An uncertainty-aware dual-path non-cooperative spacecraft pose estimation method

Abstract

This invention discloses an uncertainty-aware dual-path non-cooperative spacecraft pose estimation method, belonging to the field of spacecraft pose estimation technology. The method constructs a dual-path prediction framework based on a shared backbone network, including a geometric inference path and a global context-aware path. During the training phase, the prediction results of the two paths are aligned using a geometric consistency loss function, and joint optimization is performed using multi-task losses such as keypoint regression, classification, rotation, and translation regression. During the inference phase, the pose estimation results are adaptively fused based on the prediction variance of each path output. When the variance is low, the output of the global context-aware path is directly adopted; otherwise, the keypoint projections of the two paths are merged to recover the final pose. This invention exhibits higher estimation accuracy, robustness, and interpretability in complex space environments such as occlusion, illumination changes, and large-scale variations, providing reliable visual navigation support for on-orbit autonomous missions.

Description

Technical Field

[0001] This invention belongs to the field of spacecraft pose estimation technology, specifically relating to an uncertainty-aware dual-path non-cooperative spacecraft pose estimation method. Background Technology

[0002] With the rapid development of on-orbit servicing activities, such as autonomous rendezvous and docking of spacecraft, maintenance of failed satellites, and space debris removal, there is an urgent need for high-precision and robust pose estimation of non-cooperative spacecraft. Monocular vision-based pose estimation methods have become a research hotspot in this field due to their simplicity, low cost, and ease of integration into space platforms. However, space missions face extreme and dynamic environmental challenges: target distances vary greatly, leading to drastic changes in spacecraft scale in images; space lighting conditions are complex, with highlights, shadows, and extreme contrasts; furthermore, targets may be partially or severely occluded, or be in high-speed motion. These factors all significantly affect the stable extraction and matching of visual features.

[0003] Existing monocular vision pose estimation methods are mainly divided into two categories: one is based on end-to-end deep learning to directly regress rotation matrices and translation vectors. Although the process is simple, its interpretability is weak, it is sensitive to the distribution of training data, and its generalization ability is limited in scenarios with feature degradation; the other is an indirect method based on keypoint detection and PnP solving. It has strong geometric interpretability, but keypoint detection is prone to failure when the target appearance is blurred, texture is missing, or there is occlusion, leading to pose calculation failure. Both types of methods have obvious limitations, especially in real-world spatial scenarios such as drastic changes in lighting, low resolution of distant target imaging, and local occlusion, where their accuracy and stability decrease significantly.

[0004] More importantly, existing methods generally lack the ability to quantify the uncertainty of estimation results. In real-world systems with noise, ambiguity, and model errors, pose estimation inevitably involves uncertainty. Failure to effectively assess this uncertainty will directly impact the safety of subsequent autonomous decision-making, potentially leading to collision risks during rendezvous and docking. Therefore, developing a highly robust pose estimation method that can adapt to extreme space vision conditions and provide uncertainty metrics has become a key technological bottleneck driving the practical application and autonomy of on-orbit servicing technologies. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an uncertainty-aware dual-path non-cooperative spacecraft pose estimation method. It proposes a dual-path estimation framework that combines geometric reasoning and context awareness, and introduces an uncertainty-aware fusion mechanism, aiming to significantly improve the pose estimation performance and system decision security of non-cooperative spacecraft in complex space environments.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] An uncertainty-aware dual-path non-cooperative spacecraft pose estimation method includes:

[0008] Step 1: Construct a dual-path prediction framework based on a shared backbone network. The dual paths include two complementary prediction paths: a geometric inference path and a global context-aware path.

[0009] Step 2: Train the dual-path prediction framework jointly using multiple loss functions. A geometric consistency loss function is introduced to minimize the difference between the two-dimensional projection point sets generated by the two paths projecting the key points of the target 3D model onto the image plane based on their respective predicted poses, thereby forcing the prediction outputs of the two paths to be geometrically aligned.

[0010] Step 3: Input the image into the trained dual-path prediction framework, calculate the variance of the dual-path prediction results respectively, and adaptively select a fusion strategy based on the relationship between the variance and a preset threshold to generate the final pose estimation result and the corresponding uncertainty measure.

[0011] In a second aspect, the present invention provides an electronic device comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned uncertainty-aware dual-path non-cooperative spacecraft pose estimation method.

[0012] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned uncertainty-aware dual-path non-cooperative spacecraft pose estimation method.

[0013] The beneficial effects of this invention are as follows:

[0014] Performance Robustness and Accuracy Improvement: By coordinating geometric inference paths and global context-aware paths, this invention effectively integrates the dual advantages of geometric interpretability and context robustness. The geometric inference path ensures the stability and interpretability of the method under target scale changes; the global context-aware path enhances adaptability to complex visual conditions such as occlusion, illumination changes, and feature degradation. The two paths complement each other. Experiments on standard datasets such as SwissCube and SPEED show that this method significantly outperforms existing mainstream methods in overall estimation accuracy and stability under extremely challenging scenarios such as large-scale changes and noise interference.

[0015] Enhanced Prediction Consistency and Generalization Ability: A novel geometric consistency loss function is introduced during model training, forcibly constraining the outputs of the two prediction paths to remain consistent at the 3D model projection level. This mechanism plays an effective regularization role, not only improving the model's prediction consistency under different observation conditions, but also significantly enhancing its generalization ability and overall stability on unseen data.

[0016] Safety Decision Support and System Reliability: This invention innovatively introduces an adaptive fusion mechanism based on uncertainty awareness. This mechanism not only dynamically selects or fuses the optimal estimate based on the confidence level (variance) of each path prediction during the inference phase, providing more stable and accurate point estimation results, but also outputs the probability distribution of pose prediction and its confidence level metric. This explicit quantification of uncertainty provides crucial information support for subsequent on-orbit autonomous control, risk assessment, and safety decision-making, significantly improving the reliability and safety of the entire visual navigation system.

[0017] Practical Engineering Value: This method boasts a clear architecture, high modularity, and ease of integration into existing spacecraft visual navigation systems. Its highly interpretable geometric reasoning paths and final results with uncertain outputs meet the stringent requirements of high-reliability space missions for algorithm transparency and security, providing an efficient and reliable solution for practical space missions such as autonomous rendezvous and docking, and on-orbit servicing. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall framework of the uncertainty-aware dual-path non-cooperative spacecraft pose estimation method of the present invention.

[0019] Figure 2 This is a schematic diagram of the geometric reasoning path feature construction in this invention;

[0020] Figure 3 This is a schematic diagram of the global context-aware path feature construction in this invention;

[0021] Figure 4 This is a schematic diagram of the spatial self-attention module structure in this invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] like Figure 1 As shown, this invention provides an uncertainty-aware dual-path non-cooperative spacecraft pose estimation method, the method comprising the following steps:

[0024] Step 1: Construct a dual-path prediction framework based on a shared backbone network. The dual paths include two complementary prediction paths: a geometric inference path and a global context-aware path. Specifically:

[0025] Step 1.1: Constructing a geometric inference path to achieve geometrically interpretable pose estimation.

[0026] like Figure 2 As shown, a general feature pyramid (FPN) is first constructed based on some high-level basic features {C3, C4, C5} of the DarkNet-53 backbone network. C The first three layers of features {F1 c F2 c F3 c Specifically, firstly, a 1×1 convolution is applied to the high-level output feature C5 of the backbone network to unify the channel dimensions, and then a 3×3 convolution is used to generate the output feature F3. c Then F3 c Upsampling is performed, and the feature is then element-wise added and fused with the C4 feature after 1×1 convolution channel alignment. Finally, an intermediate feature layer F2 is generated through 3×3 convolution. c Further F2 c After upsampling, the feature is fused with the C3 feature aligned by a 1×1 convolution channel, and then the intermediate feature layer F1 is obtained through a 3×3 convolution. c Then by using F3 c Perform stepwise downsampling to generate deeper {F4} layers c F5 c Construct a complete 5-level feature pyramid system FPN C This enhances the adaptability to multi-scale changes. Then, the two-dimensional image coordinates of predefined target keypoints are obtained through keypoint prediction head regression, and the generated 2D-3D point correspondence is input into the PnP solver to indirectly calculate the target pose information, including rotation matrices and translation vectors.

[0027] Step 1.2: Global context-aware path construction, used to achieve robust pose estimation under visual degradation conditions. For example... Figure 3 As shown, an adaptive weighted feature pyramid (FPN) is first constructed based on the basic features {C1, C2, C3, C4, C5} of all layers of the DarkNet-53 backbone network. W Generate multi-scale representations {F1} for regression tasks w F2 w F3 w F4 w F5 w Specifically, such as Figure 4 As shown, for each input feature C jAll methods introduce a Spatial Self-Attention Module (SSAM) for feature enhancement. This module first performs global average pooling and global max pooling operations on the input features along the channel dimension. Specifically, it obtains an average response map by averaging the responses of each channel and a maximum response map by maximizing the response, thus forming two complementary single-channel spatial description features. Subsequently, these two single-channel features are concatenated along the channel dimension and input into a two-dimensional convolutional layer with a 7×7 kernel to model spatial correlations and generate spatial attention weights within a large receptive field. The convolutional output is normalized using a Sigmoid activation function to obtain the spatial attention map, which is then passed through the spatial attention module M. S The attention map is then element-wise multiplied with the original input features to adaptively enhance the response of key spatial regions and suppress background and irrelevant region features, achieving spatially selective enhancement of the input features to obtain the processed features. Then, for each scale feature F... i w It is obtained by adaptively fusing all backbone layer features according to learnable weights:

[0028] ,

[0029] In the formula λ i,j These are trainable parameters, i = 1, 2, …, 5, representing FPN. W The i-th layer, j=1, 2, …, 5, represents the j-th layer of the backbone network, and the function… C j Preprocess and align to F i w Spatial resolution. Finally, three prediction heads are connected: an anchor-box-based foreground / background classification head, a direct rotation regression head, and a direct translation regression head, where the translation regression head decouples in-plane translation (x, y) from depth (z) into two independent branches for estimation.

[0030] Step 2: Jointly train the dual-path prediction framework using multiple loss functions. A geometric consistency loss function is introduced to minimize the sum of squared distances between corresponding points in the two-dimensional projection point sets generated by projecting all keypoints in the target 3D model onto the image plane based on their respective predicted poses. Specifically:

[0031] Step 2.1: Based on the set of key points M = {X} of the 3D model of the given target. m ∈ | m = 1, 2, …,M},X mHere, is the 3D coordinate of the m-th key, and M is the total number of keypoints. The poses (R1, T1) and (R2, T2) predicted by the geometric inference path and the global context-aware path, respectively, are projected onto the 2D image plane:

[0032] ,

[0033] In the formula Represents the camera projection operator. This is the two-dimensional projection result of branch k.

[0034] Step 2.2: Calculate the difference between the two sets of projection points of the dual path as the geometric consistency loss. gc This loss is minimized during training to force the predictions of the two paths to be geometrically consistent, thus acting as a regularization mechanism to improve generalization ability and prediction stability. The specific formula is as follows:

[0035] ,

[0036] In the formula, SmoothL1 represents the loss function.

[0037] Step 2.3: Calculate the overall loss function by weighted sum of the losses of each component. The specific formula is as follows:

[0038] ,

[0039] In the formula, each loss l kpt , l cls , l rot , l trans , l gc The losses are keypoint regression, foreground / background classification, rotation regression, translation regression, and geometric consistency. Foreground / background classification uses the FocusLoss loss function, while keypoint regression, rotation regression, translation regression, and geometric consistency use the Smooth L1 Loss function. The weights for each loss are set as ω1 = 0.05, ω2 = 0.01, ω3 = 0.1, ω4 = 0.5, and ω5 = 0.01.

[0040] Step 3: Based on the trained dual-path prediction framework, during the inference application stage, the variance of the dual-path prediction results is calculated for each input image. Based on the relationship between the variance and a preset threshold, a fusion strategy is adaptively selected to generate the final pose estimation result and corresponding uncertainty metric. Specifically:

[0041] Step 3.1: For the input image, select N candidate anchor boxes that best match the target on both the geometric inference path and the global context-aware path, generating N pose prediction results. Assuming that each path prediction follows a Gaussian distribution, calculate its mean and variance, and determine the fusion method based on the relationship between the variance and a preset threshold;

[0042] Step 3.2: When the variance of the two branch predictions is lower than the threshold, the global context-aware path output is directly used as the final pose estimation result.

[0043] Step 3.3: If the prediction variance of any path exceeds the threshold, the two-dimensional projection points of the two paths to the key points of the model are merged into a combined corresponding point set, and the combined point set is input into the MAGSAC-PnP solver to recover the final pose estimate. The uncertainty measure of the final pose estimate is estimated by repeatedly sampling the combined point set and statistically analyzing the changes in the PnP solution results multiple times.

[0044] This invention enhances the adaptability to target scale changes and the robustness to complex conditions such as occlusion and feature degradation by coordinating geometric inference paths and global context-aware paths. It also introduces a fusion strategy of geometric consistency loss constraint and variance adaptation, which outputs probabilistic predictions and confidence scores while ensuring pose estimation accuracy. This significantly improves the stability and reliability of pose estimation in extreme space environments such as long distances, large-scale changes and noise interference, providing a high-precision, interpretable and safe visual navigation solution for missions such as autonomous rendezvous and docking and on-orbit servicing of spacecraft.

[0045] The uncertainty-aware dual-path non-cooperative spacecraft pose estimation method provided in this embodiment can be achieved by calculating the ADI-0.1d index, which is 0.1 times the object diameter of the average distance from the model point, and e. pose To assess its accuracy. The ADI-0.1d index is calculated as follows: Let the point set of the object's 3D model be Q = {X...} s ∈ | s = 1, 2, …, S}, where each X s Let R be a 3D point on the surface of the object model. The predicted pose is determined by the rotation matrix R. pred With translation vector T pred The true pose is constituted by the rotation matrix R. gt With translation vector T gt The process involves: For each model point, its predicted coordinates are obtained through a predicted pose transformation, and then nearest neighbor matching is performed with the set of model points after the actual pose transformation. The minimum Euclidean distance between the two is calculated. The symmetric ADI error is obtained by averaging the minimum distances of all points, and its specific formula is as follows:

[0046] ,

[0047] If the error is less than 10% of the object diameter d, i.e., ADI < 0.1d, then the pose prediction of the sample is considered correct. Finally, the proportion of samples that meet this condition is used as the evaluation index to quantify the pose estimation accuracy of the algorithm in different depth ranges.

[0048] e pose The index calculation process is as follows, which is based on the rotation angle error e. R With translation error e T Adding them together, we get: e pose =e R +e T The rotation angle error e R With translation error e T The calculation method is as follows:

[0049] ,

[0050] Where (q) pred ,T pred ) represents the predicted rotation quaternion and translation vector, (q) gt ,T gt () represents the true rotation quaternion and translation vector. This represents the quaternion inner product operation.

[0051] In this embodiment, based on the SwissCube and SPEED datasets, HRNet, PRTR, PoET, WDRNet, CA-SpaceNet, SCLPNet, and the reproduced WDRNet are compared with those of other datasets. CA-SpaceNet The methods were compared, and the pose estimation comparison results are shown in Tables 1 and 2. It can be seen that this embodiment, through the collaborative work of the geometric inference path and the global context-aware path, achieves a significant improvement in pose estimation accuracy compared to the methods described above on the SwissCube dataset, and obtains estimation results close to the best method described above on the SPEED dataset.

[0052] Based on the SwissCube image dataset, this embodiment tests the robustness of WDRNet and the proposed method by applying Gaussian noise (mean=0, variance=0.1) to the test images. To more clearly compare the target pose estimation accuracy at different distances, the image dataset is divided into three groups: near range, medium range, and long range. The experimental results are shown in Table 3. It can be seen that the proposed method, through the dual-path framework and geometric consistency constraints, has significantly stronger anti-interference ability than the WDRNet method, and the relative improvement is greater the distance.

[0053] Table 1

[0054]

[0055] Table 2

[0056]

[0057] Table 3

[0058]

[0059] In a second aspect, the present invention provides an electronic device comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned uncertainty-aware dual-path non-cooperative spacecraft pose estimation method.

[0060] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned uncertainty-aware dual-path non-cooperative spacecraft pose estimation method.

[0061] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for estimating the pose of a dual-path non-cooperative spacecraft with uncertainty awareness, characterized in that, include: Step 1: Construct a dual-path prediction framework based on a shared backbone network. The dual paths include two complementary prediction paths: a geometric inference path and a global context-aware path. The construction of the geometric inference path specifically includes: Extracting multi-layer basic features based on the backbone network; A feature pyramid network is constructed using some high-level basic features to obtain multi-scale geometric features; Predict the two-dimensional image coordinates of predefined target key points by regressing the head at key points; The generated 2D-3D point correspondence is input into the PnP solver to indirectly solve the target pose information; The construction of a global context-aware path specifically includes: Extracting multi-layer basic features based on the backbone network; By utilizing all the basic features, a weighted feature pyramid network is constructed through adaptive fusion of learnable weights to generate multi-scale representations for regression tasks. A spatial self-attention module is introduced on each level of the weighted feature pyramid network; Connect the anchor-frame-based classification head, the direct rotation regression head, and the direct translation regression head that decouples in-plane translation and depth estimation to perform end-to-end target pose estimation; Step 2: Train the dual-path prediction framework jointly using multiple loss functions. A geometric consistency loss function is introduced to minimize the difference between the two-dimensional projection point sets generated by the two paths projecting the key points of the target 3D model onto the image plane based on their respective predicted poses, thereby forcing the prediction outputs of the two paths to be geometrically aligned. Step 3: Input the image into the trained dual-path prediction framework, calculate the variance of the dual-path prediction results respectively, and adaptively select a fusion strategy based on the relationship between the variance and a preset threshold to generate the final pose estimation result and the corresponding uncertainty measure.

2. The uncertainty-aware dual-path non-cooperative spacecraft pose estimation method according to claim 1, characterized in that, In step 2, the poses predicted by the two paths are used to project all key points in the target 3D model onto the 2D image plane to form two projection point sets; the sum of squares of the distances between corresponding points in the two projection point sets is calculated as the geometric consistency loss.

3. The uncertainty-aware dual-path non-cooperative spacecraft pose estimation method according to claim 1, characterized in that, In step 2, training the dual-path prediction framework jointly using multiple loss functions includes training by using the weighted sum of multiple component losses as the total loss function. The component losses include at least keypoint regression loss, foreground / background classification loss, rotation regression loss, translation regression loss, and geometric consistency loss.

4. The uncertainty-aware dual-path non-cooperative spacecraft pose estimation method according to claim 1, characterized in that, In step 3, the adaptive fusion strategy specifically includes: If the variance of the prediction results of both paths is lower than the preset threshold, the output of the global context-aware path is directly used as the final pose estimation result. If the variance of the prediction result of any path exceeds the preset threshold, the two-dimensional projection points of the target key points predicted by the two paths are merged into a combined point set; the correspondence between the combined point set and the key points of the target three-dimensional model is input into the PnP solver and solved again to obtain the final pose estimation result.

5. The uncertainty-aware dual-path non-cooperative spacecraft pose estimation method according to claim 4, characterized in that, In step 3, the uncertainty measure of the final pose estimation is estimated by repeatedly sampling the combined point set and statistically analyzing the changes in the PnP solution results over multiple iterations.

6. The uncertainty-aware dual-path non-cooperative spacecraft pose estimation method according to claim 1, characterized in that, The backbone network is DarkNet-53, and the PnP solver is the MAGSAC-PnP solver.

7. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; Wherein, when one or more programs are executed by the one or more processors, the one or more processors implement the uncertainty-aware dual-path non-cooperative spacecraft pose estimation method according to any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the uncertainty-aware dual-path non-cooperative spacecraft pose estimation method as described in any one of claims 1-6.

Images