计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (13): 194-204.DOI: 10.3778/j.issn.1002-8331.2302-0319

• 图形图像处理 • 上一篇    下一篇

改进YOLOv5的交通标志检测算法

杨祥,王华彬,董明刚   

  1. 桂林理工大学 信息科学与工程学院,广西 桂林 541006
  • 出版日期:2023-07-01 发布日期:2023-07-01

Improved YOLOv5’s Traffic Sign Detection Algorithm

YANG Xiang, WANG Huabin, DONG Minggang   

  1. College of Information Science and Engineering, Guilin University of Technology, Guilin, Guangxi 541006, China
  • Online:2023-07-01 Published:2023-07-01

摘要: 如今对交通标志的检测在自动驾驶、智慧交通等领域是必不可少的关键环节,其关系到人们的驾驶安全问题。针对目前对交通标志的识别存在漏检、误检以及识别精度低、模型参数过多的问题,提出了一种基于改进YOLOv5的交通标志检测算法。增加一个小目标检测头,提高对小目标的识别精度。设计了一种由CBAM、SPConv、C3相结合的CSC3模块,引入YOLOv5主干网络中,同时减少其数量,目的是提升特征提取能力,降低参数量。将用于检测大目标的检测头删除,再把SPP替换成SPPCSPC,提高模型对交通标志的检测能力。增加跨层连接,并且通过重构Concat连接,目的是提高算法的识别精度。引入EIOU来替换CIOU损失函数,从而解决漏检、误检问题。使用DWConv替换主干网络的Conv,目的是减少模型参数,提高检测精度。实验结果表明,改进后的算法的平均准确率均值mAP@0.5:0.95为62.6%,比原YOLOv5s提高了8.3个百分点,参数量下降了10.1%,并且检测速度达到了74?FPS,能够满足实际检测需求。

关键词: YOLOv5, 交通标志检测, 特征融合, 深度可分离卷积, 目标检测

Abstract: Nowadays, the detection of traffic signs is an essential key link in automatic driving, intelligent transportation and other fields, which is related to people’s driving safety. A traffic sign detection algorithm based on improved YOLOv5 is proposed to solve the problems of missing detection, false detection, low recognition accuracy and excessive model parameters in the current traffic sign recognition. Firstly, a small target detection head is added to improve the recognition accuracy of small targets. Secondly, a CSC3 module combining CBAM, SPConv and C3 is designed, which is introduced into the YOLOv5 backbone network and reduces its number at the same time, in order to improve the feature extraction ability and reduce the number of parameters. The detection head used to detect large targets is deleted, and the SPP is replaced with SPPCSPC to improve the model’s ability to detect traffic signs. Cross-layer connections are added, and Concat connections are reconstructed to improve the recognition accuracy of the algorithm. EIOU is introduced to replace the CIOU loss function to solve the problem of missed detection and false detection. Finally, DWConv is used to replace the Conv of the backbone network to reduce the model parameters and improve the detection accuracy. The experimental results show that the average accuracy of the improved algorithm mAP @ 0.5:0.95 is 62.6%, which is 8.3 percentage points higher than the original YOLOv5s, the number of parameters has decreased by 10.1%, and the detection speed has reached 74 FPS, which can meet the actual detection requirements.

Key words: YOLOv5, traffic sign detection, feature fusion, depth-separable convolution, object detection