Online Test-Time Adaptation for On-Orbit Noncooperative Spacecraft Pose Estimation

Cross-domain noncooperative spacecraft pose estimation is crucial for autonomous on-orbit servicing, as it effectively mitigates performance degradation caused by domain gap. However, existing methods generally require access to source domain data or complete target domain data during the adaptation stage, which is incompatible with on-orbit servicing scenarios. To address these issues, this article presents an online test-time adaptation method to handle previously unseen target-domain data. Initially, to overcome catastrophic forgetting in online pose estimation, a dynamic memory bank mechanism with a first-in-first-out queue is designed to persistently store high-confidence samples and their predictions from the target domain, enabling self-training of recent historical knowledge. Subsequently, a teacher–student colearning framework is established, which facilitates cross-domain knowledge transfer and reliable pseudolabel generation through class awareness consistency and spatial heatmap consistency constraints. Notably, source-domain class prototypes constructed from the pretrained model serve as anchors to enforce prediction distribution alignment between teacher and student models in the target domain, thereby preserving domain invariance of critical semantic features. Finally, experiments demonstrate that the proposed method achieves state-of-the-art accuracy on the SPEED+ and SHIRT datasets.