Abstract
In recent years, there is an increasing demand for orbital robotic missions for various reasons such as life extension of functional satellites, reuse the unique orbital slots and to reduce the risk of orbital collision. In such robotic missions, the satellite’s autonomous navigation capability is a critical component that enables it to perform relative navigation, inspection, and repair
with minimal human-in-loop intervention. Pose estimation is an important task within autonomous GNC for spacecraft in orbit. There have been recent, new development of deep learning based pose estimation algorithms in order to meet growing demands of autonomous orbital applications. This paper presents a new keypoint-based framework using Convolutional Neural Network models for pose estimation of known non-cooperative targets in orbit, which is thoroughly compared to existing state-of-the-art algorithms also based on deep learning. Within the proposed pose estimation pipeline, a ResNet-based architecture used for object detection, a Scale-Aware High-Resolution Network (HigherHRNet) used for keypoint regression and PnP-RANSAC for computing the pose. The framework is benchmarked with the SPEED dataset as well as the Soyuz dataset from STAR LAB Orbital Visual Simulator and the results were presented.