适用于边缘设备的轻量级人体检测算法
Lightweight human body detection algorithm suitable for edge devices
周宁 1陶青川 1彭勃兴1
作者信息
- 1. 四川大学电子信息学院,成都 610065
- 折叠
摘要
针对现存的人体检测网络都比较复杂,部署到边缘设备上时表现不佳的问题,基于YOLOv7提出一种轻量级人体检测算法.该算法首先使用改进后的ShuffleNev2基本模块替换原网络ELAN模块;接着在主干网络末端添加SE注意力和SPPF池化;然后在Neck部分使用改进后的GSConv替换标准卷积,引入基于GSConv的VoVGSCSP替换ELAN-W模块.通过在GPU和Sophon SE5上的验证结果表明,该轻量级人体检测算法与YOLOv7相比损失2.6%的精度,但计算量大幅度降低,在Sophon SE5上推理速度达到了54 FPS,相比较YOLOv7提升了39 FPS.
Abstract
A lightweight human detection algorithm based on YOLOv7 is proposed to address the issue of complex human de-tection networks that perform poorly when deployed on edge devices.The algorithm first replaces the original network ELAN mod-ule with the improved ShuffleNev2 basic module;Next,add SE attention and SPPF pooling at the end of the backbone network;Then,in the Neck section,the improved GSConv is used to replace the standard convolution,and the GSConv based VoVGSCSP is introduced to replace the ELAN-W module.The validation results on GPU and Sophon SE5 show that this lightweight human detec-tion algorithm loses 2.6%accuracy compared to YOLOv7,but significantly reduces computational complexity.The inference speed on Sophon SE5 reaches 54 FPS,which is 39 FPS higher than YOLOv7.
关键词
YOLOv7/目标检测/ShuffleNet/边缘计算/Sophon/SE5Key words
YOLOv7/target detection/ShuffleNet/edge computing/Sophon SE5引用本文复制引用
出版年
2024
NETL
NSTL
万方数据