摘要
为了解决当前交通标志检测方法在小目标、模糊目标及复杂环境下检测精度低的问题,本文提出了一种改进的交通标志检测模型YOLOv8-NTS,以提升模型在复杂交通环境下的识别性能。模型在YOLOv8基础上进行了3方面改进:首先在骨干网络中设计了经过轻量化的混合注意力变换器模块SlimHAT,以增强对全局像素信息的建模能力,提高特征表示精度;其次设计了基于WTConv的WT-C2fBlock模块替代原有的C2f模块,在减少12.2%模型参数量的同时保持模型检测精度不变;最后设计了结合空间注意力机制与卷积操作的新型检测头RFAhead,优化特征提取和融合过程,进一步增强模型对目标的表达能力和鲁棒性。在TT100K交通标志数据集上的实验表明,与基线模型YOLOv8相比,改进后的YOLOv8-NTS模型在精确率、召回率、mAP50及mAP50~90指标上分别提升了6.5%、5.0%、7.3%和5.3%,显示出显著的性能优势。所提出的YOLOv8-NTS模型能够在保持较低计算成本的同时显著提升交通标志检测精度和泛化能力,验证了该方法的有效性与实用价值,可为智能交通场景中的交通标志识别提供可靠的技术支持。
To address the low detection accuracy of current traffic sign detection methods for small,blurred targets and complex environments,this paper proposes an improved traffic sign detection model YOLOv8-NTS,to enhance recognition performance in complex traffic scenarios.The model incorporates three key enhancements over YOLOv8:First,it introduces the lightweight Hybrid Attention Transformer(SlimHAT)module within the backbone network to strengthen global pixel information modeling and improve feature representation accuracy.Second,it replaces the original C2f module with the WT-C2fBlock module based on WTConv,reducing model parameters by 12.2%while maintaining detection accuracy.Finally,a novel detection head RFAhead was designed by integrating spatial attention mechanisms with convolutional operations,optimizing feature extraction and fusion processes to further enhance the model’s object representation capability and robustness.Experiments on the TT100K traffic sign dataset demonstrate that compared to the baseline YOLOv8 model,the improved YOLOv8-NTS achieves significant performance gains:6.5%increase in precision,5.0%increase in recall,7.3%improvement in mAP50,and 5.3%enhancement in mAP50~90.The proposed YOLOv8-NTS model substantially improves traffic sign detection accuracy and generalization capabilities while maintaining low computational cost,validating the method’s effectiveness and practical value.It provides reliable technical support for traffic sign recognition in intelligent transportation scenarios.
作者
李鹏飞
熊召新
王桂宝
LI Pengfei;XIONG Zhaoxin;WANG Guibao(School of Physics and Electronic Engineering,Shaanxi University of Technology,Hanzhong 723000,China;School of Electronic Information and Artificial Intelligence,Shaanxi University of Science&Technology,Xi′an 710021,China)
出处
《液晶与显示》
北大核心
2025年第12期1868-1880,共13页
Chinese Journal of Liquid Crystals and Displays
基金
陕西省重点研发计划(No.2025CY-YBXM-122)
陕西省秦创原“科学家+工程师”队伍建设项目(No.2024QCY-KXJ-168)。
