PsiNet: A Multispectral Imaging-Based Deep Learning Approach for Spacecraft Rendezvous Relative Navigation

Abstract

Accurate relative pose estimation for noncooperative spacecraft rendezvous remains challenging under spaceborne imaging conditions, where visible images can be degraded by eclipse transitions, solar glare, shadows, low target contrast, and target tumbling. Existing learning-based approaches are still largely dominated by visible-only sensing or by multispectral fusion schemes that do not fully exploit modality-specific representations and temporal pose dependencies. This article presents PsiNet, an end-to-end multispectral deep recurrent convolutional framework for six-degree-of-freedom spacecraft relative pose estimation from paired visible and long-wave infrared imagery. The proposed architecture uses modality-specific feature extractors for the RGB and thermal streams, followed by a squeeze-and-excitation late-fusion module that adaptively recalibrates cross-modal features. Pose estimation is performed through a decoupled strategy: translation is first regressed from the fused representation and then used as a geometric conditioning input to a recurrent attitude-estimation branch. A three-stage training procedure is adopted to progressively learn translation, temporal attitude dynamics, and reprojection-based attitude refinement. PsiNet is evaluated on the synthetic Astos dataset and on real multispectral sequences from the City dataset. On the Astos test sequences, PsiNet reduces the average relative translation error from 2.58% to 0.81% and the average attitude error from $6.91^\circ$ to $3.49^\circ$ compared with ChiNet, corresponding to relative reductions of approximately 69% and 49%, respectively. On the real CITY/03 trajectory, PsiNet achieves a mean translation error of $0.10\; \mathrm{m}$ and a mean attitude error of $2.43^\circ$, reducing the relative position error from approximately 6.5% to 0.27% and the mean attitude error from approximately $5.5^\circ$ to $2.43^\circ$. Ablation studies further show that the proposed late-fusion design outperforms early- and decision-level fusion, confirming the benefit of modality-specific multispectral feature extraction, adaptive feature-level fusion, and translation-conditioned temporal attitude refinement for spacecraft relative navigation.