基于骨架特征的人体跌倒检测  

Human fall detection based on skeleton features

作　　者：汤发源 赵永兴 刘晓亮[2] 赵欣[3] 王京华 TANG Fayuan;ZHAO Yongxing;LIU Xiaoliang;ZHAO Xin;WANG Jinghua(College of Mechanical and Electric Engineering,Changchun University of Science and Technology,Changchun 130022,China;Department of Hematology,the First Hospital of Jilin University,Changchun 130021,China;Department of Pediatric Respiratory,the First Hospital of Jilin University,Changchun 130021,China)

机构地区：[1]长春理工大学机电工程学院,吉林长春130022 [2]吉林大学第一医院血液科,吉林长春130021 [3]吉林大学第一医院小儿呼吸科,吉林长春130021

出　　处：《传感器与微系统》2024年第3期115-119,124,共6页Transducer and Microsystem Technologies

基　　金：国家基金“111”计划资助项目(D17017);吉林大学白求恩第一医院成果转化基金资助项目(JDYZH-2102036)。

摘　　要：针对现有基于人体骨架跌倒检测设备要求高的问题,提出了一种基于轻量级OpenPose生成骨架特征的跌倒检测方法。首先,基于轻量级OpenPose网络检测人体关键点,利用人体部分关键点生成边界框,并对关键点坐标进行标准化处理,将边界框的纵横比和标准化后的关键点坐标作为表示人体姿态的特征向量。最后,将人体姿态特征向量作为多层感知机(MLP)的输入,判断人体是否发生跌倒。实验结果表明,基于单目相机采集图片构造的自定义跌倒数据集,网络可以实现98.64%的跌倒检测准确率,并且在CoreTMi5—9300H CPU上达到20fps的检测速度。Aiming at the high requirements of existing fall detection equipment based on human skeleton,a fall detection method based on lightweight OpenPose generating skeleton features is proposed.Firstly,the keypoints of the human body are detected based on the lightweight OpenPose network.The partial keypoints of the human are used to generate the bounding box,and the coordinates of the keypoints are normalized processing.The aspect ratio of the bounding box and the standardized keypoints coordinates as feature vectors representing human pose.Finally,the human body pose feature vector is used as the input of the multilayer perceptron(MLP)to determine whether the human body falls or not.The experimental results show that the network can achieve fall detection accuracy rate of 98.64%based on the customed fall dataset constructed by the images collected by the monocular camera and can achieve a detection speed of 20 fps on the CoreTM i5-9300H CPU.

关 键 词：关键点 边界框 特征向量 多层感知机 跌倒检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP212[自动化与计算机技术—计算机科学与技术]