Airborne camera anti-shake control method and system

Abstract

The invention discloses an airborne camera anti-shake control method and system, and relates to the field of camera anti-shake control, and the method comprises the steps: building a hysteretic deformation tensor field, generating a visual hysteretic residual tensor through differential operation and lagging feature analysis, taking the visual hysteretic residual tensor as observation input, and carrying out the state estimation through a LuGre dynamic friction model. Calculating a micro bristle deformation state and an equivalent friction torque value of a blocking holder, performing visual compensation analysis based on the equivalent friction torque value of the blocking holder, judging whether the system is in a static friction zero-crossing critical region or not based on a micro bristle deformation estimated value, generating a reverse feed-forward compensation torque instruction according to the generalized all-source anti-interference torque, and determining whether the system is in a static friction zero-crossing critical region or not. The integral term of the controller is dynamically reset, vector synthesis is carried out on a reverse feed-forward compensation torque instruction and a basic instruction of a main image stabilization control loop, final driving voltage is generated and applied to a pan-tilt motor, and stepped picture lagging and jumping caused by instant release of energy in a long-focus view field are restrained.

Description

Technical Field

[0001] This invention relates to the field of camera image stabilization control, specifically to an airborne camera image stabilization control method and system. Background Technology

[0002] When airborne optoelectronic pods perform long-range target tracking tasks, especially under long-range, high-magnification zoom conditions, the airborne platform typically experiences slow and continuous attitude drift. Existing gimbal servo control systems largely rely on motor encoders for position closed-loop control. However, the inherent gear backlash and bearing friction torque in precision reduction mechanisms create nonlinear mechanical dead zones. When the gimbal is in extremely low-speed motion or zero-crossing reversal phases, due to the adsorption effect of static friction, although the motor rotor produces a slight rotation, it does not effectively drive the load across the gear backlash, resulting in the camera's optical axis not actually following the command. At this time, the traditional PID controller, unable to eliminate position errors, continuously accumulates integral terms until the integral output torque exceeds the maximum static friction threshold, causing the system to release energy instantaneously, resulting in sudden slippage and overshoot of the gimbal load. This nonlinear dynamic behavior, characterized by alternating viscosity and slippage, manifests as stepped stutters and jumps in the image as the target moves within a long-range field of view, severely disrupting the continuity and stability of visual tracking.

[0003] Furthermore, existing image stabilization control schemes typically treat visual image stabilization and motor servo as two independent, decoupled components, lacking the ability to perceive and compensate for the microscopic physical state of the friction contact surface. Traditional Coulomb friction compensation models only consider macroscopic sliding friction, ignoring the microscopic hair-like elastic deformation characteristics of the friction contact surface during the pre-sliding stage. This results in the controller being unable to provide accurate feedforward torque to counteract nonlinear resistance during the static friction adsorption stage. Simultaneously, electronic image stabilization algorithms that rely solely on image processing suffer from frame rate lag, failing to correct the physical motion of the mechanical system in real time. Traditional motor torque loops, on the other hand, cannot directly perceive pixel-level visual hysteresis residuals. This asynchrony between visual feedback and mechanical execution means that the system cannot fundamentally suppress the visual axis tremors and image blur caused by the nonlinear characteristics of the mechanical transmission chain when facing low-speed crawling or minute oscillations due to the Stribeck effect.

[0004] To address the aforementioned shortcomings, a technical solution is provided. Summary of the Invention

[0005] To address the technical problems mentioned in the background section, this invention is proposed. Embodiments of this invention provide an airborne camera image stabilization control method and system.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] The airborne camera image stabilization control method includes the following steps:

[0008] The system synchronously collects inertial attitude data of the airborne platform, timing data of the gimbal motor encoder, and lens zoom position information, and calculates the camera intrinsic parameter matrix and instantaneous rotational attitude quaternion.

[0009] Real-time video streams are acquired, and ideal rigid body motion projections are calculated using the camera intrinsic parameter matrix and instantaneous rotational attitude quaternions. A hysteretic deformation tensor field is constructed, and a visual hysteretic residual tensor is generated based on the hysteretic deformation tensor field through difference operations and hysteresis feature analysis.

[0010] Using the visual hysteresis residual tensor as the observation input, the state is estimated through the LuGre dynamic friction model, the micro-hair deformation state and the equivalent friction torque value of the gimbal are calculated, and visual compensation analysis is performed based on the equivalent friction torque value of the gimbal to obtain the generalized full-source anti-disturbance torque.

[0011] Based on the estimated value of microscopic bristle deformation, it is determined whether the system is in the static friction zero-crossing critical region. Based on the generalized full-source disturbance rejection torque, a reverse feedforward compensation torque command is generated to dynamically reset the integral term of the controller.

[0012] The reverse feedforward compensation torque command and the basic command of the main image stabilization control loop are vector-synthesized to generate the final drive voltage applied to the gimbal motor.

[0013] Furthermore, the analysis steps for generating the visual hysteresis residual tensor are as follows:

[0014] By combining spherical projection geometry and mechanical dead zone determination, a hysteretic deformation tensor field for gimbal motion is constructed.

[0015] Using the hysteresis deformation tensor field as the background motion reference, a pixel-by-pixel difference operation is performed with the measured dense optical flow field of the current frame to solve the original residual vector set;

[0016] The original residual vector set is subjected to temporal sliding window integration to extract the viscous-slip hysteresis loop area features and construct the visual hysteresis residual tensor.

[0017] Furthermore, the analytical steps for constructing the hysteresis deformation tensor field of the gimbal motion are as follows:

[0018] Based on the camera intrinsic parameter matrix, a unit physical sphere space is constructed. The discrete pixel array of the image plane is projected backward onto the sphere using the inverse perspective projection transformation to generate a rigid line-of-sight beam cluster.

[0019] Using the instantaneous rotational attitude quaternion to drive the virtual active inner spherical shell rotation within the unit physical spherical space, the ideal tangential arc path of the rigid line-of-sight beam cluster as it moves along the spherical tangential plane is calculated;

[0020] Furthermore, the analytical steps for constructing the hysteresis deformation tensor field of the gimbal motion also include:

[0021] A mechanical dead zone cone representing gear clearance is constructed at the beam endpoint. The ideal tangential arc path and the bottom surface of the mechanical dead zone cone are included and cross-boundary. The effective drag component that exceeds the boundary is extracted to generate a physical response displacement vector.

[0022] The physical response displacement vector is remapped back to the two-dimensional image plane using perspective projection forward transformation, and a hysteretic deformation tensor field is generated through spatial interpolation.

[0023] Furthermore, the analysis steps for the equivalent frictional torque value of the obstructing gimbal are as follows:

[0024] Establish a vision-mechanical mapping equation to transform the vision hysteresis residual tensor into virtual relative angular velocity observations;

[0025] An unscented Kalman filter is constructed, and the virtual relative angular velocity observation value is used as the measurement update signal to iteratively calculate the estimated value of micro-hair deformation.

[0026] Furthermore, the analysis step of the equivalent frictional torque value of the obstructing gimbal also includes:

[0027] The derivative of the bristle deformation is obtained by analytical calculation based on the equation of state, using the estimated values ​​of Coulomb friction torque and microscopic bristle deformation.

[0028] Based on the estimated value of microscopic bristle deformation, combined with the derivative of bristle deformation and the preset Coulomb friction coefficient and viscous friction coefficient, the equivalent frictional torque value that hinders the gimbal is calculated.

[0029] Furthermore, the analysis steps for the generalized all-source disturbance rejection moment are as follows:

[0030] A visual-mechanical virtual impedance mapping model is constructed, and the original residual vector set is input into the impedance mapping model to calculate the visual compensation torque.

[0031] A frequency-domain complementary fusion filter is constructed to perform high-pass filtering on the equivalent friction torque value of the obstructing gimbal and low-pass filtering on the visual compensation torque. The two are then vector-superimposed to generate a generalized full-source anti-disturbance torque.

[0032] Furthermore, the analysis steps for generating the reverse feedforward compensation torque command and dynamically resetting the integral term of the controller are as follows:

[0033] The estimated value of microscopic bristle deformation is monitored. When it does not reach the maximum static friction deformation threshold, the system is determined to be in the static friction adsorption stage.

[0034] During the static tribosorption stage, a feedforward compensation torque command with equal amplitude and opposite direction is generated by using the generalized full-source anti-disturbance torque.

[0035] During the static tribosorption stage, the integral separation logic is triggered synchronously. While outputting the feedforward compensation torque command, the accumulation of the integral term of the PID controller is paused.

[0036] Furthermore, the analysis steps for generating the final driving voltage applied to the gimbal motor are as follows:

[0037] Receive the basic output command output by the main image stabilization control loop after cascading operations of the position loop and the velocity loop;

[0038] The feedforward compensation torque command is superimposed on the current loop input terminal of the basic output command to form a torque feedforward enhancement command;

[0039] The torque feedforward enhancement command is converted into a three-phase bridge arm duty cycle signal by the space vector pulse width modulation module, thereby generating the final drive voltage.

[0040] Airborne camera image stabilization control system, including:

[0041] The acquisition and calculation module is used to synchronously acquire the inertial attitude data of the airborne platform, the timing data of the gimbal motor encoder, and the zoom position information of the lens, and to calculate the camera intrinsic parameter matrix and instantaneous rotational attitude quaternion.

[0042] The hysteresis residual module is used to acquire real-time video streams. It uses the camera intrinsic parameter matrix and instantaneous rotational attitude quaternion to solve the ideal rigid body motion projection, constructs a hysteresis deformation tensor field, and generates a visual hysteresis residual tensor based on the hysteresis deformation tensor field through difference operations and hysteresis feature analysis.

[0043] The torque analysis module is used to take the visual hysteresis residual tensor as the observation input, perform state estimation through the LuGre dynamic friction model, solve the micro-hair deformation state and the equivalent friction torque value of the gimbal, perform visual compensation analysis based on the equivalent friction torque value of the gimbal, and obtain the generalized full-source anti-disturbance torque.

[0044] The critical compensation module is used to determine whether the system is in the static friction zero-crossing critical region by estimating the micro-hair deformation value. It generates a reverse feedforward compensation torque command based on the generalized full-source disturbance rejection torque to dynamically reset the integral term of the controller.

[0045] The drive voltage module is used to vector synthesize the reverse feedforward compensation torque command and the basic command of the main image stabilization control loop to generate the final drive voltage applied to the gimbal motor.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] This invention utilizes the camera's intrinsic parameter matrix and instantaneous rotational attitude quaternions to calculate the ideal rigid body motion projection, constructing a hysteretic deformation tensor field. Based on this field, a visual hysteretic residual tensor is generated through difference operations and hysteresis characteristic analysis. Using this residual tensor as the observation input, a LuGre dynamic friction model is used for state estimation, calculating the microscopic mane deformation state and the equivalent frictional torque value hindering the gimbal. Visual compensation analysis is then performed based on this torque value to obtain the generalized full-source anti-disturbance torque. By constructing a unit physical sphere space and a mechanical dead zone conical model, the nonlinear backlash caused by gear backlash and bearing friction is geometrically and precisely quantified, effectively solving the problems of integral term accumulation and sudden overshoot caused by static friction adsorption during the extremely low speed or zero-crossing commutation phase of traditional PID controllers. Similar to existing technologies that rely on passive integration based on positional errors, this invention can actively identify whether the system is in the static friction adsorption stage based on the estimated value of microscopic hair deformation. By generating a reverse feedforward compensation torque in conjunction with integral separation logic, it can "use force to counteract force" to offset external disturbances before the motor rotor experiences macroscopic slippage. This mechanism fundamentally suppresses the stepped image stuttering and jumps caused by the instantaneous release of energy in the telephoto field of view, ensuring the smoothness and continuity of visual tracking in the critical state of mechanical dead zones.

[0048] This invention determines whether the system is in the static friction zero-crossing critical region based on the estimated value of microscopic hair deformation. It generates a reverse feedforward compensation torque command based on the generalized full-source disturbance rejection torque, dynamically resets the integral term of the controller, and vector-synthesizes the reverse feedforward compensation torque command with the basic command of the main image stabilization control loop to generate the final driving voltage applied to the gimbal motor. A direct mapping mechanism from the macroscopic visual hysteresis residual tensor to the microscopic LuGre dynamic friction model is established, overcoming the limitations of existing schemes where visual image stabilization and motor servo are decoupled and can only compensate for macroscopic Coulomb friction. Through a visual-mechanical virtual impedance model and a frequency-domain complementary fusion filter, the system can directly convert pixel-level visual hysteresis residuals into torque compensation signals. This overcomes the frame rate lag defect of pure image electronic image stabilization algorithms and compensates for the shortcoming of traditional torque loops in sensing the microscopic hair elastic deformation during the pre-sliding stage of the contact surface. This design enables the system to exhibit compliant impedance characteristics similar to biological muscles when faced with low-speed crawling or minute oscillations caused by the Stribeck effect, effectively eliminating visual axis micro-flutter caused by the nonlinear characteristics of the mechanical transmission chain, and achieving accurate image stabilization in high dynamic environments. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. The following drawings are not drawn to scale according to the actual size, but are intended to show the main idea of ​​the present invention.

[0050] Figure 1 A flowchart illustrating the method for controlling image stabilization of an airborne camera.

[0051] Figure 2 This is a system block diagram of an airborne camera image stabilization control system.

[0052] Figure 3 A schematic diagram of the construction of a mechanical dead zone cone based on a unit physical sphere space;

[0053] Figure 4 This is a schematic diagram for determining the dead zone on the tangent plane and extracting the physical response displacement vector. Detailed Implementation

[0054] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are also within the scope of protection of the present invention.

[0055] Example 1

[0056] like Figure 1 As shown, the airborne camera image stabilization control method includes the following steps:

[0057] Step S100: Synchronously collect the inertial attitude data of the airborne platform, the timing data of the gimbal motor encoder, and the zoom position information of the lens, and calculate the camera intrinsic parameter matrix and the instantaneous rotational attitude quaternion.

[0058] Using an FPGA or a high-real-time embedded processor as a synchronous trigger source, the IMU inertial measurement unit, gimbal motor encoder, and lens zoom potentiometer are simultaneously triggered at a set frequency to ensure that the three sets of data have a unified timestamp tag. The set frequency is, for example, 1kHz. The inertial attitude data includes the three-axis angular velocity and acceleration of the airborne platform, and the gimbal motor encoder timing data includes the real-time angle values ​​of each joint axis of the gimbal. A mapping lookup table between zoom position and focal length is established, and the physical focal length at the current moment is obtained by looking up the table based on the collected lens zoom position information. And combined with the pre-calibrated principal point coordinates Construct the camera intrinsic parameter matrix Its form is usually expressed as The matrix construction process is a common technique in computer vision and photogrammetry, and will not be elaborated upon in this manual. Its main function is to establish the geometric mapping relationship between the projection of three-dimensional spatial points in the camera coordinate system onto the two-dimensional image plane, providing a mathematical basis for subsequent line-of-sight back projection. Based on the principle of multibody dynamics kinematic chains, the inertial attitude data of the airborne platform and the joint angle data of the gimbal motor encoder are fused using forward kinematics. That is, through the chain multiplication of continuous transformation matrices from the body coordinate system to the gimbal coordinate system and then to the camera coordinate system, the instantaneous rotational attitude of the camera's optical center coordinate system relative to the geographic coordinate system is calculated, and this rotational attitude is represented as a unit quaternion. ,in Represents the scalar real part of a quaternion. The vector imaginary part of the quaternion is represented by these four components, which together describe the rotation state of the camera relative to the reference coordinate system at the current moment. This serves as the motion reference for driving the rotation of the virtual spherical shell in subsequent steps. The forward kinematics, coordinate system transformation process, and representation in the form of unit quaternions are existing technologies in the field of robot control, and will not be elaborated here.

[0059] Furthermore, the airborne camera image stabilization control method also includes the following steps:

[0060] Step S200: Acquire real-time video stream, use camera intrinsic parameter matrix and instantaneous rotational attitude quaternion to solve ideal rigid body motion projection, construct hysteretic deformation tensor field, and generate visual hysteretic residual tensor based on hysteretic deformation tensor field through difference operation and hysteresis feature analysis.

[0061] The construction of the visual hysteresis residual tensor in step S200 specifically includes:

[0062] Step S201: Combining spherical projection geometry and mechanical dead zone determination, construct the hysteresis deformation tensor field of the gimbal motion;

[0063] Step S2011: Construct a unit physical sphere space based on the camera intrinsic parameter matrix, and use the inverse perspective projection transformation to project the discrete pixel array of the image plane onto the sphere in the opposite direction to generate a rigid line-of-sight beam cluster.

[0064] With the optical center of the camera as the center of the sphere, and per unit length A virtual three-dimensional Euclidean space is constructed for the radius, resulting in the unit physical sphere space. Within this space, feature pixels or all pixels in each frame of the image acquired by the image sensor are selected and denoted as two-dimensional pixel coordinates. , For pixels. Using the camera intrinsic parameter matrix. inverse matrix For the two-dimensional pixel coordinates Perform the inverse perspective projection transformation, that is, through the formula Reconstruct two-dimensional points on the image plane into three-dimensional direction vectors in the camera coordinate system. The vector is then normalized so that its magnitude equals the radius of a unit physical sphere. The set of all normalized three-dimensional direction vectors forms a set of rays radiating outwards from the center of the sphere and fixed in position relative to the camera's own coordinate system, resulting in a rigid line-of-sight beam cluster. This beam cluster physically represents the original orientation of each object point within the camera's field of view relative to the optical center at the current moment.

[0065] Step S2011 is based on establishing an isomorphic space between visual imaging and mechanical motion. Camera imaging is essentially a perspective projection of a three-dimensional point onto a two-dimensional plane, while the mechanical motion of the gimbal (pitch, yaw) occurs in a three-dimensional spherical coordinate system. Directly using two-dimensional pixel coordinates for control will produce nonlinear errors due to field-of-view distortion. Therefore, the inverse transformation of the intrinsic parameter matrix is ​​used to restore the pixel to a rigid line-of-sight beam within a unit physical spherical space. The underlying logic is to treat the camera as a rigid body and transform each pixel into a three-dimensional vector with fixed angular constraints emanating from the optical center. This modeling method conforms to the principles of rigid body kinematics, making the visual information no longer flat data, but possessing a physical dimension that matches the rotation of the gimbal's mechanical axis, thus laying a rigorous geometric foundation for the subsequent accurate calculation of angular deviations in a unified spherical coordinate system.

[0066] Step S2012: Use the instantaneous rotational attitude quaternion to drive the virtual active inner spherical shell in the unit physical sphere space to rotate, and calculate the ideal tangential arc path that the rigid line-of-sight beam cluster moves along the spherical tangential plane;

[0067] A virtual spherical layer, called the virtual active inner spherical shell, is defined within the unit physical spherical space and is fixed to the camera's optical axis. This virtual active inner spherical shell is essentially a spherical manifold with a unit radius defined at the origin of three-dimensional Euclidean space. It serves as a moving reference frame, with its geometric center coinciding with the camera's optical center. Its spatial attitude is updated in real-time with the movement of the camera's optical axis, representing the angular position of the camera's field of view relative to the global coordinate system. The instantaneous rotational attitude quaternion is then used. Acting on the virtual active inner spherical shell, simulating a camera in a very short time The actual physical rotation within, specifically, for each vector in the rigid line-of-sight beam cluster. Calculate the new vector position after rotation using quaternion rotation operators. ,Right now On the surface of a unit physical sphere, local tangent planes are constructed with the intersection points of each vector with the sphere as tangent points. These local tangent planes are geometrically perpendicular to the initial vector. It is a linearized approximation of the sphere at the point of tangency, designed to simulate the geometric properties of the camera imaging sensor plane within a local infinitesimal element, and to transfer vectors from... arrive The rotational trajectory is projected onto the tangential plane. Geometrically, this projection trajectory is an arc with direction and curvature; physically, it represents the theoretically expected displacement path of image feature points under the influence of only changes in the camera's own attitude. Through this projection transformation onto the tangential plane, the geodesic motion on the three-dimensional sphere is transformed into a tangential displacement vector on the two-dimensional plane, thus establishing a mapping relationship between three-dimensional spatial rotation and two-dimensional imaging plane motion. This path is an ideal tangential arc path, reflecting the optical flow component introduced by pure rotational motion.

[0068] Step S2012 constructs a frictionless ideal kinematic model as a reference. The system uses quaternions calculated by a high-frequency IMU to drive a virtual spherical shell to perform a pure rotation at the mathematical level. This physically simulates the absolute rigidity response of the gimbal without any mechanical backlash, frictional resistance, or flexible deformation. The mathematical basis for introducing tangent plane projection lies in the concept of tangent space in Riemannian geometry, which linearizes the nonlinear geodesic motion on the sphere into Euclidean displacement on the tangent plane within a local infinitesimal element. This process establishes the theoretically perfect trajectory that image feature points should follow under pure attitude command drive, providing the unique theoretical truth value for subsequent identification and separation of actual trajectory deviations caused by mechanical defects.

[0069] Step S2013: Construct a mechanical dead zone cone representing gear clearance at the beam endpoint, determine whether the ideal tangential arc path and the bottom surface of the mechanical dead zone cone are included or exceed the boundary, extract the effective drag component that exceeds the boundary, and generate a physical response displacement vector.

[0070] like Figure 3 The figure shown is a schematic diagram illustrating the mapping relationship between the line-of-sight vector and the mechanical dead zone cone in a unit physical sphere space, according to an embodiment of this application. As shown, in order to map pixels on a two-dimensional image plane... Converted into control commands for physical space, the system constructs a system centered on a light core. Center and radius of the sphere The system defines a unit physical sphere space. Within this space, the transmission gaps of the mechanical structure are modeled as a mechanical dead zone cone emanating from the optical center. This cone represents the ineffective range of motion where the gimbal cannot generate effective displacement during reversal or startup. The line-of-sight vector is represented as a rigid line-of-sight beam emanating from the optical center; when the line of sight changes, the path swept across the sphere by this beam is the rotated vector. For computational convenience, a local tangent plane is defined at the edge of the dead zone cone, tangent to the cone at the point of tangency. On this tangent plane, the ideal motion trajectory of the gimbal is mapped as an ideal tangential arc path. This geometric modeling transforms complex nonlinear spherical motion into vector operations on a tangent plane, clarifying the spatial relationships between image pixels, the line-of-sight vector, and the mechanical dead zone boundary.

[0071] like Figure 4 The figure shows a schematic diagram of dead zone filtering and physical response vector calculation based on a tangent plane according to an embodiment of this application. As shown in the figure, in the above-mentioned local tangent plane coordinate system, a radius of is defined with the tangent point as the origin. The circular region is designated as the mechanical dead zone (invalid motion range). This region corresponds to the physical clearance in the gimbal gears or transmission mechanism. The system calculates the position of the ideal projection point based on the target motion command. During this process, the system executes the following judgment and calculation logic, for amplitude... smaller than the dead zone radius For minute motion commands, the system identifies them as minor jitters and filters them directly using the dead-zone characteristic, without generating a drive signal, thus avoiding oscillations of the gimbal near its stationary position. For motion commands exceeding the dead-zone range, the system calculates the vector from the tangent point to the ideal projection point and subtracts the mechanical backlash, i.e., the dead-zone radius. The components of the physical response displacement vector are obtained by calculating the physical response displacement vector. This physical response displacement vector is the component that drives the actual movement of the gimbal. In this way, the system compensates for mechanical backlash while effectively suppressing minor noise, ensuring the accuracy of the gimbal's movement.

[0072] Based on a rigid line-of-sight beam cluster, select any vector within it. The intersection point with the unit physical sphere's spatial sphere is taken as the tangent point, i.e., the beam endpoint, and a local tangent plane coordinate system is established. At this tangent point, a coordinate system based on vectors is constructed. The radial extension of the cone is the central axis, and the apex is located at the center of the sphere, which is the optical center of the camera. The physical meaning of this cone is to simulate the backlash present in mechanical transmissions (such as gimbal gears). In mechanical transmissions like gimbal gears, small movements within the solid angle range of this cone will not drive the load; only movements exceeding this range are effective. The base of the cone is the cross-section between the cone and the local tangent plane, and its geometric shape is circular. The radius of the base of the cone... The calculation is based on the following: Let the preset mechanical gear clearance threshold angle be... Unit: radians, this angle It is a constant predetermined or set based on the mechanical transmission characteristics of the actual physical system. Specifically, it can be obtained by consulting the technical specifications of the reducer or gearbox used, or by calibrating the mechanical transmission chain through a reverse hysteresis test. The technical specifications typically specify the backlash or hysteresis value in arcminutes or degrees, which characterizes the maximum idle angle range where the input shaft rotates but the output shaft has not yet responded. Since the distance from the tangential plane to the center of the sphere is per unit length... According to the triangular geometric mapping relationship, the dead zone radius Here, the height of the cone is numerically equal to the radius of a unit physical sphere, that is, equal to... The Euclidean geometric boundary of the dead zone on the tangent plane was determined. This is the tangent function. It projects the ideal tangential arc path onto the tangent plane containing the base of the cone. Specifically, it rotates the new vector... The projection point is obtained by projecting it radially onto the tangent plane. Connect the tangent point and the projection point The vector formed is the total displacement vector of the ideal path. Calculate the vector Length of the module This modulus represents the total displacement distance that should theoretically occur on the tangent plane. Inclusion and boundary checks are performed: if the displacement modulus... Less than or equal to the dead zone radius If the system is deemed to have fallen into a dead zone, no physical displacement is generated; if the displacement magnitude is... Larger than the dead zone radius If it is found to be out of bounds, then the valid dragged component needs to be extracted. The calculation formula is:

[0073] ;

[0074] in Let be the unit vector in the direction of displacement. The physical meaning of this formula is to remove the idle distance consumed by the dead zone, i.e., subtract the radius . The invalid displacement vector is retained, only the effective portion driving the load motion is preserved. This effective drag component is... Defined as the physical response displacement vector, it represents the actual physical displacement that the virtual spherical shell should undergo after considering the nonlinear characteristics of mechanical clearance.

[0075] Step S2013 performs topological geometric modeling of the nonlinear characteristics of the transmission backlash. The backlash of the mechanical gear is physically represented as an ineffective angular sector centered on the shaft, which, when mapped to a unit sphere, forms a dead zone cone with its optical center as the vertex. The underlying logic of this step follows the "input-output hysteresis" law of mechanical transmission: only when the cumulative displacement at the input exceeds the gear backlash (i.e., exceeds the cone boundary) will the output (load) truly receive a driving force. Subtracting the dead zone radius component through vector operations essentially simulates the physical process of "waste being consumed" in the transmission chain at a mathematical level, ensuring that the extracted physical response displacement vector is a truly effective driving component that can be transmitted to the load end, thus achieving a precise digital reproduction of the mechanical backlash characteristics.

[0076] Step S2014: Remap the physical response displacement vector back to the two-dimensional image plane using perspective projection forward transformation, and generate a hysteretic deformation tensor field through spatial interpolation.

[0077] The perspective projection matrix in the graphics rendering pipeline is used to calculate the physical response displacement vector in 3D space. Perform a coordinate transformation to project the image from the camera coordinate system back to the two-dimensional screen image plane, thus obtaining the corresponding two-dimensional pixel displacement vector. This process essentially utilizes the camera's intrinsic parameter matrix. The process of restoring three-dimensional spatial geometric quantities to two-dimensional image plane pixel quantities is a common technique in computer vision and computer graphics, namely perspective projection from three-dimensional points to a two-dimensional plane, which will not be elaborated upon in this manual. Considering that the image is a discrete pixel array, it is not enough to only calculate the displacement at the beam endpoint; it is necessary to generate a continuous field covering the relevant area. Therefore, a spatial interpolation algorithm, such as radial basis function (RBF) interpolation or Gaussian weighted interpolation, is used to calculate the two-dimensional pixel displacement vector at the beam endpoint. Using the central control point, attenuation weights are calculated based on Euclidean distance. This displacement vector is then radially attenuated and diffused to surrounding pixels, thereby constructing a vector field describing the degree of local distortion in the image. The specific mathematical operation process of generating a dense vector field based on the control point using an interpolation function is existing technology in the field of image processing and will not be elaborated here. The generated vector field is a hysteretic deformation tensor field. The term hysteresis characterizes the nonlinear delay and backlash characteristics of the image response relative to the user's input due to the existence of the aforementioned dead zone cone. The deformation tensor describes the displacement distribution of image pixels in the horizontal and vertical directions. The generated hysteretic deformation tensor field will serve as input data for subsequent image processing steps, thereby presenting a visual feedback effect with a realistic sense of mechanical damping and gaps in the user's view.

[0078] Step S2014 visualizes the forward projection of mechanical errors. Previous steps calculated the physical displacement of the beam endpoints, but to interface with image processing algorithms, it's necessary to remap this displacement back to a two-dimensional pixel coordinate system using perspective projection principles. The introduction of spatial interpolation to generate a "tensor field" is based on the assumption of the continuity of the image medium; that is, the visual axis tremors caused by mechanical gaps lead to a non-uniform overall offset of pixels across the entire image. By constructing a hysteretic deformation tensor field, the system mathematically builds a predictive model: "If only mechanical dead zones exist, what kind of distortion should the current image exhibit?" This provides full-resolution predictive data for subsequently separating the influence of simple mechanical gaps from the complex measured optical flow.

[0079] Step S202: Using the hysteresis deformation tensor field as the background motion reference, perform pixel-by-pixel difference operation with the measured dense optical flow field of the current frame to solve the original residual vector set;

[0080] The hysteresis deformation tensor field, in its physical essence, represents the ideal mechanical response displacement that an image pixel should theoretically undergo when only considering the gimbal's mechanical dead zone effect (gear clearance). This displacement is defined as the background motion reference. The current frame of video image is processed using standard dense optical flow estimation algorithms, such as the Farneback algorithm or the TV-L1 optical flow method, to calculate the optical flow of each pixel in the image. The actual displacement vector between adjacent frames yields the measured dense optical flow field, denoted as . This measured value incorporates the combined effects of all physical factors, including mechanical dead zone, friction, and wind resistance; pixel-by-pixel vector difference calculation is performed, and the calculation formula is as follows: That is, subtracting the known dead zone space component from the measured mixed motion, and using vector subtraction based on the principle of motion superposition. It can effectively separate the motion components caused by geometric gaps, thereby decoupling the nonlinear residual components caused solely by physical friction, i.e., viscous-slip characteristics; it traverses all pixels on the image plane and calculates all the difference vectors. The collection, which describes the dynamic characteristics of friction across the entire image, is the original residual vector set.

[0081] Step S202 employs a signal decoupling strategy based on the principle of motion superposition. The actually observed optical flow field (total error) is a linear superposition of geometric errors caused by mechanical dead zones, dynamic hysteresis errors caused by friction, and external disturbance errors. Since step S2014 has already predicted the dead zone error component (background motion reference) based on an accurate geometric model, this known geometric deviation can be mathematically eliminated through vector difference operations. The physical essence of this operation is signal separation, filtering out the "hard" errors determined by the gear structure, thereby purifying the "soft" friction residuals caused only by the microscopic viscous-slip characteristics of the contact surface. This allows the subsequent control model to focus on compensating for complex dynamic friction forces, rather than being disturbed by fixed mechanical clearances.

[0082] Step S203: Perform temporal sliding window integration on the original residual vector set, extract the viscous-slip hysteresis loop area features, and construct the visual hysteresis residual tensor.

[0083] Construct a length of A time-dimension first-in-first-out sliding window queue, for example Frames are used to store the most recent The original residual vector set data at each time point is used to divide the current and past data within a sliding window. The original residual vector at each time step This is represented as a residual velocity sequence, which is then accumulated in the time domain (discrete integration) to obtain the corresponding residual displacement sequence. , where n is the discrete time index within the sliding window. Let n be the original residual vector at time n. Construct a hysteresis phase plane coordinate system with the residual displacement magnitude as the horizontal axis and the residual velocity magnitude as the vertical axis. The residual displacement magnitude is the Euclidean norm of the residual displacement vector at the current time, and the residual velocity magnitude is the velocity vector obtained by differentiating the residual displacement sequence and then taking the modulus of this velocity vector. Then, the values ​​within the sliding window... The displacement-velocity data is mapped onto this coordinate system to form a motion trajectory. If the beginning and end of the trajectory do not coincide, the trajectory is forcibly closed by a virtual straight line connecting the end point and the starting point, forming a closed viscous-slip hysteresis loop. The geometric area enclosed by this hysteresis loop is calculated using the Discrete Green's Theorem. According to the principle of hysteresis dynamics, this area Physically equivalent to the energy dissipation area of ​​a frictional hysteresis loop, its magnitude is directly proportional to the strength of the nonlinear viscous drag experienced at that location. The physical mechanism lies in the fact that in classical mechanics, mechanical hysteresis loss is represented by the area enclosed by the force-displacement curve. In the visual phase plane constructed in this embodiment, the residual velocity is a characterization of the gimbal's response lag due to frictional resistance, numerically positively correlated with the uncompensated nonlinear frictional force. The residual displacement characterizes the relative slip distance. The area integral over the velocity-displacement trajectory is isomorphic to the work integral of the force-displacement curve in terms of physical dimensions. This area... The larger the value, the more energy the gimbal dissipates during its movement in that local area due to overcoming frictional damping; in other words, the stronger the viscous drag. The calculated area feature values ​​are then processed by traversing the image plane. According to its original spatial pixels The structure is reorganized to construct a single-channel matrix with the same resolution as the original image. This matrix is ​​the visual hysteresis residual tensor, which quantifies the current micro-friction state distribution of the gimbal system and provides spatial measurement basis for the subsequent controller to perform targeted reverse torque compensation.

[0084] Furthermore, the airborne camera image stabilization control method also includes the following steps:

[0085] Step S300: Using the visual hysteresis residual tensor as the observation input, state estimation is performed through the LuGre dynamic friction model to solve the micro-hair deformation state and the equivalent friction torque value of the gimbal. Based on the equivalent friction torque value of the gimbal, visual compensation analysis is performed to obtain the generalized full-source anti-disturbance torque.

[0086] The calculation of the microscopic bristle deformation state and equivalent frictional torque value in step S300 specifically includes:

[0087] Step S301: Establish the vision-mechanics mapping equation and convert the vision hysteresis residual tensor into virtual relative angular velocity observations;

[0088] The mapping equation is established based on the visual hysteresis residual tensor. ,in, This is the tensor feature extraction function, specifically the Frobenius norm operator of the tensor, and its calculation formula is: The summation operator iterates through all pixel coordinates covered by the tensor. This operator aggregates the local frictional dissipation characteristics of a two-dimensional distribution into a single scalar value characterizing the global hysteresis intensity by calculating the square root of the sum of squares of the tensor elements. The pre-calibrated visual-angular velocity conversion coefficient has the following physical dimensions: It is responsible for mapping pixel area back to angle. , which is a unit of measurement for pixel area. The dimension is angular. The mapping equation is established based on the dissipative structure theory in classical mechanics, where the area enclosed by the hysteresis loop is... Physically, this corresponds precisely to the energy density dissipated by the system overcoming nonlinear frictional forces within one motion cycle. In classical theories such as the LuGre dynamic friction model, the frictional power dissipation is positively correlated with the microscopic relative sliding velocity between the contact surfaces; that is, the larger the apparent hysteresis area, the greater the velocity lag component caused by frictional resistance within the system. Therefore, this mapping equation is established. (Coefficients) The specific calibration steps are as follows: while the system is offline, control the gimbal to operate at multiple known constant angular velocities. Rotating at a constant speed, where the subscript Indicates the first The sequence number of each calibration experiment; the tensor norm value corresponding to each experiment is calculated simultaneously. Construct a data set The set is fitted using linear regression with the least squares method, and the slope of the fitted line is the coefficient. The result obtained by calculating using this equation This refers to the quantized virtual relative angular velocity observation value. It does not represent the physical rotational speed of the gimbal motor, but rather characterizes the velocity lag component caused by frictional resistance of the gimbal, serving as the input signal for subsequent observers.

[0089] Step S302: Construct an unscented Kalman filter, use the virtual relative angular velocity observation value as the measurement update signal, and iteratively calculate to obtain the estimated value of micro-hair deformation;

[0090] Based on the LuGre dynamic friction model theory, the microscopic physical properties of the gimbal's friction contact surface are modeled as the contact of countless elastic bristles, and the system state vector is defined. ,in express The average deformation deflection of the microscopic mane at any given time was determined, and a nonlinear state equation describing the dynamic characteristics of mane deformation was constructed. In the formula The angular velocity is the commanded value of the controller. The derivative of the system state vector. This is the Stribeck effect function, used to describe the nonlinear decay characteristics of frictional force as a function of velocity. Its functional form is as follows: In the formula Represents the relative sliding angular velocity of the contact surfaces, where for This parameter was obtained by controlling the gimbal to move at a constant speed during the offline phase and measuring the stable frictional torque. The maximum static friction force is obtained by measuring the peak drag torque of the gimbal at the instant it starts from a standstill. The Stribeck velocity is a parameter determined by least-squares curve fitting based on experimental data of frictional torque at different velocities; a measurement vector is defined. Establish measurement equations ,in The observation function describes the mapping relationship between microscopic deformation states and macroscopic visual hysteresis characteristics. Specifically, the observation function is constructed as a linear mapping. ,in This is the observation coefficient matrix, used to characterize the theoretical projection component of the macroscopic relative velocity generated by the microscopic mane deformation through the equivalent elastic stiffness. The value of this matrix was determined by applying a sinusoidal excitation signal to the gimbal during the offline stage and then scaling it up by comparing the width of the hysteresis curve observed visually with the theoretical deformation amplitude calculated by the model. To measure noise, time updates are performed during the filter iteration process using an unscented transform strategy, based on the posterior state estimate from the previous time step. and its covariance matrix Generate according to deterministic sampling rules Sigma sampling points, Given that generating Sigma point sets using unscented transformations is an existing technique, the specific sampling formula will not be elaborated here. These Sigma point sets are substituted into the aforementioned nonlinear state equations for nonlinear propagation, and weighted calculations are performed to obtain the prior state estimate for the current time step. and the prior covariance matrix This process effectively avoids the truncation error introduced by linearizing the nonlinear model; measurement updates are performed using the virtual relative angular velocity calculated in step S301. Assigned as the observed true value to the measurement vector at the current time. Based on the predicted prior state estimate Through the above linear observation function Predict the current measurement estimate And calculate the observation residuals. The residual reflects the deviation between the visual observations and the model predictions, and the Kalman gain matrix is ​​then calculated. Using the formula The prior estimate is corrected to obtain the optimal posterior estimate of the microscopic mane deformation state. (Right now This refers to the estimation value of microscopic hair deformation, which solves the problem that microscopic deformation cannot be directly measured and provides accurate state variables for friction compensation.

[0091] Step S302 is based on the LuGre dynamic friction model principle. The microscopic physical interaction between the friction contact surfaces can be equivalent to the random contact and deformation of countless tiny elastic bristles. The average deformation deflection of these bristles, i.e., the state variable z, directly determines the magnitude of the friction force and the hysteresis characteristics. The aforementioned nonlinear state equation is established based on the physical conservation relationship that the bristle deformation rate equals the relative sliding velocity minus the bristle rebound slip rate caused by sliding. However, the microscopic bristle deformation belongs to the hidden state inside the system and cannot be directly measured by sensors, but its dynamic changes will act on the gimbal through friction torque, manifesting as macroscopic motion speed lag. Therefore, the estimation logic of this embodiment is to use an unscented Kalman filter to process the highly nonlinearity of the model, use the virtual relative angular velocity observed by macroscopic vision as an external observation constraint, and continuously correct the model prediction deviation through Kalman gain, thereby mathematically deducing the invisible microscopic bristle deformation state in reverse, realizing an accurate mapping from macroscopic phenomena to microscopic essence.

[0092] Step S303: Obtain the derivative of the bristle deformation using an analytical calculation method based on the equation of state, based on the estimated values ​​of Coulomb friction torque and microscopic bristle deformation.

[0093] Based on the estimated value of microscopic bristle deformation, combined with the derivative of bristle deformation and the preset Coulomb friction coefficient and viscous friction coefficient, the equivalent frictional torque value that hinders the gimbal is calculated.

[0094] Obtain preset system inherent parameters, including bristle stiffness coefficient. bristle damping coefficient viscous friction coefficient and Coulomb friction torque The Coulomb friction coefficient includes the bristle stiffness coefficient. and bristle damping coefficient Regarding the coefficient of viscous friction Coulomb friction torque The system acquires and controls the gimbal to perform multiple sets of unidirectional uniform rotational motions with different speed gradients. In each set of uniform speed states, the system's driving torque is primarily used to overcome steady-state friction, at which point the dynamic term is zero. By recording the steady-state driving torque corresponding to each speed point, a velocity-torque mapping curve is constructed. The least squares method is used to fit the high-speed linear segment of this curve, and its slope represents the viscous friction coefficient. The intercept is the Coulomb friction torque. Regarding the bristle stiffness coefficient With bristle damping coefficient To obtain the data, the gimbal is controlled to perform a small-amplitude sinusoidal frequency sweep excitation motion near zero velocity, so that the friction is in the pre-slip region. At this time, the friction behavior is mainly dominated by the elastic deformation of the bristles. By collecting the input torque and corresponding micro-displacement and velocity data, a dynamic equation based on the hysteresis loop is established. Then, the data of the pre-slip stage are nonlinearly fitted using the least squares method or genetic algorithm to calculate the stiffness coefficient characterizing the micro-properties of the contact surface. With damping coefficient To accurately calculate the micro-damping component in the frictional torque, it is necessary to obtain the derivative of the bristle deformation. Instead of numerically differentiating the deformation variables, an analytical calculation method based on state equations is used. The specific process is as follows: Constructing the Stribeck function in the moment domain. Based on the current relative sliding angular velocity of the gimbal Using the formula Calculate the current Stribeck effect value. The maximum static friction torque characterizes the maximum resistance that the system must overcome at the instant it starts from a standstill. It is obtained by executing a series of uniform motion experiments covering a very low to medium-low speed range in the control system, collecting the steady-state driving torque at each velocity point to construct a Stribeck curve, focusing on extracting the peak torque data near zero velocity, and combining this with the identified Coulomb friction torque. The exponential decay term in the Stribeck model is fitted using the nonlinear least squares method to accurately calculate the maximum static friction torque. Here, a symbol is introduced. This function, distinct from the general function in step S302, is specifically designed to characterize the nonlinear decay properties at the torque level. The aim is to define the boundary of frictional torque at low speeds. The estimated value of microscopic bristle deformation output from step S302 is used. Current relative sliding angular velocity and the above calculations Substituting the differential equation of state into the LuGre model, the rate of change of mane deformation is calculated. The differential equation of state is: ;

[0095] Calculate using this formula Because The estimated value is obtained through Kalman filtering. Direct differential calculation would introduce significant quantization noise and amplify high-frequency interference, leading to torque output fluctuations. Solving using the regression model's differential equations ensures that the damping term calculation strictly adheres to the physical dynamics of the LuGre model, guaranteeing the smoothness and physical consistency of the friction torque calculation. The torque output equation using the LuGre dynamic friction model is shown below. Perform the calculation, where the first term The elastic friction component is the estimated value of the microscopic bristle deformation output from step S302. With stiffness coefficient Multiplying them together yields the elastic restoring force, which characterizes the microscopic deformation between the contact surfaces. The second term... The micro-damping component is determined by the rate of change of hair deformation. With damping coefficient Multiplying them together gives the energy dissipation during deformation, the third term. The macroscopic viscous component is determined by the current actual angular velocity of the gimbal. With viscosity coefficient Multiplying these terms yields a linear resistance characteristic of fluid lubrication. Adding these three terms together gives the equivalent frictional torque value that hinders the gimbal. This value precisely quantifies the total drag torque of the gimbal under the influence of nonlinear friction. Based on this, the controller generates a compensating torque of equal magnitude but opposite direction to eliminate motion lag and crawling caused by friction.

[0096] Step S304: Construct a visual-mechanical virtual impedance mapping model, input the original residual vector set into the impedance mapping model, and solve for the visual compensation torque;

[0097] The vision-mechanical virtual impedance mapping model is constructed based on the principle of mechanical impedance, simulating the pixel deviation during gimbal visual tracking as the response of a spring-damped system under virtual external force. Specifically, the model is a second-order dynamic equation: ,in, The visual compensation torque to be calculated. The velocity deviation vector is the value of the original residual vector in the original residual vector set, and its value is derived from the original residual vector calculated in step S202. This represents the difference between the target pixel's moving speed and the gimbal's rotation speed, i.e., the rate of change of the error at the current moment. This is the position deviation vector in the original residual vector set, and its value comes from the residual displacement sequence obtained by performing time-domain sliding window integration on the original residual vector in step S203. This represents the cumulative pixel distance between the center of the target in the image plane and the center of the field of view. The virtual stiffness coefficient is used to characterize the system's ability to resist static drag error. Its physical meaning is the proportional gain that converts pixel position deviation into restoring torque, similar to the stiffness coefficient of a virtual spring, used to eliminate steady-state error. The virtual damping coefficient characterizes the system's ability to suppress unmodeled dynamic disturbances. Physically, it converts pixel velocity deviations into a gain of damping torque, similar to a virtual damper, used to suppress high-frequency oscillations. The virtual stiffness coefficient... With virtual damping coefficient As inherent parameters of the system, these parameters are obtained and stored in the controller through pre-conducted visual-torque step response tests. The calibration process consists of two stages: one for ensuring the accuracy of the centering process. Adjustment, in settings Furthermore, when the gimbal is stationary, a step visual error can be artificially added during image processing, such as a fixed pixel offset, at a preset reference magnification. Gradually increase Real-time monitoring of the gimbal's response curve to eliminate the static error was performed, and the steady-state error was selected when the damping ratio converged to zero. At the boundary between underdamped and critically damped (i.e., the oscillation decay rate is greater than a preset threshold, for example...) The critical gain value at which the optimal value is obtained is taken as the critical gain value. Its physical meaning is the proportional gain that converts pixel deviation into restoring torque, characterizing the system's ability to resist static drag; it is responsible for dynamic stability. Adjust and maintain optimal performance. Keeping the external pulse disturbance constant, apply it to the gimbal to simulate a gust of wind, gradually increasing the intensity. And observe the overshoot during the visual residual convergence process. When the system response reaches the critical damping state, that is, the maximum overshoot is less than the preset allowable value (e.g. When the adjustment time is shortest, record the parameters at this point as the optimal value. Its physical meaning is the system's ability to suppress dynamic mutations. Through the above calibration based on objective response indicators, it is ensured that the model can accurately call the optimal parameters for torque calculation during real-time operation. The acquired original residual vector set data (i.e., the corresponding...) and Substituting into the above equation, the result is calculated in real time. The physical meaning of this torque is to eliminate the current visual tracking residuals. The gimbal motor needs to apply an additional virtual torque to overcome environmental disturbances and system steady-state errors that are not captured by the physical model, thereby achieving direct cross-domain compensation from pixel-level visual errors to mechanical torque.

[0098] The core theoretical basis for constructing the mapping model in step S304 lies in the concept of impedance control. Traditional gimbal control loops typically rely solely on motor encoders for position closed-loop control. However, when there is gear backlash, nonlinear friction, or structural deformation in the transmission mechanism, the accuracy of the encoder's physical reading does not equate to the actual stability of the lens's visual axis, resulting in a mechanical-visual asynchrony problem. A virtual physical entity system is constructed in visual space using the above equations, where the term... This is equivalent to connecting a virtual spring between the image center and the target point; when the image shifts ( When the value increases, an elastic restoring force proportional to the deviation is generated to overcome the static dragging error caused by mechanical clearance; This is equivalent to installing a virtual damper along the motion path, which is effective when the image experiences severe jitter or abrupt changes. When the value is large, a reverse viscous drag is generated to suppress unmodeled high-frequency dynamic interference. This model is used to mitigate pixel-level visual errors through... The parameters are converted into torque signals, allowing the vision algorithm to access the motor's torque loop control without going through a cumbersome position loop conversion. Its response speed is much faster than traditional position loop correction at the physical level, reducing control latency. Unlike traditional rigid control that forces the error to be cleared instantly, the virtual impedance model allows the system to have small dynamic errors when disturbed, exhibiting compliance similar to biological muscles. This effectively avoids high-frequency oscillations in the image caused by over-correction, making the visual experience smoother and more natural while maintaining stability.

[0099] Step S305: Construct a frequency domain complementary fusion filter, perform high-pass filtering on the equivalent friction torque value of the obstructing gimbal, perform low-pass filtering on the visual compensation torque, and vector superimpose the two to generate a generalized full-source anti-disturbance torque.

[0100] The core of constructing a frequency-domain complementary fusion filter lies in utilizing the complementary characteristics of signals in the frequency domain to solve the problem that a single model cannot simultaneously consider dynamic response and steady-state accuracy. The construction of the frequency-domain complementary fusion filter is based on the following physical principles and mathematical derivations: High-pass filtering is applied to the equivalent frictional torque hindering the gimbal. This is based on the fact that the physical friction model calculated in step S303 mainly reflects the nonlinear characteristics of the motor rotation instantaneously, such as the sudden change in static friction and the direction switching of Coulomb friction. These characteristics are mainly concentrated in the high-frequency band in the frequency domain. Furthermore, since the physical model parameters are difficult to be perfectly accurate, directly using its low-frequency components may introduce accumulated modeling errors. Therefore, a high-pass filter is needed to extract its rapidly changing dynamic friction components and filter out inaccurate low-frequency deviations. Low-pass filtering is applied to the visual compensation torque. This is based on the fact that the calculated visual impedance model is based on image processing and is limited by camera frame rate and image noise. Its signal has hysteresis and high-frequency jitter noise, but it can reflect the actual offset of the image with extremely high accuracy in the low-frequency band. Therefore, a low-pass filter is needed to filter out high-frequency visual noise and retain only its accurate low-frequency steady-state compensation components. A complementary cutoff frequency is set. For example, a value of 5Hz-15Hz is used, which is the dividing point between dynamic abrupt changes and steady-state drift, and is determined by the Laplace operator. , In control theory, this represents a complex frequency domain variable used to describe the system's differential or integral response characteristics to signals of different frequencies. The high-pass filter extraction process involves constructing the transfer function of a first-order high-pass filter. The function takes the equivalent frictional torque value that hinders the gimbal as input and outputs the high-frequency frictional torque component. In this formula, the molecule has This represents the differential action, ensuring that only rapidly changing signals can pass through, while steady-state signals are blocked; the low-pass filtering extraction process involves constructing the transfer function of a first-order low-pass filter. The visual compensation torque is input into this function, and the low-frequency visual compensation component is output. This formula utilizes the inertial element in the denominator to smooth the signal and filter out high-frequency noise. Vector superposition generates a generalized full-source disturbance rejection torque. The reason why it is possible to and Perform vector superposition, that is It is based on the principle of complementary sensitivity function, because This means that, theoretically, the full-frequency torque required for control is split into two parts, processed separately, and then recombined to restore the complete physical quantity. Since both components have been uniformly converted into the same physical quantity—torque—and act on the same gimbal motor shaft, they satisfy the principle of linear superposition in physical space, generating… It possesses both the physical model's millisecond-level rapid response capability to sudden disturbances, contributed by high-frequency components, and the visual feedback's absolute correction capability for the center of the image, contributed by low-frequency components, thus achieving perfect coverage of complex interference environments.

[0101] Furthermore, the airborne camera image stabilization control method also includes the following steps:

[0102] Step S400: Based on the estimated value of micro-hair deformation, determine whether the system is in the static friction zero-crossing critical region, generate a reverse feedforward compensation torque command according to the generalized full-source disturbance rejection torque, and dynamically reset the integral term of the controller;

[0103] The generation of feedforward compensation torque command and dynamic reset of integral term in step S400 specifically includes:

[0104] Step S401: Monitor the estimated value of microscopic bristle deformation. When it does not reach the maximum static friction deformation threshold, the system is determined to be in the static friction adsorption stage.

[0105] The estimated value of microscopic bristle deformation is an internal state variable defined based on the LuGre dynamic friction model. Physically, it represents the average elastic deformation of the microscopic elastic bristles between the contact surfaces after being subjected to force. In the gimbal rotation control scenario of this embodiment, its physical unit is specifically radians, characterizing the minute elastic torsional angle of the motor rotor relative to the stator before macroscopic rotation occurs. The maximum static friction deformation threshold is a pre-calibrated critical physical quantity. It is obtained during the offline identification phase by locking the gimbal axis and gradually applying torque until the moment of macroscopic slippage, recording the maximum elastic deformation at this point. Mathematically, this is usually approximately equal to the maximum static friction torque. Equivalent stiffness coefficient of mane The ratio, i.e. This threshold defines the physical boundary between elastic deformation and macroscopic sliding. Logically, the controller calculates the absolute value of the current estimated microscopic bristle deformation in real time. If this value is less than the maximum static friction deformation threshold, it indicates that although the microscopic bristles have deformed, they have not yet broken. The motor rotor is still held in place by static friction, and the system is in the static friction adsorption stage. At this time, the friction force mainly manifests as a nonlinear elastic restoring force that hinders the motion trend.

[0106] Step S402: In the static triboelectric adsorption stage, a feedforward compensation torque command with equal amplitude and opposite direction is generated using the generalized full-source anti-disturbance torque;

[0107] After confirming that the system is in the aforementioned static friction adsorption stage, the core of the control strategy lies in using force to counteract force to maintain equilibrium in a static or slightly moving state. The generalized full-source disturbance rejection torque encompasses a comprehensive physical quantity that includes both high-frequency frictional dynamic abrupt change components and low-frequency visual steady-state deviation components, fully reflecting the total external disturbance experienced by the system at the current moment. Generating a constant-amplitude, reverse-direction feedforward compensation torque command means that the controller reads the current... The numerical value is then inverted to generate a torque command of equal magnitude but opposite direction. This is a feedforward compensation torque command. Instead of passing through the conventional error feedback adjustment loop, this command is superimposed as a feedforward quantity onto the input of the motor current controller. Its physical significance lies in actively outputting a torque that completely cancels out the estimated external disturbance within the tiny displacement range before static friction is overcome. Like a spring, it cancels out external interference forces, thereby locking the motor shaft at a microscopic level, eliminating the dead zone effect and start-up lag caused by static friction, and ensuring high rigidity of the system under extremely low or zero-speed conditions.

[0108] Step S403: During the static friction adsorption stage, the integral separation logic is triggered synchronously. While outputting the feedforward compensation torque command, the integral term accumulation of the PID controller is paused.

[0109] Step S403 pauses the accumulation of the integral term in the PID controller to prevent integral saturation overshoot caused by static friction, addressing the inherent defects of traditional PID control under static friction conditions. During the static friction adsorption stage, the motor shaft is attracted by static friction and cannot immediately generate displacement. Position error persists and cannot be eliminated by small outputs; this position error is the difference between the set position and the actual position. If the integral term (I term) of the PID controller continues to operate at this time, it will continuously accumulate this uneliminable error, causing the integral output value to increase rapidly, entering integral saturation. Once the friction torque is exceeded, entering the dynamic friction stage, the excessive integral output will be released instantaneously, causing the motor to accelerate violently and overshoot the target position, resulting in severe overshoot or oscillation. Therefore, integral separation logic is implemented. Specifically, when the judgment flag indicates the static friction adsorption stage, the control algorithm forcibly cuts off the input of the integral channel in the PID controller or sets the integral gain to zero at the software level, keeping the value of the integral register unchanged or clearing it. Control is then performed solely based on the feedforward compensation torque in step S402 and the proportional (P term) and derivative (D term) terms in the PID controller. This approach ensures that when the system breaks through static friction and enters macroscopic motion, the controller is in a clean state, enabling smooth and overshoot-free transition to subsequent motion control.

[0110] Furthermore, the airborne camera image stabilization control method also includes the following steps:

[0111] Step S500: The reverse feedforward compensation torque command and the basic command of the main image stabilization control loop are vector-synthesized to generate the final drive voltage applied to the gimbal motor.

[0112] Step S501: Receive the basic output command output by the main image stabilization control loop after cascading operation of the position loop and the velocity loop;

[0113] The main image stabilization control loop refers to the core closed-loop control structure in the gimbal control system responsible for maintaining line-of-sight stability. It typically employs a cascaded PID control architecture, consisting of an outer position loop and an inner velocity loop. Specifically, the position loop calculates the target angular velocity command based on the deviation between the gimbal's current attitude angle feedback and the target attitude angle. The velocity loop then compares this target angular velocity command with the real-time angular velocity feedback from the gyroscope. After processing by a proportional-integral regulator, it outputs a basic output command for driving the motors. This basic output command is usually physically represented as a current command value, such as the q-axis reference current. Its value directly corresponds to the electromagnetic torque that the motor needs to generate to overcome the conventional load and drive the gimbal to move. However, at this time, the instruction does not yet include a special compensation component for complex nonlinear friction. The specific control algorithm and model derivation for obtaining the basic output instruction by using the cascade operation of the position loop and the speed loop are existing technologies in the field of motor servo control, and will not be elaborated here.

[0114] Step S502: Superimpose the feedforward compensation torque command onto the current loop input terminal of the basic output command to form a torque feedforward enhancement command;

[0115] Standardized processing of execution units, utilizing motor torque constant. feedforward compensation torque command Converted to a compensation current component in amperes. The motor torque constant. This is a physical quantity characterizing the linear proportional relationship between motor current and output torque. Its value is usually obtained directly from the technical specifications provided by the motor manufacturer, or precisely measured through offline parameter identification experiments during system initialization. The specific conversion calculation involves converting the feedforward compensation torque command... Divide by the motor torque constant The system performs a superposition operation at the node after the speed loop output and before the current loop input, performing an algebraic summation operation on the basic output command and the compensation current component, i.e. The torque feedforward enhancement command is obtained. , This is the basic output command, the reference current output by the speed loop. The superposition operation is based on the physical superposition of torque and the feedforward control principle. Specifically, the total electromagnetic torque output by the motor is physically equal to the product of the current and the torque constant. This torque needs to simultaneously meet two requirements: one is the inertial torque required to drive the load to track its motion trajectory, which is... The other part is the resistance torque required to overcome the current nonlinear friction of the contact surface, provided by... The corresponding current component is provided. By superimposing the two at the current command level, the system is equivalent to pre-injecting a current component to counteract friction into the current loop setpoint before the frictional force causes a speed deviation, thereby improving the dynamic response speed and stability of the system during commutation and low-speed operation.

[0116] Step S503: The torque feedforward enhancement command is converted into a three-phase bridge arm duty cycle signal through the space vector pulse width modulation module to generate the final drive voltage.

[0117] The Space Vector Pulse Width Modulation (SVPWM) module is the core execution unit of the motor driver. Its function is to convert commands in the vector control domain into switching actions of the physical circuit. Specifically, based on the Parker inverse transform and Clark inverse transform in the field-oriented control algorithm, combined with the real-time electrical angle of the motor rotor, it converts the DC-side torque feedforward enhancement command into a voltage vector in a two-phase stationary coordinate system. The SVPWM algorithm is based on Based on the sector location, the conduction time of the six power switches on the three-phase bridge arm of the inverter is calculated, i.e., the duty cycle signal of the three-phase bridge arm. The power switches are such as MOSFETs or IGBTs. The drive circuit performs high-frequency switching control of the inverter based on these duty cycle signals, synthesizing an equivalent sinusoidal drive voltage on the motor stator winding, which is the final drive voltage. This generates an electromagnetic torque that includes precise friction compensation, driving the gimbal to achieve high-precision and stable movement. The process of converting the current command into the motor drive voltage using current loop and space vector pulse width modulation technology belongs to the existing technology in the field of motor control, and the specific underlying circuit structure or conventional algorithm formulas will not be elaborated here.

[0118] Example 2

[0119] like Figure 2 As shown, the airborne camera image stabilization control system includes:

[0120] The acquisition and calculation module is used to synchronously acquire the inertial attitude data of the airborne platform, the timing data of the gimbal motor encoder, and the zoom position information of the lens, and to calculate the camera intrinsic parameter matrix and instantaneous rotational attitude quaternion.

[0121] The hysteresis residual module is used to acquire real-time video streams. It uses the camera intrinsic parameter matrix and instantaneous rotational attitude quaternion to solve the ideal rigid body motion projection, constructs the hysteresis deformation tensor field, and generates the visual hysteresis residual tensor based on the hysteresis deformation tensor field through difference operations and hysteresis feature analysis.

[0122] The torque analysis module uses the visual hysteresis residual tensor as the observation input, performs state estimation through the LuGre dynamic friction model, solves the micro-hair deformation state and the equivalent friction torque value of the gimbal, performs visual compensation analysis based on the equivalent friction torque value of the gimbal, and obtains the generalized full-source anti-disturbance torque.

[0123] The critical compensation module is used to determine whether the system is in the static friction zero-crossing critical region by estimating the micro-hair deformation value. It generates a reverse feedforward compensation torque command based on the generalized full-source disturbance rejection torque to dynamically reset the integral term of the controller.

[0124] The drive voltage module performs vector synthesis of the reverse feedforward compensation torque command and the basic command of the main image stabilization control loop to generate the final drive voltage applied to the gimbal motor.

[0125] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0126] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0127] The foregoing description is illustrative of the invention and should not be construed as limiting it. Although several exemplary embodiments of the invention have been described, those skilled in the art will readily understand that many modifications can be made to the exemplary embodiments without departing from the novel teachings and advantages of the invention. Therefore, all such modifications are intended to be included within the scope of the invention as defined in the claims. It should be understood that the foregoing description is illustrative of the invention and should not be construed as limiting it to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The invention is defined by the claims and their equivalents.

Claims

1. An airborne camera image stabilization control method, characterized in that, Includes the following steps: The system synchronously collects inertial attitude data of the airborne platform, timing data of the gimbal motor encoder, and lens zoom position information, and calculates the camera intrinsic parameter matrix and instantaneous rotational attitude quaternion. Real-time video streams are acquired, and ideal rigid body motion projections are calculated using the camera intrinsic parameter matrix and instantaneous rotational attitude quaternions. A hysteretic deformation tensor field is constructed, and a visual hysteretic residual tensor is generated based on the hysteretic deformation tensor field through difference operations and hysteresis feature analysis. Using the visual hysteresis residual tensor as the observation input, the state is estimated through the LuGre dynamic friction model, the micro-hair deformation state and the equivalent friction torque value of the gimbal are calculated, and visual compensation analysis is performed based on the equivalent friction torque value of the gimbal to obtain the generalized full-source anti-disturbance torque. Based on the estimated value of microscopic bristle deformation, it is determined whether the system is in the static friction zero-crossing critical region. Based on the generalized full-source disturbance rejection torque, a reverse feedforward compensation torque command is generated to dynamically reset the integral term of the controller. The reverse feedforward compensation torque command and the basic command of the main image stabilization control loop are vector-synthesized to generate the final drive voltage applied to the gimbal motor.

2. The airborne camera image stabilization control method according to claim 1, characterized in that, The analytical steps for generating the visual hysteresis residual tensor are as follows: By combining spherical projection geometry and mechanical dead zone determination, a hysteretic deformation tensor field for gimbal motion is constructed. Using the hysteresis deformation tensor field as the background motion reference, a pixel-by-pixel difference operation is performed with the measured dense optical flow field of the current frame to solve the original residual vector set; The original residual vector set is subjected to temporal sliding window integration to extract the viscous-slip hysteresis loop area features and construct the visual hysteresis residual tensor.

3. The airborne camera image stabilization control method according to claim 2, characterized in that, The analytical steps for constructing the hysteresis deformation tensor field of the gimbal motion are as follows: Based on the camera intrinsic parameter matrix, a unit physical sphere space is constructed. The discrete pixel array of the image plane is projected backward onto the sphere using the inverse perspective projection transformation to generate a rigid line-of-sight beam cluster. The virtual active inner spherical shell within the unit physical sphere space is rotated using the instantaneous rotational attitude quaternion, and the ideal tangential arc path of the rigid line-of-sight beam cluster as it moves along the spherical tangential plane is calculated.

4. The airborne camera image stabilization control method according to claim 3, characterized in that, The analytical steps for constructing the hysteresis deformation tensor field of the gimbal motion also include: A mechanical dead zone cone representing gear clearance is constructed at the beam endpoint. The ideal tangential arc path and the bottom surface of the mechanical dead zone cone are included and cross-boundary. The effective drag component that exceeds the boundary is extracted to generate a physical response displacement vector. The physical response displacement vector is remapped back to the two-dimensional image plane using perspective projection forward transformation, and a hysteretic deformation tensor field is generated through spatial interpolation.

5. The airborne camera image stabilization control method according to claim 1, characterized in that, The analysis steps for the equivalent frictional torque value of the obstructing gimbal are as follows: Establish a vision-mechanical mapping equation to transform the vision hysteresis residual tensor into virtual relative angular velocity observations; An unscented Kalman filter is constructed, and the virtual relative angular velocity observation is used as the measurement update signal to iteratively calculate the estimated value of microscopic mane deformation.

6. The airborne camera image stabilization control method according to claim 5, characterized in that, The analysis steps for the equivalent frictional torque value of the obstructing gimbal also include: The derivative of the bristle deformation is obtained by analytical calculation based on the equation of state, using the estimated values ​​of Coulomb friction torque and microscopic bristle deformation. Based on the estimated value of microscopic bristle deformation, combined with the derivative of bristle deformation and the preset Coulomb friction coefficient and viscous friction coefficient, the equivalent frictional torque value that hinders the gimbal is calculated.

7. The airborne camera image stabilization control method according to claim 6, characterized in that, The analysis steps for the generalized all-source disturbance rejection moment are as follows: A visual-mechanical virtual impedance mapping model is constructed, and the original residual vector set is input into the impedance mapping model to calculate the visual compensation torque. A frequency-domain complementary fusion filter is constructed to perform high-pass filtering on the equivalent friction torque value of the obstructing gimbal and low-pass filtering on the visual compensation torque. The two are then vector-superimposed to generate a generalized full-source anti-disturbance torque.

8. The airborne camera image stabilization control method according to claim 7, characterized in that, The analysis steps for generating the reverse feedforward compensation torque command and dynamically resetting the integral term of the controller are as follows: The estimated value of microscopic bristle deformation is monitored. When it does not reach the maximum static friction deformation threshold, the system is determined to be in the static friction adsorption stage. During the static tribosorption stage, a feedforward compensation torque command with equal amplitude and opposite direction is generated by using the generalized full-source anti-disturbance torque. During the static tribosorption stage, the integral separation logic is triggered synchronously. While outputting the feedforward compensation torque command, the accumulation of the integral term of the PID controller is paused.

9. The airborne camera image stabilization control method according to claim 8, characterized in that, The analysis steps for generating the final driving voltage applied to the gimbal motor are as follows: Receive the basic output command output by the main image stabilization control loop after cascading operations of the position loop and the velocity loop; The feedforward compensation torque command is superimposed on the current loop input terminal of the basic output command to form a torque feedforward enhancement command; The torque feedforward enhancement command is converted into a three-phase bridge arm duty cycle signal by the space vector pulse width modulation module, thereby generating the final drive voltage.

10. An airborne camera image stabilization control system, characterized in that, An airborne camera image stabilization control method for performing any one of claims 1-9 includes: The acquisition and calculation module is used to synchronously acquire the inertial attitude data of the airborne platform, the timing data of the gimbal motor encoder, and the zoom position information of the lens, and to calculate the camera intrinsic parameter matrix and instantaneous rotational attitude quaternion. The hysteresis residual module is used to acquire real-time video streams. It uses the camera intrinsic parameter matrix and instantaneous rotational attitude quaternion to solve the ideal rigid body motion projection, constructs a hysteresis deformation tensor field, and generates a visual hysteresis residual tensor based on the hysteresis deformation tensor field through difference operations and hysteresis feature analysis. The torque analysis module is used to take the visual hysteresis residual tensor as the observation input, perform state estimation through the LuGre dynamic friction model, solve the micro-hair deformation state and the equivalent friction torque value of the gimbal, perform visual compensation analysis based on the equivalent friction torque value of the gimbal, and obtain the generalized full-source anti-disturbance torque. The critical compensation module is used to determine whether the system is in the static friction zero-crossing critical region by estimating the micro-hair deformation value. It generates a reverse feedforward compensation torque command based on the generalized full-source disturbance rejection torque to dynamically reset the integral term of the controller. The drive voltage module is used to vector synthesize the reverse feedforward compensation torque command and the basic command of the main image stabilization control loop to generate the final drive voltage applied to the gimbal motor.

Images