The objective of this dissertation is to develop and demonstrate autonomous ship-board landing with computer vision. The problem is hard primarily due to the unpredictable stochastic nature of deck motion. The work involves a fundamental understanding of how vision works, what are needed to implement it, how it interacts with aircraft controls, the necessary and sufficient hardware, and software, how it differs from human vision, its limits, and finally the avenues of growth in the context of aircraft landing. The ship-deck motion dataset is provided by the U.S. Navy. This data is analyzed to gain fundamental understanding and is then used to replicate stochastic deck motion in a laboratory setting on a six degrees of freedom motion platform, also called Stewart platform. The method uses a shaping filter derived from the dataset to excite the platform. An autonomous quadrotor UAV aircraft is designed and fabricated for experimental testing of vision-based landing methods. The entire structure, avionics architecture, and flight controls for the aircraft are completely developed in-house. This provides the flexibility and fundamental understanding needed for this research. A fiducial-based vision system is first designed for detection and tracking of ship-deck. This is then utilized to design a tracking controller with the best possible bandwidth to track the deck with minimum error. Systematic experiments are conducted with static, sinusoidal, and stochastic motions to quantify the tracking performance. A feature-based vision system is designed next. Simple experiments are used to quantitatively and qualitatively evaluate the superior robustness of feature-based vision under various degraded visual conditions. This includes: (1) partial occlusion, (2) illumination variation, (3) glare, and (4) water distortion. The weight and power penalty for using feature-based vision are also determined.

The results show that it is possible to autonomously land on ship-deck using computer vision alone. An autonomous aircraft can be constructed with only an IMU and a Visual Odometry software running on stereo camera. The aircraft then only needs a monocular, global shutter, high frame rate camera as an extra sensor to detect ship-deck and estimate its relative position. The relative velocity however needs to be derived using Kalman filter on the position signal. For the filter, knowledge of disturbance/motion spectrum is not needed, but a white noise disturbance model is sufficient. For control, a minimum bandwidth of 0.15 Hz is required. For vision, a fiducial is not needed. A feature-rich landing area is all that is required. The limits of the algorithm are set by occlusion(80% tolerable), illumination (20,000 lux-0.01 lux), angle of landing (up to 45 degrees), 2D nature of features, and motion blur. Future research should extend the capability to 3D features and use of event-based cameras. Feature-based vision is more versatile and human-like than fiducial-based, but at the cost of 20 times higher computing power which is increasingly possible with modern processors. The goal is not an imitation of nature but derive inspiration from it and overcome its limitations. The feature-based landing opens a window towards emulating the best of human training and cognition, without its burden of latency, fatigue, and divided attention.