Pose Estimation of Space Targets Based on Geometry Structure Features

Abstract

The pose estimation of space targets is of great significance for space target state assessment, anomaly detection, fault diagnosis, etc. With the development of adaptive optics technology, the imaging quality of ground-based optical systems has been greatly improved, and we can use the observed images to estimate the pose of space targets. However, the imaging process of the ground-based optical system is still affected by various noises and disturbances, which makes the images degrade. Aiming at the space target pose estimation with these degraded images, we propose a new pose estimation pipeline based on robust geometry structure features. By associating the corresponding geometry structure feature between consecutive frames, we can get the target pose by optimization method. This paper will explain the definition and extraction of the proposed geometry structure feature. We propose a geometry structure feature prediction method base on set prediction in a multi-task way with target components classification and segmentation. Experiments show that our structure feature prediction network achieves competitive results on the simulated photo-realistic SpaceShuttle dataset which is rendered according to the physics imaging process.