摘 要:针对靶场返回舱自动识别与跟踪的需求,采用 YOLOv5 神经网络进行自动识别的技术方案。为解决返回舱数据集较少的问题,采用 Mosaic 数据增强的方法,不仅增加了数据集数量,而且提高了网络的鲁棒性。针对返回舱目标较小的问题,提出以特征明显、目标较大的降落伞作为主要识别对象,返回舱本体为次要识别对象,对二者同时进行识别。通过数据集的建立、训练和测试,最终得出实验结论:在不同环境、不同光线条件下的降落伞和返回舱识别中,所提方法检测准确率可达90% ~ 95%,对返回舱落地过程中高效、准确识别跟踪任务具有重要意义。
关键词:返回舱;降落伞;目标识别;YOLOv5
DOI:10.19850/j.cnki.2096-4706.2021.10.005
中图分类号:TP183 文献标识码:A 文章编号:2096-4706(2021)10-0020-07
Identification of Return Cabin Based on Convolutional Neural Network
FENG Kai 1 ,ZHANG Shuya 1 ,LI Jinxuan1 ,MA Shuli 2 ,QIAN Kechang2
(1.School of Aerospace Command,PLA Strategic Support Force University of Aerospace Engineering,Beijing 101416,China; 2.School of Aerospace Information,PLA Strategic Support Force University of Aerospace Engineering,Beijing 101416,China)
Abstract:Aiming at the requirement of automatic identification and tracking of the return cabin in the shooting range,it adopts YOLOv5 neural network for automatic identification. In order to solve the problem of less data sets in the return cabin,Mosaic data enhancement method is adopted,which not only increases the number of data sets,but also improves the robustness of the network. Aiming at the problem that the target of the return cabin is small,it is proposed,taking the parachute with obvious characteristics and large target as the main identification object and the return cabin body as the secondary identification object,to identify both of them at the same time. Through the establishment,training and testing of the data sets,the experimental conclusion is finally drawn:in the identification of parachutes and return cabins under different environments and different light conditions,the detection accuracy of the proposed method can reach 90% ~ 95%,which is of great significance to identify tracking targets efficiently and accurately during the landing of return cabins.
Keywords:return cabin;parachute;target identification;YOLOv5
参考文献:
[1] 柯岩,林小竹,廖蕊,等 . 卷积神经网络层级分解研究 [J]. 计算机工程,2019,45(11):191-197.
[2] 侯笑晗,金国栋,谭力宁,等 . 基于自适应和最优特征 的合成孔径雷达舰船检测方法 [J]. 计算机应用,2021,41(7): 2150-2155.
[3] 徐德洋,王蓬勃,孙立宁 .FPN 的影响因素分析及相关应 用 [J]. 材料导报,2014,28(19):104-110.
[4] 檀结庆,朱星辰,蔡蒙琪.基于GMSD的新混合损失函数 [J]. 合肥工业大学学报(自然科学版),2021,44(4):472-477+502.
[5] 侯志强,刘晓义,余旺盛,等 . 使用 GIoU 改进非极大值 抑制的目标检测算法 [J]. 电子学报,2021,49(4):696-705.
[6] 杨有为,周刚 . 面向自然场景文本检测的改进 NMS 算法 [J/OL]. 计算机工程与应用:1-7(2021-01-13).http://kns.cnki. net/kcms/detail/11.2127.TP.20210112.1135.006.html.
作者简介:冯凯(2001—),男,汉族,山西临汾人,本科在 读,研究方向:目标检测算法;张书雅(2001—),女,汉族,陕 西宝鸡人,本科在读,研究方向:目标检测算法;李锦暄(2001—), 女,汉族,陕西宝鸡人,本科在读,研究方向:目标检测算法;马 淑丽(1989—),女,汉族,山东济宁人,讲师,博士,研究方向: 图像重建、凸优化、张量理论;钱克昌(1984—),男,汉族,江 苏邳州人,副教授,博士,研究方向:信息对抗与安全。