摘要
为了解决现有交通标志检测方法在精度与速度间难以平衡、小目标漏检率高等问题,提出一种轻量级实时交通标志检测模型(lightweight real-time traffic sign detection model, LRTD)。该模型以YOLO11n模型为基础,在主干网络中引入特征增强模块(feature enhancement block, FEB)和全局-局部协同注意力模块(global-local collaborative attention module, GLCAM);在颈部网络中设计多尺度感受野协同模块(multi-scale receptive field collaborative module, MSRFC),并优化特征融合策略构建高分辨率检测头。在CCTSDB数据集和GTSDB数据集上,将LRTD与当前先进的检测模型进行性能对比,并进行消融实验验证LRTD各模块的功能。结果表明:在CCTSDB和GTSDB数据集上,LRTD的mAP@50分别为83.1%、95.6%,mAP@50-95分别为55.6%、81.5%,较YOLO11n模型的mAP@50分别提升6.7个百分点、2.1个百分点,mAP@50-95分别提升6.0个百分点、4.5个百分点;该模型在CCTSDB数据集上保持155.0 fps的实时推理速度,参数量与计算量分别降低1.9个百分点、1.6个百分点。所提模型可有效提升复杂场景下交通标志的识别性能,并为智能交通系统中的实时目标检测任务提供可行的技术方案。
To address the challenges of balancing accuracy and speed,as well as the high missed detection rate of small targets in existing traffic sign detection methods,a lightweight real-time traffic sign detection model(LRTD)was proposed.Based on YOLO11n as the baseline,the model introduced a feature enhancement block(FEB)and a global-local collaborative attention module(GLCAM)into the backbone network;In the neck network,a multi-scale receptive field collaborative module(MSRFC)was designed and the feature fusion strategy was optimized to construct a high-resolution detection head.On the public datasets CCTSDB and GTSDB,performance comparisons between the LRTD model and state-of-the-art detection models were conducted,and ablation experiments were carried out to verify the functionality of each module.The results show that on the CCTSDB and GTSDB datasets,the LRTD model achieves mAP@50 of 83.1%and 95.6%respectively,and mAP@50-95 of 55.6%and 81.5%respectively.Compared with the YOLO11n model,it increases mAP@50 by 6.7 percentage points and 2.1 percentage points respectively,and mAP@50-95 by 6.0 percentage points and 4.5 percentage points respectively.Additionally,the model maintains a real-time inference speed of 155.0 fps on the CCTSDB dataset,with its parameter count and computational complexity reduced by 1.9 percentage points and 1.6 percentage points respectively.The proposed model can effectively improve the recognition performance of traffic signs in complex scenarios and provide a feasible technical solution for real-time object detection tasks in intelligent transportation systems.
作者
田泉泉
张杨
TIAN Quanquan;ZHANG Yang(School of Information Science and Engineering,Hebei University of Science and Technology,Shijiazhuang,Hebei 050018,China)
出处
《河北工业科技》
2026年第1期10-20,91,共12页
Hebei Journal of Industrial Science and Technology
基金
河北省自然科学基金(F2023208001)
河北省引进留学人员资助项目(C20230358)。
关键词
计算机图像处理
交通标志检测
特征增强
注意力机制
多尺度特征融合
computer image processing
traffic sign detection
feature enhancement
attention mechanism
multi-scale feature fusion
