CN-121980964-A - Spacecraft pose estimation and uncertainty modeling method based on eigenspace

CN121980964ACN 121980964 ACN121980964 ACN 121980964ACN-121980964-A

Abstract

The invention discloses a spacecraft pose estimation and uncertainty modeling method based on an eigenspace, and belongs to the technical field of spacecraft pose estimation. The method comprises the steps of directly rotating manifold SO (3) and translating space And constructing a layered Bayesian probability model, respectively using the Fisher distribution and the Gaussian distribution, introducing conjugate prior, and realizing principle decomposition and quantification of accidental uncertainty and cognitive uncertainty. And through an end-to-end multi-task neural network, pose probability model parameters, key points and segmentation information are jointly learned, and an iterative optimization module is adopted to improve estimation accuracy. In the training stage, the probability distribution of pose prediction is obtained through analysis and marginalization in the reasoning stage, and the two kinds of uncertainty are distinguished. The method can output uncertainty with good calibration while improving pose estimation precision, and provides a reliable basis for autonomous on-orbit safe operation of the spacecraft.

Inventors

PENG FULIN
DING LEI
LIU YUNMENG
XIANG QING

Assignees

中国科学院上海技术物理研究所

Dates

Publication Date: 20260505
Application Date: 20260401

Claims (10)

1. The spacecraft pose estimation and uncertainty modeling method based on the eigenspace is characterized by comprising the following steps of: Step 1, considering that mathematical spaces where rotation components and translation components are located in six-degree-of-freedom poses of a spacecraft are different, respectively forming SO (3) and Euclidean spaces in a rotation matrix Constructing a corresponding rotation probability distribution model and translation probability distribution model, and introducing conjugate prior for the distribution parameters to construct a layering Bayesian probability model; Step 2, constructing an end-to-end multi-task neural network, extracting features of an input monocular spacecraft image, and simultaneously predicting key point information, a target segmentation result and distribution parameters of the rotation probability distribution model and the translation probability distribution model through shared feature representation; Step 3, in the training stage, performing joint optimization on the edge negative log likelihood functions corresponding to the rotation component and the translation component, and inhibiting the prediction result of the multi-task neural network through evidence regularization constraint; and 4, in the reasoning stage, analyzing and marginalizing hidden variables in the hierarchical Bayesian probability model to obtain probability distribution of pose prediction and uncertainty measurement thereof, and distinguishing and quantifying accidental uncertainty and cognitive uncertainty.
2. The method for estimating pose and modeling uncertainty of spacecraft based on eigenspace according to claim 1, wherein in the step 1, the rotation probability distribution model is a matrix fischer-tropsch distribution, and the translation probability distribution model is a gaussian distribution.
3. The spacecraft pose estimation and uncertainty modeling method based on the eigenspace according to claim 2 is characterized in that the matrix Fisher distribution is obtained by predicting a natural parameter matrix through a multi-task neural network, obtaining a principal axis direction matrix, a reference axis direction matrix and three concentration parameters through singular value decomposition, wherein the three concentration parameters are used as potential random variables and endow gamma prior distribution, and analyzing and marginalizing to obtain Student-t distribution of each component of a rotation error in a tangent space.
4. The spacecraft pose estimation and uncertainty modeling method based on the eigenspace according to claim 2 is characterized in that Gaussian distribution predicts a shift mean vector and a covariance matrix by the multi-task neural network, the covariance matrix adopts a diagonal structure, each dimension variance is used as a potential random variable and is given to inverse gamma prior distribution, and the Student-t distribution of each dimension shift error component is obtained after analysis and marginalization.
5. The method for estimating pose and modeling uncertainty of spacecraft based on eigenspace according to claim 1, wherein in the step 2, the end-to-end multitasking neural network comprises feature extraction by adopting EFFICIENTNET as a backbone network, multi-scale feature fusion by a bi-directional feature pyramid network, and generating a plurality of features with different resolutions, wherein the features with different resolutions are respectively input into a key point thermodynamic diagram regression head, a target segmentation prediction head, a rotational regression head based on matrix Fisher distribution and a translational regression head based on Gaussian distribution.
6. The spacecraft pose estimation and uncertainty modeling method based on the eigenspace according to claim 1 is characterized in that in the step 3, an edge negative log likelihood function corresponding to a rotation component is obtained by negative log likelihood summation of Student-t distribution of rotation error in each component of a tangent space, the evidence regularization constraint is formed by a ratio of an absolute value of the rotation error component to a gamma priori evidence sum, and the gamma priori evidence sum is obtained by calculation of twice of a sum of a shape parameter and an inverse scale parameter of the gamma distribution.
7. The spacecraft pose estimation and uncertainty modeling method based on the eigenspace according to claim 1, wherein in the step 3, the edge negative log likelihood function corresponding to the translation component is obtained by negative log likelihood summation of Student-t distribution of each dimension of the translation error, the evidence regularization constraint is formed by a ratio of an absolute value of the translation error component to an inverse gamma prior evidence sum, and the inverse gamma prior evidence sum is obtained by calculation of twice a sum of a shape parameter and a scale parameter of the inverse gamma distribution.
8. The spacecraft pose estimation and uncertainty modeling method based on the eigenspace according to claim 1 is characterized in that in the step 4, accidental uncertainty is characterized by expected condition variance and is obtained through calculation of scale parameters and degrees of freedom of Student-t distribution, cognitive uncertainty is characterized by degrees of freedom of Student-t distribution, the smaller the degrees of freedom are, the thicker the distribution tail is, and the greater the degree of knowledge loss of a reflecting model in a corresponding input space is.
9. An electronic device, comprising: One or more processors; A memory for storing one or more programs; Wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the eigenspace-based spacecraft pose estimation and uncertainty modeling method of any of claims 1-8.
10. A computer readable storage medium having stored thereon executable instructions which when executed by a processor enable the processor to implement a spacecraft pose estimation and uncertainty modeling method based on eigenspace as claimed in any of claims 1-8.

Description

Spacecraft pose estimation and uncertainty modeling method based on eigenspace Technical Field The invention belongs to the technical field of spacecraft pose estimation, and particularly relates to a spacecraft pose estimation and uncertainty modeling method based on an eigenspace. Background With the rapid development of on-orbit service technology, tasks such as autonomous rendezvous and docking of spacecraft, on-orbit maintenance, space debris removal and the like are required to be higher for the perception of the relative pose of a non-cooperative spacecraft. The pose estimation method based on monocular vision has become a research hot spot in the field because of simple system structure, low power consumption and easy integration into an aerospace platform. However, the real space environment has significant complexity and uncertainty. The imaging scale is changed drastically due to the large range of the target distance, the space illumination condition is complex and changeable, and is often accompanied by strong contrast and deep shadow, and the target can be blocked locally or the imaging resolution is insufficient. All the factors can influence the stable extraction of visual features, so that the pose estimation is easy to have reduced precision and even failure. The existing spacecraft pose estimation method based on deep learning is mainly divided into two types, namely an indirect method based on key point detection and PnP solving, such as a scheme disclosed in Chinese patent publication No. CN114419149B, and the pose is optimized by minimizing the projection contour geometric residual error of a spacecraft CAD model in an image, so that the method has strong geometric interpretability. However, the method still needs to rely on deterministic regression to obtain an initial value before pose parameter optimization, and when the appearance of a target is fuzzy or shielding exists, key point detection is easy to fail, so that pose calculation fails. The other type is a method for directly returning pose parameters end to end, such as the scheme disclosed in Chinese patent publication No. CN117078753B, wherein the corresponding distribution relation between dense pixel points and the surface of the three-dimensional model is obtained through a progressive characteristic distribution sampling network, and then pose assumption voting is carried out. Although the method has simple flow, the method has weak interpretability, is sensitive to the distribution of training data, and has limited generalization capability in a characteristic degradation scene. In recent years, some research has begun to attempt to introduce uncertainty measures to enhance the reliability assessment capabilities of the results. For example, chinese patent publication No. CN121577053a discloses a method for estimating pose of a dual-path non-cooperative spacecraft with uncertainty perception, by using statistical variance of two predicted paths as uncertainty measure, and designing an adaptive fusion strategy according to the statistical variance. However, the method still has the following defects that firstly, uncertainty measurement is indirectly obtained by depending on statistical variances of prediction results of a plurality of paths, accidental uncertainty caused by observation noise is difficult to clearly distinguish from cognitive uncertainty caused by insufficient model cognition, SO that result interpretation is insufficient, secondly, a rotation parameter is essentially defined on a non-Euclidean rotation manifold SO (3), the existing method mainly adopts Euclidean space approximate modeling and uncertainty propagation, geometrical inconsistency is easy to introduce, and final uncertainty characterization is distorted, thirdly, most methods only consider random uncertainty caused by observation noise, lack systematic description of model cognition uncertainty, and difficulty in reflecting prediction reliability when training samples are insufficient or distribution changes. In an actual space task, if effective quantification and expression of uncertainty are lacking, the safety of subsequent autonomous decision-making is directly affected. For example, during a rendezvous docking process, excessive confidence but inaccurate pose estimation may trigger pose control instability and even collision risk. Therefore, there is a need for a spacecraft pose estimation method capable of stably working under complex spatial vision conditions, which not only has higher estimation precision, but also can realize unified and interpretable uncertainty modeling in the intrinsic space of pose parameters, distinguish uncertainty of different sources, and provide reliable basis for subsequent safety decisions. Disclosure of Invention In order to solve the technical problems, the invention provides a spacecraft pose estimation and uncertainty modeling method based on an eigenspace, which is directly implemented in a rotation