作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (9): 55-62. doi: 10.19678/j.issn.1000-3428.0062843

• 热点与综述 • 上一篇    下一篇

基于改进YOLOv5s的交通信号灯识别方法

邓天民, 谭思奇, 蒲龙忠   

  1. 重庆交通大学 交通运输学院, 重庆 400074
  • 收稿日期:2021-09-29 修回日期:2021-11-16 发布日期:2021-11-23
  • 作者简介:邓天民(1979—),男,副教授、博士,主研方向为交通大数据、自动驾驶、交通控制;谭思奇、蒲龙忠,硕士研究生。
  • 基金资助:
    国家重点研发计划(SQ2020YFF0418521);中央引导地方科技发展专项(CSTC2020JSCX-DXWTB0003);川渝联合实施重点研发项目(CSTC2020JSCX-CYLHX0007)。

Traffic Light Recognition Method Based on Improved YOLOv5s

DENG Tianmin, TAN Siqi, PU Longzhong   

  1. College of Traffic and Transportation, Chongqing Jiaotong University, Chongqing 400074, China
  • Received:2021-09-29 Revised:2021-11-16 Published:2021-11-23

摘要: 交通信号灯的检测与识别是提升无人驾驶系统安全性的关键技术,传统基于深度学习的识别方法不能在精度和速度之间达到较好的平衡,难以满足实际环境下的检测要求。YOLOv5具有网络规模小的优势,适合在交通场景下进行交通信号灯检测。对YOLOv5网络进行改进,提出TL-YOLOv5s网络用于交通信号灯识别。通过简化主干网络中卷积层的数量提高特征提取效率,同时对残差组件进行密集连接和多层次跨连接,得到2种新的CSP残差结构替换原网络中的残差结构,增强网络特征融合能力,提高识别精度。考虑到交通信号灯的小目标属性,在网络中保留中小目标检测尺度而去除大目标检测尺度,进一步提升识别速率。在法国巴黎LaRA信号灯数据集上进行实验,结果表明,TL-YOLOv5s网络mAP值达到70.1%,相比于基线网络YOLOv5提升6.3个百分点,且检测速度达到22.4 frame/s,能够满足现实环境下的实时性要求。

关键词: 交通信号灯, 目标检测, 深度学习, 图像处理, 小尺度目标

Abstract: The detection and recognition of traffic lights are critical to improving the safety of unmanned driving systems.Existing recognition methods based on deep learning cannot achieve a good balance between accuracy and speed;hence, it is difficult to satisfy the detection requirements in actual environments.YOLOv5 has the advantage of a small network scale and is suitable for traffic signal detection in traffic scenarios.This study proposes the TL-YOLOV5s network for traffic signal light recognition by improving the YOLOv5 network.By simplifying the number of convolutional layers in the backbone network, the efficiency of feature extraction is improved.In addition, the residual components are densely connected and multilevel cross-connected, and two new CSP residual structures are obtained to replace the residual structure in the original network;thus, the feature fusion ability of the network is strengthened, and the recognition accuracy is improved.Based on the small target attribute of traffic lights, the detection scales of small and medium targets are retained in the network, and the detection scale of large targets is removed to further improve the recognition rate.The experiment is conducted for a LaRA dataset of traffic lights in Paris, France.The results show that the mAP value of the TL-YOLOV5s network reaches 70.1%, which is 6.3 percentage points higher than that of baseline network, YOLOv5.Furthermore, the detection speed reaches 22.4 frame/s, which can satisfy the real-time requirements in actual environments.

Key words: traffic light, target detection, deep learning, image processing, small-scale target

中图分类号: