主办：陕西省汽车工程学会
ISSN 1671-7988  CN 61-1394/TH
创刊：1976年
汽车实用技术

汽车实用技术 ›› 2024, Vol. 49 ›› Issue (2): 67-73.DOI: 10.16638/j.cnki.1671-7988.2024.002.013

• 智能网联汽车 • 上一篇    

基于改进 YOLOX 的交通标志检测

陈涧鑫 1，甘海云 2   

  1. 1.天津职业技术师范大学 汽车与交通学院； 2.天津市智能交通技术工程中心
  • 发布日期:2024-01-30
  • 通讯作者: 陈涧鑫
  • 作者简介:陈涧鑫（1997－），男，硕士研究生，研究方向为视觉 SLAM，E-mail:19746850990@qq.com。

Traffic Sign Detection Based on Improved YOLOX

CHEN Jianxin1 , GAN Haiyun2   

  1. 1.School of Automobile and Transportation, Tianjin University of Technology and Education; 2.Tianjin Intelligent Transportation Technology Engineering Center
  • Published:2024-01-30
  • Contact: CHEN Jianxin

PDF

摘要： 现有的交通标志检测算法无法做到精准快速地检测，并且存在很大程度漏检，针对此 问题，选取 YOLOX-s 作为基础网络模型，首先在 Backbone 部分增加有效输出层，并在 PAFPN 部分改进多尺度特征融合方式，增加融合（SF）结构，使网络进一步融合图像浅层特征。其 次在 FAFPN 部分嵌入坐标注意力机制，使模型更准确定位目标区域。最后针对损失函数进行 改进，使用 EIoU 的计算边界框回归损失，使用 Polyloss 计算类别和置信度损失。改进后的模 型在 TT100K 数据集上进行实验平均精度均值（mAP）达到 92.70%，相较于原 YOLOX-s 模 型仅在参数量增加 0.2 MB 的基础上，mAP 提升了 11.43%且检测速度达到 77 frame/s，满足实 时性需求。改进后模型对交通标志识别的准确率有较大提升，交通标志检测能力的提升是实 现可持续、高效和安全的自动驾驶交通系统的关键一步。

关键词: 交通标志检测；YOLOX；特征融合；注意力机制；损失函数

Abstract: The existing traffic sign detection algorithms cannot achieve accurate and fast detection and suffer from a significant degree of missed detections. To address this problem, YOLOX-s is selected as the base network model. Firstly, effective output layers are added to the Backbone section, and the multi-scale feature fusion method is improved in the PAFPN section by adding the structure fusion (SF) component, which further incorporates shallow image features. Secondly, the FAFPN section embeds a coordinate attention mechanism to enable more precise localization of the target regions. Lastly, improvements are made to the loss function, using EIoU for bounding box regression loss and Polyloss for category and confidence losses.The improved model achieves an mean average precision (mAP) of 92.70% on the TT100K dataset. Compared to the original YOLOX-s model, the parameter count only increases by 0.2 MB, yet the mAP is improved by 11.43%, and the detection speed reaches 77 frames per second, meeting real-time requirements. The enhanced model shows significantly improved accuracy in traffic sign recognition, and the advancement in traffic sign detection capabilities is a crucial step towards achieving sustainable, efficient, and safe automated driving traffic systems.

Key words: Ttraffic sign detection; YOLOX; Feature fusion; Attention mechanism; Loss function