基于YOLOv5的微小型无人机实时探测方法  

Real-time detection method of micro UAV based on YOLOv5

作　　者：包文歧 谢立强 徐才华 刘智荣 朱敏 BAO Wenqi;XIE Liqiang;XU Caihua;LIU Zhirong;ZHU Min(College of Defense Engineering, Army Engineering University, Nanjing 210007, China)

机构地区：[1]陆军工程大学国防工程学院,南京210007

出　　处：《兵器装备工程学报》2022年第5期232-237,共6页Journal of Ordnance Equipment Engineering

基　　金：国家自然科学基金资助项目(51975584)。

摘　　要：无人机的广泛应用在给生产生活带来便利的同时,也对公共安全构成了威胁,这就需要对非法飞行的无人机进行探测和识别。然而,微小无人机因体积小运动灵活,使得传统的雷达、光电等探测手段难以应对。为此,提出了一种基于YOLOv5深度学习网络框架的微小无人机实时探测方法。通过拍摄无人机飞行姿态构建实验数据集,并进行标注。随后利用数据集对YOLOv5网络模型进行训练,测试训练效果,通过数据集测试网络模型能达到94.2%的精确率、82.8%的召回率和93.5%的平均精度。最后,对模型在真实场景下进行测试,视频流帧速率为30 FPS条件下,可在30 m范围内准确识别出特征尺寸为200 mm以上的无人机,并具有较好的实时性。While the widespread use of unmanned aerial vehicles(UAVs)brings convenience to production and life,they also poses a threat to public safety,which requires the detection and identification of illegal flying UAVs.However,due to the small size and flexible movement of micro-UAVs,traditional radar,photoelectric and other detection means are difficult to deal with.Therefore,a real-time detection method of small unmanned aerial vehicles based on YOLOv5 deep learning network framework was proposed.The experimental data set was constructed by photographing the UAV flight state,and then it was annotated.Then the YOLOv5 network model was trained using the data set and the training effect was tested.The network model was able to achieve 94.2%precision,82.8%recall and 93.5%mAP by testing on data set.Finally,the model was tested and verified in the real scene.Under the condition of the video stream frame rate of 30 FPS,the UAV with the characteristic size of over 200 mm can be accurately identified within the range of 30 meters and has good real-time performance.

关 键 词：目标检测 YOLOv5 微小型无人机 实时探测 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]