华南理工大学学报(自然科学版) ›› 2026, Vol. 54 ›› Issue (3): 65-78.doi: 10.12141/j.issn.1000-565X.250092

• 智慧交通系统 • 上一篇    下一篇

基于CGT-YOLO的小目标交通标志识别算法

邢岩1,2, 郭思豪1, 张振2,3, 潘晓东2,3, 安冬1,4   

  1. 1.沈阳建筑大学 交通与测绘工程学院,辽宁 沈阳 110168
    2.道路交通安全管控技术国家工程研究中心,辽宁 沈阳 110168
    3.沈阳市公安局交通管理支队,辽宁 沈阳 110168
    4.沈阳寒武纪交通科技有限公司,辽宁 沈阳 110168
  • 收稿日期:2025-04-01 出版日期:2026-03-25 发布日期:2025-10-31
  • 通信作者: 郭思豪(1999 —),男,硕士,主要从事智能交通研究。 E-mail:27129433@qq.com
  • 作者简介:邢岩(1985—),男,博士,教授,主要从事智能交通研究。E-mail: xingyan@sjzu.edu.cn
  • 基金资助:
    道路交通安全管控技术国家工程研究中心开放课题(2024GCZXKFKT13B)

CGT-YOLO-Based Algorithm for Small-Target Traffic Sign Recognition

XING Yan1,2, GUO Sihao1, ZHANG Zhen2,3, PAN Xiaodong2,3, AN Dong1,4   

  1. 1.School of Transportation and Geomatics Engineering,Shenyang Jianzhu University,Shenyang 110168,Liaoning,China
    2.National Engineering Research Center for Road Traffic Safety Control Technology,Shenyang 110168,Liaoning,China
    3.Traffic Management Detachment of Shenyang Public Security Bureau,Shenyang 110168,Liaoning,China
    4.Shenyang Cambrian Transportation Technology Co. ,Ltd. ,Shenyang 110168,Liaoning,China
  • Received:2025-04-01 Online:2026-03-25 Published:2025-10-31
  • Contact: 郭思豪(1999 —),男,硕士,主要从事智能交通研究。 E-mail:27129433@qq.com
  • About author:邢岩(1985—),男,博士,教授,主要从事智能交通研究。E-mail: xingyan@sjzu.edu.cn
  • Supported by:
    the Open Project of National Engineering Research Center for Road Traffic Safety Control Technology(2024GCZXKFKT13B)

摘要:

针对小目标交通标志的错检和漏检导致识别网络精度下降的问题,提出了一种基于CGT-YOLO的小目标交通标志识别算法。首先,采用上下文增强模块(CAM)替代YOLOv5s网络中的快速空间金字塔池化(SPPF)模块,通过设置不同膨胀率的并行膨胀卷积,在不降低分辨率的条件下增强小目标交通标志的多尺度特征表达与上下文信息。其次,在YOLOv5s主干网络的拼接操作后插入全局注意力机制(GAM),提取CAM增强后的特征并通过三维置换、多层感知器及卷积空间注意力,增强通道与空间之间的全局交互,从而突出小目标交通标志的特征,解决复杂背景和远距离带来的负面影响。最后,通过上下文特定任务(TSC)解耦头对分类和定位任务的特征进行解耦,通过语义上下文编码(SCE)和细节保留编码(DPE)模块,分别生成语义丰富的低分辨率分类特征图和包含边界信息的高分辨率定位特征图,从特征源头解耦分类与定位任务,解决小目标交通标志分类与定位任务之间的特征冲突。实验结果表明,在整合TT100K和CCTSDB构建的数据集上,改进后的模型在各项指标上均有提升:漏报率和误报率分别降低12.1和11.6个百分点;mAP(0.50∶0.95)提高0.026 0。与YOLOv8s、NanoDet-Plus、RT-DETR-Nano等模型对比,CGT-YOLO在多项指标上均具有优势,且在保持较高推理速度(72.5 桢/s)的同时,可减少错检与漏检,显著提升了小目标交通标志在复杂场景下的检测精度与鲁棒性。

关键词: 小目标识别, 交通标志识别, 膨胀卷积, 注意力机制, 解耦头

Abstract:

To address the degradation in recognition accuracy caused by false and missed detections of small target traffic signs, this study proposes a small traffic sign recognition algorithm based on CGT-YOLO. First, a context-aware enhancement module (CAM) is introduced to replace the spatial pyramid pooling fast (SPPF) module in the YOLOv5s network. By employing parallel dilated convolutions with different dilation rates, the CAM enhances multiscale feature representation and contextual information of small traffic signs without reducing spatial resolution. Second, a global attention mechanism (GAM) is inserted after the concatenation operation in the backbone network of YOLOv5s. The GAM extracts features enhanced by the CAM and strengthens global interaction between channel and spatial dimensions through 3D permutation, multi-layer perceptron, and convolutional spatial attention, thereby highlighting the features of small traffic signs and mitigating the negative effects of complex backgrounds and long distances. Finally, a task-specific context (TSC) decoupled head is utilized to separate features for classification and localization tasks. Through the semantic context encoder (SCE) and detail preservation encoder (DPE) modules, the head generates semantically rich low-resolution feature maps for classification and high-resolution feature maps containing boundary information for localization, respectively. This disentangles classification and localization tasks at the feature source, resolving feature conflicts between the two tasks for small target traffic signs. Experimental results on a dataset constructed by integrating TT100K and CCTSDB show that the improved model achieves enhanced performance across all metrics: the missed detection rate and false detection rate are reduced by 12.1 and 11.6 percentage points, respectively, while mAP(0.50∶0.95) increases by 0.026 0. Compared to models such as YOLOv8s, NanoDet-Plus, and RT-DETR-Nano, CGT-YOLO demonstrates superior performance across multiple metrics. While maintaining a high inference speed (72.5 FPS), it effectively reduces false and missed detections, significantly improving the detection accuracy and robustness of small target traffic signs in complex scenarios.

Key words: small target recognition, traffic sign recognition, dilated convolution, attention mechanism, decoupled head

中图分类号: