基于沙漏瓶颈模块的无人机实时检测算法  

UAV Real-time Detection Algorithm Based on SandGlass Bottleneck Block

作　　者：李鸿 杜芸彦 邵林松 雷铭 彭锦锦 杨锦辉 毛耀 LI Hong;DU Yunyan;SHAO Linsong;LEI Ming;PENG Jinjin;YANG Jinhui;MAO Yao(Key Laboratory of Optical Engineering,Chinese Academy of Sciences,Chengdu 610000,China;Institute of Optics and Electronics,Chinese Academy of Sciences,Chengdu 610000,China;University of Chinese Academy of Sciences,Beijing 100000,China)

机构地区：[1]中国科学院光束控制重点实验室,成都610000 [2]中国科学院光电技术研究所,成都610000 [3]中国科学院大学,北京100000

出　　处：《电光与控制》2022年第12期58-65,共8页Electronics Optics & Control

基　　金：国家重点研发计划(2017YFB1103002)。

摘　　要：随着无人机的快速发展与应用,无人机的普及也对公共安全、军事安全和个人隐私等造成了一定的安全隐患。无人机具有飞行速度快、体积小等特点,如何精准快速地发现并定位无人机位置具有一定的挑战。针对此问题,提出了一种基于沙漏瓶颈模块的YOLOv3无人机实时检测算法。首先,将原本3个特征尺度检测扩展为在5个特征尺度上进行检测,充分利用多尺度信息帮助提升小目标检测精度;然后,堆叠沙漏瓶颈模块作为该方法的骨干网络部分,沙漏瓶颈模块作为一种轻量化网络对模型进行加速,并使用通道注意力机制在上采样之后的拼接部分关注更重要的通道信息,抑制不利的信息。为了验证所提算法的有效性,生成基于复杂城市背景下的无人机数据集,实验结果表明,所提算法的精度能够达到98.92%,且具有98.76%的召回率,在1080Ti上达到62.37帧/s的实时速度,模型权重大小仅为5.38 MiB,为进一步在嵌入式平台和移动端实现实时目标检测提供了可能。With the rapid development and application of UAVs,the popularity of UAVs has also caused certain security risks to public security,military security and personal privacy.UAVs has the characteristics of high flying speed and small volume,so how to accurately and quickly find and locate the position of UAVs is a challenge.A real-time detection algorithm for YOLOv3 UAVs based on SandGlass Bottleneck Block is proposed.Firstly,the original three feature scales are extended to five feature scales to make full use of multi-scale information to help improve the detection accuracy of small targets.Then,the SandGlass Bottleneck Block is stacked as the backbone network part of the method,and the SandGlass Bottleneck Block is taken as a lightweight network to accelerate the model,which uses the channel attention mechanism to focus on more important channels in the splicing part after upsampling unformation and suppresses unfavorable information.In order to verify the proposed algorithm effectiveness,a UAVs data set is generated based on complex urban background.Experimental results show that the proposed algorithm can achieve 98.92% accuracy and a recall rate of 98.76%,achieves a real-time detection speed of 62.37 FPS on 1080Ti graphics card,the model weight is only 5.38 MiB,which further provides the possibility for real-time target detection on embedded platforms and mobile devices.

关 键 词：无人机检测 轻量化网络 沙漏瓶颈模块 特征金字塔网络 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]