计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (17): 169-177.DOI: 10.3778/j.issn.1002-8331.2206-0258

• 图形图像处理 • 上一篇    下一篇

基于CA-ASFF-YOLOv4的交通标志识别研究

冷坤,秦伦明,王悉   

  1. 1.上海电力大学 电子与信息工程学院,上海 201306
    2.北京交通大学 电子信息工程学院,北京 100044
  • 出版日期:2023-09-01 发布日期:2023-09-01

Research on Traffic Sign Recognition Based on CA-ASFF-YOLOv4

LENG Kun, QIN Lunming, WANG Xi   

  1. 1.College of Electronic and Information Engineering, Shanghai University of Electric Power, Shanghai 201306, China
    2.School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing 100044, China
  • Online:2023-09-01 Published:2023-09-01

摘要: 交通标志识别是智能交通系统的核心技术。针对实际情况下,交通标志总体目标小且呈现多尺度分布、图像背景复杂,造成识别精度低的问题,提出了一种基于改进YOLOv4模型的交通标志识别算法CA-ASFF-YOLOv4。算法去除用于检测大目标的深度特征层,引入高分辨率特征层,并在各个特征层后添加注意力机制CA模块,有效加强了主干网络对小目标的特征提取。在颈部网络使用自适应特征融合ASFF代替原有的路径聚合网络PAnet,通过优化特征融合解决目标尺度多变的问题。减少残差块的堆叠,抑制背景重复叠加,提高复杂背景下的检测精度。实验结果表明,CA-ASFF-YOLOv4在TT100K 交通标志数据集上的mAP@0.5达到了91.47%,比YOLOv4提升了9个百分点,显著提高了实际应用场合中交通标志的检测精度。

关键词: 交通标志识别, YOLOv4, 注意力机制, 自适应特征融合

Abstract: Traffic sign detection is the core technology of intelligent transportation system. A traffic sign recognition algorithm CA-ASFF-YOLOv4 based on improved YOLOv4 model is presented to solve the problem of low recognition accuracy due to small overall target, multiscale distribution and complex image background of traffic signs in real traffic environment. The algorithm removes the depth feature layer used to detect large targets, introduces a high-resolution feature layer, and adds the attention mechanism CA module after each feature layer, which effectively enhances the feature extraction of small targets from the backbone network. At the same time, the adaptive feature fusion ASFF is used in the neck network instead of the original route aggregation network PAnet to solve the problem of variable target scales by optimizing the feature fusion. It reduces the stacking of residual blocks, suppresses background overlap, and improves the detection accuracy in complex background. The experimental results show that mAP@0.5 of CA-ASFF-YOLOv4 on TT100K traffic sign dataset reaches 91.47%, which is 9 percentage points higher than YOLOv4, and it further improves the recognition accuracy of traffic signs in practical applications.

Key words: traffic sign recognition, YOLOv4, attention mechanism, adaptive feature fusion