单点测量数据多模态时序图像框架对人体跌倒姿态的鉴别  

Recognition of human fall posture using multi-mode time-series image frame based on single-point measurement data

作　　者：孙坚[1,2] 胡鹏程 Sun Jian;Hu Pengcheng(College of Electrical Engineering and New Energy,China Three Gorges University,Yichang 443000,China;Hubei Provincial Collaborative Innovation Center for New Energy Microgrid,China Three Gorges University,Yichang 443000,China)

机构地区：[1]三峡大学电气与新能源学院,宜昌443000 [2]新能源微电网湖北省协同创新中心三峡大学,宜昌443000

出　　处：《电子测量技术》2023年第11期83-89,共7页Electronic Measurement Technology

基　　金：湖北省自然科学基金青年项目(2020CFB248)资助。

摘　　要：为了准确识别老人跌倒姿态,及时进行医疗干预,提出一种基于多模态时序图像的人体跌倒姿态鉴别方法。首先将合加速度进行小波包分解、重构出3个子序列,利用3种时序图像算法将之转化,得到3种三通道时序图像;然后通过ResNet-18提取其高维特征,运用多模态特征融合;最后将融合结果结合改进随机森林算法,完成人体跌倒姿态的鉴别。在UMAFall和SisFall两个公开数据集进行验证,得到98.7%和99.3%的精准率。结果表明,该方法在人体跌倒鉴别中具有较高准确性,可为跌倒的老人及时提供帮助。To accurately identify the elderly's fall posture and timely carry out the medical intervention,a human fall posture identification method based on multi-modal time series images was proposed.First,the resultant acceleration is decomposed into three sub-sequences by wavelet packet,and then three time series image algorithms are used to transform the resultant acceleration into three three-channel time series images.Then,its high-dimensional features are extracted through ResNet-18,and multimodal feature fusion is used.Finally,the fusion results are combined with the improved random forest algorithm to complete the identification of human fall posture.The accuracy of 98.7%and 99.3%were verified in UMAFall and SisFall.The results show that the method has high accuracy in the identification of human falls,and can provide timely help for elderly people who fall.

关 键 词：人体跌倒姿态鉴别 加速度 时序图像 多模态融合 随机森林 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]