基于高分辨率网络的地铁人体姿态估计研究  

Research on Human Pose Estimation at Subways Based on High Resolution Network

作　　者：刘珊珊[1] 冯赛楠 田青[1] 钱付余 豆飞[3] 牛志斌 LIU Shanshan;FENG Sainan;TIAN Qing;QIAN Fuyu;DOU Fei;NIU Zhibin(School of Information,North China University of Technology,Beijing 100144,China;Traffic Control Technology Co.,Ltd.,Beijing 100070,China;Beijing Mass Transit Railway Operation Corporation Limited,Beijing 100044,China)

机构地区：[1]北方工业大学信息学院,北京100144 [2]交控科技股份有限公司,北京100070 [3]北京市地铁运营有限公司,北京100044

出　　处：《铁路技术创新》2023年第3期70-77,共8页Railway Technical Innovation

基　　金：国家重点研发计划项目(2020YFB1600702)。

摘　　要：目前,人体姿态估计从二维发展到三维,从图像发展到视频,从复杂网络发展到轻量化网络,在不断发展过程中,姿态估计又融合了深度学习的理论,采用卷积神经网络作为模型的主要构建单元,使姿态估计获得更大的发展空间。研究采用高分辨率网络为主干网络,并行处理多个分辨率网络分支,在更深的网络层级中产生高分辨率表征,并通过并行的网络各个层级进行多尺度融合来增强高分辨率表征的语义丰富程度,通过在网络中添加注意力机制模块增强特征提取能力,提高人体姿态估计的准确度。At present,human pose estimation continuously develops from 2D to 3D,from images to videos,and from complex networks to lightweight networks.In this process,pose estimation integrates deep learning theory,and uses convolutional neural network as the main building unit of the model,so that pose estimation has a greater development space.In this study,a high-resolution network is used as the backbone network and multiple resolution network branches are processed in parallel to generate high-resolution representations in deeper network levels.Multi-scale fusion is carried out at all levels of parallel networks to enhance the semantic richness of high-resolution representations,and attention mechanism modules are added to the network to enhance the ability of feature extraction and improve the accuracy of human pose estimation.

关 键 词：人体姿态估计 高分辨率网络 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]