基于深度学习的转弯车流量检测方法
DOI:
作者:
作者单位:

1.长安大学汽车运输安全保障技术交通行业重点实验室;2.西安比亚迪汽车有限公司

作者简介:

通讯作者:

中图分类号:

U495

基金项目:

国家自然科学基金面上项目(51978075)


Turning Traffic Flow Detection Method Based on Deep Learning
Author:
Affiliation:

1.Key Laboratory of Transportation Industry of Automobile Transportation Safety and Security Technology,Chang’an University;2.Xi’an BYD Auto Co., LTD

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    为方便统计转弯车流量,并提升交叉口转弯车流量的检测速度与精度,提出基于深度学习的方法对城市交叉口转弯车流量进行检测、跟踪和计数。首先,通过对比分析选用轻量高效的YOLOv5s作为目标检测框架,并采用无人机航拍方式获取城市交叉口交通流视频,自建车辆航拍图像数据集;利用预训练权重及最新权重文件完成自建数据集的训练与推理;且模型评估表明,基于YOLOv5的车辆检测模型具有较高的检测速度与精度:其中模型的box_loss值迅速下降并稳定在0.03左右,mAP_0.5值迅速上升并保持在0.9附近;之后,对接DeepSORT模型作为后端多车辆跟踪算法,通过坐标转换以简化车辆轨迹提取,并对行驶轨迹线展开有效性判断;针对检测框角点跃变现象,提出角点-质心点坐标变换以强化轨迹点的坐标信息鲁棒性,且采用六次多项式拟合车辆轨迹线,将不满足函数映射要求的轨迹线进行旋转优化,以正常拟合全部轨迹;最后根据预设的转弯角度判定阈值,实现转弯车辆的检测与计数。为验证本文提出的转弯车流量检测方法的有效性,以某一城市交叉口为例进行车辆检测实验,对比分析人工计数值和本方法检测结果。结果表明:四个流向平均检测精度为92.9%,最高可达95.7%,能够满足实际交叉口场景转弯车流量的常规检测要求。

    Abstract:

    In order to facilitate the statistical analysis of turning traffic flow at intersections and enhance the detection speed and accuracy of turning vehicle flows, we propose a deep learning-based approach for detecting, tracking, and counting turning vehicle flows at urban intersections. Firstly, through comparative analysis, we select the lightweight and efficient YOLOv5s as the target detection framework. We acquire urban intersection traffic flow videos using unmanned aerial vehicle (UAV) aerial photography and construct a custom vehicle aerial image dataset. We employ pre-trained weights and the latest weight files to train and infer on our custom dataset. Model evaluation indicates that the vehicle detection model based on YOLOv5 exhibits high detection speed and accuracy: the model"s box_loss quickly decreases and stabilizes around 0.03, while mAP_0.5 swiftly rises and remains near 0.9. Subsequently, we integrate the DeepSORT model as the backend multi-vehicle tracking algorithm. We simplify vehicle trajectory extraction through coordinate transformations and conduct effectiveness judgments on trajectory lines. To address the phenomenon of abrupt changes in detection box corners, we propose a corner-to-centroid coordinate transformation to enhance the robustness of trajectory point coordinates. Additionally, we utilize a sixth-degree polynomial to fit vehicle trajectory lines and optimize rotation for trajectory lines that do not meet the function mapping requirements, ensuring proper fitting of all trajectories. Finally, based on pre-defined turning angle threshold criteria, we achieve turning vehicle detection and counting. To validate the effectiveness of our proposed method for turning traffic flow detection, we conduct vehicle detection experiments at a selected urban intersection. Comparative analysis between manual counting and our method"s detection results demonstrates an average detection accuracy of 92.9% across four directions, with a maximum accuracy of 95.7%, meeting conventional requirements for turning traffic flow detection in real-world intersection scenarios.

    参考文献
    相似文献
    引证文献
引用本文

张韡,李永,刘涛,等. 基于深度学习的转弯车流量检测方法[J]. 科学技术与工程, , ():

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2024-03-21
  • 最后修改日期:2024-05-27
  • 录用日期:2024-06-05
  • 在线发布日期:
  • 出版日期:
×
亟待确认版面费归属稿件,敬请作者关注