Image-based spatial non-cooperative target approach tracking method and system

Abstract

The present invention discloses a method and system for approaching and tracking a spatial non-cooperative target based on an image, which relates to the field of target tracking technology, and includes: step 1, obtaining a current observed image in real time, and extracting and matching feature points based on the current observed image and an expected image; step 2, inputting the feature point matching result into a navigation function to obtain the expected image feature, and calculating the image error based on the expected image feature; step 3, calculating the control force and the control torque based on the image error using an image-based control algorithm; step 4, inputting the control force and the control torque into a simulation environment, obtaining the observed image at the next moment, and jumping to execute step 1. The present invention directly generates control instructions based on image errors, thereby avoiding relative pose estimation and reducing the degree of dependence of the control effect on the visual algorithm. A two-level control scheme based on image errors is proposed, which has good compatibility with visual servo control for first-order dynamic systems.

Description

Technical Field

[0001] The present invention relates to the field of target tracking technology, and more particularly to an image-based spatial non-cooperative target approach tracking method and system. Background Art

[0002] With the increasing frequency of human space activities, the number of space debris and failed satellites is increasing, which seriously threatens the safety of on-orbit spacecraft and the sustainability of the space environment. In order to meet this challenge, on-orbit service technology, especially the technology for repairing failed satellites and clearing space debris, has developed rapidly. The core of these technologies lies in the ability to accurately approach and capture non-cooperative targets, that is, those spacecraft or space debris that cannot actively cooperate in orbit adjustment or docking.

[0003] In the process of approaching and tracking non-cooperative targets, visual servo control technology has attracted widespread attention due to its high precision and autonomy. Visual servo control can be divided into two categories: pose-based visual servo control and image-based visual servo control.

[0004] Pose-based visual servo control: This method first estimates the relative pose through a visual algorithm, and then calculates the control error based on the estimated relative pose and inputs it into the controller. Although this method can better adapt to the rendezvous and docking control algorithm of cooperative targets, it relies on the accuracy of the pose estimation algorithm. If there is an error in the pose estimation, the control accuracy will be affected. In addition, this method has risks when the target moves out of the field of view.

[0005] Image-based visual servo control: Unlike the pose-based method, image-based visual servo control directly defines the control error in the pixel plane without estimating the pose of the target. The control goal of this method is to keep the current visual observation consistent with the visual observation at the target position. Since the control law is designed directly in the image space, the image-based method has better robustness. However, the current image-based method is mainly used for control in robotic arm scenarios, and its application to second-order dynamic systems such as spacecraft is still in the exploratory stage. In addition, the image-based method has difficulties in controller design and is difficult to be compatible with the existing attitude and orbit integrated control method.

[0006] Although image-based visual servo control has many advantages, it still faces some challenges in practical applications. For example, the image Jacobian matrix may have singularities at positions far from the desired state, resulting in unstable control. In addition, due to the lack of direct estimation of the target pose, image-based methods may have difficulty in achieving precise control in some cases.

[0007] Therefore, how to deal with the control problem in the second-order dynamic system and overcome the limitations of the prior art in controller design and compatibility is an issue that technical personnel in this field urgently need to solve. Summary of the invention

[0008] In view of this, the present invention provides an image-based spatial non-cooperative target proximity tracking method and system, which overcome the above-mentioned defects.

[0009] In order to achieve the above object, the present invention adopts the following technical solution:

[0010] An image-based spatial non-cooperative target approach tracking method, the specific steps are:

[0011] Step 1: Acquire the current observation image in real time, and extract and match feature points based on the current observation image and the expected image;

[0012] Step 2: Input the feature point matching result into the navigation function to obtain the expected image feature, and calculate the image error based on the feature point extraction result and the expected image feature;

[0013] Step 3, calculating the control force and the control torque using an image-based control algorithm according to the image error;

[0014] Step 4: input the control force and the control torque into the simulation environment, obtain the observation image at the next moment, and jump to execute step 1.

[0015] Optionally, the feature point extraction and matching in step 1 are implemented based on a feature extraction and matching model, and the steps for constructing the feature extraction and matching model are:

[0016] Select a pre-trained model for feature point extraction and matching;

[0017] Constructing a fine-tuning dataset and the fine-tuning loss function;

[0018] The pre-trained model is iteratively trained using the fine-tuning data set and the fine-tuning loss function until a preset iteration condition is reached to obtain the feature extraction and matching model.

[0019] Optionally, the steps of constructing the fine-tuning dataset are:

[0020] Select a reference pose;

[0021] Constructing a screening data subset based on the reference posture, and performing data screening based on a rotation angle difference between the shooting posture of each image in the original data set and the reference posture and a preset threshold, and incorporating the screened data into the corresponding screening data subset;

[0022] The fine-tuning data set is obtained by arranging and combining the data in each of the screening data subsets.

[0023] Optionally, the expression of the fine-tuning loss function is:

[0024] ;

[0025] in, ;

[0026] ;

[0027] In the formula, To calculate the Pearson correlation coefficient between two input vectors; Output a score for the model; is the degree to which the model satisfies the imaging constraints; For the The degree to which the feature point matching pairs satisfy the imaging constraints; Transpose the matrix; The camera 1 is used to The observation vector calculated from the feature points; The camera 2 is used to The observation vector calculated from the feature points; The vector from camera 1 to camera 2.

[0028] Optionally, the expression of the navigation function is:

[0029] ;

[0030] In the formula, is the coordinates of the image feature points observed at the initial moment; is the desired image feature point coordinates; is the number of iterations to reach the desired image features; is the number of steps currently iterated.

[0031] Optionally, the image-based control algorithm includes two levels of control laws, namely: based on the image kinematic equation, the image error is set to be a damped oscillation to obtain an "error-speed" level control law; based on the spacecraft attitude and orbit integrated dynamics equation, the tracking error is set to be a damped oscillation to obtain a "speed-command" level control law.

[0032] Optionally, the expression of the “error-speed” level control law is:

[0033] ;

[0034] In the formula, is the instruction generalized speed; All are control law hyperparameters; is the image Jacobian matrix The Moore-Penrose generalized inverse matrix of ; is the image error.

[0035] Optionally, the expression of the “speed-command” level control law is:

[0036] ;

[0037] In the formula, is a general force; All are controller parameters; is the tracking error; is a matrix; is the generalized speed.

[0038] Optionally, the simulation environment in step 4 iterates the spacecraft motion state according to the control force and the torque to achieve motion simulation; wherein, the spacecraft motion state iteration includes: iterating the spacecraft position based on the second-order dynamic equations under the gravitational field; iterating the spacecraft attitude using the spacecraft system dynamics equations and quaternion kinematic equations.

[0039] An image-based spatial non-cooperative target approach tracking system, comprising:

[0040] A feature extraction and matching module is used to obtain the current observation image in real time, and extract and match feature points based on the current observation image and the expected image;

[0041] A navigation module, used for inputting the feature point matching result into the navigation function to obtain the expected image feature, and calculating the image error according to the feature point extraction result and the expected image feature;

[0042] A control module, configured to calculate a control force and a control torque according to the image error using an image-based control algorithm;

[0043] The environment simulation module is used to input the control force and the control torque into the simulation environment, obtain the observation image at the next moment, and input the observation image at the next moment into the feature extraction and matching module.

[0044] It can be seen from the above technical solutions that, compared with the prior art, the present invention provides an image-based spatial non-cooperative target approach tracking method and system, which directly generates control instructions based on image errors, thereby avoiding relative pose estimation and reducing the control effect's dependence on the visual algorithm. A two-level control scheme based on image errors is proposed, which has good compatibility with the existing visual servo control for first-order dynamic systems. The present invention proposes a navigation function to constrain the motion trajectory of feature points during the control process to ensure the visibility of the target, while improving the stability of the control algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0046] Figure 1 A schematic diagram of the method provided by the present invention;

[0047] Figure 2 A schematic diagram of a method for constructing a fine-tuning dataset provided by the present invention;

[0048] Figure 3 A schematic diagram of the constraint relationship between the feature point matching result and the relative position and posture of the camera provided by the present invention;

[0049] Fig. 4(a) is a schematic diagram of the simulation of relative position change in a slow-spinning scenario provided by the present invention; Fig. 4(b) is a schematic diagram of the simulation of relative speed change in a slow-spinning scenario provided by the present invention; Fig. 4(c) is a schematic diagram of the simulation of relative posture change in a slow-spinning scenario provided by the present invention; Fig. 4(d) is a schematic diagram of the simulation of relative position change in a non-slow-spinning scenario provided by the present invention; Fig. 4(e) is a schematic diagram of the simulation of relative speed change in a non-slow-spinning scenario provided by the present invention; Fig. 4(f) is a schematic diagram of the simulation of relative posture change in a non-slow-spinning scenario provided by the present invention;

[0050] FIG. 5( a ) is a diagram of the present invention. FIG5( b ) is an observation image provided by the present invention; ... FIG5 (c) is an observation image provided by the present invention. FIG5( d ) is an observation image provided by the present invention. Observation image at time

[0051] Figure 6 This is a schematic diagram of the system structure provided by the present invention. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] An image-based spatial non-cooperative target approach tracking method, such as Figure 1 As shown, the specific steps are:

[0054] Step 1: Acquire the current observation image in real time, and extract and match feature points based on the current observation image and the expected image;

[0055] Step 2: Input the feature point matching result into the navigation function to obtain the expected image features, and calculate the image error based on the feature point extraction result and the expected image features;

[0056] Step 3, calculating the control force and the control torque using an image-based control algorithm according to the image error;

[0057] Step 4: Input the control force and control torque into the simulation environment, obtain the observation image at the next moment, and jump to step 1.

[0058] In one embodiment, the feature point extraction and matching in step 1 is implemented based on a feature extraction and matching model, and the steps for constructing the feature extraction and matching model are:

[0059] Step 11: Select a pre-trained model for feature point extraction and matching;

[0060] Step 12: Construct a fine-tuning dataset and a fine-tuning loss function;

[0061] Step 13: Iteratively train the pre-trained model using the fine-tuning dataset and the fine-tuning loss function until the preset iteration conditions are met to obtain a feature extraction and matching model.

[0062] Furthermore, the pre-trained model used in step 11 is a SuperPoint+SuperGlue framework trained on a ground visual positioning dataset that performs well in the field of deep learning.

[0063] In one embodiment, the steps of constructing the fine-tuning dataset are:

[0064] Select a reference pose;

[0065] A screening data subset is constructed based on the reference posture, and data is screened based on the rotation angle difference between the shooting posture of each image in the original data set and the reference posture and a preset threshold, and the screened data is included in the corresponding screening data subset;

[0066] The data in each filtered data subset are arranged and combined to obtain a fine-tuning data set.

[0067] Furthermore, when fine-tuning the dataset construction, due to the Tango dataset The training data volume is 12,000 images. Generating data based on pairing will result in too much data ( There are also two pictures with different shooting angles, which results in no feature points. Therefore, this embodiment constructs a fine-tuning data set based on the shooting posture facing the solar panel (front) and the shooting posture facing the satellite (back). The steps are as follows: Figure 2 As shown, specifically:

[0068] First, build a screening dataset , including two reference postures based on and Constructed positive subset With backface subset , calculate the Tango dataset The shooting posture of any image in the image is relative to the two reference postures and The rotation angle difference and , the rotation angle difference is less than a certain threshold (in this embodiment ) into the corresponding screening data subset. Then, according to the The two filtered data subsets in are arranged and combined internally to obtain the fine-tuning data set.

[0069] In one embodiment, random number screening is also introduced in the data screening process, specifically: when the rotation angle difference meets the preset threshold, a random number is taken. , determine the random number Whether the preset conditions are met ( ), if not, the data is included in the corresponding screening data subset, if so, further judgment is performed.

[0070] Furthermore, the above process can be used to filter the data set , which contains 2 filtered data subsets, and then the observation postures in each filtered data subset are arranged and combined to obtain fine-tuning data. Each data in the fine-tuning data set contains two images and the observation postures of these two images.

[0071] In one embodiment, the steps of constructing the fine-tuning loss function in step 12 are:

[0072] Since the Tango dataset It is a pose estimation dataset, which only provides visual observations and relative pose labels and cannot be directly used for feature point extraction and matching fine-tuning. Therefore, this embodiment derives the constraint relationship that the feature point matching result and the camera relative pose must satisfy when the camera relative pose is known, such as Figure 3 As shown, a fine-tuning loss function is constructed based on this. The specific steps are:

[0073] Feature points of camera 1 and camera 2 Observe, then the two cameras are on the feature points The observation vector The relative translation vector between the two cameras Coplanar, that is:

[0074] (1);

[0075] In the formula, is the vector pointing from camera 1 to camera 2.

[0076] If the image observation of camera 1 and the image observation of camera 2 can well achieve feature point matching, that is, accurately find the feature points The pixel coordinates in camera 1 and camera 2 are calculated based on the observation vector The relative translation between the two cameras The above formula must be satisfied, so the expression of the loss function is:

[0077] (2);

[0078] In the formula, The camera 1 is used to The observation vector calculated from the feature points; The camera 2 is used to The observation vector calculated from the feature points; To detect the total number of matched feature points, and eliminate the influence of vector modulus on the loss function. To make the model output score To better reflect the quality of feature point matching, the loss function is rewritten as the loss of each feature point and the model output score The Pearson correlation coefficient is:

[0079] (3);

[0080] in, is the Pearson correlation coefficient between the two input vectors, , For the The score of feature point matching pairs; Measured the The degree to which the feature point matching pairs satisfy the imaging constraints; The camera 1 is used to The observation vector calculated from the feature points; The camera 2 is used to The observation vector calculated from the feature points; is the vector pointing from camera 1 to camera 2; is the matrix transpose. The training target is the model output score of the score vector in each image The degree to which the model satisfies the imaging constraints Negatively correlated.

[0081] In one embodiment, a navigation function is introduced , let the feature points approach the expected feature point coordinates according to the specified trajectory. Converges to the desired image feature point coordinates In this embodiment, a straight line trajectory in the image space is used as the navigation trajectory, and the navigation trajectory is calculated based on the ideal state of reaching the destination at a uniform speed within a certain period of time. The specific expression is as follows:

[0082] (4);

[0083] In the formula, is the coordinates of the image feature points observed at the initial moment; is the desired image feature point coordinates; is the number of iterations to reach the desired image features; The image error is calculated based on the expected image features output by the navigation function at each moment and the feature point extraction results (the image features observed at the current moment). , thus generating the instruction generalized velocity.

[0084] In one embodiment, the image-based control algorithm includes two levels of control laws, namely: based on the image kinematic equation, the image error is set to be a damped oscillation to obtain the "error-speed" level control law; based on the spacecraft attitude and orbit integrated dynamics equation, the tracking error is set to be a damped oscillation to obtain the "speed-command" level control law.

[0085] Furthermore, the image-based control algorithm uses the image error as input and outputs the control force and control torque of the spacecraft. It is mainly divided into the "error-speed" level control law and the "speed-command" level control law. First, the relationship between the image error and the speed is constructed based on the image kinematic equation, and the image error is set to be a damped oscillation to obtain the "error-speed" level control law. Then, based on the spacecraft attitude and orbit integrated dynamics equation, the tracking error is set to be a damped oscillation to obtain the "speed-command" level control law. The "error-speed" level control law obtains the command generalized speed based on the image error, and the "speed-command" level control law obtains the control force and torque based on the command generalized speed and the current generalized speed.

[0086] In one embodiment, the “error-speed” level control law is:

[0087] (5);

[0088] In the formula, is the instruction generalized speed; All are control law hyperparameters; is the image Jacobian matrix The Moore-Penrose generalized inverse matrix of ; is the image error; Transpose the matrix.

[0089] Furthermore, the derivation process of the “error-speed” level control law is as follows:

[0090] The kinematic equation of the known image is:

[0091] (6);

[0092] In the formula, , For the The speed at which the coordinates of feature points change; , For the The image Jacobian matrix calculated from feature points; is the generalized speed; , where , characterizing the The pixel coordinate changes caused by the movement of feature points relative to the inertial space, for The number of pixels corresponding to the focal length in the axial direction; for The number of pixels corresponding to the focal length in the axial direction; For the The relative position of feature points to the camera in the camera coordinate system Axis component; For the The coordinates of feature points on the pixel plane; is the principal point coordinate of camera 1; For the The three-axis components of the velocity of each feature point in the camera coordinate system; is the matrix transpose. Consider the spacecraft moving at the command generalized velocity, and subtract both sides of the equation , then:

[0093] (7);

[0094] In the formula, is the rate of change of image error.

[0095] Will is regarded as a distractor and ignored, then If the image error is set to be a damped oscillation, the image error satisfies: , where the parameters should satisfy ; It is the second-order differential of the image error change, that is The second-order differential with respect to time; integrating with respect to time gives: Substitute equation (7) and ignore the interference term:

[0096] (8);

[0097] The above formula can be transformed into , which is formula (5).

[0098] In one embodiment, the expression of the “speed-command” level control law is:

[0099] (9);

[0100] In the formula, is a general force; is the speed tracking error; All are controller parameters; is the generalized speed, is a matrix.

[0101] Furthermore, the derivation process of the “speed-command” level control law is:

[0102] The compact form of the six-degree-of-freedom dynamic equations for the generalized velocity of a spacecraft is:

[0103] (10);

[0104] In the formula, is the generalized speed; is the rate of change of generalized velocity, i.e., the first-order differential with respect to time; is a general force; It is a generalized interference force.

[0105] Defining Tracking Error ,in, is the generalized speed, is the command generalized speed. Subtract both sides of the above formula from , then:

[0106] (11);

[0107] in, is the rate of change of tracking error.

[0108] Ignore distractions , and believes that , if the tracking error is allowed to vibrate with damping over time, the tracking error satisfies: , is the second-order differential of the tracking error change, that is, The second-order differential with respect to time. At this time, the control force and torque should satisfy equation (9), which is the "speed-command" level control law.

[0109] In one embodiment, in step 4, the simulation environment iterates the spacecraft motion state according to the control force and torque to realize motion simulation; wherein, the spacecraft motion state iteration includes: iterating the spacecraft position based on the second-order dynamic equation under the gravitational field; iterating the spacecraft attitude using the spacecraft system dynamic equation and quaternion kinematic equation.

[0110] Furthermore, this embodiment verifies the system in slow-spin and non-slow-spin scenarios. The parameter settings of the two scenarios are shown in Table 1.

[0111] Table 1

[0112]

[0113] Assume that both the tracking spacecraft and the target spacecraft are in geosynchronous orbit, and the true anomaly angles of the two are different. , about The tracking spacecraft maintains three-axis attitude stability relative to the ground at the initial moment. The control goal is to control the tracking spacecraft into the target spacecraft's own system. The structural parameters of the two spacecraft are shown in Table 2.

[0114] Table 2

[0115]

[0116] Considering that both spacecraft are affected by the earth's gravity, the approximate calculation is to J2000, without considering the gravitational effects of other celestial bodies such as the moon and the sun. The camera intrinsic matrix parameters are set as: .

[0117] The parameters of the “error-speed” level control law are set as: The scalar parameters in the “speed-command” level control law are expanded to a diagonal matrix , The simulation curves are shown in Figure 4 (a) to Figure 4 (f).

[0118] The changes in image observations during approach tracking are shown in Figure 5(a)-Figure 5(d) (the left side is the current image observation, and the right side is the target image observation).

[0119] In one embodiment, a mathematical derivation process of fine-tuning the loss function is also disclosed, specifically:

[0120] Set two cameras with the same intrinsic parameter matrix, namely camera 1 ( ) and Camera 1 ( ), whose camera coordinate systems are and There is a point in space , relative to The position vector of The coordinate array below is , which is relative to The position vector of The coordinate array below is , then its The coordinates on the pixel plane are:

[0121] (12);

[0122] In the formula, for Point Coordinates on the pixel plane; for The principal point coordinates, for The number of pixels corresponding to the focal length in the axial direction; for The number of pixels corresponding to the focal length in the axial direction.

[0123] remember express Point Imaged pixel coordinates Homogeneous representation of the camera internal parameter matrix , similarly The pixel coordinates of the image are , for relative to The position vector of The coordinate array below.

[0124] The coordinate array of the relative position of the point relative to the two cameras in their respective camera coordinate systems Satisfies the following relationship:

[0125] (13);

[0126] In the formula, For the coordinate system arrive The rotation matrix of Representing coordinate system The origin, relative to The position vector of the origin is The coordinate array below. The position augmented coordinates , then equations (1) and (2) can be transformed into:

[0127] (14);

[0128] (15);

[0129] In the formula, for The position augmented coordinates of

[0130] By definition, , then substitute formula (14) into formula (15) and we have:

[0131] (16);

[0132] because The position augmented coordinates With point exist The coordinates on the pixel plane satisfy the following relationship:

[0133] (17);

[0134] In the formula, for Point Coordinates on the pixel plane.

[0135] Then substitute formula (17) into formula (16):

[0136] (18);

[0137] remember ;

[0138] . Marking quantity . Write formula (18) as:

[0139] (19);

[0140] The above formula shows that The position augmented coordinates and The relative translation vector satisfies the above linear relationship, where the coefficients are related to the point is related to the depth.

[0141] On the other hand, this embodiment discloses an image-based spatial non-cooperative target approach tracking system, which applies the above method, such as Figure 6 As shown, including:

[0142] The feature extraction and matching module is used to obtain the current observation image in real time, and extract and match feature points based on the current observation image and the expected image;

[0143] A navigation module, used for inputting the feature point matching result into the navigation function to obtain the expected image feature, and calculating the image error according to the feature point extraction result and the expected image feature;

[0144] A control module, for calculating a control force and a control torque using an image-based control algorithm according to an image error;

[0145] The environment simulation module is used to input the control force and the control torque into the simulation environment, obtain the observation image at the next moment, and input the observation image at the next moment into the feature extraction and matching module.

[0146] Furthermore, the feature extraction and matching module uses the current observation image and the expected image As input, output feature point extraction and matching results . The SuperPoint model is used for feature point extraction, and the SuperGlue model is used for feature point matching. This embodiment proposes a fine-tuning loss function based on imaging constraints, which can achieve fine-tuning of the model when the homography matrix is ​​missing. The above loss function is used to fine-tune the pre-trained model of the ground dataset in the aerospace dataset.

[0147] The navigation function in the navigation module takes the feature point extraction and matching results as input, outputs the current expected feature point position, constrains the feature point to approach the target position along a straight line, performs linear interpolation to obtain the current expected feature point coordinates, and thus obtains the image error. .

[0148] The image-based control algorithm in the control module takes the image error as input and outputs the control force and torque of the spacecraft. It is mainly divided into the "error-speed" level control law and the "speed-command" level control law. First, the relationship between the image error and the speed is constructed based on the image kinematic equation, and the image error is set to be a damped oscillation to obtain the "error-speed" level control law. Then, based on the spacecraft attitude and orbit integrated dynamics equation, the tracking error is set to be a damped oscillation to obtain the "speed-command" level control law.

[0149] The environmental simulation module is used to verify the proposed control algorithm. It provides image observation of the target and the control force and torque interface input of the tracking spacecraft. After receiving the control force and torque, the spacecraft state is updated to obtain the observation image at the next moment.

[0150] The present invention is aimed at the problem of approaching and tracking non-cooperative targets in space, derives the image dynamics equation, and expands the description scenario of the image kinematics equation to the second-order dynamics system; introduces a feature point extraction and matching algorithm based on deep learning to realize the visual feature extraction of image observation in space scenes; introduces a navigation function to ensure the visibility of the target during the approach process; and proposes a two-level control law based on the image to directly obtain the control command based on the image error. Finally, it realizes the approach tracking of non-cooperative targets based only on image observations, reduces the dependence of control accuracy on visual algorithms, and can be used for on-orbit service satellites to approach and track non-cooperative targets.

[0151] The system proposed in the present invention can realize the extraction of visual features of image observation in space scenes, and directly obtain control instructions based on image features while ensuring that the target is visible, providing a solution for achieving close tracking of non-cooperative targets.

[0152] The present invention constructs a fine-tuning data set based on a real aerospace data set, and proposes a fine-tuning loss function based on imaging constraints, so that the feature point extraction and matching model can accurately extract image features in a space environment; it also introduces a navigation function to constrain the motion trajectory of the feature points during the docking process to ensure the target visibility during the approach process, while improving the stability of the controller and the smoothness of the movement; the present invention disassembles a single second-order control system into two first-order control systems, and proposes an "error-speed" level control law and a "speed-instruction" level control law, which greatly simplifies the control problem, the control law is simple in form, and the parameters contained are easy to solve; and the system is verified in a simulation environment to achieve the full process from feature extraction to control instructions. Therefore, the approach tracking task of non-cooperative targets can be achieved, providing a complete solution for the realization of this task.

[0153] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0154] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A spatial non-cooperative target approach tracking method based on an image, characterized in that: The specific steps are: Step 1: Acquire the current observation image in real time, and extract and match feature points based on the current observation image and the expected image; Step 2: Input the feature point matching result into the navigation function to obtain the expected image feature, and calculate the image error based on the feature point extraction result and the expected image feature; Step 3, calculating the control force and the control torque using an image-based control algorithm according to the image error; Step 4: input the control force and the control torque into the simulation environment, obtain the observation image at the next moment, and jump to step 1; The image-based control algorithm includes two levels of control laws, namely: based on the image kinematic equation, the image error is set to be a damped oscillation to obtain an "error-speed" level control law; based on the spacecraft attitude-orbit integrated dynamics equation, the tracking error is set to be a damped oscillation to obtain a "speed-command" level control law; The expression of the "error-speed" level control law is: Where V cd is the command generalized speed; p ,λ i All are control law hyperparameters; is the image Jacobian matrix L s The Moore-Penrose generalized inverse matrix of ; e is the image error; The expression of the "speed-command" level control law is: F u =-K p e track -K i ∫e track dt-OFF c ; In the formula, F u is the generalized force; K p , K i All are controller parameters; e track is the tracking error; A is the matrix; V c is the generalized speed.

2. The image-based spatial non-cooperative target approach tracking method according to claim 1, characterized in that: The feature point extraction and matching in step 1 are implemented based on the feature extraction and matching model. The construction steps of the feature extraction and matching model are as follows: Select a pre-trained model for feature point extraction and matching; Construct fine-tuning dataset and fine-tuning loss function; The pre-trained model is iteratively trained using the fine-tuning data set and the fine-tuning loss function until a preset iteration condition is reached to obtain the feature extraction and matching model.

3. The image-based spatial non-cooperative target approach tracking method according to claim 2, characterized in that: The steps for constructing the fine-tuning dataset are: Select a reference pose; Constructing a screening data subset based on the reference posture, and performing data screening based on a rotation angle difference between the shooting posture of each image in the original data set and the reference posture and a preset threshold, and incorporating the screened data into the corresponding screening data subset; The fine-tuning data set is obtained by arranging and combining the data in each of the screening data subsets.

4. The image-based spatial non-cooperative target approach tracking method according to claim 2, characterized in that: The expression of the fine-tuning loss function is: Loss = corr(score, dist); Where, dist = [d1,…d i …, d K ] T ; In the formula, corr(.,.) is the Pearson correlation coefficient between the two input vectors; score is the model output score; dist is the degree to which the model satisfies the imaging constraints; d i is the degree to which the i-th feature point matching pair satisfies the imaging constraint; T is the matrix transpose; X 1，i is the observation vector calculated by camera 1 for the i-th feature point; X 2，i is the observation vector calculated by camera 2 for the i-th feature point; Y 21 The vector from camera 1 to camera 2.

5. The image-based spatial non-cooperative target approach tracking method according to claim 1, characterized in that: The expression of the navigation function is: Where s(0) is the image feature observed at the initial moment; s* is the coordinate of the desired image feature point; K is the number of iterations to reach the desired image feature; and k is the number of iterations currently completed.

6. The image-based spatial non-cooperative target approach tracking method according to claim 1, characterized in that: The simulation environment described in step 4 iterates the spacecraft motion state according to the control force and the torque to realize motion simulation; wherein, the spacecraft motion state iteration includes: iterating the spacecraft position based on the second-order dynamic equation under the gravitational field; iterating the spacecraft attitude using the spacecraft system dynamic equation and quaternion kinematic equation.

7. An image-based space non-cooperative target approach tracking system, using the image-based space non-cooperative target approach tracking method according to claim 1, characterized in that: include: A feature extraction and matching module is used to obtain the current observation image in real time, and extract and match feature points based on the current observation image and the expected image; A navigation module, used for inputting the feature point matching result into the navigation function to obtain the expected image feature, and calculating the image error according to the feature point extraction result and the expected image feature; A control module, configured to calculate a control force and a control torque according to the image error using an image-based control algorithm; The environment simulation module is used to input the control force and the control torque into the simulation environment, obtain the observation image at the next moment, and input the observation image at the next moment into the feature extraction and matching module.

Images