. The actual image is affected by noise, and the precise radius of the light spot cannot be accurately obtained. Therefore, in order to improve the accuracy of the center positioning of the image feature point, different The corresponding maximum value is obtained to determine the optimal radius of the light spot .

[0034] Step 13: Perform nonlinear pre-optimization of pose parameters with the goal of minimizing image positioning deviation. This includes:

[0035] After the high-precision extraction of image feature points in step 12, constraints are established based on the mapping relationship between the far-field target feature points and the corresponding image, where the target feature point image is back-projected to the target plane, and the corresponding spatial target point distance is minimized; after the initial value of the far-field target pose parameters is obtained by the RPnP method, target points to establish the objective function by minimizing the back-projection error of the image feature points. To optimize the pose parameters, namely:

[0036] (3)

[0037] in, It is the observation The true projection coordinates of the characteristic target points, is its optimized coordinate. In the objective function middle Represents the pose parameters in the objective function:

[0038] (4)

[0039] in, translation vector, To surround The angles of rotation of the three coordinate axes. In three-dimensional space, the rotation matrix can be divided into three angles around The three basic rotations are defined by the product of the three coordinate axis rotations. , and To represent, thus defining the rotation matrix .

[0040] The present invention uses the Levenberg-Marquardt algorithm to solve the above nonlinear optimization problem. The Levenberg-Marquardt algorithm is a typical algorithm for solving nonlinear optimization problems, which combines the advantages of the steepest descent method and the Gauss-Newton method. In the nonlinear optimization model using the Levenberg-Marquardt algorithm, the objective function to be minimized is After the first-order Taylor expansion, it can be expressed as:

[0041] (5)

[0042] in, represents the posture parameter, which is , used to indicate surrounding Three angles of rotation of three coordinate axes with translation vector , Indicates the initial value of the parameter. To optimize the number of iterations, is the parameter variation, The Jacobian matrix of the objective function for the pose parameters is:

[0043] (6)

[0044] in, Represents the pose parameters No. elements, if is the current optimization target parameter. The parameter change using the Gauss-Newton algorithm for nonlinear optimization is:

[0045] (7)

[0046] The Gauss-Newton algorithm requires the Jacobian matrix The Levenberg-Marquardt algorithm used in the present invention adds a positive definite matrix to the Gauss-Newton algorithm. arrive To improve Becomes a positive definite matrix. So we can deduce that the parameter change of the Levenberg-Marquardt algorithm is:

[0047] (8)

[0048] in, is the identity matrix. is called the Levenberg-Marquardt parameter. When , the Levenberg-Marquardt algorithm tends to the Gauss-Newton algorithm; when , the Levenberg-Marquardt algorithm approaches the method of steepest descent.

[0049] After obtaining the initial values ​​of the parameters to be optimized, nonlinear optimization can be performed according to the following procedure:

[0050] 1) Enter the initial value. Initial parameters Set to 0.01;

[0051] 2) Calculate the objective function and the Jacobian matrix of the target optimization parameters ;

[0052] 3) Calculate the change in the Levenberg-Marquardt parameter , then update .

[0053] 4) If , then stop the iteration and output the result; otherwise, Set to , then return to procedure 2 to continue iterating;

[0054] 5) If , then Set to , recalculate .

[0055] Step 14: Center normalization of pose parameters, that is, center normalization of the Jacobian matrix of the optimized pose parameters.

[0056] The centralization method is used to reduce the disparity of Jacobian matrices, reduce the correlation between parameters, and make the iterative optimization results more accurate. At the same time, the properties and measurements of the parameters in the Jacobian matrix are different. After centralization, standardization is still required to ensure the consistency of the scale between parameters and improve the accuracy of iterative optimization.

[0057] Step 14.1 Centralization includes:

[0058] In the nonlinear optimization model used in the present invention, formula (3) is the objective function to be minimized After the first-order Taylor expansion, equation (6) represents the optimization step length of Jacobian matrix of coefficients , each element of the Jacobian matrix usually represents the partial derivative of the model output with respect to each input parameter. The correlation between these elements may lead to coupling effects in the parameter optimization process, thus affecting the accuracy of the optimization results. This correlation can be effectively reduced by centralizing the Jacobian matrix, that is, subtracting the mean of each element.

[0059] The present invention first centers each column. Let the Jacobian matrix No. Listed as , the centered column vector is :

[0060] (9)

[0061] in, It is The mean vector of the columns is defined as:

[0062] (10)

[0063] in, The Jacobian matrix OK Column elements. Therefore, the centered Jacobian matrix It can be expressed as:

[0064] (11)

[0065] in, is a length of A column vector of all ones, is a column containing the mean of each column. The centralization process helps to make the influence of each parameter on the model more independent, thereby reducing the interaction between parameters during the optimization process and improving the accuracy and stability of the overall solution.

[0066] Step 14.2 Standardization includes:

[0067] In nonlinear optimization, each column of the Jacobian matrix corresponds to a different parameter, and these parameters may have significant differences in properties, units, and dimensions. This difference may cause the influence of certain parameters on the model to be amplified or reduced during the optimization process, thereby affecting the convergence and accuracy of the optimization. In order to solve this problem, the present invention further performs standardization after the Jacobian matrix is ​​centralized. The standardization process adjusts the scale of each parameter to be consistent so that different parameters have the same weight during iterative optimization. This processing method can not only improve the stability of the optimization algorithm, but also accelerate convergence and improve the accuracy of the final solution.

[0068] The Jacobian matrix after centering During the standardization process, the standardized matrix Defined as:

[0069] (12)

[0070] in, and Represents matrices and matrix of OK Column elements. yes No. The standard deviation of the column is calculated as:

[0071] (13)

[0072] in, is the mean of the column vectors of the centered Jacobian matrix. The standardized Jacobian matrix The scale differences among parameters are eliminated and the correlation is weakened, thereby improving the convergence and accuracy of the nonlinear optimization algorithm.

[0073] Finally, the Jacobian matrix after center normalization is It can be expressed as:

[0074] (14)

[0075] Since the scale of the Jacobian matrix in the nonlinear optimization model has changed, the optimization step size should also be adjusted accordingly, that is, the optimization step size should be used in equation (8) To replace the original :

[0076] (15)

[0077] After the step size and Jacobian matrix are both center-normalized, subsequent optimization iterations are performed.

[0078] After center standardization, It becomes a symmetric matrix with 1 on the main diagonal. The solution obtained in this way is relatively stable and can effectively reduce the interference of accidental observation errors on the results. Incorporating the idea of ​​central normalization into nonlinear optimization methods can effectively reduce parameter coupling, avoid falling into local minima, accelerate convergence, and improve solution efficiency.

[0079] Step 15: Nonlinear optimization is used to solve the remote target pose parameters, including:

[0080] The Levenberg-Marquardt algorithm is used again for nonlinear optimization to obtain the optimal solution for the pose parameters. Depending on the optimization effect, it can be chosen whether to impose a priori parameter constraints on the depth information during the optimization process.

[0081] Step 16: Visual measurement and analysis of multiple sets of far-field target poses, including:

[0082] After successfully building the observation device of step 11, adjust the far-field target pose to be measured, and perform steps 12 to 15 in sequence to obtain multiple sets of pose parameters. Use a short-focus camera to image and measure the target at a close distance, and a long-focus camera to image and measure the target at a long distance. The former is the true value of the target motion pose, and the latter is the measured value. Move the target multiple times and count the error of the target pose change to verify the effectiveness and accuracy of the present invention.

[0083] Figure 2 This is a schematic diagram of the far-field target pose visual measurement principle of the monocular camera of the present invention; in the figure, Represents the camera coordinate system; Represents the world coordinate system; represents the far-field object coordinate system; Indicates the far field distance in km; Indicates that the far field distance is much greater than the camera focal length; Represents the characteristic points of far-field objects; Represents the projection points of feature points of far-field objects.

[0084] Figure 3 This is the principle diagram for the verification of the accuracy test of the far-field target posture visual measurement; in the figure, Represents the telephoto camera coordinate system; Represents the short-focus camera coordinate system; represents the far-field target coordinate system; The pose parameters representing the pose transformation of the target; Represents the target's pose parameters relative to the telephoto camera; Represents the pose parameters of the target relative to the short-focus camera.

[0085] Example:

[0086] 1. Simulation experiment

[0087] The present invention needs to use simulation experiments to verify the effectiveness of the proposed method. In the process of simulating far-field observation, in order to make the camera image occupy about 2 / 3 of the screen as much as possible to improve the observation accuracy, it is necessary to use different object distances to , different side lengths The far-field target (assumed to be square) is configured with different camera focal lengths The target parameters are shown in Table 1.

[0088] Table 1 Correspondence table of different object distances, focal lengths, and target side lengths

[0089] Corresponding to the different observation targets in Table 1, the far-field target to be measured in the simulation experiment is set to have a side length of of At the same time, in order to ensure the reference value of the simulation experiment, it is necessary to select a certain random posture parameter for the far-field target to be measured: Three rotation angles of the axis Set them to random angles between 5° and 30° respectively; set the translation vector of the observed target Set to arrive A random value between ; the translation vector of the observed target Set to arrive A random value between .

[0090] According to the above content, the camera focal length used in the simulation experiment It can be obtained according to the different observation targets selected in Table 1. The other internal parameters of the camera are set as follows: principal point coordinates Pixels, Pixel, radial distortion coefficient , , , tangential distortion coefficient , .

[0091] After determining all camera intrinsic parameters, target size and pose parameters, Gaussian white noise with a range of 0 to 1 pixel and a spacing of 0.1 pixel is added to each target point to simulate the far-field position error. In the simulation experiment, the target pose measurement results at each error level are measured and optimized 100 times, and noise is added again each time to enhance the randomness and credibility of the experiment. Each measurement can obtain a set of pose parameter results obtained by different methods. Next, the RMS error and reprojection error of all results are calculated, and finally displayed and analyzed through charts.

[0092] In order to accurately verify the optimization effect of the invention method, three methods are used for simulation detection, namely:

[0093] Method 1: For the simulated far-field target and the simulated target imaging image, the RPnP method is used to solve the pose parameters (RPnP);

[0094] Method 2: The posture parameters obtained by the RPnP algorithm in method 1 are used as the initial optimization values, and the Levenberg-Marquardt algorithm is used for nonlinear iterative optimization (PnP-BA);

[0095] Method 3: Consider method 2 as pre-optimization, and use the obtained pose parameters as the initial values ​​for subsequent optimization. The Levenberg-Marquardt nonlinear optimization algorithm based on the central normalization idea is used to iteratively optimize the initial values. At the same time, it is possible to choose whether to introduce appropriate parameter prior constraints (PnP-CSBA) according to the actual optimization effect.

[0096] Among them, the range of the parameter prior constraints is set based on the posture parameters obtained by pre-optimization in method 2.

[0097] In the simulation process, the three proposed methods are compared with the traditional DLT method and EPnP method at multiple noise levels to verify the effectiveness of the proposed methods and demonstrate their optimization effects. The evaluation indicators are the RMS error of the pose parameters and the reprojection error of the far-field target. The results are as follows: Figure 4 shown.

[0098] Depend on Figure 4 It can be observed from the results that both PnP+BA and PnP+CSBA methods are significantly better than the basic RPnP method in reducing the RMS error of angle and translation vector, and the optimization effect is not significantly affected as the noise level increases. In particular, the introduction of the centralized PnP+CSBA method shows the best performance under all experimental conditions, proving that the PnP+CSBA method not only performs well under low noise conditions, but also maintains a low error level in high noise environments. It is the most stable and accurate pose estimation method under various noise conditions.

[0099] 2. Physical experiment

[0100] After completing the internal calibration of the Hikvision camera and telephoto camera, a checkerboard target with a size of 12 rows and 9 columns and a side length of 45 mm per grid will be prepared as a target for subsequent observations.

[0101] After fixing the positions of the near-field Hikvision camera and the far-field telephoto camera, the target's posture and position are adjusted, and the two cameras are used to shoot the adjusted target each time. After each target adjustment, a set of pose parameters of the target in the camera coordinate system can be obtained, and the pose change in two different poses can be derived through RT transformation. Therefore, two non-repeating sets of pose parameters can be selected from the observed pose parameters to obtain their pose change. Using the pose change as the pose parameter observation value can not only improve the calculation efficiency, but also generate multiple sets of data for analysis, thereby improving the stability and accuracy of the results. The true value of the target's pose parameter is calculated by the RPnP algorithm from the target image taken by the Hikvision camera. The target's pose parameter observation value is obtained by the RPnP, PnP-BA and PnP-CSBA methods from the target image taken by the telephoto camera.

[0102] In the experiment, the target's different postures were adjusted 20 times, and two sets of posture parameters were randomly selected without repetition, and their posture changes were solved as the posture parameter observations. Since far-field imaging is susceptible to noise, in order to reduce the impact of gross errors on the observation results, the median absolute deviation (MAD) method was used to identify outliers that differed from the median by more than three times the MAD. After filtering out the outlier solutions in the observation results, the optimization effects of the three methods of RPnP, PnP-BA and PnP-CSBA for the telephoto camera were evaluated based on the true values ​​of the short-focus camera's posture parameters. The evaluation index is the RMS error of the posture parameters, and the results are shown in Table 2.

[0103] Table 2 RMS error of pose parameters of physical experiment

[0104] According to the RMS error data of the pose parameters of the three methods in Table 2, the relative error reduction percentage of the PnP-BA method and the PnP-CSBA method can be calculated:

[0105] In the right In the RMS error evaluation of parameters, the relative error reduction percentages of the PnP-BA method relative to the RPnP method were 43.53%, 47.11% and 41.92%, respectively, and the relative error reduction percentages of the PnP-CSBA method relative to the RPnP method were 43.84%, 48.09% and 42.45%, respectively. Moreover, the PnP-CSBA method reduced the relative error by 0.31%, 0.98% and 0.53% more than the PnP-BA method.

[0106] In the right In the RMS error evaluation of parameters, the relative error reduction percentages of the PnP-BA method relative to the RPnP method were 28.19%, 30.27% and 31.15%, respectively, and the relative error reduction percentages of the PnP-CSBA method relative to the RPnP method were 28.44%, 31.00% and 31.51%, respectively. In addition, the PnP-CSBA method reduced the relative error by 0.25%, 0.73% and 0.35% more than the PnP-BA method.

[0107] From the above calculation results, it can be seen that in multiple measurement results, the RMS errors of the pose parameters of the PnP-CSBA method and the PnP-BA method are much smaller than that of the RPnP method, and the RMS error of the pose parameters of the PnP-CSBA method is smaller than that of the PnP-BA method, which verifies the effectiveness of the method proposed in this paper.

[0108] Therefore, after analyzing the results of the actual experiment, the observation accuracy of the pose parameters of the PnP-BA method and the PnP-CSBA method have been greatly improved. At the same time, the observation results of the PnP-CSBA method are more accurate than those of the PnP-BA method.

[0109] In the actual experiment, a telephoto camera with a focal length of 1300 mm was selected, and the experimental results were consistent with the simulation experimental results of a camera with the same focal length, verifying the effectiveness of the method proposed in the present invention. Figure 5 is the reprojection RMS error of the far-field target in the simulation experiment.

Claims

1. A far-field target pose visual measurement method based on center standardization, characterized in that The implementation steps are as follows: Step 11, design a far-field target posture visual measurement device based on a monocular telephoto camera, fix the monocular telephoto camera on a dual-axis turntable, and connect the data acquisition and processing device to the camera; take the far-field target as the far-field target to be measured, and after powering the device, adjust the far-field target posture to achieve high-precision posture observation of the long-distance target; Step 12: Using a telephoto optical camera to image the far-field target and accurately extract image feature points; Step 13, obtaining the initial values ​​of the pose parameters of the far-field target by a non-iterative RPnP method; Step 14: Establish an objective function by minimizing the back-projection error of image feature points, perform nonlinear pre-optimization through the Levenberg-Marquardt algorithm, and obtain initial values ​​for subsequent pose parameter optimization; Step 15, the Jacobian matrix of the initial value of the subsequent posture parameter optimization is centralized and standardized, and the Levenberg-Marquardt algorithm is used again for nonlinear optimization to obtain the optimal solution of the posture parameters; Step 16: Adjust the far-field target pose, collect multiple sets of image feature points in sequence, and finally substitute them into steps 12 to 15 to finally solve and obtain multiple sets of pose parameters.

2. The method for visually measuring far-field target pose based on center standardization according to claim 1, characterized in that: In step 11, the rotation angle and position of the far-field target are adjusted, and a monocular telephoto camera is installed on a two-axis turntable in the laboratory for observation, so as to build a far-field target posture visual measurement device and system based on a monocular telephoto camera.

3. The method for visually measuring far-field target pose based on center standardization according to claim 1, characterized in that: In the step 12, for the far-field target image feature points, an image feature point extraction method based on adaptive multi-scale is adopted to achieve optimal scale extraction of all far-field target image feature points.

4. The method for visually measuring far-field target pose based on center standardization according to claim 3, characterized in that: The step 13 uses a non-iterative RPnP algorithm to obtain the initial values ​​of the pose parameters based on the coordinates of the image feature points in step 12.

5. The method for visually measuring far-field target pose based on center standardization according to claim 3, characterized in that: The step 14 includes: establishing a mapping equation between the coordinates of the feature points of the far-field target image and their corresponding world coordinates, establishing an objective function with the minimum reprojection error of the image feature points, performing nonlinear pre-optimization through the Levenberg-Marquardt algorithm, and obtaining initial values ​​for subsequent pose parameter optimization.

6. The method for visually measuring far-field target pose based on center standardization according to claim 5, characterized in that: The step 15 centers and standardizes the Jacobian matrix of the initial value of the subsequent posture parameter optimization obtained in step 14, and again uses the Levenberg-Marquardt algorithm for nonlinear optimization to obtain the optimal solution for the posture parameters.

7. The method for visually measuring far-field target pose based on center standardization according to claim 5, characterized in that: In step 16, the far-field target posture is adjusted, multiple groups of image feature point images are collected, and steps 12, 13, 14, and 15 are repeated in sequence to solve and optimize the posture parameters.

8. The method for visually measuring far-field target pose based on center standardization according to claim 7, characterized in that: Use a short-focus camera to perform imaging measurement on the target at a close distance, and a long-focus camera to perform imaging measurement on the target at a long distance. The former is the true value of the position and posture parameters of the target to be measured, and the latter is the measured value. The position and posture of the target to be measured are adjusted multiple times, and the change in the position and posture parameters of the target to be measured is taken as the observed value, and the error between the measured value and the true value is calculated.

Images