Ensuring Accuracy in Auto-Bounding Box Generation for the Autonomous Aerial Refueling Mission

Abstract

The United State Navy has a vested interest in developing methods for the certification of autonomous aerial refueling by uncrewed aircraft. For leadership to accept the risk of allowing an uncrewed platform to act as the receiver for autonomous aerial refueling there needs to be standards and methods of compliance for allowing an uncrewed platform to complete the task. The United States Naval Academy, with the support of the Office of Naval Research, has begun a line of research into developing certification evidence that will enable an uncrewed aircraft to complete the autonomous aerial refueling task. This line of research assumes the use of a deep neural network to properly identify the refueling drogue and coupler. As with most items revolving around training a neural network, they will only perform as well as the labeled data set that was used to train them. The United States Naval Academy has focused on generating large data sets for this line of research through auto-labeling techniques. This paper highlights the generation of one of those data sets and details a follow on effort for improving the technique.