An unmanned aerial vehicle target positioning method based on attitude error self-calibration

Abstract

The application discloses a kind of unmanned plane target positioning methods based on attitude error self-calibration, based on unmanned plane flight parameter and the target pixel coordinates perceived by camera, extract the observation sequence with time sequence association;Construct package multidimensional constraint optimization model;Through multi-start L-BFGS-B quasi-Newton algorithm, global nonlinear iterative optimization is carried out;Obtain the gradient information of multidimensional constraint optimization model by numerical differentiation, obtain optimal deviation operator;The optimal deviation operator is compensated into space line-of-sight projection model, and then the accurate solution of target longitude and latitude is realized.The application improves the traditional static positioning to dynamic self-calibration positioning based on space-time logic through unsupervised self-evolution calibration mechanism, effectively solves the positioning deviation problem caused by sensor system error in unmanned plane inspection, significantly improves the spatial positioning accuracy and trajectory convergence of target in complex geographical environment.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence and computer vision technology, specifically relating to a UAV target localization method based on attitude error self-calibration. Background Technology

[0002] With the rapid development of unmanned aerial vehicle (UAV) technology, the use of UAVs equipped with visible light or infrared cameras for ground target positioning and tracking has been widely applied in fields such as power line inspection, search and rescue, and traffic monitoring. However, achieving high-precision ground target positioning in practical applications still faces many challenges:

[0003] Significant sensor system errors: UAV ground target positioning relies on the comprehensive calculation of Global Navigation Satellite System (GNSS), Inertial Measurement Unit (IMU), gimbal encoder, and camera intrinsic parameters. Due to hardware manufacturing precision, installation deviations, environmental temperature variations, and physical vibrations generated during long-term flight, sensors often exhibit systematic drift. For example, zero-point drift of the barometric altimeter, slight offsets in the gimbal attitude angle, and changes in camera focal length under high magnification zoom are all significantly amplified in the spatial projection process of "flight parameters - image pixels - ground coordinates," resulting in severe offsets in the calculated position of targets hundreds of meters away.

[0004] Traditional calibration methods have significant limitations: existing high-precision calibration methods typically rely on ground control points or known geographical references for auxiliary correction. However, in complex geographical environments such as field inspections and disaster sites, it is often difficult to pre-deploy ground control points, and there is a lack of obvious geographical reference features, making it difficult to implement traditional supervised calibration methods.

[0005] In summary, how to break free from dependence on ground control points and utilize the time-series observation sequences acquired by the UAV itself to automatically eliminate parameter deviations and dynamically enhance positioning accuracy has become a key technical problem that urgently needs to be solved in the current field of UAV inspection. Summary of the Invention

[0006] This application provides a UAV target localization method based on attitude error self-calibration to solve the technical problems of existing UAV localization technologies that rely too much on ground control points and are greatly affected by sensor system errors.

[0007] To address the aforementioned technical problems, this application adopts the following technical solution: a UAV target localization method based on attitude error self-calibration, comprising:

[0008] S1. Based on the UAV flight parameters and the target pixel coordinates sensed by the camera, construct a spatial line-of-sight projection model, obtain the initial geospatial location of the target, and extract the observation sequence with temporal correlation;

[0009] S2. Based on the observation sequence, construct a multi-dimensional constraint optimization model that includes observation consistency constraint operators, trajectory smoothness constraint operators, weakly supervised mapping reference terms, and parameter regularization penalty mechanisms;

[0010] S3. Based on the deviation operator to be optimized, a global nonlinear iterative optimization is performed using the multi-starting-point L-BFGS-B quasi-Newton algorithm with the goal of minimizing the multidimensional constrained optimization model;

[0011] S4. During the iterative optimization process, the gradient information of the multidimensional constrained optimization model is obtained through numerical differentiation, and the approximate value of the inverse Hessian matrix is ​​dynamically updated. The second-order curvature information is used to correct the search step size and obtain the optimal deviation operator.

[0012] S5. The optimal deviation operator is compensated into the spatial line-of-sight projection model to eliminate the systematic projection drift caused by the distortion of the altimeter, inertial measurement unit and camera, thereby achieving accurate calculation of the target latitude and longitude.

[0013] Furthermore, the method in step S1 includes:

[0014] S11. Establish the target direction vector in the camera coordinate system. Based on the homogeneous pixel coordinates of the target within the camera, calculate the unit direction vector of the target in the camera coordinate system:

[0015] S12. Construct a multi-level composite rotation matrix, obtain the pitch angle, yaw angle and roll angle of the UAV gimbal in real time, construct the corresponding rotation component matrices respectively, and synthesize the comprehensive transformation moment by matrix multiplication;

[0016] S13. Solve the geospatial projection ray. Use the comprehensive transformation matrix to transform the direction vector in the camera coordinate system to the northeast-sky world coordinate system to obtain the normalized projection ray of the target in geospatial space.

[0017] S14. Based on the intersection of ray and plane, the horizontal displacement is deduced. Combined with the height of the UAV relative to the ground, the scale factor is calculated by using the geometric intersection relationship between the projected ray and the ground horizontal plane, and then the horizontal eastward and northward displacements of the target relative to the UAV platform are calculated.

[0018] S15. Perform geodetic coordinate system transformation, introduce the Earth's semi-major axis and flattening parameters, calculate the actual physical distance corresponding to each degree of latitude and longitude at the current latitude of the UAV, convert the horizontal displacement into latitude and longitude deviation, and compensate it into the current GNSS coordinates of the UAV to obtain the target's geographical location.

[0019] Furthermore, the method in step S2 includes:

[0020] S21. For multiple observation points under the same target identifier, calculate the centroid of the geographic coordinates of the observation points after calibration. By calculating the physical distance from each observation point to the centroid of the geographic coordinates and evaluating the accuracy of the parameters, design a piecewise penalty function and construct an observation consistency constraint operator.

[0021] S22. Utilizing the prior knowledge of the continuity of ground target motion, step size constraints and velocity smoothing constraints are applied to the time series observation sequence to construct a trajectory smoothness evaluation function;

[0022] S23. Calculate the deviation between the calibrated target geographic coordinates and the original projection model output coordinates, introduce a small-weight Euclidean distance constraint term, and establish a weakly supervised mapping reference term;

[0023] S24. Apply regularization constraints to the five-dimensional deviation operator to be optimized, allocate heterogeneous weight factors according to the sensitivity of sensor parameters, and construct a parameter regularization penalty mechanism;

[0024] S25. Combine the various weighting factors by linear weighting coefficients to construct a multidimensional constrained optimization model.

[0025] Furthermore, the method in step S3 includes:

[0026] S31. Based on the sensor hardware specifications and installation error range, set the search boundary for the five-dimensional deviation operator to be optimized, and limit the parameters to be optimized within a reasonable physical order.

[0027] S32. Within the search boundary, use uniform sampling or random perturbation strategies to generate multiple sets of initial deviation vectors as multiple starting positions for optimization;

[0028] S33. For each set of initial bias vectors, the L-BFGS-B algorithm is started independently to iterate and optimize, avoiding local minima in the non-convex loss field;

[0029] S34. After all starting points converge, compare and optimize the model values, and select the deviation parameter corresponding to the minimum value as the global optimal deviation operator.

[0030] Furthermore, the method in step S4 includes:

[0031] S41. Apply small perturbations to each dimension of the five-dimensional deviation operator based on numerical differentiation, calculate the rate of change of the optimized model, and obtain the gradient vector;

[0032] S42. Maintain a fixed-length historical observation window and store the shift vector and gradient change vector of the decision variable in the most recent iteration;

[0033] S43. Iteratively update the approximate value of the inverse Hessian matrix using the BFGS formula with limited memory, and calculate the quasi-Newton search direction by combining the gradient vector;

[0034] S44. Perform a line search in the search direction to determine the optimal step size, project it to the physical boundary to complete the parameter update, and output the optimal deviation operator after the convergence condition is met.

[0035] Furthermore, the method in step S5 includes:

[0036] S51. Extract the optimal deviation operator into height compensation, attitude correction and field of view correction, and inject them into the spatial line-of-sight projection model to obtain the correction model;

[0037] S52. Based on the modified model, the time series observation sequence is recalculated to obtain the reconstructed geographic latitude and longitude coordinates;

[0038] S53. Evaluate the dispersion of the reconstructed geographic latitude and longitude coordinate sequence, verify the effect of deviation elimination, and output high-precision target positioning results.

[0039] The beneficial effects of this application are as follows: This application establishes a spatial line-of-sight projection model based on the line-of-sight principle, accurately representing the geometric mapping relationship between the "pixel coordinate system, camera coordinate system, and world coordinate system"; combining the prior knowledge of the physical continuity of ground target motion, it constructs a multi-dimensional constrained optimization model covering observation consistency, trajectory smoothness, and weakly supervised mapping reference, transforming sensor system errors into optimizable bias operators; utilizing a multi-starting-point L-BFGS-B quasi-Newton algorithm to perform nonlinear global optimization within a preset physical boundary, and dynamically updating the approximate Hessian matrix using the finite-memory BFGS formula, it achieves accurate convergence of bias parameters in complex non-convex spaces; finally, through adaptive parameter compensation and coordinate reconstruction, it effectively eliminates systematic projection drift caused by altimeter, inertial measurement unit, and camera distortion. This method achieves unsupervised self-evolutionary calibration free from ground control points, combining the advantages of both spatiotemporal logical constraints and numerical optimization algorithms, resulting in a significant improvement. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating an embodiment of the UAV target localization method based on attitude error self-calibration of this application;

[0041] Figure 2 This is a schematic diagram of the relevant coordinate system of a spatial line-of-sight projection model of an embodiment of the UAV target localization method based on attitude error self-calibration of this application;

[0042] Figure 3 This is a comparison diagram of the results before and after optimization using the L-BFGS-B algorithm in an embodiment of the UAV target localization method based on attitude error self-calibration of this application. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0044] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0045] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the UAV target localization method based on attitude error self-calibration according to this application. The method includes:

[0046] S1. Obtain the UAV flight parameters and the target pixel coordinates perceived by the camera. Based on the line-of-sight principle, establish the transformation relationship of "pixel coordinate system - camera coordinate system - world coordinate system", i.e., the spatial line-of-sight projection model, deduce the initial geospatial position of the target, and extract the observation sequence with temporal correlation.

[0047] For details, please refer to Figure 2 The method of step S1 includes:

[0048] S11. Establish the target direction vector in the camera coordinate system, based on the camera intrinsic parameters K and the homogeneous pixel coordinates of the target. Calculate the unit direction vector of the target in the camera coordinate system. The specific formula is as follows:

[0049] (1);

[0050] S12. Construct a multi-level composite rotation matrix to obtain the pitch angle of the UAV gimbal in real time. Yaw angle and roll angle Construct the corresponding rotation component matrices respectively. The comprehensive transformation matrix is ​​synthesized through matrix multiplication. The specific formula is as follows:

[0051] (2);

[0052] S13. Solve for the geospatial projection ray. Use the comprehensive transformation matrix to transform the direction vector in the camera coordinate system to the Northeast-East-South (ENU) world coordinate system to obtain the normalized projection ray of the target in geospatial space. The specific formula is as follows:

[0053] (3);

[0054] in, is the scale factor.

[0055] S14. Determine the horizontal displacement based on ray-plane intersection, combined with the UAV's altitude relative to the ground. Using projection rays Calculate the scale factor based on the geometric orthogonality with the ground horizontal plane. Then, the horizontal eastward displacement of the target relative to the UAV platform was calculated. and northward displacement The specific formula is as follows:

[0056] (4);

[0057] (5);

[0058] (6);

[0059] S15. Perform geodetic coordinate system transformation and introduce the Earth's semi-major axis. and flatness parameters Calculate the current latitude of the drone The actual physical distance corresponding to each degree of latitude and longitude is used to convert horizontal displacement into latitude and longitude deviation. The coordinates are then compensated and added to the UAV's current GNSS coordinates to complete the target's geographical location calculation; the specific formula is as follows:

[0060] (7);

[0061] (8);

[0062] in, The radius of curvature of the circle is denoted as . The radius of curvature of the meridian:

[0063] (9);

[0064] (10);

[0065] in, , , .

[0066] Then the final geographical location of the target is obtained. for:

[0067] (11);

[0068] (12).

[0069] S2. Utilizing the prior knowledge of the physical continuity and smoothness of ground target motion, a multi-dimensional constrained optimization model is constructed, which includes an observation consistency constraint operator, a trajectory smoothness evaluation index, a weakly supervised mapping reference term, and a parameter regularization penalty mechanism. This is used to quantify the consistency of the system's positioning logic.

[0070] Specifically, the method in step S2 includes:

[0071] S21. Establish observation consistency constraint operators For n observation points under the same target identifier, calculate their calibrated geographic coordinate centroids. By calculating the physical distance from each observation point to the centroid To evaluate the accuracy of the parameters, a piecewise function is designed as a penalty: when the distance exceeds a preset physical limit threshold... When the distance is within a reasonable range, a high quadratic penalty is applied to force the discrete points to converge; when the distance is within a reasonable range, a small penalty is applied to fine-tune the accuracy. This term aims to constrain the solution results for the same target from different perspectives to tend to be uniform. The specific formula is as follows:

[0072] (13);

[0073] (14);

[0074] (15);

[0075] (16);

[0076] S22. Establish a trajectory smoothness evaluation function. By utilizing the prior knowledge of the continuity of ground target motion, kinematic constraints are applied to the time-series observation sequence.

[0077] Step size constraint: Calculate the physical displacement between adjacent observation frames. The penalty exceeds the target's normal step size threshold. The coordinate jumps are encouraged, and the displacement is kept within a reasonable range.

[0078] (17);

[0079] Velocity smoothing constraint: Calculate the average motion velocity for each time period using timestamps. If the speed changes drastically and exceeds the set threshold, Then increase the punishment:

[0080] (18);

[0081] (19);

[0082] S23. Establish weakly supervised mapping reference items The deviation between the calibrated target geographic coordinates and the original spatial line-of-sight projection model output coordinates is calculated. Using the original calculated values ​​as a weak benchmark reference, a small-weight Euclidean distance constraint term is introduced into the optimization model to ensure that the optimized parameters, while correcting for systematic errors, do not deviate from the reasonable range of basic physical observations.

[0083] (20);

[0084] S24. Establish a parameter regularization penalty mechanism L2 regularization constraints are applied to the five-dimensional deviation operator to be optimized, based on different sensor parameters. Heterogeneous weighting factors are assigned to the sensitivity to positioning accuracy. This assigns a higher constraint weight to the pitch angle deviation, ensuring that the calibrated parameters conform to the actual performance characteristics of the sensor hardware. The specific formula is as follows:

[0085] (twenty one);

[0086] S25. Multidimensional constraint indices are comprehensively integrated. Following the evaluation logic of "spatial convergence - smooth motion - physical rationality", each sub-constraint is linearly weighted and fused through weight coefficients {w1, w2, w3, w4} to construct the final evaluation function. The specific formula is as follows:

[0087] (twenty two);

[0088] Where N is the total number of target observations.

[0089] Therefore, based on the above formula definition, the multi-dimensional optimization objective function is obtained:

[0090] (twenty three);

[0091] in, For the five-dimensional deviation operator to be optimized, and The set search boundaries.

[0092] S3. Define the systematic error of the sensor as a five-dimensional deviation operator to be optimized. Within the preset physical boundary constraints, the multi-starting-point L-BFGS-B quasi-Newton algorithm is used to perform global nonlinear iterative optimization with the goal of minimizing the multidimensional constraint optimization model.

[0093] For details, please refer to Figure 3 The method of step S3 includes:

[0094] S31. Set the variable search space and physical boundaries. Based on the hardware specifications and installation error range of each sensor, determine the five-dimensional deviation operator to be optimized. Set the search boundary B=[l i , u i ],in and Let i represent the horizontal field of view and the vertical field of view, i∈{1,2,3,4,5}, and let i ∈{1,2,3,4,5}. The parameters are optimized within a reasonable physical order.

[0095] S32. Multi-starting point initialization solution vector deployment: Within the search boundary, n sets of heterogeneous initial deviation vectors are generated using a uniform sampling or random perturbation strategy. As multiple starting points for the optimization engine, among which .

[0096] S33. Iterative search and local minima avoidance: For each set of initial deviation vectors, the L-BFGS-B algorithm is started to perform independent iterative optimization. By starting from different regions in space, gradient detection is performed to avoid local minima pitfalls in the non-convex loss energy field.

[0097] S34. Evaluation of Optimization Results and Selection of Global Optimum: After the iterative process of all starting points converges, compare the final multidimensional constraint optimization model values ​​corresponding to each starting point, and select the set of deviation parameters with the smallest value as the global optimal deviation operator. .

[0098] S4. During the iterative optimization process, the gradient information of the multidimensional constrained optimization model is obtained through numerical differentiation, and the approximate value of the inverse Hessian matrix is ​​dynamically updated using the BFGS formula with limited memory. The search step size is corrected using second-order curvature information to achieve convergent solution of the bias operator in non-convex space.

[0099] Specifically, the method in step S4 includes:

[0100] S41. Gradient probing based on numerical differentiation: In each iteration, a small perturbation is applied to each dimension of the current five-dimensional bias operator vector. Calculate the rate of change of the multidimensional constrained optimization model to obtain the gradient vector in the current parameter space. Determine the steepest descent direction of the multidimensional constrained optimization model:

[0101] (twenty four);

[0102] in, Corresponding five-dimensional deviation operator .

[0103] S42. Limited memory storage for spatiotemporal evolution data, maintaining a fixed-length historical observation window of length n in real time to store the shift vectors of decision variables during the most recent k iterations. With gradient change vector The specific formula is as follows:

[0104] (25);

[0105] (26);

[0106] S43. Dynamic approximation update of the Hessian matrix inverse: Utilizing stored variable shift and gradient change data, the approximation operator of the Hessian matrix inverse is iteratively updated using the BFGS formula with limited memory. and combined with the current gradient vector Solving the quasi-Newtonian search direction ;

[0107] (27);

[0108] (28);

[0109] S44. Search step size optimization and constraint projection update: Perform a line search in the search direction to determine the optimal search step size. The generated search step size is projected onto the physical boundary described in step S31, and parameter updates are performed. When the reduction of the multidimensional constrained optimization model in two adjacent iterations is less than a preset threshold, or the gradient norm tends to zero, the parameters are determined to have reached convergence, and the optimal deviation operator is output. .

[0110] S5. The optimal deviation operator obtained through optimization is compensated into the spatial line-of-sight projection model to eliminate the systematic projection drift caused by the distortion of the altimeter, inertial measurement unit and camera, thereby achieving accurate calculation of the target latitude and longitude.

[0111] Specifically, the method in step S5 includes:

[0112] S51. Instantiation compensation of optimal deviation parameters, using the optimal deviation operator The error is extracted into height compensation, attitude correction and field of view correction respectively, and then injected into the spatial line-of-sight projection model as a compensation term to complete the static elimination of the original sensor system error.

[0113] S52. Using the corrected input and spatial line-of-sight projection model, the target pixel coordinates of each frame in the time-series observation sequence are recalculated to obtain the reconstructed geographic latitude and longitude coordinates.

[0114] S53. Positioning accuracy and consistency verification: The dispersion of the reconstructed geographic latitude and longitude coordinate sequence is evaluated. By calculating the convergence ratio of the calibrated trajectory relative to the original trajectory, the effectiveness of the system deviation elimination is verified, and the final high-precision target positioning result is output.

[0115] This application establishes a spatial line-of-sight projection model based on the line-of-sight principle to accurately depict the three-dimensional coordinate transformation logic of "pixel-camera-world". It also utilizes the prior physical continuity of ground target motion to construct a multi-dimensional constrained optimization model, transforming systematic sensing errors into a deviation operator optimization problem in a high-dimensional parameter space. Through the multi-starting-point L-BFGS-B algorithm, nonlinear iterative optimization is performed, and the search path is dynamically corrected using second-order curvature information. This achieves accurate convergence and adaptive compensation of deviation parameters in non-convex space, effectively eliminating sensor coupling errors, thereby improving the accuracy of UAV ground observation target positioning coordinates and trajectory consistency in complex environments.

[0116] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A UAV target localization method based on attitude error self-calibration, characterized in that, include: S1. Based on the UAV flight parameters and the target pixel coordinates sensed by the camera, construct a spatial line-of-sight projection model, obtain the initial geospatial location of the target, and extract the observation sequence with temporal correlation; S2. Based on the observation sequence, construct a multi-dimensional constraint optimization model that includes an observation consistency constraint operator, a trajectory smoothness constraint operator, a weakly supervised mapping reference term, and a parameter regularization penalty mechanism; S3. Based on the deviation operator to be optimized, a global nonlinear iterative optimization is performed using the multi-starting-point L-BFGS-B quasi-Newton algorithm with the goal of minimizing the multidimensional constrained optimization model; S4. During the iterative optimization process, the gradient information of the multidimensional constrained optimization model is obtained through numerical differentiation, and the approximate value of the inverse Hessian matrix is ​​dynamically updated. The search step size is corrected using second-order curvature information to obtain the optimal deviation operator. S5. The optimal deviation operator is compensated into the spatial line-of-sight projection model to eliminate the systematic projection drift caused by the distortion of the altimeter, inertial measurement unit and camera, thereby achieving accurate calculation of the target latitude and longitude.

2. The method according to claim 1, characterized in that, The method of step S1 includes: S11. Establish the target direction vector in the camera coordinate system. Based on the homogeneous pixel coordinates of the target within the camera, calculate the unit direction vector of the target in the camera coordinate system: S12. Construct a multi-level composite rotation matrix, obtain the pitch angle, yaw angle and roll angle of the UAV gimbal in real time, construct the corresponding rotation component matrices respectively, and synthesize the comprehensive transformation moment by matrix multiplication; S13. Solve the geospatial projection ray, and use the comprehensive transformation matrix to transform the direction vector in the camera coordinate system to the northeast-sky world coordinate system to obtain the normalized projection ray of the target in geospatial space; S14. Based on the intersection of ray and plane, the horizontal displacement is deduced. Combined with the height of the UAV relative to the ground, the scale factor is calculated by using the geometric intersection relationship between the projected ray and the ground horizontal plane, and then the horizontal eastward and northward displacements of the target relative to the UAV platform are calculated. S15. Perform geodetic coordinate system transformation, introduce the Earth's semi-major axis and flattening parameters, calculate the actual physical distance corresponding to each degree of latitude and longitude at the current latitude of the UAV, convert the horizontal displacement into latitude and longitude deviation, and compensate it into the current GNSS coordinates of the UAV to obtain the target's geographical location.

3. The method according to claim 1, characterized in that, The method of step S2 includes: S21. For multiple observation points under the same target identifier, calculate the calibrated geographic coordinate centroid of the observation points, and design a piecewise penalty function and construct an observation consistency constraint operator by calculating the physical distance from each observation point to the geographic coordinate centroid and evaluating the accuracy of the parameters. S22. Utilizing the prior knowledge of the continuity of ground target motion, step size constraints and velocity smoothing constraints are applied to the time-series observation sequence to construct a trajectory smoothness evaluation function; S23. Calculate the deviation between the calibrated target geographic coordinates and the original projection model output coordinates, introduce a small-weight Euclidean distance constraint term, and establish a weakly supervised mapping reference term; S24. Apply regularization constraints to the five-dimensional deviation operator to be optimized, allocate heterogeneous weight factors according to the sensitivity of sensor parameters, and construct a parameter regularization penalty mechanism; S25. The various weighting factors are linearly weighted and fused through weighting coefficients to construct the multidimensional constrained optimization model.

4. The method according to claim 1, characterized in that, The method of step S3 includes: S31. Based on the sensor hardware specifications and installation error range, set the search boundary for the five-dimensional deviation operator to be optimized, and limit the parameters to be optimized within a reasonable physical order. S32. Within the search boundary, a uniform sampling or random perturbation strategy is used to generate multiple sets of initial deviation vectors as multiple starting positions for optimization; S33. For each set of initial bias vectors, the L-BFGS-B algorithm is started independently to iterate and optimize, avoiding local minima in the non-convex loss field; S34. After all starting points converge, compare and optimize the model values, and select the deviation parameter corresponding to the minimum value as the global optimal deviation operator.

5. The method according to claim 4, characterized in that, The method of step S4 includes: S41. Apply small perturbations to each dimension of the five-dimensional deviation operator based on numerical differentiation, calculate the rate of change of the optimized model, and obtain the gradient vector; S42. Maintain a fixed-length historical observation window and store the shift vector and gradient change vector of the decision variable in the most recent iteration; S43. Iteratively update the approximate value of the inverse Hessian matrix using the BFGS formula with limited memory, and calculate the quasi-Newton search direction by combining the gradient vector; S44. Perform a line search in the search direction to determine the optimal step size, project it onto the physical boundary to complete the parameter update, and output the optimal deviation operator after the convergence condition is met.

6. The method according to claim 1, characterized in that, The method of step S5 includes: S51. Extract the optimal deviation operator into height compensation, attitude correction and field of view correction, and inject them into the spatial line-of-sight projection model to obtain the correction model; S52. Based on the modified model, the time series observation sequence is recalculated to obtain the reconstructed geographic latitude and longitude coordinates; S53. Evaluate the dispersion of the reconstructed geographic latitude and longitude coordinate sequence, verify the effect of deviation elimination, and output high-precision target positioning results.

Images