Tiny satellites, big challenges: A feasibility study of machine vision pose estimation for PocketQubes during conjunctions

This paper assesses the feasibility of estimating the position and attitude of a target satellite using computer vision, not during proximity operations, but during conjunctions at distances greater than 1 km. The selected target is a small 5 × 5 × 5 cm 3 PocketQube. By exploring such a challenging limit case, the results have great implications for general in-orbit pose estimation applications such as Active Debris Removal and On-Orbit Servicing. The workflow of this paper can be summarized as (1) the analysis of feasible orbits for conjunction based observation; (2) the creation of an annotated image dataset including proportional blurring artifacts; and (3) the evaluation of a machine learning-based algorithm to perform pose estimation. Pose estimation was found to be feasible at 4 km distances (target resolved at 28.9 pixels on average). For targets at a distance of 1.5 km, in idealized cases (without blurring or noise), position errors less than 50 m were demonstrated 60% of the time. Attitude estimation errors less than 20 deg were demonstrated 70% of the time. When incorporating realistic image artifacts, pose estimation results were found to be largely insensitive to Earth backgrounds, highly sensitive to zoom errors (Depth of Field blurring) and moderately sensitive to motion blur. In nonidealized conjunction scenarios at 1.5 km, accounting for blur and with backgrounds included, less than 50 m position and 20 deg attitude errors were achieved more than 30% of the time. • Studying pose estimation for tiny satellites at > 1 km during conjunctions. • Conducted mission design analyzing the feasibility of long-distance satellite observation. • Proved pose estimation of PocketQubes up to 4 km using machine vision is feasible. • Pose estimation remains robust despite blur and background artifacts. • Created and shared a 45,000+ image dataset for satellite pose estimation research.