A key point domain adaptive spacecraft attitude estimation method

Abstract

The application discloses a spacecraft attitude estimation method in a key point domain, and the method comprises the following steps: training a main key point regressor on a source domain; introducing an adversarial regressor to identify domain differences in confidence perception; and introducing a geometric constraint regressor, after the adversarial regressor captures the domain differences, the geometric constraint regressor is used to enforce rigid geometry and projection consistency during feature adaptation; then, an alternating optimization strategy is used to train the entire network; after the training is completed, the feature extractor and the main key point regressor are retained; for a real image of a target domain to be measured, the 2D key point coordinates of the real image are predicted by the main key point regressor, and the known 3D key point coordinates and the camera intrinsic parameters are combined to solve the final six-degree-of-freedom attitude of the spacecraft by using a PnP algorithm. The application can significantly improve the attitude estimation accuracy and robustness in a real scene, and has the advantages of light weight and strong universality.

Description

Technical Field

[0001] This invention belongs to the technical fields of spacecraft attitude estimation, pattern recognition and machine learning, and relates to visible light image analysis and space target attitude estimation. Specifically, it relates to a key point domain adaptive spacecraft attitude estimation method. Background Technology

[0002] Currently, with the increasing number of space missions, spacecraft attitude estimation is a very important task. It can obtain the relative attitude and position of the target and itself, and play a series of roles such as assisting in close-range navigation and rendezvous and docking operations, providing initialization parameters for accurate attitude tracking, and providing target satellite attitude information in space-based surveillance.

[0003] In recent years, with the popularization and exponential growth of deep learning methods, deep learning-based approaches have driven many new developments in the field of attitude estimation, and have become the preferred solution for solving the attitude estimation problem of non-cooperative spacecraft. However, training deep models requires large-scale annotated datasets, and collecting real spacecraft images with accurate six-DOF attitude labels is very time-consuming and technically challenging. To overcome this challenge, synthetic datasets generated from computer-aided design models (CAD) have been widely adopted. For example, the widely used SPEED dataset provides tens of thousands of labeled synthetic images of spacecraft under different attitude and lighting conditions, enabling engineers to effectively train and deploy attitude estimation models. While synthetic data greatly accelerates model training and facilitates deployment testing, models trained solely on synthetic images often suffer severe performance degradation when applied to real images due to domain disparities. These domain disparities stem from variations in sensor noise, lighting, background, and other environmental factors, which are further exacerbated in the on-orbit space environment, where lighting changes drastically, reflective materials on spacecraft surfaces create complex visual effects, and imaging conditions for different missions are highly dynamic, making it difficult for models trained on ground-based synthetic data under single imaging conditions to be truly applicable in orbit.

[0004] In terms of domain adaptation techniques, most existing pre-adaptation methods have not fully considered the inherent, imaging domain-independent geometric constraints of satellites as a specific target. In fact, these geometric constraints determine that, given a particular attitude, satellite keypoints exhibit natural consistency across different imaging domains. This cross-domain stable structural characteristic itself contains strong prior information that can guide the domain adaptation process, but it has often been ignored or underutilized in existing research. Meanwhile, current mainstream domain adaptation methods are mostly designed for dense prediction or classification tasks. Their modeling assumptions do not align with the sparse and semantically clear distribution of keypoints in satellite attitude estimation, lacking targeted modeling of the structural relationships and physical meanings of keypoints. Ignoring these factors not only weakens the effectiveness of cross-domain knowledge transfer but may also further restrict the improvement of attitude estimation accuracy. Furthermore, in domain adaptation research specifically for spacecraft attitude estimation, existing methods typically rely on manually designed complex adaptation strategies or specific heuristic rules. The model structure is relatively complex and highly sensitive to parameters and training processes. When imaging conditions, lighting environments, or background interference change significantly, the stability and robustness of the methods are still difficult to guarantee.

[0005] Unsupervised domain adaptation (UDA) offers a promising solution by leveraging labeled synthetic images and unlabeled real images to reduce the performance gap between the source and target domains. Existing UDA methods in computer vision, such as self-training and adversarial alignment, have been extensively explored, but directly applying these methods to spacecraft attitude estimation still presents considerable limitations.

[0006] In summary, existing related technologies generally suffer from insufficient versatility, strong dependence on missions and scenarios, and high actual deployment costs, making it difficult to achieve efficient and stable promotion and application in complex and ever-changing real aerospace application environments.

[0007] Therefore, there is an urgent need to propose a key point domain-adaptive spacecraft attitude estimation technique, which adapts the domain of satellite key points based on the confidence level of the key points and the geometric constraints of the satellite, thereby improving the accuracy of spacecraft attitude estimation under different imaging conditions. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention proposes a key-point domain adaptation method for spacecraft attitude estimation. The spacecraft attitude estimation domain adaptation problem is formulated as an unsupervised domain adaptation problem based on key points, and a general and flexible training framework is adopted, which can reduce the difficulty of implementation and deployment, and improve the accuracy and robustness of spacecraft attitude estimation under different imaging conditions.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] In a first aspect, the present invention provides a spacecraft attitude estimation method that adapts to the key point domain, the method comprising the following steps: Source domain supervised training: The feature extractor and the master keypoint regressor are trained on the source domain synthesized image to perform keypoint detection by heatmap regression; Confidence-aware domain difference recognition: An adversarial regressor is introduced and trained using unlabeled real images of the target domain. The adversarial regressor is trained to make its keypoint predictions in the target domain differ from those of the master keypoint regressor, while remaining consistent in the source domain, thus identifying domain gaps. The target of the difference is a confidence-aware error distribution. Geometric constraint feature adaptation: Fix the master keypoint regressor and the adversarial regressor, update the feature extractor; and introduce a geometric constraint regressor that enforces rigid geometry and projection consistency during feature adaptation after the adversarial regressor captures domain differences. Alternating optimization and attitude estimation: The entire network is trained using an alternating optimization strategy. After training, the feature extractor and master keypoint regressor are retained. For the real image of the target domain to be tested, the 2D keypoint coordinates are predicted by the master keypoint regressor. Combined with the known 3D keypoint coordinates and camera intrinsic parameters, the PnP algorithm is used to solve the final six-degree-of-freedom attitude of the spacecraft.

[0011] Furthermore, in the source domain supervised training step, the heatmap regression uses KL divergence as the supervised loss.

[0012] Furthermore, in the confidence-aware domain difference recognition step, the error distribution is a weighted aggregation of the predicted heatmap distributions of other key points in the same image by the main key point regressor.

[0013] Furthermore, in the geometric constraint feature adaptation step, the 2D key points of the target domain image are predicted using the master key point regressor, the temporary pose is estimated by the PnP solver with RANSAC, and the known 3D key point coordinates are reprojected to generate geometrically consistent 2D key point coordinates, which serve as the supervision signal for the geometric constraint regressor.

[0014] Furthermore, in the alternating optimization and pose estimation step, the training process employs an alternating optimization strategy, and the execution process includes: The feature extractor and the master keypoint regressor are jointly trained in the source domain, and the adversarial regressor is constrained to conform to its behavior. With the feature extractor and master keypoint regressor fixed, the adversarial regressor is updated to maximize its difference from the master keypoint regressor in the target domain; Fix the master keypoint regressor and the adversarial regressor, and update the feature extractor to minimize the prediction difference between the two in the target domain; Geometric constraints are generated based on the prediction of the master keypoint regressor, and the feature extractor and geometric constraint regressor are jointly optimized.

[0015] Furthermore, the principal keypoint regressor, adversarial regressor, and geometric constraint regressor share the same feature extractor, and only the feature extractor and principal keypoint regressor need to be deployed during the inference phase, while the other regressors can be discarded, achieving lightweight deployment.

[0016] Secondly, the present invention also provides a keypoint domain-adaptive spacecraft attitude estimation system, which includes a feature extractor, a master keypoint regressor, an adversarial regressor, and a geometric constraint regressor, and uses the above-described methods to perform spacecraft attitude estimation.

[0017] In a second aspect, the present invention also provides an electronic device, characterized in that it includes a processor and a memory, the memory storing machine-executable instructions executable by the processor, the processor executing the machine-executable instructions to implement the above-described method.

[0018] As can be seen from the above technical solution, compared with the prior art, the beneficial effects of the present invention include: 1. This invention defines the spacecraft attitude estimation domain adaptation problem as an unsupervised domain adaptation problem based on key points, and designs a general and flexible training framework, which can reduce the difficulty of implementation and deployment, and improve the accuracy and robustness of spacecraft attitude estimation under different imaging conditions.

[0019] 2. This invention designs a confidence-aware error distribution to support domain gap identification, enabling the model to emphasize information-rich object regions while mitigating the impact of on-orbit background clutter and noise on attitude estimation.

[0020] 3. This invention combines geometrically consistent constraints to regulate key point prediction, thereby improving the stability of adaptation and the physical rationality of attitude estimation results.

[0021] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0025] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention.

[0026] Figure 2 This is a schematic diagram of the construction geometric constraint signals provided in an embodiment of the present invention.

[0027] Figure 3 This is a schematic diagram of the predicted pose visualization results provided in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 some embodiments of the present invention, but not all embodiments.

[0029] In the description of this invention, it should be noted that some processes described in this application specification and drawings include multiple operations that appear in a specific order. However, it should be clearly understood that these operations may be performed in any order or in parallel. Furthermore, various numbers are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0030] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0031] See Figure 1As shown, this invention discloses a keypoint-domain adapted spacecraft attitude estimation method, a spacecraft attitude estimation technique for monocular visible light images, applied to space target image analysis and space target attitude estimation. It comprises three main components: supervised keypoint detection in the source domain, gap identification in the confidence-aware domain, and geometrically constrained keypoint adjustment, all of which are jointly optimized in a unified training process. Specifically, this invention employs a shared feature extractor, then uses multiple regressors to assist training; after training, only the regressors need to be retained. The feature extractor can be any mature deep feature extraction backbone network, such as HRNet, ResNet, or MobileNet, to extract discriminative multi-scale feature representations from the original image. Different backbone networks can flexibly balance computational complexity and feature representation capabilities to adapt to different computing power and application scenarios. To adapt to the heatmap regression setting of the keypoint detection task, all regressors can adopt a 1×1 convolutional structure, linearly mapping the high-dimensional feature maps output by the feature extractor to the heatmap space corresponding to the target keypoints. This structure can achieve feature compression and recombination in the channel dimension without introducing additional spatial information mixing, ensuring the accuracy of keypoint localization while significantly reducing the number of model parameters and computational overhead. In this method, to ensure flexible adaptation to various backbone networks, the regressor is not designed with complex features; it only requires the backbone network to output feature maps at a specific resolution, thus possessing good backbone network replaceability.

[0032] The specific implementation of this method is as follows: Figure 1 As shown, the feature extractor and keypoint regressor First, training is performed on the labeled source domain, and then an additional adversarial regressor is introduced. To capture domain prediction discrepancies, keypoint confidence is incorporated into the discrepancy objective to adaptively weight the contribution of samples from the target domain. Finally, after the adversarial regressor captures the domain discrepancies, the geometrically constrained regressor... Rigid geometry and projection consistency are enforced during feature adaptation. The entire framework is trained using an alternating optimization strategy, and the implementation steps and specific details of each part are as follows: 1. Source Domain Training: This invention employs a heatmap-based paradigm for keypoint detection because it provides dense pixel-level supervision and facilitates uncertainty modeling. Given an input image... Network prediction Heatmap of key points , H Indicates altitude, W The width is represented, and softmax normalization is performed on the spatial dimension to form a probability distribution:

[0033] in: The key point heatmap for network prediction is located in ( h , w The value of the position.

[0034] For the source domain, each real keypoint is used as a 2D image coordinate. Provided. By placing on the image plane to The target spatial distribution is constructed using a normalized Gaussian kernel centered on the center. :

[0035] in: It is a normalization constant that ensures the sum of the probability distributions equals 1. Represent the probability distribution. Express the spatial probability of the prediction as... , This represents a heatmap output by the network. Represents the source domain image; the supervised keypoint loss is defined as the average Kullback-Leibler (KL) divergence between the predicted distribution and the target distribution:

[0036] in: For key point losses, This represents the spatial probability distribution of the k-th keypoint in the source domain image predicted by the master regressor. This probabilistic modeling preserves pixel-level supervision while achieving a divergence-based numerical stability objective, which is crucial for maximizing differences in subsequent domain adaptation.

[0037] 2. Confidence-aware domain difference recognition: Training the keypoint regressor on the source domain Subsequently, this invention trains an adversarial regressor. This is used to characterize the domain gap between the source and target domains. The goal is to encourage... Generates from and in the source domain To maintain consistent predictions while exhibiting maximum discrepancies across the target domain, thereby revealing domain-specific variations, this invention constructs an erroneous spatial distribution representing seemingly reasonable but incorrect target domain predictions and trains... Align with this distribution. For a given keypoint Error distribution Defined as a weighted aggregation of the predicted heatmap distributions of all other key points:

[0038] in: This represents the spatial probability distribution of the j-th keypoint in the target domain image predicted by the master regressor (j is not the same as k). It is a normalization constant that ensures the sum of the probability distributions equals 1. It is the weight of the spatial distribution of each key point, and the calculation formula is:

[0039] in: The peak response of its predicted spatial distribution is weighted by its variance to represent the key points. The confidence score, The numerical balance parameter is set to 0.1. Indicates the first l The confidence scores of keypoints are given, and T is a temperature parameter controlling the weighted sharpness, set to 10. This softmax-based normalization ensures that keypoints with higher confidence contribute more to the error distribution, while suppressing noise and uncertain predictions. By using the predictions of other keypoints as the optimization objective, the auxiliary regressor is forced to modify the spatial location of its response, thereby identifying domain differences:

[0040] in: Indicates domain difference loss, This represents the spatial probability distribution of the k-th keypoint in the target domain image predicted by the adversarial regressor. By employing confidence-aware weighting, reliable but inaccurate predictions are emphasized in inter-domain gap identification, while noise and uncertain responses are suppressed. This enables the auxiliary regressor to capture structured, geometrically correlated differences between domains, providing effective guidance for subsequent feature adaptation.

[0041] 3. Feature adaptation of geometric constraints: Optimizing the adversarial regressor To maximize the regressor in the target domain with respect to the source training regressor After identifying the prediction discrepancies, the next step is to adjust the feature extractor to make it more accurate and accurate. and The predictions become consistent across the target samples. This process aligns the source and target feature distributions in a statistical sense for keypoint predictions, thereby reducing the domain gap. However, this difference-driven feature adaptation operates purely at the distribution level and does not explicitly consider the geometry behind the pose estimation task. Therefore, this invention introduces an additional geometrically constrained regression branch. This guides feature adaptation in the target domain. The supervision signal for this branch originates from the original trained regressor. This allows for predictions that avoid any reliance on target domain labels.

[0042] like Figure 2As shown, for the target image First from Obtain 2D keypoint predictions and estimate the temporary pose using a PnP solver with RANSAC. The RANSAC process effectively filters out geometrically inconsistent keypoints and outliers, producing a pose estimate that is globally consistent with the rigid object model. Given the estimated pose... and known 3D keypoint coordinates Reprojecting 3D keypoints onto image coordinates yields geometrically consistent 2D positions:

[0043] in: This represents the coordinates of the k-th 3D keypoint. Indicates a temporary rotation. Indicates temporary displacement; Let represent the camera projection function. These key points of reprojection serve as geometric constraints to ensure that rigid geometry and projection consistency are satisfied. The loss is defined as:

[0044] in: For geometric loss, This represents the spatial distribution of the reprojection points. Represents the target domain image; Encouraging learning features encodes representations that preserve geometry, which not only improves the robustness of pose estimation over the target domain, but also complements distribution-based differential targets by adapting to physically meaningful solutions.

[0045] 4. Training process: This invention follows an alternating optimization strategy, in which different network components are updated in separate steps to achieve stable and efficient domain adaptation. Each training iteration consists of four consecutive steps, denoted as steps 1-4, specifically: Step 1: Train the feature extractor using labeled source domain data and principal keypoint regressor To minimize the loss at key monitoring points. Simultaneously, an auxiliary regressor... Restricted to the source domain To maintain consistency and ensure that the two regressors share the same source domain behavior, the optimization objective is:

[0046] in: Indicates the total loss in the source domain. This indicates the monitoring loss at key points in the source domain. This represents the balancing parameter, which is set to 0.5 during training. This represents the spatial probability distribution of the k-th keypoint in the source domain image predicted by the adversarial regressor.

[0047] Step 2: Fix the feature extractor and master regressor Only update the auxiliary regressor For each target image, based on The prediction constructs a confidence-aware error distribution and optimizes it:

[0048] This step encourages Offset from the target domain in a structured manner This allows for the explicit capture of domain-specific differences.

[0049] Step 3: After the auxiliary regressor identifies the domain gaps, the third step adjusts the feature extractor to reduce this discrepancy. In this step, the regressor... and It is fixed and only the feature extractor is updated. To ensure that their predictions are consistent across the target domain. The corresponding optimization objective is:

[0050] in: This represents the feature adaptation loss. This represents the spatial probability distribution of the k-th keypoint in the target domain image predicted by the main regressor. This represents the spatial probability distribution of the k-th keypoint in the target domain image predicted by the adversarial regressor.

[0051] Step 4: Further standardize feature adaptation by incorporating geometric constraints. Using... For prediction on target domain samples, this invention estimates the temporary pose using PnP and RANSAC, generates geometrically consistent constraints, and then jointly optimizes the geometry-aware regressor. and feature extractor The optimization objective of step 4 is: .

[0052] 5. Attitude estimation: After target domain adaptation is complete, the model trained only in the source domain can perform pose estimation and prediction in the target domain. Since this invention is a framework based on keypoint domain adaptation, pose estimation utilizes the estimated 2D keypoint coordinates and the corresponding known 3D target coordinates to estimate perspective projection, i.e., the PnP algorithm.

[0053] Indicates the predicted attitude. This indicates the location of 2D key points in the image.

[0054] in This refers to the camera's intrinsic parameters. In this invention, the regressor during the domain adaptation process is used for attitude estimation. and All can be discarded, only used and This design is compatible with most network structures and makes the model lightweight and easy to deploy.

[0055] 6. Experimental verification: This invention relates to a spacecraft attitude estimation technique for monocular visible light images, applicable to space target image analysis and space target attitude estimation. To verify the advantages of this invention, the publicly available spacecraft attitude estimation dataset SPEED+ was collected and trained on the dataset's synthetic training set. The trained model, after being adapted using this technique, was experimentally tested on two real test sets.

[0056] The SPEED+ dataset consists of synthetic and real images. Primarily designed for the Tango spacecraft, it comprises three distinct partitions corresponding to different data acquisition modes and lighting conditions: Synthetic, Lightbox, and Sunlamp. The significant domain gap between the synthetic and real partitions stems from differences in lighting modes, background composition, sensor noise, and appearance realism.

[0057] First, the domain adaptation effectiveness of the proposed method was tested on a dataset, with rotation error selected as the test metric. Translation error and total error The test results are shown in Table 1.

[0058] Table 1. Statistics of attitude estimation errors in different domains

[0059] This paper showcases the performance of the pure source domain model, the proposed technique, and the pure target domain model on the SPEED+ dataset, including results for both the target domain (Lightbox and Sunlamp) and the source domain (synthetic). The results show a significant performance drop in the target domain for the source domain only, indicating a substantial domain gap between synthetic training images and real hardware-in-the-loop images. In contrast, the proposed keypoint domain adaptation technique achieves consistent improvements across both backbones. For example, the ResNet-based model reduces the total error score on Lightbox from 0.553 to 0.134 and on Sunlamp from 1.441 to 0.143, while maintaining comparable performance in the source domain. Similarly, the HRNet-based model achieves scores of 0.113 and 0.078 on Lightbox and Sunlamp, respectively, while preserving source domain performance. These results demonstrate that the adaptation process effectively mitigates the domain gap without compromising source domain accuracy, and that the technique can be flexibly implemented in mainstream deep learning networks.

[0060] Subsequently, the inference time for generating single-point predictions from the prediction set was tested on the dataset when the method was applied to different feature extraction networks. The test results are shown in Table 2.

[0061] Table 2. Statistics on inference time and number of parameters

[0062] The statistical results show that the framework of this invention is compatible with backbone networks of different capacities, from lightweight ResNet-18 to higher-capacity ResNet-101 and HRNet. This proves that the proposed domain adaptation strategy does not depend on increasing model capacity and can be easily applied to different backbone architectures. This allows for flexible selection of networks and use of this technology for domain adaptation when deploying in orbit, taking into account the resource constraints in orbit.

[0063] Finally, the poses predicted by this method are visualized, and the results are as follows: Figure 3 As shown in the image, red represents the predicted attitude, and blue represents the true attitude. It can be seen that even in real images with complex lighting variations, planetary background interference, and strong reflections from the satellite surface, this method can still stably and accurately estimate the spacecraft's attitude. This visualization not only intuitively verifies the method's accuracy but also demonstrates its robustness in complex environments and its practical application value, providing a reliable basis for subsequent on-orbit deployment.

[0064] As described in the above embodiments, those skilled in the art will understand that the present invention provides a spacecraft attitude estimation method based on keypoint domain adaptation. This method trains a master keypoint regressor in the source domain; introduces an adversarial regressor to identify confidence-perceived domain differences; introduces a geometrically constrained regressor, using PnP+RANSAC to generate 2D keypoints that satisfy rigid body structural consistency as self-supervised signals, guiding the feature extractor to perform geometrically reasonable domain alignment; and jointly trains each module through an alternating optimization strategy. After training, only the backbone network is retained for inference. This invention can significantly improve the accuracy and robustness of attitude estimation in real-world scenarios, while also possessing the advantages of lightweight design and high versatility.

[0065] Furthermore, the present invention also provides a keypoint domain-adaptive spacecraft attitude estimation system, which includes a feature extractor, a master keypoint regressor, an adversarial regressor, and a geometric constraint regressor, and applies the above-described methods to perform spacecraft attitude estimation.

[0066] The system provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in the aforementioned method embodiment, and will not be repeated here.

[0067] Furthermore, the present invention also provides an electronic device, including a processor and a memory, the memory storing machine-executable instructions executable by the processor, the processor executing the machine-executable instructions to implement the above-described method.

[0068] Those skilled in the art will understand that embodiments of the present invention can be provided as products such as methods, software systems, or electronic devices. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0069] It should be noted that the word "comprising" does not exclude the presence of components or steps not listed in the claims. The words "a" or "an" preceding a component do not exclude the presence of a plurality of such components. This invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer.

[0070] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0071] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A key-point domain adaptive spacecraft attitude estimation method, characterized in that, The method includes the following steps: Source domain supervised training: The feature extractor and the master keypoint regressor are trained on the source domain synthesized image to perform keypoint detection by heatmap regression; Confidence-aware domain difference recognition: An adversarial regressor is introduced and trained using unlabeled real images of the target domain. The adversarial regressor is trained to make its keypoint predictions in the target domain differ from those of the master keypoint regressor, while remaining consistent in the source domain, thus identifying domain gaps. The target of the difference is a confidence-aware error distribution. Geometric constraint feature adaptation: Fix the master keypoint regressor and the adversarial regressor, update the feature extractor; and introduce a geometric constraint regressor that enforces rigid geometry and projection consistency during feature adaptation after the adversarial regressor captures domain differences. Alternating optimization and attitude estimation: The entire network is trained using an alternating optimization strategy. After training, the feature extractor and master keypoint regressor are retained. For the real image of the target domain to be tested, the 2D keypoint coordinates are predicted by the master keypoint regressor. Combined with the known 3D keypoint coordinates and camera intrinsic parameters, the PnP algorithm is used to solve the final six-degree-of-freedom attitude of the spacecraft.

2. The method according to claim 1, characterized in that, In the source domain supervised training step, the heatmap regression uses KL divergence as the supervised loss.

3. The method according to claim 1, characterized in that, In the confidence-aware domain difference recognition step, the error distribution is a weighted aggregation of the predicted heatmap distributions of other key points in the same image by the master key point regressor.

4. The method according to claim 3, characterized in that, In the geometric constraint feature adaptation step, the 2D key points of the target domain image are predicted using the master key point regressor, the temporary pose is estimated by the PnP solver with RANSAC, and the known 3D key point coordinates are reprojected to generate geometrically consistent 2D key point coordinates, which serve as the supervision signal for the geometric constraint regressor.

5. The method according to claim 1, characterized in that, In the alternating optimization and pose estimation step, the training process employs an alternating optimization strategy, and the execution process includes: The feature extractor and the master keypoint regressor are jointly trained in the source domain, and the adversarial regressor is constrained to conform to its behavior. With the feature extractor and master keypoint regressor fixed, the adversarial regressor is updated to maximize its difference from the master keypoint regressor in the target domain; Fix the master keypoint regressor and the adversarial regressor, and update the feature extractor to minimize the prediction difference between the two in the target domain; Geometric constraints are generated based on the prediction of the master keypoint regressor, and the feature extractor and geometric constraint regressor are jointly optimized.

6. The method according to claim 1, characterized in that, The principal keypoint regressor, adversarial regressor, and geometric constraint regressor share the same feature extractor. During the inference phase, only the feature extractor and the principal keypoint regressor need to be deployed, while the other regressors can be discarded, achieving lightweight deployment.

7. A key-point domain adaptive spacecraft attitude estimation system, characterized in that, The system includes a feature extractor, a master keypoint regressor, an adversarial regressor, and a geometric constraint regressor, and performs spacecraft attitude estimation using the method described in any one of claims 1–6.

8. An electronic device, characterized in that, It includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to perform the method as described in any one of claims 1-6.

Images