基于YOLO的航管一次雷达目标检测方法  

Target detection method of primary surveillance radar based on YOLO

作　　者：施端阳 林强[1] 胡冰[1] 杜小帅 SHI Duanyang;LIN Qiang;HU Bing;DU Xiaoshuai(Air-Defense Early Warning Equipment Department,Air Force Early Warning Academy,Wuhan 430019,China;Unit 95174 of the PLA,Wuhan 430040,China;Unit 94005 of the PLA,Jiuquan 735000,China)

机构地区：[1]空军预警学院防空预警装备系,湖北武汉430019 [2]中国人民解放军95174部队,湖北武汉430040 [3]中国人民解放军94005部队,甘肃酒泉735000

出　　处：《系统工程与电子技术》2024年第1期143-151,共9页Systems Engineering and Electronics

摘　　要：针对传统恒虚警率(constant false alarm rate,CFAR)检测方法检测率低的问题,提出一种基于YOLO(you only look once)的深度学习雷达目标检测方法。首先,利用同相正交(in-phase/quadrature,I/Q)数据匹配滤波后形成的雷达原始图像自建雷达目标图像数据集。然后,改进YOLO检测模型的网络结构、特征融合策略和损失函数以提高模型的精度,并引入迁移学习思想,利用预训练的深度学习网络提取图像特征,降低了检测模型对训练样本量的要求。最后,在自建数据集上对YOLO目标检测方法进行了实验验证。航管一次雷达实测数据的实验证明:与传统CFAR检测方法和两阶段的快速区域卷积神经网络(region convolutional neural networks,R-CNN)检测方法相比,所提方法的目标检测率大幅提高,虚警率明显降低,且实现了实时检测。Aiming at the low detection rate of traditional constant false alarm rate(CFAR)detection methods,a deep learning radar target detection method based on you only look once(YOLO)is proposed.Firstly,the radar target image dataset is constructed by using the original radar image formed by in-phase/quadrature(I/Q)data matching filtering.Then,the network structure,feature fusion strategy,and loss function of the YOLO detection model are improved to improve the accuracy of the model.And the idea of transfer learning is introduced to extract image features using the pre-trained deep learning network,which reduced the requirement of the detection model on the training sample size.Finally,the YOLO target detection method is experimentally verified on the self-built dataset.The experimental results on the measured data of the primary surveillance radar show that,compared with the traditional CFAR detection method and the two-stage faster region convolutional neural networks(R-CNN)detection method,the target detection rate of the proposed method is greatly improved,the false alarm rate is significantly reduced,and the real-time detection is realized.

关 键 词：航管一次雷达 深度学习 目标检测 YOLO 

分 类 号：TN957[电子电信—信号与信息处理]