面向自主空中加油任务的目标检测技术研究  

Object Detection Technology for Autonomous Air to Air Refueling

作　　者：张西林 何亚坤 张恪易 王君秋 Zhang Xilin;He Yakun;Zhang Keyi;Wang Junqiu(Chinese Aeronautical Establishment,Beijing 100029,China)

机构地区：[1]中国航空研究院,北京100029

出　　处：《航空科学技术》2023年第2期64-71,共8页Aeronautical Science & Technology

摘　　要：针对软式空中加油任务自主对接过程中的视觉需求,本文进行了基于深度学习的自主空中加油锥套目标检测技术研究。首先,构造了一组包含不同目标尺度、背景并经数据增广后的软式自主空中加油锥套图像数据集。其次,结合当前主流的智能检测算法,对比分析了不同单双阶段检测算法在加油锥套数据集中的检测效果。在Faster RCNN网络添加了自适应置信度筛选模块,降低了其错检概率。为满足工程化应用需求,对YOLOv5网络进行轻量化改造,在几乎不降低检测精度的情况下提高了YOLOv5的检测速度,并大大降低了模型复杂度与运算资源消耗。该研究验证工作为后续展开锥套跟踪及位姿解算等研究提供了良好基础,可以为相关算法在自主空中加油任务中的进一步工程化应用提供重要的先验信息和技术参考。Based on the visual requirements during the process of autonomous air to air refueling,this paper conducts a research on drogue detection technology using deep learning methods.Firstly,a dataset of drogue is constructed and augmented,which contains different target scales and background.Secondly,the detection performance of different detection algorithms in the refueling dorgue dataset is compared and analyzed.An adaptive confidence filtering module is added to the Faster RCNN network to reduce its error detection probability.In order to meet the requirements of engineering applications,the YOLOv5 network is lightened and modified to improve the detection speed with little reduction in detection accuracy,and greatly reduce the model complexity and computing resource consumption.This paper provides a good basis for the subsequent research on dorgue tracking and attention solution,and can provide important information and technical references for further engineering applications of the relevant algorithms in AAAR missions.

关 键 词：自主空中加油 目标检测 深度学习 软管锥套 计算机视觉 

分 类 号：V219[航空宇航科学与技术—航空宇航推进理论与工程]