Robust Relative Pose Estimation for Non-Cooperative Long-Distance Spacecraft Based on SAM2 Model

Estimating the relative pose of unknown, non-cooperative space targets is a critical challenge for autonomous in-orbit operations, as traditional methods rely on pre-existing models or labeled data. To address this gap, we propose a zero-shot learning architecture that integrates a foundation model with geometric optimization for model-free perception. Our method leverages the Segment Anything Model 2 (SAM2) to perform template-free, pixel-level segmentation of the target. The resulting semantic regions guide a hierarchical registration process, which combines semantic matching with a stepwise alignment of local and global features to establish robust 2D-2D correspondences. This strategy improves the robustness of the matching while bypassing the computational cost of direct global registration. The final pose is then efficiently calculated through geometric optimization. Comprehensive validation on realistic synthetic scenes and a hardware-in-the-loop platform demonstrates a significant improvement in algorithmic robustness for interacting with unknown space objects.