Improved yolopose-based space target pose estimation method

Abstract

The invention relates to a space target pose estimation method based on improved yolopose. The method comprises the steps of constructing an experimental data set through simulation and actual satellite images, extracting and labeling satellite feature points to form a pre-training data set, and calculating 3D feature point coordinates in combination with a multi-view geometric algorithm. The method comprises the following steps: constructing an improved YOLOPose network containing a backbone network, a neck network, a detection head and a 3D task loss function, training the network by using pre-training data, capturing long-range dependence by combining a deep branch and a Swin Transformer module through backbone network cross-stage feature aggregation and attention optimization, neck network size recovery and multi-stage feature fusion, and outputting a two-dimensional feature point by the detection head. And constructing a projection equation through key point matching, solving three-dimensional feature points by a least square method, and finally completing target satellite pose estimation based on the two-dimensional-three-dimensional feature points to realize pose perception under a space task. By adopting the method, lightweight model architecture and real-time reasoning can be realized.

Description

Technical Field

[0001] This application relates to the field of spatial target pose estimation technology, and in particular to a spatial target pose estimation method based on an improved YOLOpose algorithm. Background Technology

[0002] In space operations, the six-degree-of-freedom pose estimation of the target spacecraft (including three-dimensional spatial position and attitude information) constitutes the basic perception layer of the navigation system, and its accuracy directly determines the success or failure of critical tasks such as automatic docking and debris removal. Compared with ground-based application scenarios, the space operation environment presents unique engineering challenges: First, the lack of atmospheric scattering in the vacuum environment leads to high dynamic range imaging difficulties for optical sensors under extreme lighting conditions, such as high-contrast shadows against a deep-space background and strong interference from direct sunlight; second, the spacecraft platform is limited by the limited computing power and stringent power consumption constraints of the onboard computing unit, making it difficult to directly deploy traditional large-scale models based on deep learning. These dual constraints necessitate that algorithm design must balance robustness and computational efficiency, overcoming the challenges of feature extraction under complex space lighting conditions while achieving a lightweight model architecture and real-time inference capabilities. Therefore, developing pose estimation methods adapted to the special working conditions of space has become a key technological breakthrough direction for ensuring the reliable execution of on-orbit service missions.

[0003] Currently, deep learning-based spacecraft pose estimation methods can be categorized into: two-stage methods, end-to-end methods, domain adaptation methods, and lightweight methods. Two-stage methods involve using a deep convolutional network to extract key points from spacecraft images in the first stage, typically in conjunction with a target recognition network. After spacecraft target detection, a key point prediction network is used. The second stage uses the PnP method to solve for the 3D-2D pose relationship. End-to-end methods are single-stage, hierarchical, end-to-end trainable methods. They use only a single deep learning model to directly regress the spacecraft pose from the image in an end-to-end manner, without relying on intermediate stages. Finally, a loss function calculated from the pose error is used to train the model. Domain adaptation methods map the data features of the target and test domains to the same feature space to complete the transfer task of spacecraft pose estimation. Lightweight methods improve the speed of pose estimation by reducing network parameters, given the limitations of onboard computing power and memory.

[0004] The main limitation of the two-stage approach is that current pose estimation models are generally based on the assumption of cooperative targets, i.e., assuming that the spacecraft has a known geometric configuration, pre-installed optical markings, or highly reflective features. Such assumptions significantly deviate from the real-world scenarios of on-orbit servicing missions. Therefore, considering pose estimation methods for non-cooperative targets is necessary. Due to the limited computing power of onboard computers, space missions impose dual rigid requirements on the computing system: existing commonly used deep learning-based pose estimation models rely on high-precision heatmap prediction or iterative optimization, resulting in a generally large number of parameters and slow single-frame inference speed, making it difficult to meet on-orbit real-time requirements. Therefore, lightweight network design for existing pose models is also a key issue in realizing on-orbit network applications. Summary of the Invention

[0005] Therefore, it is necessary to provide an improved YOLOpose-based spatial target pose estimation method that can achieve lightweight model architecture and real-time inference to address the above-mentioned technical problems.

[0006] A spatial target pose estimation method based on an improved YOLOpose algorithm, the method comprising: A satellite experimental dataset was constructed using simulated satellite images and actual satellite model images. Feature points were extracted from different satellite targets in the dataset, and these feature points were selected and labeled to obtain a pre-training dataset. Based on 2D keypoints, multi-view geometric algorithms were used to calculate the coordinates of 3D satellite feature points. An improved YOLOpose network was constructed, which includes a backbone network, a neck network, a detection head, and a loss function designed for 3D spatial tasks. The network was trained on the pre-training dataset, and 2D feature points of the target satellite were extracted. The backbone network was used to extract the initial feature map of the target satellite through cross-stage feature aggregation and dynamic optimization of channels and spatial attention. The neck network was used to restore the size of the initial feature map and fuse multi-level features. Then, deep branch expansion of channels was combined with the SwinTransformer module to capture long-range dependencies, resulting in candidate features. The candidate features were input into the detection head, and the 2D feature points of the target satellite were output. Key points are extracted from the pre-training dataset, and projection equations are constructed by matching the key points. The projection equations are expanded to obtain independent equations. The independent equations from multiple perspectives are solved using the least squares method to obtain the three-dimensional feature points of the target satellite. The target satellite's pose is estimated by using two-dimensional and three-dimensional feature points.

[0007] The aforementioned spatial target pose estimation method based on the improved YOLOpose algorithm integrates simulation and actual satellite imagery to construct a dataset. This dataset covers the complex features of non-cooperative targets and provides accurate supervision signals through labeled feature points, reducing reliance on large-scale data and lowering model complexity. An improved YOLOpose network is constructed. The backbone network reduces redundant computation through cross-stage feature aggregation and dynamically selects key features using channel and spatial attention mechanisms, avoiding the overhead of full feature processing. The neck network utilizes a feature pyramid structure, fusing P3-P5 multi-level features through bilinear upsampling and lateral connections. Shallow layers preserve details, middle layers balance accuracy and speed, and deep layers combine with the Swin Transformer module to capture long-range dependencies, achieving fine-grained allocation of computational resources. Simultaneously, the windowed self-attention mechanism of the Swin Transformer is introduced into the deep branches, capturing global information while limiting computational complexity, significantly reducing the number of parameters and improving inference speed compared to traditional Transformers. The detection head focuses on outputting two-dimensional feature points and retains key satellite feature points through key point extraction. It reduces the computational cost of PnP solving by leveraging the sparsity of the three-dimensional point cloud, replaces complex model iterative optimization with multi-view constraints, and uses the least squares method and mathematical analytical solutions of projection equations to avoid the iterative overhead of multiple forward propagations required for pose estimation based on heatmaps in traditional deep learning models. This improves computational efficiency to adapt to the limited computing power on satellites. It solves for three-dimensional feature points based on geometric projection relationships, reducing dependence on massive training data and avoiding overfitting caused by model complexity. This achieves synergistic optimization of computational efficiency, data dependence, and model generalization, ultimately improving the single-frame inference speed. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating an improved YOLOpose-based spatial target pose estimation method in one embodiment. Figure 2 This is a schematic diagram of an improved YOLOpose network in one embodiment. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0010] In one embodiment, such as Figure 1 As shown, an improved YOLOpose-based spatial target pose estimation method is provided, including the following steps: Step 102: Construct a satellite experimental dataset using simulated satellite images and actual satellite model images; extract feature points from different satellite targets in the satellite experimental dataset, select satellite feature points and label them to obtain a pre-training dataset.

[0011] This application primarily constructs a dataset that integrates simulated images and actual captured images. The simulated dataset is derived from the Soyuz satellite in the URSO dataset. The URSO dataset simulation utilizes Unreal Engine 4 (UE4) features to render realistic images, such as physically based materials, halos, and lens halos. Illumination in the simulated environment is simply represented by a directional light and a spotlight, simulating sunlight and Earth's gravity, respectively. The actual captured images use Dongfanghong-1 as the target satellite.

[0012] Feature point extraction and annotation were performed separately for the two types of satellite targets: for the Soyuz satellite, eight corner points on the satellite's solar panels were selected as feature points; for the Dongfanghong satellite, 11 key points were selected from both ends of the satellite antenna and the bottom of the satellite. This approach covers the complex features of non-cooperative targets while providing accurate monitoring signals through the annotation of feature points, reducing reliance on large-scale data and lowering model complexity.

[0013] Step 104: Construct a satellite experimental dataset using simulated satellite images and actual satellite model images; extract feature points from different satellite targets in the satellite experimental dataset, select and label satellite feature points to obtain a pre-training dataset; calculate the coordinates of satellite 3D feature points based on 2D keypoints using a multi-view geometric algorithm; construct an improved YOLOpose network; the improved YOLOpose network includes a backbone network, a neck network, a detection head, and a loss function designed for 3D space tasks; train the network based on the pre-training dataset and extract the 2D feature points of the target satellite; extract the initial feature map of the target satellite based on the backbone network through cross-stage feature aggregation and dynamic optimization of channels and spatial attention; use the neck network to restore the size of the initial feature map and fuse multi-level features; then expand the channels through deep branches and combine with the Swin Transformer module to capture long-range dependencies to obtain candidate features; input the candidate features into the detection head and output the 2D feature points of the target satellite.

[0014] The improved YOLOpose network directly optimizes the object key point similarity index through end-to-end training, abandoning the limitation of traditional two-stage heatmap methods that rely on alternative L1 loss, and achieving a high degree of unity between model objectives and evaluation criteria.

[0015] Constructing an improved YOLOpose network, such as Figure 2As shown, the main backbone network consists of 13 layers. It begins with fast downsampling using 3×3 convolutions with a stride of 2 (levels P1-P9), and incorporates a C3k2 module for cross-stage feature aggregation. C3k2 employs a 25% bottleneck compression rate to control computational load, and introduces an SPPF spatial pyramid pooling module and a C2PSA dual-path attention mechanism at deeper layers. The former enhances the receptive field through multi-scale pooling, while the latter dynamically optimizes the feature response through channel and spatial attention. The final output includes feature maps from three levels: P3 (1 / 8 downsampling), P4 (1 / 16), and P5 (1 / 32).

[0016] The neck region employs a feature pyramid structure, restoring the feature map size through bilinear upsampling, and laterally concatenating (Concat) the P3-P5 multi-level features of the backbone network. Each fusion branch is followed by a C3k2 module for feature refinement. The shallow P3 branch retains 256 channels of detailed information for small object detection, the middle P4 layer uses 512 channels to balance accuracy and speed, and the deep P5 layer expands to 1024 channels and combines with the Swin Transformer module, utilizing its windowed self-attention mechanism to capture long-range dependencies.

[0017] Finally, the detection results are output simultaneously at three scales (P3-P5) through the Pose detection layer. Each detection head includes one object classification branch and eight key point coordinate regression branches. The design is characterized by the organic combination of traditional convolutional structures with attention mechanisms and Transformer modules. While maintaining the real-time performance of the YOLO series, it improves the robustness of detecting occluded targets and complex poses through multi-path feature fusion and global context modeling.

[0018] In this application, a loss function is designed for the 3D spatial target task, including two types: classification loss and regression loss. In the spatial target pose estimation task, the core task of classification loss is to accurately distinguish between the target foreground and background. Accurate target-background discrimination is mainly achieved through a triple mechanism. This loss function addresses the extreme sample imbalance problem in spatial pose estimation tasks by constructing an adaptive adjustment mechanism. The classification loss calculation formula is as follows:

[0019] The formula uses a two-way weighted cross-entropy variant, with the positive sample loss term being... The negative sample loss term is .in α The parameters are specifically designed to address class imbalance by increasing the weight of positive samples to alleviate the problem of scarce foreground targets. γThe parameters are designed to address the differences in sample difficulty, significantly increasing the loss weight for difficult samples. In the early stages of training, the focus is primarily on basic target recognition, ensuring the model quickly establishes target perception capabilities by increasing the weight of positive samples; in the later stages of training, the focus shifts to refining the detection of difficult examples, enhancing the ability to recognize partially occluded targets by adjusting the γ value.

[0020] For the keypoint estimation module, a separate regression loss function needs to be designed, which plays a crucial role in providing geometric constraints. This loss function optimizes the prediction of 2D keypoints by establishing a 2D-3D projective geometric model, coupling the predicted 2D keypoints with the 3D position of the object in the camera coordinate system. Error calculation formula:

[0021] in For the target's true 3D coordinates in the camera coordinate system, the projection operator is: (Projection matrix of ray direction), back-projected ray: (3D ray direction vector), 2D keypoint prediction .

[0022] Step 106: Extract key points from the pre-trained dataset, construct projection equations based on key point matching, expand the projection equations to obtain independent equations, and solve the independent equations from multiple perspectives using the least squares method to obtain the three-dimensional feature points of the target satellite.

[0023] Keypoints are extracted from the training set containing ground truth pose values, and then matched based on these keypoints to construct the projection equation.

[0024]

[0025] In the formula, λ represents the scale factor of the spatial target keypoint perpendicular to the image plane in the camera coordinate system, e(u,v,1) represents the second coordinate form of the 2D keypoint, and P(x,y,z,1) represents the homogeneous form of the 3D keypoint. Expanding this formula yields two independent equations (eliminating the scale factor λ).

[0026]

[0027] The above equations can be listed for each viewpoint, so a total of 2n equations can be obtained using n viewpoints, and then the coordinates of the 3D feature points can be obtained by using the least squares method.

[0028] Step 108: Estimate the pose of the target satellite based on its two-dimensional and three-dimensional feature points to obtain the target pose.

[0029] The aforementioned spatial target pose estimation method based on the improved YOLOpose algorithm integrates simulation and actual satellite imagery to construct a dataset. This dataset covers the complex features of non-cooperative targets and provides accurate supervision signals through labeled feature points, reducing reliance on large-scale data and lowering model complexity. An improved YOLOpose network is constructed. The backbone network reduces redundant computation through cross-stage feature aggregation and dynamically selects key features using channel and spatial attention mechanisms, avoiding the overhead of full feature processing. The neck network utilizes a feature pyramid structure, fusing P3-P5 multi-level features through bilinear upsampling and lateral connections. Shallow layers preserve details, middle layers balance accuracy and speed, and deep layers combine with the Swin Transformer module to capture long-range dependencies, achieving fine-grained allocation of computational resources. Simultaneously, the windowed self-attention mechanism of the Swin Transformer is introduced into the deep branches, capturing global information while limiting computational complexity, significantly reducing the number of parameters and improving inference speed compared to traditional Transformers. The detection head focuses on outputting two-dimensional feature points and retains key satellite feature points through key point extraction. It reduces the computational cost of PnP solving by leveraging the sparsity of the three-dimensional point cloud, replaces complex model iterative optimization with multi-view constraints, and uses the least squares method and mathematical analytical solutions of projection equations to avoid the iterative overhead of multiple forward propagations required for pose estimation based on heatmaps in traditional deep learning models. This improves computational efficiency to adapt to the limited computing power on satellites. It solves for three-dimensional feature points based on geometric projection relationships, reducing dependence on massive training data and avoiding overfitting caused by model complexity. This achieves synergistic optimization of computational efficiency, data dependence, and model generalization, ultimately improving the single-frame inference speed.

[0030] In one embodiment, the backbone network has a 13-layer network structure, including convolutional layers, a cross-stage local attention mechanism module, an SPPF spatial pyramid pooling module, and a C2PSA dual-path attention mechanism. The backbone network extracts the initial feature map of the target satellite through cross-stage feature aggregation and dynamic optimization of channel and spatial attention, including: A 3×3 convolution with a stride of 2 is used to quickly downsample the acquired satellite image, and a cross-stage local attention mechanism module is interspersed to achieve cross-stage feature aggregation. The cross-stage local attention mechanism module uses a bottleneck compression rate of 25% to control the amount of computation, and introduces an SPPF spatial pyramid pooling module and a C2PSA dual-path attention mechanism at a deep layer. The former enhances the receptive field through multi-scale pooling, and the latter dynamically optimizes the feature response through channel and spatial attention, outputting an initial feature map containing three layers: P3 layer 1 / 8 downsampling, P4 layer 1 / 16 downsampling, and P5 layer 1 / 32 downsampling.

[0031] In one embodiment, the neck network is a feature pyramid structure. The neck network is used to restore the size of the initial feature map and fuse multi-level features. Then, deep branches are used to expand channels, combined with the Swin Transformer module to capture long-range dependencies, resulting in candidate features, including: The initial feature map is restored by bilinear upsampling using the neck network. Multi-level features from layers P3 to P5 of the backbone network are then fused laterally. Each fusion branch is followed by a cross-stage local attention mechanism module for feature refinement. The shallow P3 layer retains 256 channels of detail information for small object detection, the middle P4 layer uses 512 channels to balance accuracy and speed, and the deep P5 layer is expanded to 1024 channels. Combined with the windowed self-attention mechanism of the Swin Transformer module, long-range dependencies are captured to obtain candidate features.

[0032] In one embodiment, the projection equation is constructed by matching key points, including: Based on the matching of key points, the projection equation is constructed as follows:

[0033]

[0034] Where λ represents the scale factor of the spatial target keypoints perpendicular to the image plane in the camera coordinate system. The secondary coordinate form representing 2D keypoints. Represent the homogeneous form of the 3D keypoints. This represents the camera intrinsic parameters, where gt represents the truth value. Represents the truth value of the rotation matrix. Represents the truth value of the translation vector. and This represents two equivalent focal lengths in the orthogonal direction of the camera. and This represents the coordinates of the principal point in the direction orthogonal to the camera.

[0035] In one embodiment, the projection equations are expanded to obtain independent equations, including: Expanding the projection equations, we obtain the independent equations as follows:

[0036]

[0037] in, This represents the x-coordinate of a 2d point. This represents the y-coordinate of a 2d point. ~ Rotation matrix The 9 elements, Represents the x-coordinate of the translation vector. Represents the y-coordinate of the translation vector. Represents the z-coordinate of the translation vector. and This represents two equivalent focal lengths in the orthogonal direction of the camera, and X, Y, and Z represent the x, y, and z coordinates of a 3d point, respectively.

[0038] In a specific embodiment, the above equations can be listed for each viewpoint, thus a total of 2n equations can be obtained using n viewpoints:

[0039] Then, the coordinates P of the 3D feature point are obtained by using the least squares method: .

[0040] In one embodiment, the pre-set 3D spatial task loss function is:

[0041] in, Represents the positive sample loss term. Represents the negative sample loss term. The true 3D coordinates of the target in the camera coordinate system. The projection matrix representing the ray direction. Represents the 3D ray direction vector. This indicates 2D keypoint prediction. N Indicates the number of key points. Indicates the true label of the sample. This indicates that the sample is identified as background. This indicates that the sample was identified as the target.

[0042] In one embodiment, the target satellite's two-dimensional and three-dimensional feature points include multiple control points; the pose estimation of the target satellite based on the target satellite's two-dimensional and three-dimensional feature points includes: From multiple control points, select four non-coplanar control points to construct the world coordinate system and the measurement system coordinate system. Based on the relationship that the two coordinate systems need to satisfy, represent the relationship of feature points in the measurement coordinate system. Based on the feature point representation relationship in the measurement coordinate system, the centroids and decentroid coordinates of two sets of three-dimensional feature points in two coordinate systems are obtained by iteratively solving the nearest point; the pose estimate is obtained by solving the centroids and decentroid coordinates of the two sets of three-dimensional feature points.

[0043] In a specific embodiment, four non-coplanar control points are selected from multiple control points to construct a world coordinate system as follows:

[0044] in, Represents the control points in the world coordinate system. Represents the characteristic points of the world coordinate system. This represents the weight of the control point corresponding to each feature point. Indicates the control point number. Indicates the index of the feature point.

[0045] From multiple control points, four non-coplanar control points are selected to construct the coordinate system of the measurement system as follows:

[0046] in, Represents the control points of the measurement system coordinate system. Characteristic points representing the coordinate system of the measurement system. This represents the weight of the control point corresponding to each feature point. Indicates the control point number. Indicates the index of the feature point.

[0047] In one embodiment, the world coordinate system and the measurement system coordinate system satisfy the following relationship:

[0048] in, Represents the control points of the measurement system coordinate system. Represents the control points in the world coordinate system. Indicates the world coordinate system. Represents the rotation matrix. Indicates the world coordinate system. This represents the translation vector.

[0049] The relationship between feature points in the measurement coordinate system is as follows

[0050] in, and Let be the initial value of the pose relation to be determined. Represents the control points in the world coordinate system. This represents the weight of the control point corresponding to each feature point. Indicates the control point number. Indicates the index of the feature point.

[0051] In one embodiment, the centroids and decentroid coordinates of two sets of 3D feature points in two coordinate systems are obtained by iterative nearest point solving:

[0052] in, This represents the total number of feature points. Indicates the index of the feature point. and These are the centroid coordinates in the world coordinate system and the centroid-free coordinates, respectively. and These are the centroid coordinates in the camera coordinate system.

[0053] In one embodiment, pose estimation is obtained by solving for the centroids and decentroided coordinates of two sets of three-dimensional feature points, including: set up ,right H Perform SVD decomposition. ,set up Rotation matrix R Translation vector T for:

[0054] in, and These are the centroid coordinates in the world coordinate system and the centroid-free coordinates, respectively. and These are the centroid coordinates in the camera coordinate system. and Both represent orthogonal matrices. This indicates the transpose operation.

[0055] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0056] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0057] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0058] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A spatial target pose estimation method based on an improved YOLOpose algorithm, characterized in that, The method includes: A satellite experimental dataset was constructed using simulated satellite images and actual satellite model images. Feature points were extracted from different satellite targets in the satellite experimental dataset, and satellite feature points were selected and labeled to obtain a pre-training dataset. An improved YOLOpose network is constructed, comprising a backbone network, a neck network, and a detection head. The improved YOLOpose network is trained using the pre-trained dataset and a pre-defined 3D spatial task loss function. The trained network is used to extract two-dimensional feature points of the target satellite. The backbone network extracts an initial feature map of the target satellite through cross-stage feature aggregation and dynamic optimization of channels and spatial attention. The neck network restores the size of the initial feature map and fuses multi-level features. Then, deep branching expands the channels, and a Swin Transformer module is used to capture long-range dependencies, yielding candidate features. These candidate features are input into the detection head, which outputs two-dimensional feature points of the target satellite. Key points are extracted from the pre-trained dataset, and projection equations are constructed by matching the key points. The projection equations are expanded to obtain independent equations. The independent equations from multiple perspectives are solved using the least squares method to obtain the three-dimensional feature points of the target satellite. The target satellite's pose is estimated by using two-dimensional and three-dimensional feature points.

2. The method according to claim 1, characterized in that, The backbone network has a 13-layer structure, including convolutional layers, a cross-stage local attention mechanism module, an SPPF spatial pyramid pooling module, and a C2PSA dual-path attention mechanism. Based on the backbone network, it extracts the initial feature map of the target satellite through cross-stage feature aggregation and dynamic optimization of channel and spatial attention, including: A 3×3 convolution with a stride of 2 is used to quickly downsample the acquired satellite image, and a cross-stage local attention mechanism module is interspersed to achieve cross-stage feature aggregation. The cross-stage local attention mechanism module uses a bottleneck compression rate of 25% to control the amount of computation, and introduces an SPPF spatial pyramid pooling module and a C2PSA dual-path attention mechanism at a deep layer. The former enhances the receptive field through multi-scale pooling, and the latter dynamically optimizes the feature response through channel and spatial attention, outputting an initial feature map containing three layers: P3 layer 1 / 8 downsampling, P4 layer 1 / 16 downsampling, and P5 layer 1 / 32 downsampling.

3. The method according to claim 2, characterized in that, The neck network is a feature pyramid structure. It is used to restore the size of the initial feature map and fuse multi-level features. Then, deep branching expands the channels, combined with the Swin Transformer module to capture long-range dependencies, resulting in candidate features, including: The initial feature map is restored by bilinear upsampling using the neck network. Multi-level features from layers P3 to P5 of the backbone network are then fused laterally. Each fusion branch is followed by a cross-stage local attention mechanism module for feature refinement. The shallow P3 layer retains 256 channels of detail information for small object detection, the middle P4 layer uses 512 channels to balance accuracy and speed, and the deep P5 layer is expanded to 1024 channels. Combined with the windowed self-attention mechanism of the Swin Transformer module, long-range dependencies are captured to obtain candidate features.

4. The method according to claim 1, characterized in that, The projection equation is constructed based on the matching of key points, including: Based on the matching of key points, the projection equation is constructed as follows: Where λ represents the scale factor of the spatial target keypoints perpendicular to the image plane in the camera coordinate system. The secondary coordinate form representing 2D keypoints. Represent the homogeneous form of the 3D keypoints. This represents the camera intrinsic parameters, where gt represents the truth value. Represents the truth value of the rotation matrix. Represents the truth value of the translation vector. and This represents two equivalent focal lengths in the orthogonal direction of the camera. and This represents the coordinates of the principal point in the direction orthogonal to the camera.

5. The method according to claim 4, characterized in that, Expanding the projection equations yields independent equations, including: Expanding the projection equations, we obtain the independent equations as follows: in, This represents the x-coordinate of a 2d point. This represents the y-coordinate of a 2d point. ~ Rotation matrix The 9 elements, Represents the x-coordinate of the translation vector. Represents the y-coordinate of the translation vector. Represents the z-coordinate of the translation vector. and This represents two equivalent focal lengths in the orthogonal direction of the camera, and X, Y, and Z represent the x, y, and z coordinates of a 3d point, respectively.

6. The method according to claim 4, characterized in that, The pre-set 3D spatial task loss function is in, Represents the positive sample loss term. Represents the negative sample loss term. The true 3D coordinates of the target in the camera coordinate system. The projection matrix representing the ray direction. Represents the 3D ray direction vector. This indicates 2D keypoint prediction. N Indicates the number of key points. Indicates the true label of the sample. This indicates that the sample is identified as background. This indicates that the sample was identified as the target.

7. The method according to claim 1, characterized in that, The target satellite's two-dimensional and three-dimensional feature points include multiple control points; pose estimation of the target satellite based on these feature points includes: Four non-coplanar control points are selected from the plurality of control points to construct a world coordinate system and a measurement system coordinate system. The relationship between the feature points in the measurement coordinate system is represented according to the relationship that the two coordinate systems need to satisfy. Based on the feature point representation relationship in the measurement coordinate system, the centroids and decentroid coordinates of two sets of three-dimensional feature points in the two coordinate systems are obtained by iteratively solving the nearest point; the pose estimate is obtained by solving the centroids and decentroid coordinates of the two sets of three-dimensional feature points.

8. The method according to claim 7, characterized in that, The world coordinate system and the measurement system coordinate system satisfy the following relationship: in, Represents the control points of the measurement system coordinate system. Represents the control points in the world coordinate system. Indicates the world coordinate system. Represents the rotation matrix. Indicates the world coordinate system. This represents the translation vector.

9. The method according to claim 4, characterized in that, By iteratively solving for the nearest point, the centroids and centroid-free coordinates of two sets of 3D feature points in two coordinate systems are obtained as follows: in, This represents the total number of feature points. Indicates the index of the feature point. and These are the centroid coordinates in the world coordinate system and the centroid-free coordinates, respectively. and These are the centroid coordinates in the camera coordinate system.

10. The method according to claim 4, characterized in that, The pose estimate is obtained by solving for the centroids and decentroided coordinates of two sets of 3D feature points, including: set up ,right H Perform SVD decomposition. ,set up Rotation matrix R Translation vector T for: in, and These are the centroid coordinates in the world coordinate system and the centroid-free coordinates, respectively. and These are the centroid coordinates in the camera coordinate system. and Both represent orthogonal matrices. This indicates the transpose operation.

Images