基于U-Net定位优化的交通标志识别网络  

Traffic Sign Recognition Network Based on U-Net Localization Refinement Approach

作　　者：刘胜 马社祥 孟鑫 李啸 LIU Sheng;MA Shexiang;MENG Xin;LI Xiao(College of electrical and electronic engineering,Tianjin University of Technology,Tianjin 300384,China;College of TUT Maritime,Tianjin University of Technology,Tianjin 300384,China;College of Computer Science and Engineering,Tianjin University of Technology,Tianjin 300384,China)

机构地区：[1]天津理工大学电气电子工程学院,天津300384 [2]天津理工大学海运学院,天津300384 [3]天津理工大学计算机科学与工程学院,天津300384

出　　处：《重庆理工大学学报（自然科学）》2020年第11期147-155,共9页Journal of Chongqing University of Technology：Natural Science

基　　金：国家自然科学基金项目(61601326,61371108)。

摘　　要：针对当前方法不能同时满足高精度和高速识别的需求,提出一种新型交通标志识别网络。该网络包含3部分,即检测网络、定位优化网络和分类网络。通过改进YOLOv3网络实现交通标志快速和精确检测,利用U-Net网络对交通标志定位优化,选用空间变换网络完成交通标志分类任务。实验结果表明:检测网络在TT100K数据集和GTSDB数据集上,获得平均精度均值分别为87.57%和97.66%,运行时间分别为44.4ms和53.0ms,达到了当前先进的交通标志检测水平;定位优化网络在GTSDB数据集上,提升了分类网络的分类精度,也提高了识别网络的整体性能。To realize high-accuracy and high-speed recognition at the same time,a novel traffic sign recognition network is proposed,which consists of three parts:detection network,localization refinement network,and classification network.First,traffic signs can be detected quickly and accurately by improving YOLOv3 network.Then,U-Net is used to refine the location of traffic signs.Finally,the Spatial Transformer Network is used to complete the classification task of traffic signs.The detection network is evaluated on TT100 K dataset and GTSDB dataset.The results show that the proposed detection network achieves state-of-the-art performance by obtaining m AP of 87.57% and 97.66% with average execution time of 44.4 ms and 53.0 ms in two datasets.The localization refinement network is evaluated on GTSDB dataset.The experimental results show that the network improves the classification accuracy of the classification network,and also improves the overall performance of the recognition network.

关 键 词：交通标志检测 交通标志识别 YOLOv3 定位优化 U-Net 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]