Uncertainty Modeling and Calibration for Spacecraft Image-Based Pose Estimation Toward On-Orbit Autonomous Measurement

Abstract Spacecraft pose estimation is a key enabling technology for on-orbit servicing, space rendezvous and docking, and related missions. However, existing deep learning-based pose estimation methods generally lack reliable uncertainty quantification capabilities, and their performance degrades significantly under the domain gap between synthetic training data and real on-orbit imagery. To address these challenges, this paper proposes the DETU-AC (Deep Evidential-based Transformer with Uncertainty-aware Adaptive Calibration) framework for uncertainty modeling and adaptive calibration in spacecraft pose estimation. The core contributions of this framework include: (1) a deep evidential learning-based keypoint detection head that employs the Normal-Inverse-Gamma (NIG) distribution to model each keypoint coordinate, enabling effective decomposition of aleatoric and epistemic uncertainties; (2) the adoption of the Matrix Fisher distribution to model rotation uncertainty while fully accounting for the non-Euclidean geometry of the SO(3) manifold; and (3) a Domain-Aware adaptive Temperature Scaling (DATS) module that achieves online calibration of prediction confidence through feature distribution monitoring. Experiments on the SPEED+ benchmark dataset demonstrate that DETU-AC achieves or surpasses state-of-the-art methods across three dimensions: pose estimation accuracy, uncertainty calibration quality, and computational efficiency. Notably, the Expected Calibration Error (ECE) is reduced by over 50% compared to traditional Bayesian methods. This work provides reliable uncertainty-aware capabilities for on-orbit autonomous systems and offers significant engineering value.