Determining Relative Positioning for Aerial Refueling Through DNN Bounding Box Geometry
Akshat Maheshwari, Ryan Lowe, Violet Mwaffo, Danielle M. Clement, Donald H. Costello
Abstract
The Office of Naval Research (ONR) has initiated the Advanced Autonomous Air-to-Air Refueling System (A4RS) project to establish standards and explore the use of deep neural networks (DNN) for autonomous refueling tasks. This research aims to develop a monovision-based camera system integrated with a DNN to achieve accurate bearing and range measurements to a refueling drogue within the final 15-20 feet before contact. As part of this study, a pre-trained DNN was used to identify the drogue by placing a bounding box around it. A similar-triangle-based algorithm was then applied to determine the relative range to the center of the drogue (x, y, and z translations) from the camera. The derived ranges were then compared with the actual ranges to determine if the errors would be acceptable for use in autonomous aerial refueling of an uncrewed system. The research found that as the range decreased, the error decreased to acceptable levels (less than four inches when the camera was five feet from the drogue). The results also showed that the more misalignment with the drogue, the larger the errors were to reality.