自然场景下的中国交通标志检测算法  

Chinese Traffic Sign Detection Algorithm in Natural Scenes

作　　者：王翰文 葛青 朱宁可 余鹏飞[1] WANG Hanwen;GE Qing;ZHU Ningke;YU Pengfei(School of Information,Yunnan University,Kunming 650504,Yunnan,China;Kunming Public Security Traffic Management Information Application Center,Kunming 650000,Yunnan,China)

机构地区：[1]云南大学信息学院,云南昆明650504 [2]昆明市公安交通管理信息应用中心,云南昆明650000

出　　处：《计算机工程》2024年第11期327-337,共11页Computer Engineering

基　　金：国家自然科学基金(62066046)。

摘　　要：当前在自然场景下对中国交通标志进行检测时通常存在检测精度和检测速度不平衡的问题,为此,提出一种基于YOLOv5的改进算法。首先根据ShuffleNet V2网络的设计理念,提出改进型轻量化卷积块来代替YOLOv5中的卷积块,以减少网络整体的计算量和参数量;其次提出一个扩大感受野的轻量化模块ASPC来代替原网络中的空间金字塔池化模块SPP,从而降低网络计算量和参数量同时提升网络的检测精度;最后把颈部特征融合网络中的上采样模块替换为CARAFE上采样算子,并提出多尺度通道混洗注意力机制MCSA,将其添加在CARAFE算子之后,让网络融合全局与局部的特征信息,更有效地减少颈部特征融合网络对交通标志特征信息的丢失。在自制的中国多类交通标志数据集CMTSD上进行实验,结果表明,改进后的算法模型大小相较于原模型减少了41%,每秒检测帧数(FPS)提高了9.37,平均检测识别精度mAP@0.5提升了2.91%,达到94.76%。改进算法在不同的自然场景下均能达到较好的检测效果,可以满足实际场景中对交通标志检测的需求。This study proposes an improved YOLOv5 traffic sign detection algorithm to address detection accuracy and speed imbalance problems for Chinese traffic sign detection in natural scenes.First,according to the ShuffleNet V2 network design concept,an improved lightweight convolutional block is proposed as a substitute for the YOLOv5 convolutional block.This improvement aims to decrease both the computational load and number of network parameters.Second,a lightweight module,the Atrous Spatial Pyramid Convolution(ASPC),is introduced to enlarge the receptive field,to replace the Spatial Pyramid Pooling(SPP)module in the original network.This decreases the computational and parameter complexity of the network and enhances its detection accuracy.Finally,the upsampling module in the neck is replaced by Content-Aware ReAssembly of FEatures(CARAFE),and a Multi-scale Channel Shuffle Attention(MCSA)mechanism is incorporated next to the CARAFE.This modification enables the network to effectively fuse global and local feature information,thereby more efficiently reducing the loss of traffic sign feature information in the neck.Experimental results on the Chinese Multi-class Traffic Sign Dataset(CMTSD)demonstrate that the optimized algorithm model achieves a significant size reduction of 41%compared with YOLOv5.Concurrently,a notable detection speed increase of 9.37 Frames Per Second(FPS)is observed.Moreover,the mAP@0.5 exhibits a substantial increase of 2.91%,reaching a remarkable value of 94.76%.This enhanced performance enables the algorithm to deliver excellent detection results spanning various natural scene environments and effectively meets traffic sign detection requirements for real-world scenarios.

关 键 词：深度学习 交通标志检测 YOLOv5网络 轻量化网络 全局与局部融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]