An online adaptive method for domain shift compensation in pose tracking of non-cooperative spacecraft in orbit

Abstract

This invention belongs to the field of on-orbit non-cooperative spacecraft pose tracking technology, specifically involving an online adaptive method to compensate for domain offset in on-orbit non-cooperative spacecraft pose tracking. Step 1: Represent the target spacecraft as 11 three-dimensional key points; extract the key point bounding boxes after projection, and crop to obtain the input RoI; train an asymmetric encoder-decoder network to regress the key point heatmap; Step 2: Design a dynamic memory library M based on a first-in-first-out queue; Step 3: The student model receives complete samples from M, and the teacher model receives masked samples from M; consistency constraints are implemented by minimizing the weighted MSE loss between the student and teacher predicted heatmaps; Step 4: Calculate the average value of the [CLS] token features of all training images as the source domain global class prototype, and align the source domain prototype with the newly calculated class prototypes of the teacher and student models. This invention makes it possible to achieve high-precision, highly robust, and highly practical real-time online tracking of non-cooperative spacecraft pose.

Description

Technical Field

[0001] This invention belongs to the field of on-orbit non-cooperative spacecraft attitude tracking technology, specifically relating to an online adaptation method for compensating for domain offset in on-orbit non-cooperative spacecraft attitude tracking. Background Technology

[0002] In practical applications, especially in on-orbit servicing missions, computer vision models often face challenges due to the dual constraints of limited onboard resources and the scarcity of real-world spatial visual data. To address this, researchers typically use rendering tools and game engines to generate synthetic images and build ground-based robot testbeds to simulate the space environment. However, because the simulated environment differs significantly from real-world scenes in terms of lighting conditions, material properties, and sensor noise distribution (collectively referred to as "domain offset"), models trained solely on synthetic data experience a significant performance degradation when deployed in real-world scenarios.

[0003] To compensate for domain offset, unsupervised domain adaptation (UDA) and source-free domain adaptation (SFDA) have become research frontiers in recent years. These methods mainly achieve cross-domain feature alignment through adversarial training or self-supervised learning, with adversarial training showing state-of-the-art performance. Achieving effective and stable knowledge transfer hinges on imposing cross-domain alignment constraints that preserve semantic invariance. However, existing methods...

[0004] UDA / SFDA methods typically require multiple rounds of iterative training using source domain data or large amounts of target domain data during the adaptation process, which is incompatible with on-orbit servicing scenarios. Direct access to raw source domain data is impractical due to onboard storage and privacy constraints. Furthermore, opportunities to observe non-cooperative spacecraft are limited, and onboard acquisition, transmission, and processing of data equivalent to a complete test set would consume significant time, bandwidth, and computational resources.

[0005] Against this backdrop, Online Test-Time Adaptation (OTTA), a technique for real-time model updates under extremely limited observation samples in the target domain, offers a potential solution to the aforementioned challenges. However, traditional OTTA methods are susceptible to catastrophic forgetting and error accumulation. Furthermore, the choice of model architecture is crucial. Unlike the local convolutional operations of Convolutional Neural Networks (CNNs), which are easily affected by input perturbations such as texture changes and occlusion, the Vision Transformer (ViT), with its global attention mechanism, can effectively capture robust high-level semantic features shared across domains, such as the overall shape of an object and the relationships between its components. This has led to ViT gradually replacing CNNs as the mainstream visual backbone network. However, existing mature OTTA strategies are mainly designed around CNN architectures, and their effectiveness and optimization on ViT still require further exploration.

[0006] Furthermore, researchers are actively utilizing self-supervised learning strategies to achieve feature learning on unlabeled data, enhancing the model's generalization ability by designing diverse pre-training tasks. In particular, ViT self-supervised pre-trained weights obtained on large datasets, such as ImageNet-21K, have been shown to significantly enhance the model's robustness in perceiving geometric structures, providing a strong initial starting point for subsequent tasks. However, simply fine-tuning these pre-trained weights on source domain data can easily lead to overfitting the model to the source domain distribution, thereby weakening its generalization performance in the target domain.

[0007] Therefore, a robust and efficient online adaptation method is urgently needed to address the domain offset challenges faced in critical missions such as pose tracking of non-cooperative spacecraft in orbit. To this end, this invention proposes an online adaptation method to compensate for domain offset in pose tracking of non-cooperative spacecraft in orbit. This method must be able to achieve rapid and robust online optimization of the model using limited real target domain observations, while meeting onboard resource constraints and data privacy requirements, avoiding forgetting and error accumulation, thereby ensuring high-precision tracking performance in complex and ever-changing space scenarios. Summary of the Invention

[0008] The purpose of this invention is to provide an online adaptive method to compensate for domain offsets in pose tracking of non-cooperative spacecraft in orbit. By integrating a dynamic memory library mechanism with a mentor-apprentice collaborative learning framework, this method achieves robust online adaptability. Specifically, a dynamic memory library based on a first-in, first-out queue is introduced to store newly arrived high-confidence target domain samples and their model inference results in real time. The problem of catastrophic forgetting is mitigated through self-training on historical samples. A source domain class prototype is constructed using a pre-trained model, and the mentor-apprentice collaborative learning framework is initialized. Prototype alignment and space heat dissipation are then used to address this issue.Figure 1 Through consistency constraints, this framework achieves cross-domain semantic alignment and generates reliable pseudo-labels.

[0009] The specific technical solution adopted by this invention is as follows:

[0010] An online adaptation method for compensating for domain offsets during pose tracking of non-cooperative spacecraft in orbit includes the following steps:

[0011] Step 1: Represent the target spacecraft as 11 3D keypoints. Establish a 2D-3D correspondence based on the camera projection model, extract the keypoint bounding boxes after projection, and crop them to obtain the input RoI. Train an asymmetric encoder-decoder network to regress the keypoint heatmap, thereby initializing the teacher-student distillation framework. To address spacecraft image domain differences, apply data augmentation (flare, histogram equalization, blur, noise) during source domain pre-training. The loss function for heatmap training is the MSE of the predicted heatmap versus the ground truth.

[0012] Step 2: To mitigate the catastrophic forgetting problem in online adaptation, a dynamic memory library M based on a first-in, first-out (FIFO) queue is designed. Initially, M is filled once with source domain samples and their ground truth values. When target domain data is input in a streaming manner, a student model is used to infer on single or mini-batch samples. Only the sample with the highest confidence in the batch and its heatmap pseudo-label are stored in M. In each adaptation iteration, all samples in M ​​are used for self-training. If the model's confidence in a newly generated pseudo-label for a sample in M ​​is higher than the old value in the library, then the pseudo-label is updated.

[0013] Step 3: The student model receives the complete samples in M, and the teacher model receives the masked samples in M. A consistency constraint is implemented by minimizing the weighted MSE loss between the student and teacher predicted heatmaps. This loss is used to update the student parameters, while the teacher updates its parameters using an exponential moving average.

[0014] Step 4: After source domain pre-training, align the source domain prototype with the newly calculated class prototypes of the teacher and student models; specifically, calculate the average value of the [CLS] token features of all training images as the source domain global class prototype. During online adaptation in the target domain, the teacher and student networks output [CLS] token feature vectors for the complete and masked target images, respectively. By minimizing the cosine distance loss between these feature vectors and the source domain global class prototype, the model is encouraged to generate global feature representations in the target domain that are semantically consistent with the source domain.

[0015] The technical effects achieved by this invention are as follows:

[0016] This invention addresses the severe domain migration challenge in on-orbit non-cooperative spacecraft pose tracking missions by proposing an innovative OTTA method. Its core advantage lies in achieving online optimization without accessing the original source domain data, utilizing only a very limited stream of target domain observations. Specifically, the effects of this invention are reflected in the following aspects:

[0017] In this invention, the designed dynamic memory library creates an online, high-quality knowledge replay mechanism through its unique historical sample and pseudo-label dynamic update mechanism. In each adaptation iteration, the model self-trains on the "reliable memory" stored in M, continuously reinforcing the learned stable knowledge, thereby significantly suppressing the loss of old knowledge caused by adapting to new samples.

[0018] In this invention, a pre-trained model is used to construct source domain class prototypes, while simultaneously initializing a teacher-student collaborative learning framework. This is achieved through prototype alignment and spatial heat dissipation. Figure 1 Consistency constraints force the model to pay more attention to high-level geometric structures shared across domains (such as rigid body shapes and spatial relationships between parts) rather than volatile low-level textures or noise, thereby aligning cross-domain semantics and generating reliable pseudo-labels.

[0019] This method is not designed around the traditional CNN architecture. Its core modules (mask distillation and class prototype alignment) fully adapt to and deeply explore the advantages of the ViT model: it utilizes its global attention mechanism to capture high-level geometric invariance and leverages its [CLS] tokens for efficient global feature aggregation and alignment. The entire solution can be smoothly applied to other visual localization tasks (such as ground vehicle and UAV pose estimation) and even broader domain adaptation and online learning tasks, demonstrating significant scalability value.

[0020] In summary, the greatest comprehensive advantage of this invention lies in its combination of theories such as domain-invariant representation and anti-forgetting mechanisms with engineering requirements such as privacy constraints, streaming processing, and low resource consumption. This successfully overcomes the shortcomings of existing methods in on-orbit scenarios, ultimately making it possible to achieve high-precision, high-robustness, and high-practicability real-time online tracking of non-cooperative spacecraft attitude in real, complex, and resource-constrained space environments. Attached Figure Description

[0021] Appendix Figure 1 This is a flowchart of the present invention;

[0022] Appendix Figure 2 This is a diagram showing the spacecraft pose estimation results in the Lightbox domain of the SPEED+ dataset in this invention.

[0023] Appendix Figure 3 This is a diagram showing the spacecraft pose estimation results in the Sunlamp domain of the SPEED+ dataset in this invention.

[0024] Appendix Figure 4 This is the loss curve of the Lightbox domain in the SPEED+ dataset during the online adaptation process in this invention;

[0025] Appendix Figure 5 This is the loss curve of the Sunlamp domain in the SPEED+ dataset during the online adaptation process in this invention;

[0026] Appendix Figure 6 This is a diagram showing the translation estimation result of trajectory ROE1 in the SHIRT dataset in this invention;

[0027] Appendix Figure 7 This is a diagram showing the rotation estimation result of trajectory ROE1 in the SHIRT dataset in this invention;

[0028] Appendix Figure 8 This is a diagram showing the translation estimation result of trajectory ROE2 in the SHIRT dataset in this invention;

[0029] Appendix Figure 9 This is a diagram showing the rotation estimation effect of trajectory ROE2 in the SHIRT dataset in this invention. Detailed Implementation

[0030] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0031] Please refer to the attached document. Figure 1 As shown, this invention provides an online adaptation method for compensating for domain offsets in pose tracking of non-cooperative spacecraft in orbit, characterized by the following steps:

[0032] Step 1: Represent the target spacecraft as 11 3D key points; establish a 2D-3D correspondence based on the camera projection model, extract the key point bounding boxes after projection, and crop them to obtain the input RoI; train an asymmetric encoder-decoder network to regress the key point heatmap, thereby initializing the teacher-student distillation framework; apply data augmentation (flare, histogram equalization, blur, noise) in the source domain pre-training to address spacecraft image domain differences; the loss function for heatmap training is the MSE of the predicted heatmap and the ground truth.

[0033] The specific process of step one is as follows:

[0034] The target spacecraft is represented as a set containing N=11 predefined three-dimensional key points; a two-dimensional pixel coordinate system (u) is established. i ,v i ) and its corresponding 3D key point homogeneous coordinates The correspondence between them is based on the monocular camera projection model:

[0035] s i [u i v i 1] T =K[R∣t]P i

[0036] Among them, s i K is the scaling factor, R is the camera intrinsic parameter matrix, and R and t are the rotation matrix and translation vector, respectively. The 3D keypoints are projected onto the image plane according to the labeled relative pose to form the corresponding 2D keypoint set. The minimum and maximum values ​​of the x and y coordinates of the point set are taken to form the 2D bounding box. The image region captured by the bounding box is the RoI, which is used as the input of the subsequent network model.

[0037] Following the ViTPose architecture, an asymmetric encoder-decoder network is constructed to learn regression keypoint heatmaps in the source domain. The encoder uses a pre-trained ViT model, specifically the DeiT-III architecture, with weights obtained through self-supervised learning on the ImageNet-21K dataset as initialization. The decoder consists of two transposed convolutional layers with 256 channels each.

[0038] ConvTranspose2d is followed by a 1×1 convolutional layer Conv2d with 11 output channels, corresponding to the heatmap prediction of the N=11 key points; to improve the model's generalization ability and alleviate domain shift, data augmentation is applied during the source domain model training stage; given that differences in texture information and changes in lighting conditions in spacecraft images are the main causes of domain differences, a method is adopted...

[0039] The Albumentations library implements enhancement operations, including: solar flare simulation, histogram equalization, Gaussian blur, contrast adjustment, and Gaussian noise; each enhancement operation is randomly applied during training with a probability of 0.5.

[0040] Given N input samples, the encoder maps the samples to high-level semantic features, and the decoder outputs a heatmap tensor. The corresponding real heatmap H true A two-dimensional Gaussian kernel with a standard deviation of σ = 3 is applied to the two-dimensional coordinates of each keypoint to generate the heatmap; the loss function for source domain training is the mean square error (MSE) between the predicted and actual heatmaps.

[0041] L src =L mse (H pred H true )

[0042] During the inference phase, a non-maximum suppression algorithm is applied to the predicted heatmap to determine the peak response position of each key point as the two-dimensional detection result. Using the 2D-3D corresponding point relationship established above, the rotation matrix R and translation vector t can be solved by using the Efficient Perspective-n-Point (EPnP) algorithm.

[0043] Step 2: To mitigate the catastrophic forgetting problem in online adaptation, a dynamic memory library M based on a first-in-first-out queue is designed. Initially, M is filled once with source domain samples and their true values. When the target domain data is input in a streaming manner, a student model is used to infer on single or mini-batch samples. Only the samples with the highest confidence in the batch and their heatmap pseudo-labels are stored in M. In each adaptation iteration, all samples in M ​​are used for self-training. If the model's confidence in a newly generated pseudo-label for a sample in M ​​is higher than the old value in the library, then the pseudo-label is updated.

[0044] The specific process of step two is as follows:

[0045] Step 2: To effectively mitigate the catastrophic forgetting problem commonly encountered during online adaptation, a dynamic memory library M based on a first-in-first-out queue is designed. M stores key-value pairs consisting of preprocessed samples and their corresponding heatmap pseudo-labels, which are obtained during the online inference phase. Before the online adaptation process of the target domain begins, M is initialized by randomly selecting preprocessed samples from the source domain and their true labels. This initialization operation is performed only once when the source domain data is available, ensuring that subsequent target domain adaptation strictly follows the source-free domain setting.

[0046] When target domain data is provided via streaming input, the model infers on a single sample or mini-batch; after inference, sample confidence is calculated based on the peak response value in the heatmap; only the sample with the highest confidence and its pseudo-label are stored in M; let (x,y) represent the i-th heatmap in the online mini-batch. The coordinates on the graph, with a confidence score of 100%.

[0047]

[0048] In each online adaptation iteration, self-training is performed using all current samples and pseudo-labels stored in M; given the i-th sample... and pseudo-tags The self-training loss is

[0049]

[0050] During self-training, for samples stored in M, if the model newly generates a heatmap S... s The confidence score of the k-th key point exceeds the corresponding stored pseudo-label heatmap S. bIf the confidence score in M ​​is used, the pseudo-label of the k-th key point of the sample in M ​​is updated to the new predicted value. This dynamic update mechanism uses reliable historical information to continuously optimize the model, which is one of the core methods to enhance model stability and resist forgetting.

[0051] Step 3: The student model receives the complete samples in M, and the teacher model receives the masked samples in M; consistency constraints are implemented by minimizing the weighted MSE loss between the predicted heatmaps of the student and the teacher, which is used to update the student parameters, while the teacher updates the parameters using an exponential moving average.

[0052] The specific process of step three is as follows:

[0053] Step 3: To improve the model's feature discrimination ability in the target domain, this invention proposes a teacher-student distillation framework based on spatial heatmap response distribution alignment; for each sample, the student network processes a complete view X. b The randomly sampled discontinuous regions are masked by zeros; specifically, the sparse binary mask M is generated through a discrete uniform distribution to achieve feature-level masking.

[0054] X m =M⊙X b

[0055] The teacher model acts as a momentum encoder, processing the masked sample X. m The network parameter φ is updated in the following way:

[0056] φ (t) =λφ (t-1) +(1-λ)θ (t)

[0057] Where θ (t) This represents the state of the student model parameters at training step t; the momentum coefficient λ = 0.999 controls the strength of feature space smoothing; based on the above formula, the teacher model gradually aggregates the domain-invariant geometric constraints learned by the student model, such as rigid geometric correlations, thereby improving contextual representation capabilities; the pseudo-labels generated by the student model are used to apply consistency loss to the teacher model to reconstruct the complete heatmap from locally visible regions and perform regularization; heat... Figure 1 The loss of efficacy is:

[0058]

[0059] Wherein, the quality weight q k Assignment is performed using a confidence-guided mechanism based on a student model; the normalized confidence score for each keypoint is...

[0060]

[0061] Where τ = 0.7 represents the temperature parameter that controls the sharpness of the distribution; this mechanism improves positioning accuracy by applying stronger supervision to high-confidence key points, while also reducing noise in low-confidence predictions. The two work together to significantly improve the accuracy and robustness of target domain attitude estimation.

[0062] Step 4: After the source domain pre-training is completed, the source domain prototype is aligned with the newly calculated class prototype of the teacher and student models; specifically, the average value of the [CLS] token features of all training images is calculated as the global class prototype of the source domain; during online adaptation of the target domain, the teacher and student networks output [CLS] token feature vectors for the complete and masked target images, respectively; by minimizing the cosine distance loss between these feature vectors and the global class prototype of the source domain, the model is encouraged to generate a global feature representation in the target domain that is semantically consistent with the source domain.

[0063] Step four involves the following steps:

[0064] Step 4: Based on the domain-invariant representation learning theory, the model should extract semantically consistent deep features from similar visual samples across different domains. In the ViT architecture, the [CLS] token serves as a global feature aggregation node, integrating the semantic information of all image patches through a self-attention mechanism. Its embedding not only contains discriminative information but also exhibits robustness to local perturbations, making it an ideal carrier for cross-domain feature alignment. Although this application scenario only involves a single spacecraft category (multi-class classification is omitted), its basic principles apply. After pre-training in the source domain, the average value of the [CLS] token embeddings calculated from all training images D is used as the global class prototype for the source domain category.

[0065]

[0066] During the online adaptation process in the target domain, the encoders of the teacher and student models output [CLS] token feature vectors for each image. To guide the model in learning domain-invariant features, class-aware consistency loss is implemented by minimizing the cosine distance between the target domain feature embedding and the source domain prototype; let f ξ ,f θ Let these be the encoders for the teacher and student models, respectively. The class-aware consistency loss is:

[0067]

[0068] in This represents the join operation, and the cosine similarity formula is:

[0069]

[0070] In summary, the overall loss function for OTTA in the target domain is a weighted sum of the above loss terms:

[0071] Ltgt =L mh +αL st +βL ca

[0072] Where α = 10, β = 0.01;

[0073] Both the source domain pre-training and target domain adaptation phases employ mixed-precision training. During pre-training, the model is trained for 30 epochs using the AdamW optimizer with a base learning rate of 5×10⁻⁶. -4 The model underwent a 500-step warm-up with a weight decay of 0.1 and a batch size of 32, with step size decay applied at the 10th and 15th epochs; the factor was 0.1. During the online adaptation phase, the model was trained for one epoch using the AdamW optimizer with a learning rate of 10. -5 Only the student model is used in the online inference phase.

[0074] In this invention, the entire experiment was conducted on a single NVIDIA RTX 6000Ada 48GB GPU, using CUDA version 11.8 and Ubuntu 18.04 as the operating system.

[0075] In this invention, Figure 1 This is a flowchart of the present invention. Figure 2 and Figure 3 The images show the spacecraft pose estimation results in the Lightbox and Sunlamp domains of the SPEED+ dataset, respectively. Figure 4 and Figure 5 The images show the loss curves for the Lightbox and Sunlamp domains in the SPEED+ dataset during the online adaptation process, respectively. Figure 6 and Figure 7 The images show the translation and rotation estimation results of trajectory ROE1 in the SHIRT dataset. Figure 8 and Figure 9 The images show the translation and rotation estimation results of trajectory ROE2 in the SHIRT dataset.

[0076] This invention addresses the severe domain migration challenge in on-orbit non-cooperative spacecraft pose tracking missions by proposing an innovative OTTA method. Its core advantage lies in achieving online optimization without accessing the original source domain data, utilizing only a very limited stream of target domain observations. Specifically, the effects of this invention are reflected in the following aspects:

[0077] In this invention, the designed dynamic memory library creates an online, high-quality knowledge replay mechanism through its unique historical sample and pseudo-label dynamic update mechanism. In each adaptation iteration, the model self-trains on the "reliable memory" stored in M, continuously reinforcing the learned stable knowledge, thereby significantly suppressing the loss of old knowledge caused by adapting to new samples.

[0078] In this invention, a pre-trained model is used to construct source domain class prototypes, while simultaneously initializing a teacher-student collaborative learning framework. This is achieved through prototype alignment and spatial heat dissipation. Figure 1 Consistency constraints force the model to pay more attention to high-level geometric structures shared across domains (such as rigid body shapes and spatial relationships between parts) rather than volatile low-level textures or noise, thereby aligning cross-domain semantics and generating reliable pseudo-labels.

[0079] This method is not designed around the traditional CNN architecture. Its core modules (mask distillation and class prototype alignment) fully adapt to and deeply explore the advantages of the ViT model: it utilizes its global attention mechanism to capture high-level geometric invariance and leverages its [CLS] tokens for efficient global feature aggregation and alignment. The entire solution can be smoothly applied to other visual localization tasks (such as ground vehicle and UAV pose estimation) and even broader domain adaptation and online learning tasks, demonstrating significant scalability value.

[0080] In summary, the greatest comprehensive advantage of this invention lies in its combination of theories such as domain-invariant representation and anti-forgetting mechanisms with engineering requirements such as privacy constraints, streaming processing, and low resource consumption. This successfully overcomes the shortcomings of existing methods in on-orbit scenarios, ultimately making it possible to achieve high-precision, high-robustness, and high-practicability real-time online tracking of non-cooperative spacecraft attitude in real, complex, and resource-constrained space environments.

[0081] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. An online adaptive method for mitigating domain shift in on-orbit non-cooperative spacecraft pose tracking, characterized in that: Includes the following steps: Step 1: Represent the target spacecraft as 11 three-dimensional key points; A 2D-3D correspondence is established based on the camera projection model. After projection, key point bounding boxes are extracted and cropped to obtain the input RoI. An asymmetric encoder-decoder network is trained to regress the key point heatmap, thereby initializing the teacher-student distillation framework. Data augmentation is applied in the source domain pre-training to address the differences in the spacecraft image domain. The loss function for heatmap training is the MSE between the predicted heatmap and the ground truth. Step two: design a dynamic memory bank based on a first-in-first-out queue ; initially, filled with source domain samples and their true values; when the target domain data is input in a streaming manner, the student model is used to infer single or small batch samples; only the sample with the highest confidence and its heatmap pseudo label in the batch are stored in ; all samples in are used for self-training in each adaptive iteration, and if the model's confidence in the newly generated pseudo label of a sample in M is higher than the old value in the library, the pseudo label is updated; Step 3: Student Model Reception The complete sample in the model is received by the teacher. Masked samples in the data; consistency constraints are implemented by minimizing the weighted MSE loss between student and teacher predicted heatmaps, which is used to update student parameters, while teachers update their parameters using an exponential moving average; Step 4: After the source domain pre-training is completed, the source domain prototype is aligned with the newly calculated class prototype of the teacher and student models; specifically, the average value of the [CLS] token features of all training images is calculated as the global class prototype of the source domain; during online adaptation of the target domain, the teacher and student networks output [CLS] token feature vectors for the complete and masked target images, respectively; by minimizing the cosine distance loss between these feature vectors and the global class prototype of the source domain, the model is encouraged to generate a global feature representation in the target domain that is semantically consistent with the source domain.

2. The online adaptation method for compensating for domain offsets in pose tracking of non-cooperative spacecraft in orbit according to claim 1, characterized in that: The specific process of step one is as follows: The target spacecraft is represented as a set of N=11 predefined three-dimensional key points; two-dimensional pixel coordinates are established. Its corresponding 3D key point homogeneous coordinates The correspondence between them is based on the monocular camera projection model: ； in, Scaling factor For the camera intrinsic parameter matrix, and These are the rotation matrix and translation vector, respectively. The 3D keypoints are projected onto the image plane according to their labeled relative poses to form a corresponding 2D keypoint set. The minimum and maximum values ​​of the x and y coordinates of the point set are taken to form a 2D bounding box. The image region captured by this bounding box is the RoI, which is used as the input to the subsequent network model. Referring to the ViTPose architecture, an asymmetric encoder-decoder network is constructed to learn regressive keypoint heatmaps in the source domain. The encoder uses a pre-trained ViT model, specifically the DeiT-III architecture, with weights obtained through self-supervised learning on the ImageNet-21K dataset as initialization. The decoder consists of two transposed convolutional layers (ConvTranspose2d) with 256 channels each, followed by a 1×1 convolutional layer (Conv2d) with 11 output channels, corresponding to the heatmap predictions of the N=11 keypoints. To improve the model's generalization ability and alleviate domain bias, data augmentation is applied during the source domain model training phase. The Albumentations library is used to implement augmentation operations, including: solar flare simulation, histogram equalization, Gaussian blur, contrast adjustment, and Gaussian noise. Each augmentation operation is randomly applied with a probability of 0.5 during training. Given Given an input sample, the encoder maps the sample to high-level semantic features, and the decoder outputs a heatmap tensor. The corresponding real heatmap By applying the standard deviation to the two-dimensional coordinates of each key point, Two-dimensional Gaussian kernel generation; the loss function for source domain training is the mean square error (MSE) between the predicted and actual heatmaps: ； During the inference phase, a non-maximum suppression algorithm is applied to the predicted heatmap to determine the peak response position of each key point as the two-dimensional detection result. Using the 2D-3D corresponding point relationship established above, the rotation matrix R and translation vector t can be solved by using the Efficient Perspective-n-Point (EPnP) algorithm.

3. The online adaptation method for compensating for domain offsets in pose tracking of non-cooperative spacecraft in orbit, as described in claim 2, is characterized in that: The specific process of step two is as follows: Step 2: Design a dynamic memory library based on a first-in-first-out queue. ; Key-value pairs consisting of preprocessed samples and their corresponding heatmap pseudo-labels are stored; these key-value pairs are obtained during the online inference phase; before the online adaptation process of the target domain begins. Initialization is performed by randomly selecting source domain preprocessed samples and their true labels; this initialization operation is only performed once when source domain data is available to ensure that subsequent target domain adaptation strictly follows the source-free domain setting. When target domain data is provided via streaming input, the model infers on single samples or mini-batches; after inference, sample confidence is calculated based on peak response values ​​in the heatmap; only samples with the highest confidence and their pseudo-labels are stored. In the middle; set Indicates the first in the online mini-batch heatmap The coordinates on the graph, with a confidence score of 100%. ； In each online adaptation iteration, the stored... Self-training is performed using all current samples and pseudo-labels; given the first... Sample and pseudo-tags The self-training loss is ； During the self-training process, for If the samples stored in the model generate a new heatmap, The Middle The confidence scores of each key point exceeded the corresponding storage pseudo-tag heatmap. The confidence score in the data is then updated. The first of the samples The pseudo-labels of each key point are used as new predicted values.

4. The online adaptation method for compensating for domain offsets in pose tracking of non-cooperative spacecraft in orbit, as described in claim 3, is characterized in that: The specific process of step three is as follows: Step 3: Propose a teacher-student distillation framework based on spatial heatmap response distribution alignment; for each sample, the student network processes a complete view. The randomly sampled discontinuous regions are masked by zeros; specifically, sparse binary masks. Feature-level masking is achieved by generating a discrete uniform distribution: ； The teacher model acts as a momentum encoder to process masked samples. Network parameters Update via the following methods: ； in Indicates training steps The state of the student model parameters; momentum coefficient The strength of feature space smoothing is controlled; based on the above formula, the teacher model gradually aggregates the domain-invariant geometric constraints learned by the student model, thereby improving the context representation ability; the pseudo-labels generated by the student model are used to apply consistency loss to the teacher model to reconstruct the complete heatmap from the locally visible region and perform regularization; the heatmap consistency loss is: ； Among them, quality weight Assignment is performed using a confidence-guided mechanism based on a student model; the normalized confidence score for each keypoint is... ； in The temperature parameter that controls the sharpness of the distribution.

5. The online adaptation method for compensating for domain offset in pose tracking of non-cooperative spacecraft in orbit according to claim 4, characterized in that: Step four involves the following steps: Step 4: Based on the domain-invariant representation learning theory, the model should extract semantically consistent deep features from similar visual samples across different domains; after pre-training in the source domain, the model will use all training images... The calculated average of the [CLS] token embeddings is used as the global class prototype for the source domain class: ； During the online adaptation process in the target domain, the encoders of the teacher and student models output [CLS] token feature vectors for each image. To guide the models in learning domain-invariant features, class-aware consistency loss is implemented by minimizing the cosine distance between the target domain feature embedding and the source domain prototype. Let these be the encoders for the teacher and student models, respectively. The class-aware consistency loss is: ； in This represents the join operation, and the cosine similarity formula is: ； In summary, the overall loss function for OTTA in the target domain is a weighted sum of heatmap consistency loss, self-training loss, and class-aware consistency loss terms: ； in ; Both the source domain pre-training and target domain adaptation phases employ mixed-precision training. During pre-training, the model is trained for 30 epochs using the AdamW optimizer, with a base learning rate of [missing information]. The model underwent a 500-step warm-up with a weight decay of 0.1 and a batch size of 32, with step size decay applied at the 10th and 15th epochs. The factor was 0.

1. During the online adaptation phase, the model was trained for one epoch using the AdamW optimizer with a learning rate of [missing information]. ; Only the student model is used in the online inference phase.

Images