基于CNN和高速通信技术的医用人体姿态识别方法  

Medical Human Posture Recognition Method Based on CNN and High Speed Communication Technology

作　　者：金玮[1] 孟君[1] 黄宇飞[1] 何萍 JIN Wei;MENG Jun;HUANG Yufei;HE Ping(Department of Information Management,Xinhua Hospital Affiliated to School of Medicine of Shanghai Jiaotong University,Shanghai 200025,China)

机构地区：[1]上海交通大学医学院附属新华医院,信息管理部,上海200025

出　　处：《微型电脑应用》2022年第7期20-22,26,共4页Microcomputer Applications

基　　金：上海市卫生与健康发展研究中心(202029114)。

摘　　要：为了快速、准确地获取人体跌倒信息,使跌倒病人及时获得救助,基于卷积神经网络(CNN)提出了一种人体姿态识别方法。该方法通过高速通信技术获取医院高清摄像头实时视频数据,利用OpenPose提取人体关键点并结合XGBoost分类器进行人体姿态估计。文中模拟医院场景进行单人与多人情况下的跌倒、正常行走和半蹲测试实验,再对分类结果建立一个状态序列集进行平滑处理,预测出是否有跌倒事件发生并及时进行告警。检测结果显示,所提出的方法准确度为99.75%、敏感度为100%、特异度为99.68%,可以准确地实现人体跌倒自动检测,且应用方便、实时性好。In order to obtain the fall information quickly and accurately,and make the fall patients get timely rescue,this paper proposes a human posture recognition method based on convolutional neural network(CNN),which obtains the real-time video data of hospital high-definition camera through high-speed communication,extracts the key points of human body using OpenPose,and estimates the human posture combined with XGBoost classifier.In this paper,we simulate the hospital scene,and carry out the fall,normal walking and squatting test experiments under the condition of single person and multiple people.Then we establish a state sequence set to smooth the classification results,predict whether there is a fall event and give an alarm in time.The detection results show that the accuracy,sensitivity and specificity of the proposed method are 99.75%,100%and 99.68%,respectively,which indicates that the method can realize automatic detection of human fall accurately,and it is convenient to use and has good real-time performance.

关 键 词：人体姿态识别 跌倒检测 分类模型 高速通信 卷积神经网络 

分 类 号：TN91[电子电信—通信与信息系统] TP311[电子电信—信息与通信工程]