Confidence-Gated Fusion for RGB-Thermal-Depth Trimodal 6DoF Pose Estimation of Non-Cooperative Spacecraft

With the increasing demand for autonomous perception in on-orbit servicing and space debris removal missions, this paper proposes a trimodal confidence-gated fusion pose estimation network for six-degree-of-freedom (6DoF) pose estimation of non-cooperative spacecraft under simulated normal operating conditions. The proposed framework integrates RGB, thermal infrared, and depth data into a unified end-to-end learning architecture, so that appearance texture, thermal radiation, and three-dimensional geometry can be exploited jointly. The Confidence-Gated Fusion (CGF) module is introduced as the core component of the network to generate modality-level confidence scores and perform confidence-weighted aggregation of trimodal features. Experiments on a self-constructed synthetic dataset show that the proposed TCGP-Net achieves the lowest translation error among the compared fusion strategies, reducing the translation L2 error from 0.852 m for the plain-attention baseline to 0.797 m. Meanwhile, its rotation error remains comparable but is slightly higher than that of plain attention, indicating a trade-off between translation-oriented reliability modeling and rotation accuracy. The results demonstrate the effectiveness of confidence-gated fusion for trimodal spacecraft pose regression, and the added degradation experiments further evaluate its behavior under illumination, thermal, depth, and missing-modality conditions.