[1]张 震,王晓杰,晋志华,等.基于轻量化 YOLOv5 的交通标志检测[J].郑州大学学报(工学版),2024,45(02):12-19.[doi:10.13705/j.issn.1671-6833.2023.05.041]
 ZHANG Zhen,WANG Xiaojie,JIN Zhihua,et al.Traffic Sign Detection Based on Lightweight YOLOv5[J].Journal of Zhengzhou University (Engineering Science),2024,45(02):12-19.[doi:10.13705/j.issn.1671-6833.2023.05.041]
点击复制

基于轻量化 YOLOv5 的交通标志检测()
分享到:

《郑州大学学报(工学版)》[ISSN:1671-6833/CN:41-1339/T]

卷:
45卷
期数:
2024年02期
页码:
12-19
栏目:
出版日期:
2024-03-06

文章信息/Info

Title:
Traffic Sign Detection Based on Lightweight YOLOv5
作者:
张 震 王晓杰 晋志华 马继骏
1. 郑州大学 电气与信息工程学院,河南 郑州 450001;2. 河南省交通调度指挥中心,河南 郑州 450001
Author(s):
ZHANG Zhen 1 WANG Xiaojie 1 JIN Zhihua 1 MA Jijun
1. School of Electrical and Information Engineering, Zhengzhou University, Zhengzhou 450001, China; 2. Henan Province Transportation Dispatching Command Center, Zhengzhou 450001, China
关键词:
交通标志检测 轻量化 YOLOv5 SIoU 损失函数 Ghost 卷积 TT100K BiFPN
Keywords:
traffic sign detection lightweight YOLOv5 SIoU loss function Ghost convolution TT100K BiFPN
DOI:
10.13705/j.issn.1671-6833.2023.05.041
文献标志码:
A
摘要:
为了提高道路交通标志的检测速度,提出一种基于轻量化 YOLOv5 的改进模型。 首先,使用 Ghost 卷积和 深度分离卷积(DWConv)构建新的主干模块,减少计算量和参数量;引入加权特征融合网络( BiFPN) 结构,增强特 征融合能力;将 CIoU 损失函数替换为 SIoU 损失函数,关注真实锚框与预测的角度信息,提升检测精度。 其次,对 TT100K 数据集进行优化,筛选出标签个数大于 200 的交通标志图片和标注信息共 24 类。 最后,实验结果取得 84%的准确率、81. 2%的召回率和 85. 4%的所有类别平均精确率的平均值 mAP@ 0. 5,相比原始 YOLOv5,参数量减 少 29. 0%,计算量减少 29. 4%,mAP@ 0. 5 仅下降 0. 1 百分点,检测帧率提升了 34 帧 / s。 使用改进后的模型进行检 测,检测速度有了明显提升,基本达到了在保持检测精度的基础上压缩模型的目的。
Abstract:
In order to improve the detection speed of road traffic signs, an improved model based on lightweight YOLOv5 was proposed. Firstly, Ghost convolution and depthwise convolution were used to build a new Bottleneck, which could reduce the amount of computation and parameters. Then the BiFPN structure was introduced, which could enhance the feature fusion ability. CIoU loss function was replaced by SIoU loss function, which focused on the angle information of ground true box and prediction one, so that it would improve the detection accuracy. Secondly, the TT100K dataset was optimized, and 24 categories of traffic sign pictures and labels with more than 200 were screened out. Finally, the experiment achieved 84% accuracy, 81. 2% recall and 85. 4% mAP@ 0. 5. Compared with the original YOLOv5 model, the number of parameters was reduced by 29. 0%, the amount of computation was reduced by 29. 4%, but the mAP@ 0. 5 was only reduced by 0. 1 percentages, and the detection frame rate was improved by 34 frames/ s. Using the improved model for detection, the detection speed could be significantly improved, could basically achieve the goal of compression model on the basis of maintaining the detection accuracy.
更新日期/Last Update: 2024-03-08