Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (13): 229-237.DOI: 10.3778/j.issn.1002-8331.2302-0163

• Graphics and Image Processing • Previous Articles     Next Articles

Improved Traffic Sign Detection Method for YOLOv5

WEI Qiang, HU Xiaoyang, ZHAO Hongxin   

  1. 1.School of Automation and Electrical Engineering, Shenyang Ligong University, Shenyang 110159, China
    2.School of Equipment Engineering, Shenyang Ligong University, Shenyang 110159, China
  • Online:2023-07-01 Published:2023-07-01

改进YOLOv5的交通标志检测方法

韦强,胡晓阳,赵虹鑫   

  1. 1.沈阳理工大学 自动化与电气工程学院,沈阳 110159
    2.沈阳理工大学 装备工程学院,沈阳 110159

Abstract: Traffic sign detection is of great significance to autonomous driving and vehicle safety, but the scale of traffic signs is affected by illumination and changes greatly, and there are occlusion situations, resulting in low model detection accuracy, false detection, missed detection and other problems. In this paper, a new traffic sign detection method is proposed by improving the YOLOv5 object detection algorithm. This method introduces recursive gated convolution, SOCA attention mechanism and regression loss function, and a large number of experiments are carried out on TT100K and CCTSDB datasets. The experimental results show that the improved YOLOv5 has an average accuracy(mAP) increase of by 43.7?percentage points on the TT100K dataset, 34.6 percentage points on the mAP@0.5:0.95, 2 percentage points on the CCTSDB dataset, and 1 percentage point on the mAP@0.5:0.95.

Key words: traffic sign detection, recursive gated convolution, attention mechanism, regression loss function

摘要: 交通标志检测对自动驾驶和车辆安全具有重要意义,但交通标志受光照影响尺度变化较大,存在遮挡等情况导致模型检测精度较低,有误检、漏检等问题。基于YOLOv5目标检测算法,提出了一种改进的交通标志检测方法。该方法引入递归门控卷积、SOCA注意力机制和回归损失函数,在TT100K和CCTSDB数据集上进行了大量实验。实验结果表明,改进的YOLOv5在TT100K数据集上平均准确率(mAP)提高了43.7个百分点,mAP@0.5:0.95提高了34.6个百分点,在CCTSDB数据集上平均准确率(mAP)提高了2个百分点,mAP@0.5:0.95提高了1个百分点。

关键词: 交通标志检测, 递归门控卷积, 注意力机制, 回归损失函数