Machine-Learning Based Pose and Attitude Estimation for Enhanced Spacecraft Proximity Operations

Stereo-camera-based navigation is highly used in broad environments due to its compact size, low cost, and efficiency in capturing data in dynamic environments. This system is gaining ground in space applications that deal with non-cooperative targets like proximity operations, in-orbit assembly, and rendezvous and docking maneuvers. All of these processes require a high accuracy level to avoid mission performance degradation. However, due to the camera's occlusion nature, capturing the entire object for navigation estimation becomes a challenging task. This paper presents an innovative approach for stereo-camera-based point cloud data enhancement of a non-cooperative, known spacecraft. The methodology integrates a 3D point cloud reconstruction technique using Generative Adversarial Networks (GANs) with an Iterative Closest Point (ICP) algorithm for position and orientation estimation. Additionally, an Extended Kalman Filter (EKF) is employed for pose refinement, simultaneously estimating the target's linear velocities and angular rates. Numerical simulations demonstrate an effective point cloud reconstruction from the object's partial view which further improved the accurate estimation of the object's relative attitude and position.