The monocular 6D pose estimation of a noncooperative satellite is essential in various on-orbit automatic operations. Nevertheless, the application of this technology still faces challenges, such as heavy model parameters and high latency. In this study, we comprehensively explored the critical components of satellite pose estimation, including landmark description, localisation, and the PnP algorithm, and we introduce a high-precision, real-time, and robust pipeline. Existing methods adopt external convex landmarks, such as 11 landmarks for the SPEED dataset, which are insufficient to describe the geometric structure of the satellite. This study adopted a wireframe-based landmark description, encompassing all visually prominent satellite joints connected directly into the satellite’s wireframe model. In addition, unlike most existing methods that rely on heatmap-based landmark localisation, which incurs high computational costs and quantisation errors, we adopt direct coordinate classification for satellite landmark localisation, which is lightweight, fast, and accurate. Experimental results on the SPEED dataset demonstrate that our method outperforms the current state-of-the-art technique, with an accuracy increase of 28% and parameter number decrease of 78%. In particular, our method is very lightweight and robust, and it runs in real-time, making it well-suited for on-orbit real-time satellite pose estimation.
Read full abstract