Spacecraft pose estimation and uncertainty modeling method based on intrinsic space

Abstract

This invention discloses a spacecraft pose estimation and uncertainty modeling method based on intrinsic space, belonging to the field of spacecraft pose estimation technology. The method involves directly constructing a hierarchical Bayesian probability model on the rotating manifold SO(3) and translation space, using Fisher and Gaussian distributions respectively, and introducing conjugate priors to achieve the fundamental decomposition and quantification of accidental and cognitive uncertainties. An end-to-end multi-task neural network is used to jointly learn the pose probability model parameters, key points, and segmentation information, and an iterative optimization module is employed to improve estimation accuracy. During the training phase, marginal negative log-likelihood and evidence regularization are jointly optimized; during the inference phase, the probability distribution of pose prediction is obtained through analytical marginalization, and the two types of uncertainty are distinguished. This invention improves pose estimation accuracy while outputting well-calibrated uncertainties, providing a reliable basis for the autonomous and safe operation of spacecraft in orbit.

Description

Technical Field

[0001] This invention belongs to the field of spacecraft pose estimation technology, specifically relating to a spacecraft pose estimation and uncertainty modeling method based on intrinsic space. Background Technology

[0002] With the rapid development of on-orbit servicing technologies, missions such as autonomous rendezvous and docking, on-orbit maintenance, and space debris removal have placed higher demands on the relative pose perception of non-cooperative spacecraft. Monocular vision-based pose estimation methods have become a research hotspot in this field due to their simple system structure, low power consumption, and ease of integration into space platforms.

[0003] However, the actual space environment is significantly complex and uncertain. Large variations in target distance lead to drastic changes in imaging scale; spatial lighting conditions are complex and variable, often accompanied by strong contrasts between light and dark and deep shadows; targets may also be partially occluded or have insufficient imaging resolution. All these factors can affect the stable extraction of visual features, making pose estimation prone to decreased accuracy or even failure.

[0004] Existing deep learning-based spacecraft pose estimation methods mainly fall into two categories: one is an indirect method based on keypoint detection and PnP (Problem-to-Pose) solving, such as the scheme disclosed in Chinese Patent Publication No. CN114419149B. This type of method optimizes the pose by minimizing the geometric residual of the projected contour of the spacecraft CAD model in the image, exhibiting strong geometric interpretability. However, this type of method still relies on deterministic regression to obtain initial values ​​before optimizing the pose parameters, and keypoint detection is prone to failure when the target appearance is blurred or occluded, leading to pose calculation failure.

[0005] Another type is the end-to-end direct regression of pose parameters method, such as the scheme disclosed in Chinese Patent Publication No. CN117078753B. This method obtains the corresponding distribution relationship between dense pixels and the surface of the 3D model through a progressive feature distribution sampling network, and then performs pose hypothesis voting. Although this type of method has a simple process, it has weak interpretability, is sensitive to the distribution of training data, and has limited generalization ability in feature degradation scenarios.

[0006] In recent years, some studies have begun to introduce uncertainty measures to improve the reliability assessment of results. For example, Chinese Patent Publication No. CN121577053A discloses an uncertainty-aware dual-path non-cooperative spacecraft pose estimation method, which uses the statistical variance of two predicted paths as an uncertainty measure and designs an adaptive fusion strategy accordingly. However, this type of method still has the following shortcomings: First, the uncertainty measure is indirectly obtained by relying on the statistical variance of the prediction results of multiple paths, making it difficult to clearly distinguish between the random uncertainty caused by observation noise and the cognitive uncertainty caused by insufficient model cognition, resulting in insufficient interpretability of the results; Second, the rotation parameter is essentially defined on the non-Euclidean rotating manifold SO(3), while existing methods mostly use Euclidean space approximation modeling and uncertainty propagation, which easily introduces geometric inconsistencies and distorts the final uncertainty representation; Third, most methods only consider the random uncertainty caused by observation noise and lack a systematic characterization of the model's cognitive uncertainty, making it difficult to reflect the prediction credibility when the training samples are insufficient or the distribution changes.

[0007] In real-world space missions, the lack of effective quantification and representation of uncertainties directly impacts the safety of subsequent autonomous decision-making. For example, during rendezvous and docking, overconfident but inaccurate pose estimation can lead to attitude control instability or even collision risks. Therefore, there is an urgent need for a spacecraft pose estimation method that can operate stably under complex space vision conditions. This method should not only possess high estimation accuracy but also achieve unified and interpretable uncertainty modeling within the intrinsic space of pose parameters, distinguishing uncertainties from different sources to provide a reliable basis for subsequent safety decisions. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a spacecraft pose estimation and uncertainty modeling method based on intrinsic space, directly in the rotation matrix manifold SO(3) and Euclidean space. A probability distribution model of pose is constructed, and based on a hierarchical Bayesian framework, the random uncertainty caused by observation noise and the cognitive uncertainty caused by insufficient model cognition are decoupled. High-precision pose estimation and quantitative evaluation of reliability are achieved simultaneously through an end-to-end neural network.

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

[0010] A spacecraft pose estimation and uncertainty modeling method based on intrinsic space includes:

[0011] Step 1: Considering that the rotation and translation components in the six-degree-of-freedom pose of the spacecraft reside in different mathematical spaces, respectively in the rotation matrix manifold SO(3) and Euclidean space. The corresponding rotation probability distribution model and translation probability distribution model are constructed internally, and conjugate priors are introduced for the distribution parameters to construct a hierarchical Bayesian probability model;

[0012] Step 2: Construct an end-to-end multi-task neural network to extract features from the input monocular spacecraft image, and simultaneously predict the key point information of the spacecraft, the target segmentation result, and the distribution parameters of the rotation probability distribution model and the translation probability distribution model by sharing feature representation.

[0013] Step 3: During the training phase, the prediction results of the multi-task neural network are suppressed by jointly optimizing the marginal negative log-likelihood functions corresponding to the rotation and translation components and by using evidence regularization constraints.

[0014] Step 4: In the inference stage, the latent variables in the hierarchical Bayesian probability model are analyzed and marginalized to obtain the probability distribution of pose prediction and its uncertainty measure, and to distinguish and quantify accidental uncertainty and cognitive uncertainty.

[0015] Furthermore, in step 1, the rotation probability distribution model is a matrix Fisher distribution, and the translation probability distribution model is a Gaussian distribution.

[0016] Furthermore, the Fisher distribution matrix is ​​obtained by predicting the natural parameter matrix from the multi-task neural network, and then obtaining the principal axis direction matrix, the reference axis direction matrix, and three concentration parameters through singular value decomposition. The three concentration parameters are used as latent random variables and assigned a gamma prior distribution. After analytical marginalization, the Student-t distribution of each component of the rotation error in the tangent space is obtained.

[0017] Furthermore, the Gaussian distribution is predicted by the multi-task neural network to shift the mean vector and covariance matrix, and the covariance matrix adopts a diagonal structure; the variance of each dimension is used as a latent random variable and given an inverse gamma prior distribution, and after analytical marginalization, the Student-t distribution of each dimension's shift error component is obtained.

[0018] Furthermore, in step 2, the end-to-end multi-task neural network includes: using EfficientNet as the backbone network for feature extraction, performing multi-scale feature fusion through a bidirectional feature pyramid network to generate multiple features of different resolutions, and inputting the multiple features of different resolutions into a key point heatmap regression head, a target segmentation prediction head, a rotation regression head based on matrix Fisher distribution, and a translation regression head based on Gaussian distribution, respectively.

[0019] Furthermore, in step 3, the negative log-likelihood function of the edge corresponding to the rotation component is obtained by summing the negative log-likelihoods of the Student-t distribution of each component of the rotation error in the tangent space; the evidence regularization constraint is composed of the ratio of the absolute value of the rotation error component to the sum of the gamma prior evidence, which is calculated as twice the sum of the shape parameter and the inverse scaling parameter of the gamma distribution.

[0020] Furthermore, in step 3, the marginal negative log-likelihood function corresponding to the translation component is obtained by summing the negative log-likelihoods of the Student-t distribution of each dimension of the translation error; the evidence regularization constraint is composed of the ratio of the absolute value of the translation error component to the sum of the inverse gamma prior evidence, and the sum of the inverse gamma prior evidence is calculated as twice the sum of the shape parameter and scale parameter of the inverse gamma distribution.

[0021] Furthermore, in step 4, the random uncertainty is characterized by the expected conditional variance, which is calculated using the scale parameter and degrees of freedom of the Student-t distribution; the cognitive uncertainty is characterized by the degrees of freedom of the Student-t distribution. The smaller the degrees of freedom, the thicker the tail of the distribution, reflecting the greater the degree of knowledge deficiency in the corresponding input space of the model.

[0022] 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 method for spacecraft pose estimation and uncertainty modeling based on intrinsic space.

[0023] 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 method for spacecraft pose estimation and uncertainty modeling based on intrinsic space.

[0024] The beneficial effects of this invention are as follows:

[0025] First, this invention performs probabilistic modeling directly in the eigenspace of the pose parameters, specifically in the rotating manifold SO(3) and Euclidean space. By constructing a Fisher and Gaussian distribution matrix, we avoid the geometric inconsistencies introduced by approximating rotation parameters in Euclidean space in traditional methods, fundamentally improving the mathematical rigor and accuracy of pose estimation. Experimental results on the SPEED+ standard dataset show that our method outperforms existing state-of-the-art methods in both translation and rotation errors under synthetic data and real-world lighting conditions.

[0026] Second, this invention achieves the fundamental decomposition and quantification of accidental and cognitive uncertainties through hierarchical Bayesian probabilistic modeling. Accidental uncertainty accurately responds to observational noise and image quality degradation, while cognitive uncertainty effectively indicates the model's knowledge gaps in out-of-distribution scenarios. Experiments demonstrate that the uncertainty interval predicted by this invention is highly consistent with the actual error distribution and can clearly characterize the heavy-tailed characteristics of the error distribution, providing a credible representation with clear physical meaning and statistical basis for subsequent autonomous decision-making.

[0027] Third, this invention employs an end-to-end multi-task neural network architecture, jointly learning keypoints, segmentation masks, and pose probability distribution parameters based on shared feature representations. Through joint optimization using edge negative log-likelihood and evidence regularization, it effectively suppresses overconfident prediction results. In the inference stage, the probability distribution of pose prediction is obtained through parsing edge detection, eliminating the need for complex post-processing or multi-model integration, resulting in high computational efficiency and significant engineering practical value.

[0028] Fourth, the decomposable uncertainty information output by this invention can provide crucial risk assessment basis for safety-critical tasks such as autonomous rendezvous and docking of spacecraft and on-orbit servicing. When cognitive uncertainty increases, the system can promptly detect unknown scenarios faced by the model, triggering conservative control strategies or manual intervention, significantly improving the safety and reliability of on-orbit operations. Attached Figure Description

[0029] Figure 1 This is a flowchart of a spacecraft attitude estimation and uncertainty modeling method based on intrinsic space according to the present invention;

[0030] Figure 2 This is the end-to-end multi-task neural network architecture in this invention;

[0031] Figure 3 These are the consistency results between the rotation and displacement prediction confidence intervals and the actual error distribution in the embodiments of the present invention, where (a) is the displacement prediction result and (b) is the rotation prediction result;

[0032] Figure 4 These are the accurate response results of the data uncertainty of rotation and displacement to noise in the embodiments of the present invention, wherein (a) is the displacement prediction result and (b) is the rotation prediction result;

[0033] Figure 5 This is a characterization of the cognitive uncertainty of external rotation and displacement in the embodiments of the present invention, where (a) is displacement error and (b) is rotation error. Detailed Implementation

[0034] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0035] like Figure 1 As shown, this invention provides a spacecraft pose estimation and uncertainty modeling method based on intrinsic space, comprising the following steps:

[0036] Step 1: Considering that the rotation and translation components in the six-degree-of-freedom pose of the spacecraft reside in different mathematical spaces, respectively in the rotation matrix manifold SO(3) and Euclidean space. Internally, corresponding rotation and translation probability distribution models are constructed, and conjugate priors are introduced for the distribution parameters to construct a hierarchical Bayesian probability model; specifically:

[0037] Step 1.1: Model the Matrix Fisher distribution for the rotation component.

[0038] Given an input monocular image x, model the rotation component R∈SO(3) as a random variable following a Matrix Fisher distribution:

[0039] ,

[0040] in, Here is the matrix of natural parameters predicted by the neural network; θ represents the network parameters. Let the Matrix Fisher distribution be represented by the following probability density function:

[0041] ,

[0042] in, Let F denote the trace of the matrix; c(F) is the normalization constant, ensuring that the integral of the probability density function is 1; the superscript T denotes the transpose. Perform singular value decomposition on the natural parameter matrix F:

[0043] ,

[0044] Where U, V ∈ SO(3) are rotation matrices, representing the principal axis direction and reference axis direction of the distribution, respectively; κ1 ≥ κ2 ≥ κ3 > 0 is a concentration parameter, controlling the concentration of the distribution along each principal axis, with larger values ​​indicating a more concentrated distribution; diag( Let M denote a diagonal matrix. The distribution modulus, i.e., the most likely rotation estimate M, is:

[0045] .

[0046] Step 1.2: Hierarchical Bayesian extension for rotational components.

[0047] To model cognitive uncertainty, the concentration parameter κ is used. j As a potential random variable, and given a Gamma prior distribution:

[0048] ,

[0049] Where, α r,j > 0 and β r,j > 0 represents the shape parameter and inverse scaling parameter of the Gamma distribution predicted by the neural network; Gamma( ) represents the Gamma distribution.

[0050] Utilizing the geometric properties of SO(3), the rotational error is mapped to the tangent space at the model rotation point. For the high-concentration region (i.e., the typical working range for pose estimation), the tangent space error component... Given the concentration parameter κ j The following approximate conditions are satisfied:

[0051] ,

[0052] in, This represents the tangential space error component of the rotational error along the j-th principal axis; Indicates a Gaussian distribution; This represents the reciprocal of the concentration parameter, corresponding to the variance of the spatial error.

[0053] Utilizing the conjugate property of the Gamma distribution and the Gaussian distribution, for κ... j After analytical marginalization, the predicted distribution of the tangent space error components is obtained as a Student-t distribution:

[0054] ,

[0055] in, Represents the tangent space error component under a given input monocular image x. Conditional probability density; St( |μ, , ν) represents a position parameter of μ and a scale parameter of A Student-t distribution with ν degrees of freedom.

[0056] Step 1.3: Model the Gaussian distribution of the translation component.

[0057] For the translation component t∈ A Gaussian distribution is used for modeling:

[0058] ,

[0059] Where, μ(x; θ)∈ Let Σ(x; θ) be the translated mean vector predicted by the neural network. Let be the covariance matrix. To maintain numerical stability and interpretability, a diagonal structure is used:

[0060] ,

[0061] Where, σ i 2 Let represent the variance along the i-th spatial dimension, where i = 1, 2, 3.

[0062] The translation error vector is represented as:

[0063] ,

[0064] in, ∈ denoted as the translation error vector; μ is the predicted translation mean vector.

[0065] Step 1.4: Hierarchical Bayesian extension for translation components.

[0066] The variance σ of each spatial dimension i 2 As a potential random variable, and given an Inverse-Gamma prior distribution:

[0067] ,

[0068] Where, α t,i > 0 and β t,i > 0 represents the shape and scale parameters of the Inverse-Gamma distribution predicted by the neural network; Inv-Gamma( ) represents the inverse gamma distribution.

[0069] Utilizing the conjugate property of the inverse-Gamma distribution and the Gaussian distribution, for σ i 2 After analytical marginalization, the predicted distribution of the translation error components in each dimension is obtained as a Student-t distribution:

[0070] ,

[0071] in, p( represents the component of the translation error in the i-th dimension; |x) represents the translation error component given an input monocular image x. The conditional probability density.

[0072] Step 2: Construct an end-to-end multi-task neural network to extract features from the input monocular spacecraft image, and simultaneously predict key point information, target segmentation results, and distribution parameters of the rotation probability model and translation probability model of the spacecraft through shared feature representation; specifically:

[0073] Step 2.1: Design a shared feature extraction network.

[0074] like Figure 2 As shown, EfficientNet is used as the backbone network to extract features from the input monocular image x. Then, a bidirectional feature pyramid network (BiFPN) is used for multi-scale feature fusion, generating five feature levels {P3, P4, P5, P6, P7} with different resolutions for subsequent multi-task prediction. Where P... l Let l represent the feature map of the l-th layer, where l = 3, 4, 5, 6, 7; the resolution decreases sequentially, and the number of channels increases accordingly.

[0075] Step 2.2: Design a multi-task prediction head.

[0076] The features fused by BiFPN are input into the following four parallel task heads:

[0077] Key Point Heatmap Regression Head: Predicts a 2D heatmap H of N predefined key points on a spacecraft. kpt ∈ Where N represents the number of key points, H and W represent the height and width respectively, and H kpt The value of each pixel in the image represents the probability that a key point exists at that location, and is used to provide geometric supervision signals;

[0078] Target segmentation prediction head: Pixel-level segmentation mask S∈ predicts the spacecraft The value of each pixel represents the probability that the pixel belongs to the spacecraft target, enhancing the ability to perceive the target's geometry;

[0079] A rotated regression head based on the Fisher distribution: predicting the parameters of the Matrix Fisher distribution. Specifically, the network outputs a rotated 6-dimensional continuous representation r. 6d ∈ The module rotation M∈SO(3) is obtained by Gram-Schmidt orthogonalization; at the same time, the prior hyperparameter α of Gamma is predicted. r = [α r,1 , α r,2 , α r,3 ] T ∈ With β r = [β r,1 , β r,2 , β r,3 ] T ∈ ;

[0080] Gaussian distribution-based translation regression head: predicts the translation mean vector μ = [μ1, μ2, μ3] T ∈ and the inverse-Gamma prior hyperparameter α t = [α t,1 , α t,2 , α t,3 ] T ∈ With β t = [β t,1 , β t,2 , β t,3 ] T ∈ .

[0081] Step 3: During the training phase, by jointly optimizing the marginal negative log-likelihood functions corresponding to the rotation and translation components, a synergistic improvement in pose prediction accuracy and uncertainty characterization is achieved. Furthermore, evidence regularization constraints are used to suppress overconfident predictions from the multi-task neural network. Specifically:

[0082] Step 3.1: Derive the rotation loss function.

[0083] The rotation loss consists of the negative log-marginal likelihood and the regularization term:

[0084] ,

[0085] in, Represents rotational loss; log p( |x) is the sum of the negative log-likelihoods of the Student-t distribution derived in step 1.2; λ r,reg This is the weighting coefficient for the rotation regularization term (taken as 0.1 in this implementation). This is the evidence regularization term. Specifically:

[0086] ,

[0087] Among them, | | represents the absolute value of the rotational error component; (2α) r,j + β r,j The sum of evidence is used to punish networks that generate excessively high confidence levels even when observation errors are large.

[0088] Step 3.2: Derive the translation loss function.

[0089] ,

[0090] in, Represents translation loss; logp( |x) is the sum of the negative log-likelihoods of the Student-t distribution derived in step 1.4; λ t,regThis is the weighting coefficient for the rotation regularization term (taken as 0.1 in this implementation). This is the evidence regularization term. Specifically:

[0091] ,

[0092] Among them, | | represents the absolute value of the rotational error component; (2α) t,i + β t,i () represents the sum of evidence.

[0093] Step 3.3: Derive the total loss function.

[0094] The total loss function is the weighted sum of the losses of each task head:

[0095] ,

[0096] in, Indicates the total loss; The mean square error loss of the key point heatmap; Binary cross-entropy loss for segmentation; For rotational losses; ω1, ω2, ω3, and ω4 are the weight coefficients of each loss term. In this embodiment, all weight coefficients are set to 1.0, and a balanced multi-task training strategy is adopted.

[0097] Step 4, the inference stage, involves analytically marginalizing the latent variables in the hierarchical Bayesian model to obtain the probability distribution of pose prediction and its uncertainty measure, and to distinguish and quantify accidental uncertainty and cognitive uncertainty; specifically:

[0098] Step 4.1: Derivation of the results of rotation uncertainty quantification.

[0099] The random uncertainty of the rotation component is characterized by the expected conditional variance:

[0100] ,

[0101] Among them, Var ale ( ) represents the random uncertainty of the rotation component in the j-th principal axis direction; It expresses expectation.

[0102] Cognitive uncertainty of rotational components Var epi ( The degrees of freedom ν of the Student-t distribution r,j = 2α r,j Representation. The smaller the degrees of freedom, the thicker the distribution tail, reflecting a greater knowledge gap in the corresponding input space.

[0103] Step 4.2: Derivation of the results of translation uncertainty quantification.

[0104] The random uncertainty of the translation component is characterized by the expected conditional variance:

[0105] ,

[0106] Among them, Var ale ( ) represents the random uncertainty of the translation component in the i-th principal axis direction.

[0107] Cognitive uncertainty of translation components Var epi ( The degrees of freedom ν of the Student-t distribution t,i = 2α t,i Characterization.

[0108] Through the above implementation process, this invention achieves collaborative optimization of spacecraft attitude estimation and uncertainty modeling within a unified framework. It can effectively solve the modeling difficulties caused by the inconsistency of the mathematical spaces of rotation and translation components, improve the accuracy, stability and reliability of monocular spacecraft attitude estimation, and has good engineering application value.

[0109] The uncertainty-aware dual-path non-cooperative spacecraft pose estimation method provided in this embodiment can be calculated by e pose To assess its accuracy. 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:

[0110] ,

[0111] Among them, (q) pred , T pred ) represents the predicted rotation quaternion and translation vector, (q) gt , T gt Let be the true rotation quaternion and translation vector, and arccos denote the inverse cosine function. This represents the quaternion inner product operation.

[0112] In this embodiment, the pose estimation accuracy of the synthesized, lightbox, and sunlamp data domains of the SPEED+ dataset is compared with that of KRN, SPNv2 and its variants, and EKRN and its variants, respectively, and the results are shown in Table 1. It can be seen that this embodiment, through multi-task learning and intrinsic space modeling, achieves a significant improvement in pose estimation accuracy on the SPEED+ dataset compared to the methods described above.

[0113] Table 1

[0114]

[0115] To evaluate the reliability of the uncertainty output in this embodiment, a systematic analysis was conducted on the consistency between the predicted confidence interval and the actual error distribution, based on the synthetic data domain of the SPEED+ dataset. For example... Figure 3 As shown, (a) represents the displacement prediction result, and (b) represents the rotation prediction result. In this embodiment, the predicted uncertainty range is highly consistent with the coverage probability of the actual error, and its overall performance is close to the ideal calibration state. A good consistency relationship is shown between the predicted confidence level and the actual observation confidence level for both rotation and translation components.

[0116] In further validation, the model's response to data uncertainty was tested by artificially introducing image noise and imaging degradation. Experimental results are as follows: Figure 4 As shown, (a) represents the displacement prediction result, and (b) represents the rotation prediction result. When the quality of the input image deteriorates, the method of the present invention can stably improve the prediction uncertainty level without significantly changing the prediction mean, and the trend of uncertainty change is consistent with the degree of image degradation. This result shows that the present invention can effectively perceive information degradation caused by observation noise and reflect changes in visual information quality through uncertainty modeling, thereby enhancing the robustness of the system in complex environments.

[0117] In addition to characterizing the random uncertainties caused by observation noise, the method of this invention can also effectively model the cognitive uncertainties caused by insufficient model cognition. In out-of-distribution scenarios, the probability distribution of spacecraft pose prediction errors exhibits a significant heavy-tailed characteristic, while... Figure 5 As shown, (a) represents displacement error and (b) represents rotation error. The error distribution predicted by the method of the present invention can cover the heavy-tailed region well, and its overall trend is consistent with the actual error distribution.

[0118] 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 method for spacecraft pose estimation and uncertainty modeling based on intrinsic space.

[0119] 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 method for spacecraft pose estimation and uncertainty modeling based on intrinsic space.

[0120] 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 spacecraft pose estimation and uncertainty modeling based on intrinsic space, characterized in that, include: Step 1: Considering that the rotation and translation components in the six-degree-of-freedom pose of the spacecraft reside in different mathematical spaces, respectively in the rotation matrix manifold SO(3) and Euclidean space. The corresponding rotation probability distribution model and translation probability distribution model are constructed internally, and conjugate priors are introduced for the distribution parameters to construct a hierarchical Bayesian probability model; Step 2: Construct an end-to-end multi-task neural network to extract features from the input monocular spacecraft image, and simultaneously predict the key point information of the spacecraft, the target segmentation result, and the distribution parameters of the rotation probability distribution model and the translation probability distribution model by sharing feature representation. Step 3: During the training phase, the prediction results of the multi-task neural network are suppressed by jointly optimizing the marginal negative log-likelihood functions corresponding to the rotation and translation components and by using evidence regularization constraints. Step 4: In the inference stage, the latent variables in the hierarchical Bayesian probability model are analyzed and marginalized to obtain the probability distribution of pose prediction and its uncertainty measure, and to distinguish and quantify accidental uncertainty and cognitive uncertainty.

2. The method of claim 1, wherein, In step 1, the rotation probability distribution model is a matrix Fisher distribution, and the translation probability distribution model is a Gaussian distribution.

3. The method of claim 2, wherein, The Fisher distribution matrix is ​​obtained by predicting the natural parameter matrix from the multi-task neural network, and then obtaining the principal axis direction matrix, the reference axis direction matrix, and three concentration parameters through singular value decomposition. The three concentration parameters are used as latent random variables and assigned a gamma prior distribution. After analytical marginalization, the Student-t distribution of each component of the rotation error in the tangent space is obtained.

4. The method of claim 2, wherein, The Gaussian distribution is predicted by the multi-task neural network to shift the mean vector and covariance matrix. The covariance matrix adopts a diagonal structure. The variance of each dimension is used as a latent random variable and given an inverse gamma prior distribution. After analytical marginalization, the Student-t distribution of each dimension's shift error component is obtained.

5. The method of claim 1, wherein, In step 2, the end-to-end multi-task neural network includes: using EfficientNet as the backbone network for feature extraction, performing multi-scale feature fusion through a bidirectional feature pyramid network to generate multiple features of different resolutions, and inputting the multiple features of different resolutions into the key point heatmap regression head, the target segmentation prediction head, the rotation regression head based on matrix Fisher distribution, and the translation regression head based on Gaussian distribution, respectively.

6. The method of claim 1, wherein, In step 3, the negative log-likelihood function of the edge corresponding to the rotation component is obtained by summing the negative log-likelihoods of the Student-t distribution of each component of the rotation error in the tangent space; the evidence regularization constraint is composed of the ratio of the absolute value of the rotation error component to the sum of the gamma prior evidence, which is calculated as twice the sum of the shape parameter and the inverse scaling parameter of the gamma distribution.

7. The method of claim 1, wherein, In step 3, the marginal negative log-likelihood function corresponding to the translation component is obtained by summing the negative log-likelihoods of the Student-t distribution of each dimension of the translation error; the evidence regularization constraint is composed of the ratio of the absolute value of the translation error component to the sum of the inverse gamma prior evidence, and the sum of the inverse gamma prior evidence is calculated as twice the sum of the shape parameter and the scale parameter of the inverse gamma distribution.

8. The method of claim 1, wherein, In step 4, the random uncertainty is characterized by the expected conditional variance, which is calculated using the scale parameter and degrees of freedom of the Student-t distribution; the cognitive uncertainty is characterized by the degrees of freedom of the Student-t distribution. The smaller the degrees of freedom, the thicker the tail of the distribution, reflecting the greater the degree of knowledge deficiency in the corresponding input space of the model.

9. An electronic device, comprising: include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the spacecraft pose estimation and uncertainty modeling method based on intrinsic space as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the spacecraft attitude estimation and uncertainty modeling method based on intrinsic space as described in any one of claims 1-8.

Images