基于改进的多算法融合地铁站内乘客行为识别  

Passenger action recognition in subway stations based on improved multi-algorithm fusion

作　　者：章宇翔 李先旺[1] 贺德强[1] 贺岁球 陈彦君 李琴 ZHANG Yuxiang;LI Xianwang;HE Deqiang;HE Suiqiu;CHEN Yanjun;LI Qin(School of Mechanical Engineering,Guangxi University,Nanning 530004,China)

机构地区：[1]广西大学机械工程学院,广西南宁530004

出　　处：《铁道科学与工程学报》2023年第11期4096-4106,共11页Journal of Railway Science and Engineering

基　　金：国家自然科学基金资助项目(U22A2053);广西创新驱动发展专项(桂科AA20302010);广西研究生教育创新计划资助项目(YCBZ2022043)。

摘　　要：乘客的行为识别在保障乘客安全方面发挥着重要作用,它能提高地铁站对乘客安全的管理能力。然而,由于地铁车站内乘客人数众多,在出现照明变化和人员遮挡时会严重影响识别的准确率。为了解决准确率低的问题,在时空图卷积神经网络(STGCN)的基础上结合有效通道注意力网络(ECANet),加强了不同节点的连接,提出一种STEGCN节点注意力算法。此外,为了进一步提高准确率,采用双流结构,更进一步提出一种2s-STEGCN算法。应用Alphapose框架,结合YOLOv5_m目标检测算法、SPPE单人姿态估计算法和2s-STEGCN算法,提出一种改进的多算法融合行为识别方法用于地铁站内乘客的行为识别。首先,利用YOLOv5_m对乘客进行框定;然后,通过SPPE对框定的乘客进行骨骼关键点提取;最后,将提取到的骨骼关键点以坐标的形式输入2s-STEGCN,完成乘客的行为识别。为了验证2s-STEGCN算法的有效性,使用南宁地铁1号线的客流数据集分别在单人场景与多人场景下开展实验。实验结果表明:在损失值方面,2s-STEGCN具有最低的损失值,它的损失值仅为0.244,比STGCN的损失值低约0.025,这表明了2s-STEGCN具有更强的模型构建能力。在准确率方面,单人场景下的2s-STEGCN拥有最高的准确率,它的准确率达到96.13%,比STGCN高3.69%。此外,2sSTEGCN的准确率在多人场景下也有明显提升。该实验结果可为地铁乘客行为识别提供参考和理论支持。Passenger action recognition plays a vital role in safeguarding passenger safety,which improves the capability of managing passenger safety in subway stations.However,the accuracy of the passenger action recognition is seriously affected by the large number of passengers in the subway stations when there are lighting changes and obstructions.Thus,to solve this problem,a STEGCN algorithm was proposed based on a spatio-temporal graphic convolutional neural network(STGCN)combined with an effective channel attention network(ECANet),with enhanced connectivity of different nodes.In addition,an algorithm of 2s-STEGCN was further proposed by using a dual-stream structure to achieve further improvement in the accuracy of passenger action recognition.Specifically,the Alphapose framework was applied in combining the YOLOv5_m object detection algorithm,the single-person pose estimation(SPPE)algorithm and the 2s-STEGCN algorithm to present an improved multi-algorithm fusion action recognition method for the passenger action recognition in subway stations.Firstly,the passenger was framed using the YOLOv5_m algorithm.Secondly,the skeletal key points of the framed passenger were detected by the SPPE algorithm.Finally,the detected skeletal key points were input into the 2s-STEGCN algorithm as coordinates to realize passenger action recognition.To verify the validity of the 2s-STEGCN algorithm,experiments were conducted using the passenger flow dataset of Nanning Rail Transit Line 1 in single-person scenarios and multi-person scenarios,respectively.The results show that the 2s-STEGCN algorithm has the lowest loss value,which is only 0.244,lower than that of the STGCN algorithm by about 0.025.The results indicate that the 2s-STEGCN algorithm has a stronger model-building ability.Regarding the accuracy rate,the 2s-STEGCN algorithm possesses the highest accuracy rate in single-person scenarios,which reaches 96.13%,3.69%higher than that of the STGCN algorithm.Additionally,the accuracy rate of the 2s-STEGCN algorithm in multi-person sc

关 键 词：行为识别 时空图卷积 目标检测 姿态估计 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]