Non-cooperative spacecraft targets—such as those in autonomous rendezvous and docking, failed missions, or on-orbit maintenance—require rapid and accurate 6D relative pose estimation for service spacecraft guidance and control. Conventional CNN and Transformer methods often lose high-frequency features (e.g., edges) during downsampling, leading to feature distortion. Multi-head loss functions and adversarial learning for transfer learning are complex and time-consuming. This study proposes a wavelet feature enhancement approach for cross-domain spacecraft pose estimation.
Key takeaways
- Multi-scale features are decomposed into high- and low-frequency bands; a Transformer performs spatial and contextual attention in the frequency domain to enhance feature representation.
- Wavelet decomposition on processed features boosts high-frequency details in low-resolution images, mitigating downsampling-induced distortion.
- A source-domain-trained model is directly applied to target-domain pose prediction, with iterative self-training using pseudo-labels filtered by PnP inlier counts.
- Experiments on SwissCube, SPEED, and SPEED+ datasets show improved pose accuracy with competitive inference speed.
For manufacturers and integrators, this method offers a robust solution for vision-based pose estimation in extreme lighting and cross-domain conditions, potentially applicable to autonomous robotic servicing and inspection tasks where precise object localization is critical.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2026-07-15 · “基于小波特征增强与自训练的域迁移航天器目标6D位姿估计”