摘要
为提高无人机在复杂环境下对地面未爆弹(UXO)目标的辨识精度,提出了一种改进YOLOv5的UXO目标检测方法。该方法在YOLOv5的基础上,改进原YOLOv5网络的损失函数以提高对UXO目标的识别精度,同时,通过添加注意力机制、改进马赛克数据增强、改进预测框筛选机制提高对UXO目标的识别效率,实现了空对地场景下对UXO目标的检测,并具有较好的精度和速度。实验选取多组不同复杂背景的UXO数据集进行标注并训练,得到UXO目标模型,然后从模型训练结果和目标检测结果的角度评估方法和模型的正确性。实验结果表明:NGG-YOLOv5所得模型检测准确性和检测速度对比原YOLOv5有明显的提升,准确率从78%提高至91%,平均精度均值(mAP)从50%提高至56%,在所用4种复杂背景下均可有效检测出UXO目标,且漏警率低。
In order to improve the recognition accuracy of Unexploded Ordnance(UX0)targets on the ground by UAVs in complex environments,a UX0 target detection method based on the improved YOLOv5 is proposed.On the basis of YOL0v5,the method improves the loss function of the original YOLOv5 network to improve the recognition accuracy of UX0 targets.At the same time,the method adds an attention mechanism,improves mosaic data enhancement,and improves the prediction frame screening mechanism to improve the recognition efficiency of UX0 targets,and realizes the detection of UXO targets in air-to-ground scenarios with better accuracy and speed.Experimentally,multiple UX0 datasets in different complex backgrounds are selected,labeled and trained to obtain UX0 target models.Then,the correctness of the algorithm and model is evaluated from the perspectives of model training results and target detection results.The experimental results show that:1)The model obtained by NGG-YOLOv5 has a significant improvement in detection accuracy and detection speed in comparison with that obtained by the original YOLOv5,with an increase in accuracy from 78%to 91%and an increase in mean Average Precision(mAP)from 50%to 56%;and 2)It can effectively detect UX0 targets in four kinds of complex backgrounds,with a low missed alarm rate.
作者
刘子玉
赵旭
李连鹏
代牮
LIU Ziyu;ZHAO Xu;LI Lianpeng;DAI Jian(Beijing Key Laboratory of High Dynamic Navigation Technology,Beijing Information Science and Technology University,Beijing 100000,China;Beijing Honda Hechuang Defense Technology Research Institute Co.Ltd.,Beijing 100000,China)
出处
《电光与控制》
CSCD
北大核心
2024年第3期70-74,共5页
Electronics Optics & Control
基金
国家重点研发计划(2020YFC1511702)
北京市科技计划课题(Z221100005222024)
高动态导航技术北京市重点实验室资助
“慧眼行动”创新成果转化应用项目(×××新型多模智能探测系统)。
关键词
空对地探测
YOLOv5
未爆弹目标
深度学习
复杂环境
air-to-ground detection
YOLOv5
unexploded ordnance target
deep learning
complex environment
