Real-Time Image-Based Relative Pose Estimation and Filtering for Spacecraft Applications
Siddarth Kaki, Jacob Deutsch, Kevin P. Black, Asher Cura-Portillo, Brandon A. Jones, Maruthi R. Akella
Abstract
The problem of estimating relative pose for uncooperative space objects has garnered great interest, especially within applications such as on-orbit assembly and satellite servicing. This paper presents a full end-to-end open-source pose estimation and filtering pipeline using monocular camera images for space systems applications. The algorithm pipeline consists of three main components: 1) a set of neural networks to perform keypoint regression; 2) a pose estimation component, implementing both nonlinear least-squares and perspective-[Formula: see text]-point solvers; and 3) a full-pose tracking component, implementing a multiplicative extended Kalman filter. While this software pipeline is designed to be a general-purpose solution, its development was motivated and driven by the size, weight, power, and cost requirements of the NASA Seeker CubeSat program. A combination of real and simulated results is presented to evaluate the neural network components, and simulated time-series results are presented to evaluate the performance of the full pipeline on flightlike hardware in real time.