面向空中目标拦截的轻量化弱小目标检测方法与应用  

Lightweight dim and small objects detection method and application for aerial target interception

作　　者：卢田雨 秦闻远 化永朝 潘成伟[2] 李清东[1] 董希旺[1,2,3] LU Tianyu;QIN Wenyuan;HUA Yongzhao;PAN Chengwei;LI Qingdong;DONG Xiwang(School of Automation Science and Electrical Engineering,Beihang University,Beijing 100191,China;School of Artificial Intelligence,Beihang University,Beijing 100191,China;Institute of Unmanned System,Beihang University,Beijing 100191,China)

机构地区：[1]北京航空航天大学自动化科学与电气工程学院,北京100191 [2]北京航空航天大学人工智能学院,北京100191 [3]北京航空航天大学无人系统研究院,北京100191

出　　处：《导航定位与授时》2024年第5期66-81,共16页Navigation Positioning and Timing

基　　金：国家自然科学基金(U2241217,62103023,62103016,62203032);北京市自然科学基金(JQ23019,4232046);中国科协青年人才托举工程(2021QNRC001);北京市科协青年人才托举工程(BYESS2022186);航空科学基金(2022Z071051015);中央高校基本科研业务费专项资金(YWF-23-T-207,YWF-23-L-1138)。

摘　　要：空中目标拦截技术依赖于目标检测与跟踪技术,而弱小目标检测作为其中的难点,其效果将直接影响整个系统的性能。针对此问题,提出了一种面向空中目标拦截的轻量化弱小目标检测方法。首先,针对弱小目标全局信息较少的问题,以YOLOv5作为基础网络,使用Swin Transformer替代其结构中的C3模块,从而增强网络的局部信息捕获能力。随后,为了补偿被稀释的语义信息,在此基础上提出具备跨连接策略的特征融合网络,通过额外融合不同尺度的特征图,解决这个问题。最后,在特征融合网络上使用一次额外的上采样并融合高分辨率特征图,进一步提升网络对弱小目标检测的能力。此外,引入目标跟踪神经网络DaSiamRPN实现对动态弱小目标长时间跟踪。为保证无人飞行器上的边缘计算设备可以实时进行模型推理,在上述基础上进行模型轻量化,剔除了模型的大尺度目标检测头,以减少模型的参数量。通过计算,改进后的算法较原YOLOv5模型参数量减少了21.5%。在VisDrone2019数据集上的实验结果表明,提出的基于YOLOv5s的轻量化目标检测算法能更好地完成弱小目标检测任务,在准确率、召回率及平均精度均值上分别达到了96.3%、59%和40.2%,各指标均明显高于原始YOLOv5s算法,且优于目前主流目标检测算法。同时在TinyPerson数据集上进行泛化实验,实验结果表明,改进后算法的弱小目标检测性能得到明显提高。为了进一步验证所提方法的有效性,在无人机平台上进行了空中目标拦截的飞行测试,结果表明该方法能很好地完成目标检测跟踪任务并成功拦截目标,为空中目标拦截提供有力的支持。Aerial target interception technology relies on target detection and tracking techniques,with the detection of dim and small objects being a challenging aspect that directly affects the overall system performance.To address this issue,a lightweight dim and small object detection method is proposed for aerial target interception.Firstly,given the problem of limited global information for dim and small objects,the method is based on the YOLOv5 network,and the Swin Transformer is introduced to replace the C3 module in its architecture,thereby enhancing the network s ability to capture local information.Then,to compensate for diluted semantic information,a feature fusion network with cross-connection strategies is introduced to facilitate the fusion of feature maps at different scales to mitigate this problem.Finally,an additional upsampling is applied to the feature fusion network and high-resolution feature maps are fused to further improve the network s ability to detect dim and small objects.Furthermore,the DaSiamRPN neural network is incorporated for long-term tracking of dynamically dim and small objects.To ensure that the edge computing devices on unmanned aerial vehicles can perform model inference in real time,the model has been lightweighted on the basis of the aforementioned,and the large-scale object detection head of the model is removed to reduce the number of model parameters.Calculations show that the improved algorithm reduces the number of parameters by 21.5%compared to the original YOLOv5 model.The experimental results on VisDrone2019 show that the proposed lightweight object detection algorithm performs better in detecting dim and small objects,achieving precision,recall,and mean average precision(mAP)of 96.3%,59%,and 40.2%,respectively.These metrics are significantly higher than those of the original YOLOv5s algorithm and surpass those of current mainstream object detection algorithms.Meanwhile,generalization experiments are carried out on the TinyPerson datasets,and experimental results indicate

关 键 词：无人机拦截 弱小目标 轻量化目标检测模型 YOLOv5 深度学习 自注意力机制 

分 类 号：V249.3[航空宇航科学与技术—飞行器设计]