Development and Validation of 3D Feature-Based Vision Algorithm for Autonomous Ship-Deck Landing

In this paper, we develop and validate a 3D feature-based algorithm for tracking stochastic ship-deck motion at high sea states, specifically Sea-State 6 using data from the Navy SCONE dataset. The new vision algorithm was developed from the structure-from-motion technique, which recovers the 3D structure of an object from a series of 2D images, and was validated using a simulated 3D ship-deck attached to a moving Stewart platform. Algorithm performance with different feature detectors and image resolutions was compared. In hand-held tests, the vision algorithm was demonstrated to accurately estimate the pose of a moving ship-deck using a quadrotor. Visually degraded conditions were also evaluated; the algorithm is robust to occlusion and low illumination, but performance reduces in severe glare. The vision algorithm was then validated in a simple free-flight test. All results were compared with Vicon ground-truth data. Additionally, as the 3D algorithm is computationally demanding, we develop and validate a method to improve the computational speed of the vision algorithm.