摘要
[目的]旨在通过目标检测方法,识别并减少车门夹人事件对城市轨道交通运营的影响,以提升车站的运营效率与安全性。基于YOLOv5改进算法设计屏蔽门夹人检测系统。[方法]基于现场收集含人不同姿态、角度图片共计5384张,将其按8∶2分为训练集和测试集;对比YOLOv5n、YOLOv5s、YOLOv5m、YOLOv5l模型训练效果,选取YOLOv5m为基准模型并进行改进;在YOLOv5m模型的基础上引入自注意力机制CoTNet网络,并将Neck网络中的FPN(特征金字塔网络)+PAN(路径聚合网络)结构优化为BiFPN(加权双向特征金字塔网络)结构。[结果及结论]改进后的YOLOv5m算法比原算法在测试精度、召回率、平均精度上都有所提高。同时该系统可以实现对单张图片、单个视频、摄像头、视频流以及整个文件夹图片进行目标检测,并在识别到目标物后自动启动报警机制。
[Objective]It is aimed to identify and reduce passenger entrapment incidents at platform screen doors in urban rail transit through object detection methods,thereby enhancing station operational efficiency and safety.A platform screen door passenger entrapment detection system is designed based on an improved YOLOv5 algorithm.[Method]A dataset of 5,384 images capturing various human postures and angles is collected on-site and split into a training set and a test set in an 8:2 ratio.The training performance of YOLOv5n,YOLOv5s,YOLOv5m,and YOLOv5l models is compared,with YOLOv5m selected as the baseline model for improvement.Enhancements are made by integrating the self-attention mechanism CoTNet into the YOLOv5m model and optimizing the Neck network structure by replacing the original FPN(feature pyramid network)+PAN(path aggregation network)with BiFPN(bi-directional weighted feature pyramid network).[Result&Conclusion]The improved YOLOv5m algorithm demonstrates higher accuracy,recall,and mean average precision compared to the original one.The system can perform object detection on single images,individual videos,live camera feeds,video streams,and entire image folders.Additionally,it automatically initiates an alarm mechanism upon detecting a passenger entrapment incident.
作者
陈修忻
CHEN Xiuxin(Track Works Branch of Shanghai Metro Maintenance Support Co.,Ltd.,200233,Shanghai,China)
出处
《城市轨道交通研究》
北大核心
2025年第S1期128-132,共5页
Urban Mass Transit
