Download files
Abstract
Landing an Unmanned Aerial Vehicle (UAV) autonomously and safely on a ship's flight deck is a challenging task for robotic researchers. The difficulties include large deck motion, disturbances caused by gusts and turbulence, and decreased visibility resulting from bad weather conditions such as rain, fog and sun reflection. Existing techniques for landing area localization and pose determination during approach and landing in these conditions tend to rely on ship-deck infrastructure based sensing units or artificial landing markers. Compared to the other works, this thesis develops a more robust mean of segmenting the landing marker which is robust to occlusion and is able to use the same international landing marker used for manned helicopter operations to finalize target recognition and pose estimation, such that no additional infrastructure is required on the ship.
The three major tasks developed in this thesis are locating a viable landing area on the ship, accurately determining the relative pose between the UAV and the shipdeck and finally implementation of an algorithm for deciding when to safely land based on calm period prediction. A self-contained, on-board real-time system, with vision sensors is developed, which makes use of the edge information from the international landing marker to perform line segment detection, feature point mapping and clustering. A cascade filtering scheme formed by a series of coarse-to-fine criteria is adopted to facilitate target recognition. Meanwhile, the vision system is designed to adapt to the challenges of visual occlusion and contamination, such that the landing marker can be reconstructed by processing the information of a partially missing marker. For the second task, the full 6 degree-of-freedom (DoF) relative pose is estimated based upon the extracted keypoint information as well as the prior knowledge of the landing marker by monocular vision. To increase the system redundancy, a secondary 3D range sensor is introduced to extend the operational range, especially when the landing marker is out of the field of view of the monocular camera. An optic flow (OF) based motion estimation method is also used to help stabilize the UAV. A mission planner is proposed to let the system switch between different tasks during landing according to the availability of measurements. Multiple state estimators are used to fuse measurements from different sources to obtain better estimation, so that the UAV can continue its current task even when some sensor outputs fail or are degraded. An on-board controller with the associated measurement-based switching scheme is designed to close the control loop. For the third task, the proposed system is adopted to capture the relative pose between the UAV and an imitated moving shipdeck, whose motion is simulated using a 3DoF moving platform based on different generated ship motion data sets. To prove the concept, a classic timer-series predictor is introduced to predict the ship motion in the future in a short period, with a proposed classifier to determine the opportunities for safe autonomous landing.
To validate the system, the vision system is evaluated by both pre-captured and real-time imagery in the presence of challenges such as like occlusions and illumination variations. The precision of relative pose estimation is quantitatively analyzed. The integrated system is then examined and demonstrated by conducting real flight tests, whose measurements are compared against a VICON motion capture system benchmark. In order to seize the most favorable landing conditions, a study on short-term ship motion prediction is carried out based on the simulated 3DoF shipdeck motion measured by vision.