改进YOLOv5的无人机影像多目标检测方法研究
Research on the method of improved YOLOv5 in the UAV images multi-target detection
欧先萍1
作者信息
- 1. 广东省国土资源测绘院,广东 广州 510700
- 折叠
摘要
针对无人机机载硬件部署目标检测模型问题,提出一种基于改进YOLOv5的轻量模型.在骨干网络中使用瓶颈结构卷积核组实施特征提取来降低计算量,同时,引入通道注意力模块让模型充分学习正样本特征.在多头检测端,使用改进NMS算法更精准地筛选出最优检测框.使用增广数据集训练模型并在低功耗硬件上完成模型测试.结果表明:本研究模型能够对多类目标进行高精度、高鲁棒性检测,并且能够在低功耗硬件环境下实时输出检测结果,适合部署在无人机机载硬件中对地面多类别目标实施检测.
Abstract
A lightweight model based on improved YOLOv5 is proposed to address the issue of target detection models for unmanned aerial vehicle hardware deployment.In the backbone network,bottleneck structure convolutional kernels are used to implement feature extraction so as to reduce computational complexity,meanwhile the channel attention modules are introduced to make the model fully learn positive sample features.At the multi-head detection end,an improved NMS algorithm is used to more accurately filter out the optimal detection boxes.The model is trained with an augmented dataset and the model testing is completed on the low-power hardware.The results show that the model proposed in this paper can detect multiple types of targets with high accuracy and robustness,and can output detection results in real-time in the low-power hardware environments.It is suitable for deployment in unmanned aerial vehicle hardware to detect ground multiple types of targets.
关键词
无人机影像/轻量检测模型/YOLOv5/瓶颈结构/通道注意力Key words
unmanned aerial vehicle images/lightweight detection model/YOLOv5/bottleneck structure/channel attention引用本文复制引用
出版年
2024
NETL
NSTL
万方数据