Satellite on-orbit observation system and method for weak and small target based on intelligent closed-loop feedback

Abstract

The application provides a satellite weak small target on-orbit observation system and method based on intelligent closed-loop feedback, and the system comprises: a detection module, which performs image enhancement and time sequence consistency enhancement on an infrared image, obtains enhanced features, and obtains a candidate target set based on the enhanced features; a tracking module, which performs target matching in a search area; when the matching is successful, a candidate position is taken as an observation value and input into a Kalman filter to obtain a target state vector; when the matching fails, a predicted state of the Kalman filter is taken as the target state vector, a search window is expanded with the predicted position as the center, and re-identification is performed; a control module, which maps the target state vector into an attitude error and generates a control instruction by using a single neuron adaptive PID controller; and a closed-loop scheduling module, which dynamically adjusts the operation parameters of at least one module according to an execution error and a detection confidence. The application improves the continuous tracking precision of a weak small target and the attitude control stability.

Description

Technical Field

[0001] This invention relates to the field of aerospace remote sensing technology, specifically to an on-orbit observation system and method for weak satellite targets based on intelligent closed-loop feedback. Background Technology

[0002] Small targets in space, such as space debris, microsatellites, and stealth aircraft, which are 1-3 pixels in size, pose a great challenge to on-orbit detection and tracking due to their weak signals, extremely small size, and rapidly changing trajectories.

[0003] Existing research largely focuses on optimizing individual aspects of detection, tracking, and control: the detection stage emphasizes image denoising and enhancement, the tracking stage relies on filtering or depth matching, and the control stage employs fixed-parameter PID control. Traditional detection algorithms are susceptible to noise interference in complex dynamic environments, resulting in insufficient accuracy and robustness; tracking algorithms lack upper-level feedback when occlusion or target disappearance occurs, leading to drift; attitude control systems have fixed parameters and lag in response, making them difficult to adapt to on-orbit disturbances and mission changes. Furthermore, existing systems generally employ a decentralized architecture for detection, tracking, and control, lacking a unified information feedback and adaptive mechanism. This prevents satellites from adjusting their attitude in real time based on observation errors and from utilizing control feedback to optimize the perception model, thus limiting the autonomy and accuracy of on-orbit operation.

[0004] Therefore, there is an urgent need for a satellite weak target on-orbit observation system and method based on intelligent closed-loop feedback, which can achieve high-precision, autonomous on-orbit observation of weak targets in space. Summary of the Invention

[0005] In view of this, it is necessary to provide a satellite weak target on-orbit observation system and method based on intelligent closed-loop feedback, so as to solve the technical problems of independent detection, tracking and attitude control links, information fragmentation and inability to form an on-orbit adaptive closed loop in the existing technology.

[0006] To address the aforementioned technical problems, in a first aspect, the present invention provides a satellite-based on-orbit observation system for weak targets based on intelligent closed-loop feedback, comprising:

[0007] The detection module is used to perform single-frame image enhancement and temporal consistency enhancement on the infrared images acquired on the satellite, obtain enhancement features, and obtain a set of candidate targets based on the enhancement features. The set of candidate targets includes multiple candidate targets and the candidate position, detection confidence and depth representation vector of each candidate target.

[0008] The tracking module is used to perform target matching within a preset search area based on the similarity between the depth representation vector and the tracking template, as well as the detection confidence. When a match is successful, the matched candidate position is used as an observation and input into a Kalman filter for state update to obtain an updated target state vector. When a match fails, the predicted state of the Kalman filter is used as the target state vector, and the search window is expanded with the predicted position corresponding to the predicted state as the center to perform re-identification and resume tracking.

[0009] The control module is used to map the target state vector into an attitude error, generate control instructions based on the attitude error and a single-neuron adaptive PID controller, and obtain the execution error after executing the control instructions.

[0010] A closed-loop scheduling module is used to dynamically adjust the operating parameters of at least one of the detection module, the tracking module, and the control module based on the execution error and the detection confidence level.

[0011] In one possible implementation, the detection module includes an image enhancement unit, a temporal consistency enhancement unit, and a detection unit;

[0012] The image enhancement unit is used to perform image enhancement processing on the infrared image based on a generative adversarial network model to obtain image enhancement features;

[0013] The temporal consistency enhancement unit is used to perform temporal consistency enhancement processing on the image enhancement features based on the state-space model to obtain the enhancement features;

[0014] The detection unit is used to perform target detection based on the enhanced features using the detection head to obtain the candidate target set.

[0015] In one possible implementation, the detection module further includes an image preprocessing unit;

[0016] The image preprocessing unit is used to perform radiometric correction and local contrast normalization on the infrared image.

[0017] In one possible implementation, the tracking module includes a similarity determination unit, a target state vector determination unit, and a re-identification unit;

[0018] The similarity determination unit is used to calculate the cosine similarity between the depth representation vector and the tracking template within a preset search area. When the cosine similarity is greater than the similarity threshold and the detection confidence is greater than the confidence threshold, the match is determined to be successful. When the cosine similarity is less than or equal to the similarity threshold or the detection confidence is less than or equal to the confidence threshold, the match is determined to be unsuccessful.

[0019] The target state vector determination unit is used to, when a match is successful, take the matched candidate position as the observation value and input it into the Kalman filter to update the state and obtain the updated target state vector; when a match fails, the predicted state of the Kalman filter is used as the target state vector.

[0020] The re-identification unit is used to expand the search window centered on the predicted position corresponding to the predicted state when the matching fails, and to determine the joint score of each candidate target in the expanded search window, determine the optimal candidate target with the highest joint score, and determine whether the joint score of the optimal candidate target is greater than the score threshold. If it is greater, the position of the optimal candidate target is used as the observation value and input into the Kalman filter for state update to obtain the updated target state vector; if it is less than or equal to, a predefined security policy is triggered.

[0021] In one possible implementation, the joint score is:

[0022]

[0023] In the formula, For joint scoring; Candidate positions and the predicted position of the Kalman filter The Euclidean distance between them; Depth representation vector and tracking template Cosine similarity; To test the confidence level; , , These are the weighting coefficients.

[0024] In one possible implementation, the control module includes an attitude error determination unit, a control law determination unit, and a control command generation unit;

[0025] The attitude error determination unit is used to map the position information in the target state vector into the attitude angle error of the satellite;

[0026] The control law determination unit is used to obtain the neuron weight vector of the current control cycle, take the attitude angle error, error integral and error derivative as input, map the neuron weight vector to obtain the proportional, integral and derivative parameters of the single neuron adaptive PID controller, and determine the control torque according to the proportional, integral and derivative parameters;

[0027] The control command generation unit is used to perform amplitude limiting processing on the control torque and generate the control command.

[0028] In one possible implementation, the control module further includes a neuron weight vector update unit;

[0029] The neuron weight vector update unit is used to obtain the attitude angle error and control increment of the current control cycle, construct a sliding window quadratic cost function based on the attitude angle error and the control increment, update the neuron weight vector based on the sliding window quadratic cost function, and use the updated neuron weight vector as the neuron weight vector for the next control cycle.

[0030] In one possible implementation, the sliding window quadratic cost function is:

[0031]

[0032]

[0033] In the formula, The cost is the cost of a quadratic form for a sliding window; The length of the sliding window; The current moment; The summation index variable represents each moment within the sliding window; This refers to the attitude angle error; To control the increment; , These are the weighting coefficients; The control torque for the current control cycle; This is the control torque of the previous control cycle.

[0034] In one possible implementation, the loss function used in the detection module is a weighted sum of adversarial loss, reconstruction loss, and feature loss; the closed-loop scheduling module is specifically used for:

[0035] When the detection confidence level is consistently lower than the detection confidence level threshold, or the execution error is consistently greater than the execution error threshold, the search area of ​​the tracking module is increased, and / or the weight of the reconstruction loss is increased;

[0036] When the amplitude of the control torque of the control module is greater than the saturation limit, the learning rate of the single neuron adaptive PID controller is reduced, and / or the control parameters of the single neuron adaptive PID controller are switched to preset conservative control parameters.

[0037] When a long-term time-varying environment is detected, the network structure parameters in the detection module are adjusted.

[0038] Secondly, the present invention also provides a method for on-orbit observation of weak satellite targets based on intelligent closed-loop feedback, applicable to the on-orbit observation system for weak satellite targets based on intelligent closed-loop feedback described in any of the above possible implementations, the method comprising:

[0039] The infrared images acquired on the satellite are enhanced with single-frame image enhancement and temporal consistency enhancement to obtain enhancement features. Based on the enhancement features, a set of candidate targets is obtained. The set of candidate targets includes multiple candidate targets and the candidate position, detection confidence and depth representation vector of each candidate target.

[0040] Within a preset search area, target matching is performed based on the similarity between the depth representation vector and the tracking template, as well as the detection confidence. When a match is successful, the matched candidate position is used as an observation and input into a Kalman filter for state update to obtain the updated target state vector. When a match fails, the predicted state of the Kalman filter is used as the target state vector, and the search window is expanded with the predicted position corresponding to the predicted state as the center to perform re-identification and resume tracking.

[0041] The target state vector is mapped to an attitude error, and control instructions are generated based on the attitude error and a single-neuron adaptive PID controller. The execution error after executing the control instructions is then obtained.

[0042] The operating parameters of at least one of the detection module, tracking module, and control module are dynamically adjusted based on the execution error and the detection confidence level.

[0043] The beneficial effects of this invention are as follows: The satellite weak target on-orbit observation system based on intelligent closed-loop feedback provided by this invention deeply couples the three originally independent links of detection, tracking, and control by introducing a closed-loop scheduling module, constructing an integrated operation architecture of "perception-estimation-decision-execution-feedback". Information flows bidirectionally between the detection module, tracking module, and control module, enabling the system to self-optimize based on the final execution effect and the front-end perception confidence level, thereby significantly reducing dependence on external intervention (such as ground commands), improving the satellite's long-term autonomous observation capability in complex dynamic scenarios, and thus significantly enhancing on-orbit autonomy. Experiments have shown that this system provides better guarantee of mission continuity in long-term operation scenarios than the separate scheme, and its performance degradation rate is reduced by 15%-25% in long-term missions.

[0044] Furthermore, the detection module sequentially performs image enhancement and temporal enhancement on the infrared image, suppressing noise at the pixel level and temporally, respectively, thereby improving the signal-to-noise ratio from both spatial and temporal dimensions. This collaborative enhancement mechanism effectively suppresses background noise and maintains the consistency of the appearance of weak targets across multiple frames, thus providing subsequent stages with a more confident and reliable set of candidate targets, ensuring the success rate of the entire observation mission from the source.

[0045] Furthermore, the tracking module employs a joint strategy of deep feature matching and Kalman filtering, and includes a pre-defined extended search and re-identification mechanism for matching failures, cleverly balancing high-precision positioning with occlusion robustness. This design ensures that the target can be accurately locked when visible, maintains its trajectory based on motion prediction when briefly lost, and achieves rapid and accurate re-acquisition, fundamentally solving the tracking interruption problem caused by target occlusion and weak signals.

[0046] Meanwhile, the control module employs a single-neuron adaptive PID controller, whose parameters can be dynamically adjusted in real time based on attitude error, rather than remaining fixed. This self-learning capability enables the controller to automatically adapt to changes in the satellite platform's dynamic characteristics and external space disturbances, effectively suppressing overshoot and abrupt changes in control inputs while maintaining rapid maneuver response, thereby achieving higher steady-state control accuracy and stronger robustness. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A schematic diagram of the structure of the satellite weak target on-orbit observation system based on intelligent closed-loop feedback provided by the present invention;

[0049] Figure 2 A schematic diagram of an embodiment of the detection module provided by the present invention;

[0050] Figure 3 A schematic diagram of an embodiment of the tracking module provided by the present invention;

[0051] Figure 4 A schematic diagram of an embodiment of the control module provided by the present invention;

[0052] Figure 5 A comparison of the control effects of the single-neuron adaptive PID controller and the fixed-parameter PID controller provided by this invention on the steady-state attitude angle;

[0053] Figure 6 This is a comparison chart of the performance indicators of the system of the present invention and the traditional distributed system;

[0054] Figure 7 This is a flowchart illustrating the on-orbit observation method for weak satellite targets based on intelligent closed-loop feedback provided by the present invention. Detailed Implementation

[0055] The technical solutions of 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 a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0056] It should be understood that the illustrative drawings are not drawn to scale. The flowcharts used in this invention illustrate operations implemented according to some embodiments of the invention. It should be understood that the operations in the flowcharts may be implemented out of order, and steps without logical contextual relationships may be reversed or performed simultaneously. Furthermore, those skilled in the art, guided by the content of this invention, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor systems and / or microcontroller systems.

[0057] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0058] This invention provides a satellite-based on-orbit observation system and method for weak targets based on intelligent closed-loop feedback, which will be described in detail below.

[0059] Figure 1 The structural schematic diagram of one embodiment of the satellite weak target on-orbit observation system based on intelligent closed-loop feedback provided by the present invention is as follows: Figure 1As shown, the satellite weak target on-orbit observation system 10 based on intelligent closed-loop feedback includes a detection module 100, a tracking module 200, a control module 300, and a closed-loop scheduling module 400 connected in a sequential closed loop. The closed-loop scheduling module 400 sends scheduling signals to the tracking module 200 and the control module 300, respectively.

[0060] The detection module 100 is used to perform single-frame image enhancement and temporal consistency enhancement on the infrared images acquired on the satellite, obtain enhancement features, and obtain a set of candidate targets based on the enhancement features. The set of candidate targets includes multiple candidate targets and the candidate position, detection confidence and depth representation vector of each candidate target.

[0061] Specifically, the candidate target set can be represented as:

[0062]

[0063] In the formula, Let be the set of candidate targets corresponding to the infrared image at time t; Let be the two-dimensional coordinates of the candidate position of the i-th candidate target; Let be the candidate confidence level of the i-th candidate target; Let be the depth representation vector of the i-th candidate target; The total number of candidate targets.

[0064] The tracking module 200 is used to perform target matching within a preset search area based on the similarity between the depth representation vector and the tracking template, as well as the detection confidence. When a match is successful, the matched candidate position is used as an observation and input into a Kalman filter for state update to obtain the updated target state vector. When a match fails, the predicted state of the Kalman filter is used as the target state vector, and the search window is expanded with the predicted position corresponding to the predicted state as the center to perform re-identification to resume tracking.

[0065] Among them, the state vector in the Kalman filter and observation vector They are respectively:

[0066]

[0067]

[0068] In the formula, The two-dimensional position of the candidate target on the image plane as output by the Kalman filter; The velocity components of the candidate target on the x and y axes are the output of the Kalman filter. These are the candidate position coordinates of the candidate targets obtained by the detection module.

[0069] The state transition model and observation model in the Kalman filter are as follows:

[0070]

[0071]

[0072] In the formula, The predicted state at time t is obtained based on the state estimate at time t-1; State estimation at time t-1; This is the state transition matrix; This is process noise, used to represent uncertainties in state transitions; This is the observation vector, which is usually the candidate location of the candidate target; This is the observation matrix, used to map the state to the observation space; This represents the true state at time t; Observation noise is used to represent the uncertainty in observations.

[0073] Among them, the commonly used and The format is:

[0074]

[0075] .

[0076] It is a time variable.

[0077] Kalman filtering estimates the state through two steps: prediction and update.

[0078] 1. Prediction steps: Determine the predicted state at time t based on the state transition model;

[0079] 2. Update steps: Correct the prediction based on the observation vector to obtain the estimated state.

[0080] The control module 300 is used to map the target state vector into attitude error, generate control commands based on the attitude error and a single-neuron adaptive PID controller, and obtain the execution error after executing the control commands.

[0081] Among them, the single-neuron adaptive PID controller can improve the steady-state error by about 30%–45% compared with the traditional fixed PID controller, and shows better robustness and convergence in terms of disturbance response (such as inertia parameter ±10% to ±20%) and control overshoot control, with the steady-state time shortened by an average of about 20%–30%.

[0082] The closed-loop scheduling module 400 is used to dynamically adjust the operating parameters of at least one of the detection module 100, tracking module 200 and control module 300 based on the execution error and detection confidence.

[0083] It should be noted that: in the embodiments of the present invention, the satellite too small target refers to the target with a scale of 1-3 pixels in the infrared image, which is usually space debris, microsatellite, or stealth aircraft.

[0084] Compared with existing technologies, the satellite weak target on-orbit observation system 10 based on intelligent closed-loop feedback provided in this embodiment of the invention deeply couples the three originally independent links of detection, tracking, and control by introducing a closed-loop scheduling module 400, constructing an integrated operation architecture of "perception-estimation-decision-execution-feedback". Information flows bidirectionally between the detection module 100, the tracking module 200, and the control module 300, enabling the system to self-optimize based on the final execution effect and the front-end perception confidence level, thereby significantly reducing dependence on external intervention (such as ground commands), improving the satellite's long-term autonomous observation capability in complex dynamic scenarios, and thus significantly enhancing on-orbit autonomy. Experiments have shown that this system is superior to the separate scheme in ensuring mission continuity in long-term operation scenarios, and its performance degradation rate is reduced by 15%-25% in long-term missions.

[0085] Furthermore, the detection module 100 sequentially performs image enhancement and temporal enhancement on the infrared images, suppressing noise at the pixel level and temporal level respectively, thereby improving the signal-to-noise ratio from both spatial and temporal dimensions. This collaborative enhancement mechanism effectively suppresses background noise and maintains the consistency of the appearance of weak targets across multiple frames, thus providing subsequent stages with a more confident and reliable set of candidate targets, ensuring the success rate of the entire observation mission from the source.

[0086] Furthermore, the tracking module 200 employs a joint strategy of deep feature matching and Kalman filtering, and includes a pre-defined extended search and re-identification mechanism for matching failures, cleverly balancing high-precision positioning with occlusion robustness. This design ensures that the target can be accurately locked when visible, maintains its trajectory based on motion prediction when briefly lost, and achieves rapid and accurate re-acquisition, fundamentally solving the tracking interruption problem caused by target occlusion and weak signals.

[0087] Meanwhile, the control module 300 employs a single-neuron adaptive PID controller, whose parameters can be dynamically adjusted in real time based on attitude error, rather than remaining fixed. This self-learning capability enables the controller to automatically adapt to changes in the satellite platform's dynamic characteristics and external space disturbances, effectively suppressing overshoot and abrupt changes in control inputs while maintaining rapid maneuver response, thereby achieving higher steady-state control accuracy and stronger robustness.

[0088] In some embodiments of the present invention, such as Figure 2 As shown, the detection module 100 includes an image enhancement unit 110, a temporal consistency enhancement unit 120, and a detection unit 130 connected in sequence.

[0089] The image enhancement unit 110 is used to perform image enhancement processing on infrared images based on the Generative Adversarial Network (GAN) model to obtain image enhancement features.

[0090] Specifically, a Generative Adversarial Network (GAN) model consists of a generator and a discriminator. A GAN model requires a training phase before deployment. Both the generator and discriminator participate in the training phase. The generator learns to map noisy images to sharp images, while the discriminator learns to distinguish between real, sharp images and images generated by the generator. In the deployment phase, only the generator is used; the discriminator is not used and its purpose is to train a high-quality generator. The generator is used in the deployment phase to perform image enhancement processing on the infrared images.

[0091] The generator uses the existing U-Net with skip connections and adds a super-resolution sub-module to the upsampling path to improve the detail discernibility of small targets. This super-resolution sub-module can adopt existing super-resolution network structures, such as an upsampling module based on residual structures. Its function is to improve spatial resolution and texture details, making the structure of small targets, such as those with a resolution of 1–3 pixels, clearer. The intermediate features (bottleneck layer features between the U-Net encoder and decoder) output dimension is 256, which provides the generator with higher-dimensional feature representation capabilities to enhance the texture and structural information of small targets. The discriminator uses PatchGAN.

[0092] It should be noted that, to ensure the generated images are more realistic, suppress pixel structure differences from the original images, and preserve high-level semantic features, in a specific embodiment of this invention, the total loss function used during the training of the generative adversarial network model is a weighted sum of adversarial loss, reconstruction loss, and feature loss, specifically:

[0093]

[0094] In the formula, This is the total loss function; To combat the losses; For reconstruction losses; For feature loss; yes abbreviation, yes abbreviation, yes abbreviation; , , Weighting coefficients for adversarial loss, reconstruction loss, and feature loss; For the discriminator to distinguish real images The discrimination probability, This represents the probability that the discriminator classifies the image generated by the generator. Images generated by the generator With input infrared image The L1 distance between them; For feature extraction operators, This represents the L2 distance between the extracted features obtained by performing feature extraction on the image generated by the generator and the input infrared image respectively. This represents the set of all learnable parameters in the generator, used to learn the mapping from noisy, weak target images to sharper, enhanced images. This represents all learnable parameters of the discriminator, used to distinguish between real, sharp images and enhanced images generated by the generator; This refers to the distribution of "true, clear images". Expectations This indicates the distribution of the "enhanced input or generator output image". Expectations Output image of generator Features after feature extraction and the real image The characteristic differences are taken as expected.

[0095] It's important to note that the adversarial loss term drives the generator to produce an output that is difficult to distinguish from a real, clear image in terms of overall appearance and texture, ensuring that the generated image meets visual fidelity requirements while suppressing unnatural artifacts and noise. The reconstruction loss ensures that the generator does not alter the image's fundamental content and structure when enhancing it, and constrains the generator's output from deviating too far from the input, preventing the generation of content irrelevant to the input. The feature loss, beyond pixel-level reconstruction, ensures that the generated image maintains consistency with the input image in high-level semantic features, and drives the generator to produce an image that is visually perceptually consistent with the input, rather than simply a pixel-level match.

[0096] The temporal consistency enhancement unit 120 is used to perform temporal consistency enhancement processing on image enhancement features based on the structured space model (SSM) to obtain enhanced features.

[0097] In a specific embodiment of the present invention, the state-space model is the Mamba model, and its working principle is as follows:

[0098] At each moment, the feature input of the current frame is received, and combined with the historical hidden state, it is updated to a new hidden state. The output feature is a temporally smoothed feature containing contextual information. This output feature can sense whether the current target has continued to appear in the past few frames, thus helping to filter out noise points that only appear in one frame.

[0099] Specifically, the working principle of the Mamba model can be expressed as follows:

[0100]

[0101] In the formula, Input the features of the current frame; Let be the hidden state at time t; This represents the hidden state at time t+1; For output features; , , , This represents the learnable matrix corresponding to each feature and state.

[0102] In a specific embodiment of the present invention, the hidden state dimension is 64-128, and the Mamba timing window is 12 frames.

[0103] The detection unit 130 is used to perform target detection based on the enhanced features of the detection head to obtain a set of candidate targets.

[0104] On the one hand, the embodiments of the present invention utilize a generative adversarial network model to unify background noise in a single frame and enhance pixel-level details of weak targets. On the other hand, a state-space model is introduced to suppress sudden noise using temporal context, smooth the detection jitter caused by short-term noise suppression, maintain the temporal consistency of the target across multiple frames, and provide temporal support for the existence of the target, thereby providing stable and high-confidence observations for tracking.

[0105] To further improve the accuracy of candidate target identification, in some embodiments of the present invention, such as Figure 2 As shown, the detection module 100 also includes an image preprocessing unit 140, which is connected to the image enhancement unit 110;

[0106] The image preprocessing unit 140 is used to perform radiometric correction and local contrast normalization on the infrared image.

[0107] Among them, radiometric correction can eliminate the uneven response caused by the image acquisition equipment itself, and local contrast normalization can enhance the difference between weak targets and local background. By performing radiometric correction and local contrast normalization on infrared images, the image signal-to-noise ratio can be further improved, providing a more solid data foundation for subsequent target detection and improving the accuracy of subsequent candidate targets.

[0108] In some embodiments of the present invention, such as Figure 3 As shown, the tracking module 200 includes a similarity determination unit 210, a target state vector determination unit 220, and a re-identification unit 230 connected in sequence.

[0109] The similarity determination unit 210 is used to calculate the cosine similarity between the depth representation vector and the tracking template within a preset search area. When the cosine similarity is greater than the similarity threshold and the detection confidence is greater than the confidence threshold, the match is determined to be successful. When the cosine similarity is less than or equal to the similarity threshold or the detection confidence is less than or equal to the confidence threshold, the match is determined to be unsuccessful.

[0110] Specifically, the cosine similarity is:

[0111]

[0112] In the formula, Depth representation vector and tracking template Cosine similarity; The magnitude of the depth representation vector; To track the template's modulus.

[0113] Both the similarity threshold and the confidence threshold can be set or adjusted according to the actual application scenario.

[0114] The target state vector determination unit 220 is used to update the state by taking the matched candidate position as the observation value when the matching is successful and inputting it into the Kalman filter; when the matching fails, the predicted state of the Kalman filter is used as the target state vector.

[0115] The re-identification unit 230 is used to expand the search window centered on the predicted position corresponding to the predicted state when the matching fails, and determine the joint score of each candidate target in the expanded search window, determine the optimal candidate target with the highest joint score, and determine whether the joint score of the optimal candidate target is greater than the score threshold. If it is greater, the position of the optimal candidate target is used as the observation value and input into the Kalman filter for state update to obtain the updated target state vector; if it is less than or equal to, a predefined security policy is triggered.

[0116] The search window can be expanded by a factor of 1.5-2.0, which means the search window is enlarged by 1.5-2 times.

[0117] Specifically, the predefined security strategy is: to enter passive tracking or report ground instructions.

[0118] This invention employs a strategy of switching between deep feature matching and target state vector determination modes. This strategy prioritizes high-confidence observations when they are available, and prioritizes the Kalman filter's predictions when no high-confidence observations exist. In other words, when a match is successful, the system enters an observation-dominated mode; when a match fails, it enters a prediction-dominated mode. Furthermore, in subsequent frames, the system attempts to re-identify using an expanded search window and a scoring threshold. This approach balances matching accuracy and occlusion robustness, avoiding frequent loss during brief occlusions and reducing the risk of mismatches.

[0119] Furthermore, after a match fails, the system re-identifies the target and resumes tracking by expanding the search window, further improving the accuracy of target tracking.

[0120] It should be noted that target matching is based on the matching degree between the depth representation vector and the tracking template. In order to adapt to changes in the appearance of the tracked target, the online representation of the tracking template needs to be refreshed after a successful match to adapt to the changes in the appearance of the tracked target and further improve the accuracy of target tracking.

[0121] In a specific embodiment of the present invention, the joint scoring is as follows:

[0122]

[0123] In the formula, For joint scoring; Candidate positions and the predicted position of the Kalman filter The Euclidean distance between them; Depth representation vector and tracking template Cosine similarity; To test the confidence level; , , These are the weighting coefficients.

[0124] In one specific embodiment of the present invention, , , The values ​​are 0.6, 0.3, and 0.1, respectively.

[0125] To verify the effectiveness of mode switching, an experiment was conducted. The experiment showed that the tracking success rate of the embodiment of the present invention was about 92% when the cross-union ratio was greater than 0.5, and the average re-identification delay was shortened by about 30-40%.

[0126] In some embodiments of the present invention, such as Figure 4 As shown, the control module 300 includes an attitude error determination unit 310, a control law determination unit 320, and a control command generation unit 330 connected in sequence.

[0127] The attitude error determination unit 310 is used to map the position information in the target state vector into the attitude angle error of the satellite.

[0128] Specifically, the position information in the target state vector is mapped to the desired line-of-sight angle (or camera pointing), and the difference between the desired line-of-sight angle and the current actual line-of-sight angle is taken as the attitude angle error.

[0129] The control objective of the control module 300 is to make the attitude angle error converge in the shortest possible time while suppressing sudden changes in the control quantity.

[0130] The control law determination unit 320 is used to obtain the neuron weight vector of the current control cycle, take the attitude angle error, error integral and error derivative as input, map the neuron weight vector to obtain the proportional, integral and derivative parameters of the single neuron adaptive PID controller, and determine the control torque based on the proportional, integral and derivative parameters.

[0131] The control torque can be expressed as:

[0132]

[0133] In the formula, To control the torque; This is a proportional parameter; For integration parameters; For differential parameters; This refers to the attitude angle error; For the error integral; The derivative of the error; For the current moment, This is the initial moment when integration begins, i.e., the point in time when the entire control process starts.

[0134] in, Weight vector of a single neuron Through mapping The result of the conversion is:

[0135] .

[0136] The control command generation unit 330 is used to perform amplitude limiting processing on the control torque and generate control commands.

[0137] Specifically, the amplitude limiting process is as follows:

[0138]

[0139] In the formula, For control commands; To control the torque; This is the maximum torque; This is the limit operator, specifically: when When, set ,when When, set .

[0140] The embodiments of the present invention, by performing amplitude limiting processing on the control torque, can prevent control commands from exceeding the actuator's capabilities and improve the rationality of control command generation.

[0141] Due to the control parameters of the single-neuron adaptive PID controller Obtained by mapping neuron weight vectors, therefore, in order to improve the adaptability of the single neuron adaptive PID controller, in some embodiments of the present invention, the control module 300 further includes a neuron weight vector update unit 340;

[0142] The neuron weight vector update unit 340 is used to obtain the attitude angle error and control increment of the current control cycle, construct a sliding window quadratic cost function based on the attitude angle error and control increment, update the neuron weight vector based on the sliding window quadratic cost function, and use the updated neuron weight vector as the neuron weight vector for the next control cycle.

[0143] In a specific embodiment of the present invention, the sliding window quadratic cost function is:

[0144]

[0145]

[0146] In the formula, The cost is the cost of a quadratic form for a sliding window; The length of the sliding window; The current moment; The summation index variable represents each moment within the sliding window; This refers to the attitude angle error; To control the increment; , These are the weighting coefficients; The control torque for the current control cycle; This is the control torque of the previous control cycle.

[0147] The weights of a single neuron are then updated according to the gradient descent direction:

[0148]

[0149]

[0150] In the formula, This is the neuron weight vector for the next control cycle; This represents the neuron weight vector for the current control cycle; The descent gradient of the quadratic cost function of the sliding window; The learning rate for the current control cycle; The learning rate for the next control cycle; It is a scaling factor; This represents the average error of the most recent N frames; This represents the average error change over the most recent N frames; For experience or design functions.

[0151] The embodiments of the present invention introduce a gain-recursive learning rate adjustment method, which can avoid oscillations or slow convergence due to fixed learning rate noise. At the same time, by adaptively updating the neuron weight vector, the accuracy and rationality of the single neuron adaptive PID controller can be improved.

[0152] To verify the effectiveness of the single-neuron adaptive PID controller proposed in this embodiment of the invention, experimental verification was conducted, and the verification results are as follows: Figure 5 As shown, Figure 5 The solid red line represents the simulation results of the single-neuron adaptive PID controller in this embodiment of the invention, while the dashed blue line represents the simulation results of the traditional fixed-parameter PID controller. Figure 5 As can be seen, the single-neuron adaptive PID controller of this invention can complete convergence in about 18 seconds, with a small oscillation amplitude and a steady-state time shortened by about 20%-30%.

[0153] In a specific embodiment of the present invention, the closed-loop scheduling module 400 is specifically used for:

[0154] When the detection confidence level is consistently below the detection confidence level threshold, or the execution error is consistently above the execution error threshold, the search area is increased, and / or the weight of the reconstruction loss is increased to enhance the perception capability.

[0155] When the amplitude of the control torque of the control module is greater than the saturation limit, the learning rate of the single neuron adaptive PID controller is reduced, and / or the control parameters of the single neuron adaptive PID controller are switched to preset conservative control parameters.

[0156] The saturation limit value is the product of the maximum torque and the preset percentage.

[0157] When a long-term time-varying environment is detected, the network structure parameters in the detection module 100 are adjusted.

[0158] Specifically, when the background noise changes significantly or re-identification fails repeatedly, it is determined that a long-term variable environment has occurred.

[0159] The closed-loop scheduling module 400 of this invention automatically adjusts sensing and estimation parameters based on execution feedback and can perform limited online fine-tuning, reducing ground intervention and improving long-term on-orbit continuous observation capabilities. During long-term observation, the detection accuracy degradation rate decreases by approximately 15%–20%, and the trajectory loss rate decreases by approximately 30%–35%.

[0160] To verify the superiority of the satellite weak target on-orbit observation system based on intelligent closed-loop feedback proposed in this invention in terms of detection, tracking, and control, the system of this invention is compared with the performance indicators of traditional distributed control systems in terms of weak target detection signal-to-noise ratio, detection accuracy, tracking success rate, and control steady-state error. The comparison results are as follows: Figure 6 As shown, Figure 6 The black rectangles in the diagram represent the performance indicators of the system proposed in this invention, while the gray rectangles represent the performance indicators of traditional distributed systems. Figure 6 As can be seen, the signal-to-noise ratio is improved by about 20%, the detection accuracy is improved by about 20%, the tracking success rate is improved by about 15%, and the control steady-state error is reduced by about 46%. That is, in terms of performance indicators in multiple dimensions, the system of the present invention outperforms the traditional distributed system, verifying the superiority of the satellite weak target on-orbit observation system based on intelligent closed-loop feedback proposed in this invention.

[0161] In summary, the embodiments of this invention: 1. Integrate GAN-Mamba-based temporal enhanced detection, deep representation-driven joint tracking, and single-neuron online adaptive PID control into a unified closed-loop framework. Information is fed back in real time between detection, tracking, and control, significantly improving the real-time performance and autonomy of the overall system. Experiments demonstrate that this closed-loop system outperforms separate solutions in ensuring task continuity during long-term operation. 2. GAN enhances noise suppression at the pixel level, while Mamba suppresses sudden noise at the temporal level. Together, they significantly improve the detection confidence and stability of weak targets at the 1-3 pixel level. 3. Deep feature matching ensures high-precision localization when visible, Kalman prediction maintains trajectory continuity during occlusion or short-term loss, and joint decision and re-identification strategies achieve fast and low-mismatch re-capture. Tracking success rate reaches approximately 92% in occluded scenarios, and re-identification latency is reduced by an average of approximately 30%–40%. 4. The single-neuron adaptive PID combined with quadratic cost and gain recursion can suppress overshoot and control variable mutations while maintaining fast maneuver response, improving steady-state accuracy and robustness under inertial parameter drift or complex disturbances. Steady-state error is improved by approximately 30%–45%, and steady-state time is shortened by approximately 20%–30%. 5. The closed-loop scheduling module automatically adjusts sensing and estimation parameters based on execution feedback and can perform limited online fine-tuning, reducing ground intervention and improving long-term on-orbit continuous observation capabilities. During long-term observation, the detection accuracy degradation rate decreases by approximately 15%–20%, and the trajectory loss rate decreases by approximately 30%–35%.

[0162] On the other hand, this embodiment of the invention also provides a method for on-orbit observation of weak satellite targets based on intelligent closed-loop feedback, applicable to any of the on-orbit observation systems for weak satellite targets based on intelligent closed-loop feedback in this embodiment, such as... Figure 7 As shown, the on-orbit observation method for weak targets on satellites based on intelligent closed-loop feedback includes:

[0163] S701. Perform single-frame image enhancement and temporal consistency enhancement on the infrared images acquired on the satellite to obtain enhancement features, and obtain a set of candidate targets based on the enhancement features. The set of candidate targets includes multiple candidate targets and the candidate position, detection confidence and depth representation vector of each candidate target.

[0164] S702. Within a preset search area, target matching is performed based on the similarity between the depth representation vector and the tracking template, as well as the detection confidence. When a match is successful, the matched candidate position is used as an observation and input into a Kalman filter for state update to obtain the updated target state vector. When a match fails, the predicted state of the Kalman filter is used as the target state vector, and the search window is expanded with the predicted position corresponding to the predicted state as the center to perform re-identification and resume tracking.

[0165] S703. Map the target state vector to attitude error, generate control commands based on attitude error and single-neuron adaptive PID controller, and obtain the execution error after executing the control commands.

[0166] S704. Dynamically adjust the operating parameters of at least one of the detection module, tracking module, and control module based on the execution error and detection confidence level.

[0167] The satellite weak target on-orbit observation method based on intelligent closed-loop feedback provided in the above embodiments can realize the technical solution described in the above embodiments of the satellite weak target on-orbit observation system based on intelligent closed-loop feedback. The specific implementation principle of each module or unit can be found in the corresponding content in the above embodiments of the satellite weak target on-orbit observation method based on intelligent closed-loop feedback, and will not be repeated here.

[0168] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.), and the computer program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0169] The above provides a detailed description of the satellite weak target on-orbit observation system and method based on intelligent closed-loop feedback provided by the present invention. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A satellite-based on-orbit observation system for weak targets based on intelligent closed-loop feedback, characterized in that, include: The detection module is used to perform single-frame image enhancement and temporal consistency enhancement on the infrared images acquired on the satellite, obtain enhancement features, and obtain a candidate target set based on the enhancement features. The candidate target set includes multiple candidate targets and the candidate position, detection confidence and depth representation vector of each candidate target. The tracking module is used to perform target matching within a preset search area based on the similarity between the depth representation vector and the tracking template, as well as the detection confidence. When a match is successful, the matched candidate position is used as the observation value and input into the Kalman filter for state update to obtain the updated target state vector; when a match fails, the predicted state of the Kalman filter is used as the target state vector, and the search window is expanded with the predicted position corresponding to the predicted state as the center to perform re-identification to resume tracking. The control module is used to map the target state vector into an attitude error, generate control instructions based on the attitude error and a single-neuron adaptive PID controller, and obtain the execution error after executing the control instructions. A closed-loop scheduling module is used to dynamically adjust the operating parameters of at least one of the detection module, the tracking module, and the control module based on the execution error and the detection confidence level.

2. The satellite weak target on-orbit observation system based on intelligent closed-loop feedback according to claim 1, characterized in that, The detection module includes an image enhancement unit, a temporal consistency enhancement unit, and a detection unit; The image enhancement unit is used to perform image enhancement processing on the infrared image based on a generative adversarial network model to obtain image enhancement features; The temporal consistency enhancement unit is used to perform temporal consistency enhancement processing on the image enhancement features based on the state-space model to obtain the enhancement features; The detection unit is used to perform target detection based on the enhanced features using the detection head to obtain the candidate target set.

3. The satellite weak target on-orbit observation system based on intelligent closed-loop feedback according to claim 2, characterized in that, The detection module also includes an image preprocessing unit; The image preprocessing unit is used to perform radiometric correction and local contrast normalization on the infrared image.

4. The satellite weak target on-orbit observation system based on intelligent closed-loop feedback according to claim 1, characterized in that, The tracking module includes a similarity determination unit, a target state vector determination unit, and a re-identification unit; The similarity determination unit is used to calculate the cosine similarity between the depth representation vector and the tracking template within a preset search area. When the cosine similarity is greater than the similarity threshold and the detection confidence is greater than the confidence threshold, the match is determined to be successful. When the cosine similarity is less than or equal to the similarity threshold or the detection confidence is less than or equal to the confidence threshold, the match is determined to be unsuccessful. The target state vector determination unit is used to, when a match is successful, take the matched candidate position as the observation value and input it into the Kalman filter to update the state and obtain the updated target state vector; when a match fails, the predicted state of the Kalman filter is used as the target state vector. The re-identification unit is used to expand the search window centered on the predicted position corresponding to the predicted state when the matching fails, and to determine the joint score of each candidate target in the expanded search window, determine the optimal candidate target with the highest joint score, and determine whether the joint score of the optimal candidate target is greater than the score threshold. If it is greater, the position of the optimal candidate target is used as the observation value and input into the Kalman filter for state update to obtain the updated target state vector; if it is less than or equal to, a predefined security policy is triggered.

5. The satellite weak target on-orbit observation system based on intelligent closed-loop feedback according to claim 4, characterized in that, The joint score is: In the formula, For joint scoring; Candidate positions and the predicted position of the Kalman filter The Euclidean distance between them; Depth representation vector and tracking template Cosine similarity; To test the confidence level; , , These are the weighting coefficients.

6. The satellite weak target on-orbit observation system based on intelligent closed-loop feedback according to claim 1, characterized in that, The control module includes an attitude error determination unit, a control law determination unit, and a control command generation unit; The attitude error determination unit is used to map the position information in the target state vector into the attitude angle error of the satellite; The control law determination unit is used to obtain the neuron weight vector of the current control cycle, take the attitude angle error, error integral and error derivative as input, map the neuron weight vector to obtain the proportional, integral and derivative parameters of the single neuron adaptive PID controller, and determine the control torque according to the proportional, integral and derivative parameters; The control command generation unit is used to perform amplitude limiting processing on the control torque and generate the control command.

7. The satellite weak target on-orbit observation system based on intelligent closed-loop feedback according to claim 6, characterized in that, The control module also includes a neuron weight vector update unit; The neuron weight vector update unit is used to obtain the attitude angle error and control increment of the current control cycle, construct a sliding window quadratic cost function based on the attitude angle error and the control increment, update the neuron weight vector based on the sliding window quadratic cost function, and use the updated neuron weight vector as the neuron weight vector for the next control cycle.

8. The satellite weak target on-orbit observation system based on intelligent closed-loop feedback according to claim 7, characterized in that, The sliding window quadratic cost function is: In the formula, The cost is the cost of the sliding window quadratic form; The length of the sliding window; The current moment; The summation index variable represents each moment within the sliding window; This refers to the attitude angle error; To control the increment; , These are the weighting coefficients; The control torque for the current control cycle; This is the control torque of the previous control cycle.

9. The satellite weak target on-orbit observation system based on intelligent closed-loop feedback according to claim 1, characterized in that, The loss function used in the detection module is a weighted sum of adversarial loss, reconstruction loss, and feature loss; the closed-loop scheduling module is specifically used for: When the detection confidence level is consistently lower than the detection confidence level threshold, or the execution error is consistently greater than the execution error threshold, the search area of ​​the tracking module is increased, and / or the weight of the reconstruction loss is increased; When the amplitude of the control torque of the control module is greater than the saturation limit, the learning rate of the single neuron adaptive PID controller is reduced, and / or the control parameters of the single neuron adaptive PID controller are switched to preset conservative control parameters. When a long-term time-varying environment is detected, the network structure parameters in the detection module are adjusted.

10. A method for on-orbit observation of weak targets on satellites based on intelligent closed-loop feedback, characterized in that, The method applicable to the satellite-based on-orbit observation system for weak targets based on intelligent closed-loop feedback as described in any one of claims 1-9, comprises: The infrared images acquired on the satellite are enhanced with single-frame image enhancement and temporal consistency enhancement to obtain enhancement features. Based on the enhancement features, a set of candidate targets is obtained. The set of candidate targets includes multiple candidate targets and the candidate position, detection confidence and depth representation vector of each candidate target. Within a preset search area, target matching is performed based on the similarity between the depth representation vector and the tracking template, as well as the detection confidence. When a match is successful, the matched candidate position is used as an observation and input into a Kalman filter for state update to obtain the updated target state vector. When a match fails, the predicted state of the Kalman filter is used as the target state vector, and the search window is expanded with the predicted position corresponding to the predicted state as the center to perform re-identification and resume tracking. The target state vector is mapped to an attitude error, and control instructions are generated based on the attitude error and a single-neuron adaptive PID controller. The execution error after executing the control instructions is then obtained. The operating parameters of at least one of the detection module, tracking module, and control module are dynamically adjusted based on the execution error and the detection confidence level.

Images