Daigo Kobayashi, Purdue University; Alexander Burton, Purdue University; Carolin Frueh, Purdue University
Keywords: near-field attitude determination, space debris
Abstract:
Accurately estimating the pose of human-made spacecraft is indispensable for various space applications such as in-orbit servicing, active debris removal, and health monitoring. However, achieving precise attitude estimation remains a challenging task. Resolved imaging via optical sensors can be achieved in the near field. Among the various methods, monostatic monocular imaging is the most practical, yet it presents significant challenges for pose estimation since only a single viewing direction per image is available. One part of that challenge is the fact that pose includes both the attitude of and distance to the object, as both are fully coupled. Existing research has addressed this issue by focusing on attitude estimation in the near field, leveraging machine learning techniques at various stages of the process.
In this paper, the pose of a spacecraft is estimated from a single image. In the first step, keypoint detection in a single image is performed. It is assumed that the keypoint locations in the spacecraft’s body-fixed frame are fully known, which corresponds to perfect knowledge of the observed spacecraft a priori, for example, via a CAD model. Four deep neural network models are tested and compared against each other from two different detection classes: regression and heatmap. Keypoint detection is the precursor to the subsequent pose estimation. Next, the position and attitude of the spacecraft relative to the observer are estimated. The baseline method implemented is the Efficient Perspective-n-Point (EPnP) algorithm. The EPnP has been explored in performance, and implementation has been optimized for our given dataset. The second method used is a particle swarm optimizer (PSO), which has not been applied to the monostatic spacecraft pose estimation problem before. Although the keypoint detection methods rely on machine learning, neither EPnP nor PSO are learning-based. The PSO is shown to produce more accurate pose estimates than the EPnP and is also expected to be more durable against imaging errors.
Date of Conference: September 19-22, 2023
Track: Satellite Characterization