Volume 40, Number 4, August 2022
|787 - 795
|30 September 2022
YOLO network-based drogue recognition method for autonomous aerial refueling
School of Automation, Northwestern Polytechnical University, Xi’an 710072, China
With the development of aerial refueling technology, autonomous aerial refueling(AAR) has become an important technology in the future battlefield, which is a promising prospective and challenging topic. Since the relative position between the receiver and the drogue is important to accomplish the AAR task, a neural network-based image recognition method is proposed to acquire the required information. Firstly, a C language-based YOLO network is used as the initial network, which meets the requirements of the on-board VxWorks system and can be run directly on the hardware. Then, considering the physical characterizes of the drogue, a multi-dimensional anchor box is designed and the network structure is optimized to adapt to the multi-dimensional situations. Finally, to address the problem of results shifts, feature maps with various sizes and the optimized loss function are used to optimize the network, where the pyramid structure suggests the design of feature maps. The experimental results indicate that the presented method can recognize the drogue more accurately and quickly.
随着空中加油技术的发展, 自主空中加油(autonomous aerial refueling, AAR)成为未来战场上的重要技术, 是具有前瞻性和挑战性的前沿课题。受油机和锥套之间的位置关系对于AAR十分重要, 故此提出一种基于神经网络的锥套图像识别方法。针对硬件要求, 使用以C语言为基础的YOLO网络作为初始网络, 使其符合机载操作系统VxWorks的要求, 可直接在嵌入式系统上运行。针对锥套的物理特点, 设计了多维度的anchor box, 优化了网络结构以适应锥套的多尺寸情况。针对识别结果漂移的问题, 参考金字塔结构使用了多种大小的特征图, 优化了网络的损失函数。测试结果表明, 经过优化设计的卷积神经网络模型在锥套图像数据集上能够更准确、更快速地识别所需目标。
Key words: target recognition / convolutional neural network / aerial refueling / YOLO
关键字 : 目标识别 / 卷积神经网络 / 空中加油 / YOLO
© 2022 Journal of Northwestern Polytechnical University. All rights reserved.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.