基于YOLOv5s改进的铁道异物侵限检测算法研究
Research on the improved railway foreign body intrusion limit detection algorithm based on YOLOv5s
刘越 1王亚飞 2王远 3刘士淼 2徐林琚 2张佳乐2
作者信息
- 1. 长春工业大学电气与电子工程学院,吉林长春 130012;长春昌鼎电子科技有限责任公司,吉林长春 130012
- 2. 长春工业大学电气与电子工程学院,吉林长春 130012
- 3. 长春昌鼎电子科技有限责任公司,吉林长春 130012
- 折叠
摘要
提出一种基于YOLOv5s改进的铁道异物侵限检测算法NH-YOLOv5.先通过列车头装载的视频检测设备实时采集相关图像,建立铁路异物入侵数据集RD.使用已建立的数据集RD对NH-YOLOv5模型进行仿真实验和在线评估.实验结果表明,所提出的NH-YOLOv5算法的精准率和召回率分别为92.6％、78.3％,均值平均精度mAP@0.5和mAP@0.5∶0.95分别为85.3％、61.5％,相比YOLOv5s原模型,精准率提高1.4个百分点,召回率提高7.4个百分点,均值平均精度mAP@0.5和mAP@0.5∶0.95分别提高6.3、7.7个百分点.可见,文中算法比YOLOv5s算法具有更好的检测效果和泛化性能,有效改善了漏检、误检等问题,且提高了小目标的检测能力,验证了该方法的准确性和可用性.
Abstract
Therefore,this paper proposes an improved railway foreign object intrusion detection algorithm NH-YOLOv5 based on YOLOv5s.Because there is no public railway invasion data set,this article uses a camera installed on the locomotive to perform real time cameras,and collects foreign objects invading images on the spot to establish an railway intrusion data set RD.Use the established data set RD to evaluate the NH-YOLOv5 model.The experimental results show that the accuracy and memories of the NH-YOLOv5 algorithm proposed are 92.6％and 78.3％,respectively,with the average average accuracy of mAP@0.5 and mAP@0.5∶0.95,respectively 85.3％and 61.5％,respectively,respectively,respectively,respectively.61.5％.Compared with the original YOLOv5s model,the precision is increased by 1.4 percentage points,the recall rate is increased by 7.4 percentage points,and the average average precision is increased by mAP@0.5 and mAP@0.5∶0.95,respectively,and the average accuracy is increased by 6.3 and 7.7 percent points.The results show that the proposed algorithm has a better detection effect than the YOLOv5s algorithm.This algorithm effectively improves the problems of missions detection and erro detection,and improves the detection ability of small targets,which verifies the accuracy and usability of the method.
关键词
铁道异物侵限检测/YOLOv5s/NAM注意力机制/NH-YOLOv5Key words
railway foreign bodies/YOLOv5s/NAM attention mechanism/NH-YOLOv5引用本文复制引用
出版年
2024
NETL
NSTL
维普
万方数据