针对低空微小型无人机的轻量型YOLOv5检测算法
Lightweight YOLOv5 detection algorithm for low-altitude micro UAV
魏峰 1周建平 1谭翔 2林静 3田莉 3王虎1
作者信息
- 1. 新疆大学智能制造现代产业学院,新疆维吾尔自治区乌鲁木齐 830000
- 2. 新疆大学智能制造现代产业学院,新疆维吾尔自治区乌鲁木齐 830000;中国科学院地理科学与资源研究所,北京 100101;中国科学院无人机应用与管控研究中心,北京 100101
- 3. 中国科学院地理科学与资源研究所,北京 100101
- 折叠
摘要
针对低空微小型无人机对公共安全造成威胁的问题,本文基于YOLOv5(you only look once v5)网络提出了 一种适用于移动端的轻量型目标检测模型YOLOv5_SS.该模型以轻量型网络ShuffleNetv2 替换 YOLOv5 原有的主干网络,引入 SENet(squeeze-and-excitation networks)注意力机制,并采用Soft-NMS(soft non-maximum suppression)算法提升对密集重叠目标的检测效果.实验结果表明,该模型在数据集上对低空微小无人机进行检测的平均精确率均值(mean average precision@0.5,mAP50)为92.75％,精度为90.49％,参数量为0.237 4 M,浮点运算数为0.9千兆浮点运算(giga floating-point operations,GFLOPS).具有检测精度高、内存占用率低的特点,有利于在移动终端上部署且在复杂背景及密集目标的场景下均有较好的检测效果.
Abstract
Aiming at the problem that low-altitude micro-UAVs pose a threat to public safety,this paper proposes a lightweight target detection model YOLOv5_SS suitable for mobile terminals based on the you only look once v5(YOLOv5)network.In this model,the lightweight network ShuffleNetv2 replaces the original backbone network of YOLOv5,introduces squeeze-and-excitation networks(SENet)attention mechanism,and uses soft non-maximum suppression(Soft-NMS)algorithm to improve the detection effect of dense overlapping targets.The experimental results show that the mean average precision@0.5(mAP50)of the model for the detection of low-altitude micro-UAV on the dataset is 92.75％,the accuracy is 90.49％,and the number of parameters is 0.237 4 M.The number of floating-point operations is 0.9GFLOPS(giga floating-point operations).
关键词
无人机检测/深度学习/轻量型网络/注意力机制/非极大值抑制(NMS)Key words
UAV detection/deep learning/lightweight network/attention mechanism/non-maximum suppression(NMS)引用本文复制引用
出版年
2024
NETL
NSTL
万方数据