基于改进YOLOv5s的轻量级车辆检测系统
Lightweight Vehicle Detection System Based on the Improved YOLOv5s
林小涵 1侯典立 1郑红霞2
作者信息
- 1. 鲁东大学信息与电气工程学院,山东 烟台 264039
- 2. 鲁东大学交通学院,山东 烟台 264039
- 折叠
摘要
基于深度学习的车辆检测在智慧交通中起着重要作用,现有模型结构复杂、计算量大,难以布置在嵌入式系统的边缘设备.本文提出一种基于YOLOv5s的轻量级车辆检测算法,通过GhostNet和剪枝优化策略对YOLOv5s进行了改进,实现系统的轻量化和检测的实时性;损失函数Focal-EIoU Loss的引入解决了样本不平衡和纵横比模糊定义问题,从而提高目标检测的性能.在UA-DETRAC数据集上的实验结果表明,所提算法与原YOLOv5s算法相比参数量、模型体积和浮点运算次数分别减少了 63%、50.6%和 64.7%,检测速度提升了50%,同时保持了较高精度、召回率,为空间、能源、资源有限的边缘检测设备提供了一种实时性算法选择.
Abstract
Vehicle detection based on deep learning plays an important role in smart transportation.The existing model structure which is complex and computationally intensive makes it difficult to deploy on edge devices of embedded systems.This paper proposed a lightweight vehicle detection algorithm based on YOLOv5s.YOLOv5s was improved through GhostNet and pruning optimization strategies to achieve lightweight system and real-time detection;and the introduction of the loss function Focal-EIoU Loss solved the problem of sample imbalance and aspect ratio blur define the problem,thereby improving the performance of object detec-tion.Experimental results on the UA-DETRAC data set are as follows.Compared with the original YOLOv5s al-gorithm,the proposed algorithm reduces the number of parameters,model volume and FLOPs by 63%,50.6%and 64.7%respectively,and the detection speed increases by 50%,while maintaining high precision and recall rate.This paper provides a real-time algorithm choice for edge detection equipment with limited space,energy,and resources.
关键词
YOLOv5s/GhostNet/车辆检测/轻量级模型/EIoUKey words
YOLOv5s/GhostNet/vehicle detection/lightweight model/EIoU引用本文复制引用
基金项目
山东省高等学校科学技术计划项目(J16LN31)
烟台市智慧城市创新实验室科研课题项目(SDGP370600000202302000504)
出版年
2024
NETL
NSTL
万方数据