Visual dynamic hover control method and system

Abstract

The application discloses a visual dynamic hovering control method and system, and belongs to the technical field of autonomous control of underwater robots and machine vision, and comprises the following steps: constructing an image kinematics prediction model based on a Gaussian process according to an acquired visual feature error, obtaining a corrected prediction model and outputting a posterior prediction variance; constructing an adaptive field of view constraint set; solving a Gaussian process model prediction control instruction according to the corrected prediction model and the adaptive field of view constraint set, and obtaining a kinematics layer expected body speed; and adopting acceleration level model-free adaptive control to perform dynamics tracking according to the kinematics layer expected body speed. The application compensates model uncertainty through online learning of the Gaussian process, adjusts the field of view constraint adaptively by using the prediction variance, and realizes smooth tracking of the speed instruction in combination with the model-free adaptive control, so that the visual hovering precision, safety and robustness of the underwater robot in an unknown image depth and a strong uncertain disturbance environment are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of underwater robot autonomous control and machine vision technology, and particularly relates to a visual dynamic hovering control method and system. Background Technology

[0002] In near-structure precision operations (such as subsea pipeline inspection, drilling platform maintenance, and underwater archaeology), underwater robots (ROVs) need to maintain stable relative attitude and position with the target object for extended periods, achieving visual dynamic hovering. Existing technologies largely rely on high-precision inertial navigation systems, underwater acoustic positioning devices, or multi-sensor fusion schemes to acquire relative pose information between the robot and the target. However, these solutions suffer from high costs, complex system integration, and susceptibility to multipath interference and acoustic noise in shallow water or near-structure areas. Vision-based servo control can directly utilize target image features extracted by a monocular camera to construct a visual closed loop, promising to achieve relative hovering control at a lower cost, and thus has gradually become a research hotspot in the field of near-structure underwater robot operations.

[0003] While visual servo control has shown potential in underwater hovering tasks, existing technologies still face the following key challenges: First, the inherent lack of depth information in target images under monocular imaging conditions leads to the inability to accurately obtain the image depth term in the image Jacobian matrix, causing mismatch in the visual kinematics model and resulting in degraded or even unstable control accuracy. Second, the underwater environment presents strong uncertainties such as illumination attenuation, scattering, water flow disturbances, and thruster nonlinearity. Traditional fixed field-of-view boundaries or fixed safety margin constraints are insufficient to balance feature point visibility (preventing features from exceeding the field of view) and control flexibility (avoiding excessive conservatism leading to slow convergence) under random disturbance conditions. Third, underwater robot dynamics models contain time-varying nonlinear parameters such as added mass, Coriolis force, and fluid damping, making it difficult for the inner-loop controller to smoothly and accurately track the velocity commands planned by the visual outer loop. Therefore, there is an urgent need for a visual dynamic hovering control method that can online correct the visual kinematics model, adaptively adjust the field-of-view safety constraints, and achieve high-precision velocity tracking at the acceleration level under conditions of unknown image depth and strong uncertain disturbances. Summary of the Invention

[0004] The main objective of this invention is to address the uncertainties caused by model mismatch and external disturbances in ROV visual dynamic hovering under complex underwater environments and model uncertainties, thereby achieving a unified approach to online learning and real-time optimization.

[0005] To achieve the above objectives, the present invention provides a visual dynamic hovering control method and system.

[0006] One of the visual dynamic hovering control methods includes:

[0007] Based on the obtained visual feature errors, an image kinematic prediction model based on Gaussian processes is constructed to obtain the corrected prediction model;

[0008] Based on the posterior prediction variance output by the modified prediction model, an adaptive vision constraint set is constructed.

[0009] Based on the modified prediction model and the adaptive field constraint set, the Gaussian process model predicts the control command to obtain the expected body velocity at the kinematic layer.

[0010] Based on the desired body velocity of the kinematic layer, dynamic tracking is performed using acceleration-level model-free adaptive control.

[0011] Preferably, the process of constructing an image kinematic prediction model based on Gaussian processes includes:

[0012] Establish a nominal image Jacobian matrix that includes the nominal image depth, and model the image depth approximation error, camera mounting bias, and unmodeled dynamics into a total uncertainty residual.

[0013] The total uncertainty residual is learned online using a Gaussian process, and the posterior prediction mean and posterior prediction variance are output in the current state.

[0014] By combining the nominal image Jacobian matrix and the posterior prediction mean, the corrected image kinematic prediction model is obtained.

[0015] Preferably, the process of learning the residuals online using a Gaussian process includes:

[0016] The Gaussian process input vector is defined as a combination of the image feature stacking vector, the kinematic layer velocity input, and the nominal image depth vector.

[0017] Residual observation labels are defined using a one-step forward differencing method;

[0018] A scalar Gaussian process model is established for each pixel velocity residual component, and the stacked posterior prediction mean vector and the diagonal covariance matrix composed of the variances of each independent output channel are output.

[0019] Preferably, the process of constructing an adaptive view constraint set includes:

[0020] Extract the posterior prediction variance of the Gaussian process output, quantify the uncertainty level of the current environment and the model based on the posterior prediction variance, and calculate the perturbation envelope set;

[0021] The original visual feasible set of the image plane is dynamically shrunk using the perturbation envelope set to generate an adaptive field-of-view constraint set.

[0022] Preferably, the process of generating the adaptive view constraint set further includes:

[0023] Define a two-sided opportunity constraint to ensure that feature points are within the field of view boundary with a preset high confidence level probability;

[0024] Transforming probabilistic chance constraints into a deterministic tightening form yields the tightening boundary;

[0025] Based on the tightened boundary, a perturbation envelope set is constructed, and the Minkowski difference between the original visual feasible set and the perturbation envelope set is calculated to obtain the adaptive visual constraint set.

[0026] Preferably, the process of solving the Gaussian process model to predict control commands includes:

[0027] The nominal image Jacobian matrix and the Gaussian process posterior prediction mean are locally linearized in the current prediction time domain to construct a local affine prediction model.

[0028] A quadratic cost function is constructed using the state error prediction sequence, control input sequence, and terminal error as penalty terms.

[0029] The pixel state of each prediction step is constrained to be within the adaptive field constraint set. The control input sequence that minimizes the cost function is solved, and the first element of the control input sequence is taken as the expected body velocity of the kinematic layer at the current moment.

[0030] Preferably, the process of dynamic tracking using acceleration-level model-free adaptive control includes:

[0031] Differential filtering is performed on the desired body velocity of the kinematic layer to obtain the feedforward acceleration, and the target acceleration of the dynamic layer is calculated in combination with the current body velocity;

[0032] The pseudo-Jacobi matrix is ​​estimated online based on the system input and output data, and the control law is calculated based on the acceleration error to drive the underwater robot to track the desired body velocity.

[0033] Preferably, the process of calculating the target acceleration includes:

[0034] The feedforward acceleration is obtained by constructing a smooth derivative of the visual desired velocity;

[0035] The target acceleration is calculated based on the feedforward acceleration, the current body velocity, and the desired body velocity, wherein the expression for the target acceleration includes a gain matrix and a transient change compensation coefficient.

[0036] The present invention also provides a visual dynamic hovering control system, comprising:

[0037] The error extraction module is used to obtain visual feature errors;

[0038] The kinematic prediction update module is used to construct an image kinematic prediction model based on Gaussian processes, output the posterior prediction mean and posterior prediction variance, and obtain the corrected image kinematic prediction model.

[0039] The adaptive vision constraint module is used to construct an adaptive vision constraint set based on the posterior prediction variance.

[0040] The velocity planning module is used to combine the modified image kinematics prediction model with the adaptive field constraint set to solve the Gaussian process model predictive control commands and generate the expected body velocity at the kinematic layer.

[0041] An adaptive dynamics tracking module is used to perform dynamics tracking using acceleration-level model-free adaptive control based on the desired body velocity of the kinematic layer.

[0042] Compared with the prior art, the present invention has the following advantages and technical effects:

[0043] First, by uniformly modeling the image depth approximation error, camera mounting bias, and unmodeled dynamics as the total uncertainty residual and using Gaussian process online learning compensation, the image Jacobian model can be continuously corrected without the need for precise image depth measurement. This solves the model mismatch problem caused by the lack of image depth in monocular vision servoing and significantly improves hovering accuracy and control consistency.

[0044] Second, the uncertainty level of the current environment and model is quantified by the posterior prediction variance output by the Gaussian process, and the field constraint is dynamically contracted and adaptively adjusted. When the uncertainty is high, the boundary is automatically tightened to reduce the risk of feature points going out of the field of view, and the constraint is relaxed when the uncertainty is low to improve the convergence speed and control flexibility. This solves the contradiction between safety and performance that is difficult to balance with fixed field constraints.

[0045] Third, by using acceleration-level modelless adaptive control to smoothly map the desired velocity command to the thruster input, high-precision velocity tracking and reduced chattering can be achieved under time-varying hydrodynamic parameters and external disturbances without the need for an accurate dynamic model. This effectively solves the problem that the visual outer loop command is difficult to be accurately tracked by the inner loop, thereby improving the overall hovering accuracy, disturbance recovery capability and operational robustness of the system. Attached Figure Description

[0046] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0047] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0048] Figure 2This is a schematic diagram of the underwater robot according to an embodiment of the present invention. Detailed Implementation

[0049] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0050] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0051] Example 1

[0052] like Figure 1 As shown, this embodiment provides a visual dynamic hovering control method, including:

[0053] Based on the obtained visual feature errors, an image kinematic prediction model based on Gaussian processes is constructed to obtain the corrected prediction model;

[0054] Based on the posterior prediction variance output by the corrected prediction model, an adaptive vision constraint set is constructed.

[0055] Based on the modified prediction model and the adaptive field constraint set, the Gaussian process model predicts the control command and obtains the expected body velocity at the kinematic level.

[0056] Based on the desired body velocity at the kinematic level, a model-free adaptive control at the acceleration level is used for dynamic tracking.

[0057] Furthermore, the process of constructing an image kinematic prediction model based on Gaussian processes includes:

[0058] Establish a nominal image Jacobian matrix that includes the nominal image depth, and model the image depth approximation error, camera mounting bias, and unmodeled dynamics into a total uncertainty residual.

[0059] The total uncertainty residual is learned online using a Gaussian process, and the posterior prediction mean and posterior prediction variance are output for the current state.

[0060] By combining the nominal image Jacobian matrix and the posterior prediction mean, the corrected image kinematic prediction model is obtained.

[0061] Furthermore, the process of learning residuals online using Gaussian processes includes:

[0062] The Gaussian process input vector is defined as a combination of the image feature stacking vector, the kinematic layer velocity input, and the nominal image depth vector.

[0063] Residual observation labels are defined using a one-step forward differencing method;

[0064] A scalar Gaussian process model is established for each pixel velocity residual component, and the stacked posterior prediction mean vector and the diagonal covariance matrix composed of the variances of each independent output channel are output.

[0065] Furthermore, such as Figure 2 As shown, in this embodiment, the visual feature error is obtained by extracting the pixel coordinates of target feature points in the current image and the desired image from the monocular camera mounted on the underwater robot, establishing the corresponding coordinate set, and calculating the visual feature error between the current pixel state and the desired pixel state.

[0066] Specifically, the underwater robot is equipped with a monocular camera, and the world coordinate system is defined as the NE-G coordinate system. Both the ROV's body coordinate system and the camera coordinate system adopt the front-right-bottom coordinate system. Let the Euclidean coordinates of the i-th target feature point in the current view frame and the desired view frame be represented as follows: and Extract the pixel coordinates of target feature points from the monocular camera in the current image and the desired image, and establish the corresponding coordinate set: Calculate the current pixel state With the desired pixel state Visual feature error between .

[0067] Furthermore, due to the lack of accurate image depth measurement information in the actual system, and because the camera is not mounted at the geometric center of the robot (assuming it is offset along the forward axis), This embodiment constructs an image kinematics prediction model based on Gaussian processes, including:

[0068] Construct a nominal image Jacobian matrix that includes the nominal image depth. The image depth approximation error, camera mounting bias, and unmodeled dynamics are uniformly modeled as the total uncertainty residual, and GP is used for online learning and compensation.

[0069] The residuals are learned online using Gaussian processes. The specific process is as follows:

[0070] Define the Gaussian process input vector at the k-th sampling time as:

[0071] ;

[0072] in, Stacked vectors of image features Input for kinematic layer velocity, The component of the body velocity in the horizontal plane , The nominal image depth vector;

[0073] Residual observation labels are defined using one-step forward differencing. Its expression is:

[0074] ;

[0075] in, The sampling period is The nominal image Jacobian matrix;

[0076] A scalar Gaussian process model is established for each pixel velocity residual component, and trained using a squared exponential kernel function to output the posterior prediction mean vector of the current state. The diagonal covariance matrix composed of the variances of each independent output channel .

[0077] By using the posterior prediction mean vector as a real-time compensation term and combining it with the nominal image Jacobian matrix, a corrected discrete image kinematic prediction model is obtained.

[0078] More specifically, a multi-output GP model with independent outputs is used to build a scalar GP model for each pixel velocity residual component, trained using a squared exponential kernel function. Hyperparameters are updated online using sliding window data, and the posterior prediction mean vector for the current state is output. The diagonal covariance matrix composed of the variances of each independent output channel After fusing GP correction, the discrete image kinematic prediction model is obtained as follows: ,in .

[0079] This embodiment explicitly constructs the image Jacobian mismatch caused by unknown image depth as a residual and learns it online. This allows for continuous correction of the visual kinematics model without the need for precise image depth measurement, thereby improving hovering accuracy and consistency.

[0080] This embodiment uses Gaussian process prediction variance to adaptively tighten the field constraints, which can be more conservative when uncertainty is high to reduce the risk of going out of bounds, and relax the constraints when uncertainty is low to improve convergence speed and control flexibility.

[0081] Furthermore, the process of constructing the adaptive view constraint set includes:

[0082] Extract the posterior prediction variance of the Gaussian process output, quantify the uncertainty level of the current environment and the model based on the posterior prediction variance, and calculate the perturbation envelope set;

[0083] The original visual feasible set of the image plane is dynamically shrunk using the perturbation envelope set to generate an adaptive field constraint set.

[0084] Furthermore, the process of generating the adaptive view constraint set also includes:

[0085] Define a two-sided opportunity constraint to ensure that feature points are within the field of view boundary with a preset high confidence level probability;

[0086] Transforming probabilistic chance constraints into a deterministic tightening form yields the tightening boundary;

[0087] By constructing a perturbation envelope set based on the tightened boundary, the Minkowski difference between the original visual feasible set and the perturbation envelope set is calculated to obtain the adaptive vision constraint set.

[0088] Furthermore, to prevent water flow disturbance from causing target features to leave the camera's field of view, this embodiment constructs an adaptive field-of-view constraint based on the GP prediction variance.

[0089] Specifically, in this embodiment, the adaptive vision constraint is constructed by extracting the posterior prediction variance (i.e., the diagonal covariance matrix) output by the Gaussian process. This can be used to quantify the level of uncertainty in the current environment and model.

[0090] Let the original visual feasible set on the image plane be... Define bilateral opportunity constraints to ensure that feature points are... The probability is within the field of view boundary;

[0091] bilateral opportunity constraints The standard deviation of each component is calculated based on the diagonal elements of the predicted covariance matrix. By utilizing Gaussian distributions to distribute the confidence bounds, the probabilistic chance constraints are transformed into a deterministic tightening form:

[0092] ;

[0093] in, and For pixel boundaries, To predict the pixel state in name, for Standard deviation of corresponding components is the confidence level coefficient.

[0094] The above-mentioned tightened boundaries constitute the perturbation envelope set. Calculate the original visual feasible set With perturbation envelope set Minkowski difference, dynamically generating adaptive vision constraint sets This allows the feasible region to shrink automatically when the prediction variance increases and the control degrees of freedom to be restored when the variance decreases, thus ensuring system security.

[0095] Furthermore, the process of solving the Gaussian process model to predict control commands includes:

[0096] The nominal image Jacobian matrix and the Gaussian process posterior prediction mean are locally linearized in the current prediction time domain to construct a local affine prediction model.

[0097] A quadratic cost function is constructed using the state error prediction sequence, control input sequence, and terminal error as penalty terms.

[0098] The pixel state of each prediction step is constrained to be within the adaptive field constraint set. The control input sequence that minimizes the cost function is solved, and the first element of the control input sequence is taken as the expected body velocity of the kinematic layer at the current moment.

[0099] Furthermore, in this embodiment, the prediction control command of the Gaussian process model is solved by locally linearizing the nominal image Jacobian matrix and the Gaussian process posterior prediction mean in the current prediction time domain to construct a local affine prediction model.

[0100] A quadratic cost function is constructed using the state error prediction sequence, control input sequence, and terminal error as penalty terms.

[0101] In the finite-time optimization problem, a modified image kinematics prediction model is used, and the pixel state of each prediction step is strictly constrained to be within the generated adaptive field-of-view constraint set. Inside;

[0102] The system finds the control input sequence that minimizes the cost function online, and uses the first element of this sequence as the expected body velocity of the kinematic layer at the current moment that satisfies the feature point visibility constraint. .

[0103] Specifically, at each sampling time, the nominal image Jacobian matrix and the posterior prediction mean are frozen at the current working point to construct a local affine prediction model.

[0104] Define the prediction time domain as Construct a quadratic cost function that includes state error, control input, and terminal error: ;

[0105] in, These are positive definite weighted matrices.

[0106] During the optimization process, the pixel state of each prediction step is strictly constrained to satisfy the following conditions: Solve this finite-time optimization problem online, taking the first element of the optimal control sequence as the desired body velocity output of the kinematic layer. .

[0107] Furthermore, the process of using acceleration-level model-free adaptive control for dynamic tracking includes:

[0108] Differential filtering is performed on the desired body velocity at the kinematic layer to obtain the feedforward acceleration, and the target acceleration at the dynamic layer is calculated in combination with the current body velocity.

[0109] The pseudo-Jacobi matrix is ​​estimated online based on the system input and output data, and the control law is calculated based on the acceleration error to drive the underwater robot to track the desired body velocity.

[0110] Furthermore, the process of calculating the target acceleration includes:

[0111] The feedforward acceleration is obtained by constructing a smooth derivative of the visual desired velocity;

[0112] The target acceleration is calculated based on the feedforward acceleration, the current body velocity, and the desired body velocity. The expression for the target acceleration includes a gain matrix and transient change compensation coefficients.

[0113] Furthermore, the dynamic tracking based on acceleration-level model-free adaptive control in this embodiment includes:

[0114] Considering the time-varying and large inertia characteristics of underwater robot dynamic parameters, the desired output body velocity is... Differential filtering is performed to obtain smooth visual feedforward acceleration. ;

[0115] Combined with the current speed of the machine Calculate the target acceleration of the dynamic layer:

[0116] ;

[0117] in, It is a positive definite diagonal gain matrix. This is the transient change compensation coefficient;

[0118] Define an unknown pseudo-Jacobi matrix Assume that it satisfies an equivalent linear mapping between the increment of body acceleration and the increment of thruster control input:

[0119] ;

[0120] Based on the system input and output data, an online update law for the pseudo-Jacobi matrix is ​​designed to obtain the estimated value. And based on acceleration error Calculate the control law: ;

[0121] in, To control the step size, As the regularization factor, output As the actual thrust and torque control quantity of the underwater robot, the drive thruster achieves high-precision and smooth tracking of the visually desired body speed; This is the acceleration error.

[0122] More specifically, dynamic tracking based on acceleration-level model-free adaptive control is difficult to accurately obtain due to the nonlinear hydrodynamic effects of added mass, Coriolis force, and fluid damping on underwater robots. This embodiment employs acceleration-level model-free adaptive control to achieve tracking... Tracking:

[0123] 1) Introduce filter coefficients Construct the smooth derivative of the desired velocity as the feedforward acceleration:

[0124] ;

[0125] 2) Calculate the target acceleration of the dynamic layer: ;

[0126] in Here is the gain matrix. This is the transient change compensation coefficient;

[0127] 3) Assume that an equivalent linear increment relationship exists within the sampling period: Based on the concept of compact model-free adaptive control, the update law of the pseudo-Jacobi matrix is ​​designed as follows: ,in To adaptively update parameters.

[0128] 4) Calculate the actual control inputs for the underwater robot's thrusters: in To control the step size, As a regularization factor, This is the acceleration error.

[0129] This embodiment uses acceleration-level model-free adaptive control to smoothly map the outer loop velocity command to the thruster input, which can improve speed tracking accuracy and reduce chattering under conditions of time-varying hydrodynamic parameters and significant external disturbances, thereby enhancing the system robustness.

[0130] Example 2

[0131] Based on the same inventive concept, this embodiment also provides a visual dynamic hovering control system, including:

[0132] The error extraction module is used to obtain visual feature errors;

[0133] The kinematic prediction update module is used to construct an image kinematic prediction model based on Gaussian processes, output the posterior prediction mean and posterior prediction variance, and obtain the corrected image kinematic prediction model.

[0134] The adaptive vision constraint module is used to construct an adaptive vision constraint set based on the posterior prediction variance.

[0135] The velocity planning module is used to combine the modified image kinematics prediction model with the adaptive field constraint set to solve the Gaussian process model to predict control commands and generate the expected body velocity at the kinematic layer.

[0136] The adaptive dynamics tracking module is used to perform dynamics tracking using acceleration-level model-free adaptive control based on the desired body velocity at the kinematic level.

[0137] The visual dynamic hovering control system provided in this embodiment has all the advantages of the visual dynamic hovering control method provided in Embodiment 1.

[0138] Example 3

[0139] This embodiment also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of Embodiment 1.

[0140] Example 4

[0141] This embodiment also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method of Embodiment 1.

[0142] Example 5

[0143] This embodiment also discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the method of Embodiment 1.

[0144] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A visual dynamic hovering control method, characterized in that, include: Based on the obtained visual feature errors, an image kinematic prediction model based on Gaussian processes is constructed to obtain the corrected prediction model; Based on the posterior prediction variance output by the modified prediction model, an adaptive vision constraint set is constructed. Based on the modified prediction model and the adaptive field constraint set, the Gaussian process model predicts the control command to obtain the expected body velocity at the kinematic layer. Based on the desired body velocity of the kinematic layer, dynamic tracking is performed using acceleration-level model-free adaptive control. The process of constructing an image kinematic prediction model based on Gaussian processes includes: Establish a nominal image Jacobian matrix that includes the nominal image depth, and model the image depth approximation error, camera mounting bias, and unmodeled dynamics into a total uncertainty residual. The total uncertainty residual is learned online using a Gaussian process, and the posterior prediction mean and posterior prediction variance are output in the current state. By combining the nominal image Jacobian matrix and the posterior prediction mean, the corrected image kinematic prediction model is obtained. The process of learning the residuals online using Gaussian processes includes: The Gaussian process input vector is defined as a combination of the image feature stacking vector, the kinematic layer velocity input, and the nominal image depth vector. Residual observation labels are defined using a one-step forward differencing method; A scalar Gaussian process model is established for each pixel velocity residual component, and the stacked posterior prediction mean vector and the diagonal covariance matrix composed of the variances of each independent output channel are output. The process of dynamic tracking using acceleration-level model-free adaptive control includes: Differential filtering is performed on the desired body velocity of the kinematic layer to obtain the feedforward acceleration, and the target acceleration of the dynamic layer is calculated in combination with the current body velocity; The pseudo-Jacobi matrix is ​​estimated online based on the system input and output data, and the control law is calculated based on the acceleration error to drive the underwater robot to track the desired body velocity.

2. The method according to claim 1, characterized in that, The process of constructing an adaptive view constraint set includes: Extract the posterior prediction variance of the Gaussian process output, quantify the uncertainty level of the current environment and the model based on the posterior prediction variance, and calculate the perturbation envelope set; The original visual feasible set of the image plane is dynamically shrunk using the perturbation envelope set to generate an adaptive field-of-view constraint set.

3. The method according to claim 2, characterized in that, The process of generating an adaptive view constraint set also includes: Define a two-sided opportunity constraint to ensure that feature points are within the field of view boundary with a preset high confidence level probability; Transforming probabilistic chance constraints into a deterministic tightening form yields the tightening boundary; Based on the tightened boundary, a perturbation envelope set is constructed, and the Minkowski difference between the original visual feasible set and the perturbation envelope set is calculated to obtain the adaptive visual constraint set.

4. The method according to claim 1, characterized in that, The process of solving the Gaussian process model to predict control commands includes: The nominal image Jacobian matrix and the Gaussian process posterior prediction mean are locally linearized in the current prediction time domain to construct a local affine prediction model. A quadratic cost function is constructed using the state error prediction sequence, control input sequence, and terminal error as penalty terms. The pixel state of each prediction step is constrained to be within the adaptive field constraint set. The control input sequence that minimizes the cost function is solved, and the first element of the control input sequence is taken as the expected body velocity of the kinematic layer at the current moment.

5. The method according to claim 1, characterized in that, The process of calculating the target acceleration includes: The feedforward acceleration is obtained by constructing a smooth derivative of the visual desired velocity; The target acceleration is calculated based on the feedforward acceleration, the current body velocity, and the desired body velocity, wherein the expression for the target acceleration includes a gain matrix and a transient change compensation coefficient.

6. A visual dynamic hovering control system, characterized in that, include: The error extraction module is used to obtain visual feature errors; The kinematic prediction update module is used to construct an image kinematic prediction model based on Gaussian processes, output the posterior prediction mean and posterior prediction variance, and obtain the corrected image kinematic prediction model. The adaptive vision constraint module is used to construct an adaptive vision constraint set based on the posterior prediction variance. The velocity planning module is used to combine the modified image kinematics prediction model with the adaptive field constraint set to solve the Gaussian process model predictive control commands and generate the expected body velocity at the kinematic layer. An adaptive dynamics tracking module is used to perform dynamics tracking using acceleration-level model-free adaptive control based on the desired body velocity of the kinematic layer.

Images