基于注意力机制的交通标志识别  

Traffic sign recognition based on attention mechanism

作　　者：李宇琼 周永军 蒋淑霞[1] 梁杨 Li Yuqiong;Zhou Yongjun;Jiang Shuxia;Liang Yang(School of Mechanical and Electrical Engineering,Central South University of Forestry and Technology,Changsha 410000,China)

机构地区：[1]中南林业科技大学机电工程学院,长沙410000

出　　处：《电子测量技术》2022年第8期116-120,共5页Electronic Measurement Technology

基　　金：湖南省自然科学基金(2019JJ60076);湖南省科技创新项目(2018NK2065);湖南省教育厅重点研究项目(16A220);湖南省自然科学基金面上项目(2017JJ2403)资助。

摘　　要：针对实际场景中的交通标志大多小而密集,导致小目标交通标志识别准确度较低的问题,提出一种改进YOLOv5算法。首先将CBAM同时嵌入YOLOv5网络的Backbone和Head部分,以提升网络特征提取能力。其次为解决GIoU Loss可能造成的模型收敛速度较慢问题,改用DIoU Loss作为网络回归损失函数。实验结果表明,改进后的算法对于交通标志图像的识别平均准确率达到96.40%,相较于原算法有了6.83%的提升。最后为验证模型的实时可行性,在TX2嵌入式系统中利用本文改进YOLOv5算法对实景视频中的交通标志进行识别,结果表明本文改进算法能在嵌入式系统中流畅运行。Aiming at the problem of low accuracy of small target detection in traffic sign recognition tasks, which caused by that most of traffic signs in actual scene are small and dense, this paper proposes an improved YOLOv5 algorithm.firstly, embedding the CBAM into the Backbone and Neck of YOLOv5 network to improve the network feature extraction ability.and in order to solve the problem of slow network converge caused by GIOU Loss, DIoU Loss was used as the regression Loss function of the network. Experimental results show that the improved algorithm reaches 96.40% mAP in traffic sign recognition task, which is 6.83% higher than the original YOLOv5 algorithm. Finally, sending the improved network into TX2 embedded system to recognize traffic signs in real video, the result shows that the improved algorithm can run smoothly in embedded system.

关 键 词：交通标志识别 YOLOv5 CBAM DIoU Loss 嵌入式系统 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]