Publication Date


Document Type


Committee Members

Thomas Wischgoll, Ph.D. (Advisor); Michael L. Raymer, Ph.D. (Committee Member); Scott Nykl, Ph.D. (Committee Member); Lingwei Chen, Ph.D. (Committee Member)

Degree Name

Doctor of Philosophy (PhD)


Automated vehicles pose challenges in various research domains, including robotics, machine learning, computer vision, public safety, system certification, and beyond. These vehicles autonomously handle navigation and locomotion, often requiring minimal user interaction, and can operate on land, in water, or in the air. In the context of aircraft, one specific application is Automated Aerial Refueling (AAR). Traditional aerial refueling involves a "tanker" aircraft using a mechanism, such as a rigid boom arm or a flexible hose, to transfer fuel to another aircraft designated as the "receiver". For AAR, the boom arm may be maneuvered automatically, or in certain instances the tanker may remotely pilot the receiver to ensure station keeping during the refueling process. Due to latency issues, ground control is impractical. Any AAR system must perform rapid and precise relative pose calculations, operating in near real-time (approximately 10 Hz) with an accuracy of around 10 centimeters when the receiver is 30 meters from the tanker. Utilizing stereo vision for relative pose estimation necessitates high-resolution images to achieve this level of accuracy. Given the large resolution, data processing must be extremely fast. The Iterative Closest Point (ICP) method estimates a pose from images by establishing point correspondences and calculating the relative pose between two point clouds. To accelerate ICP, two enhancements contribute to the runtime reduction: using Delaunay triangulation for finding correspondences and parallelizing each step of the ICP process. Since ICP is iterative, caching correspondences at each iteration enables the nearest neighbor search to start from a point likely to be close to the true closest point. Experiments with accelerated ICP demonstrate a speedup on the order of 103 compared to an O(n2) nearest neighbor method and 102 compared to a k-d tree. When tested against other parallel ICP implementations, the accelerated Delaunay ICP matches or outperforms them in terms of timing results. Notably, Delaunay ICP does not require parameter tuning, leading to more accuracy pose estimations compared to other parallel ICP implementations. To further validate the correctness of Delaunay ICP, real and virtual simulations of an AAR scenario provide a robust test-bed for stereo vision and point registration techniques.

Page Count


Department or Program

Department of Computer Science and Engineering

Year Degree Awarded