第13卷第3期 智能系统学报 Vol.13 No.3 2018年6月 CAAI Transactions on Intelligent Systems Jun.2018 D0:10.11992/tis.201710013 网络出版地址：http:/kns.cnki.net/kcms/detail/23.1538.TP.20180408.1522.018.html 基于递归神经网络的跌倒检测系统 牛德姣，刘亚文，蔡涛，彭长生，詹永照，梁军 (江苏大学计算机科学与通信工程学院，江苏镇江212013) 摘要：针对现有跌倒检测方法存在适应性差和功能较单一等问题，引入递归神经网络，通过发掘位置传感器数据之 间的内在联系提高检测跌倒行为的效果。首先，设计了传感器、训练与检测输入数据的序列化表示方法，为发掘其中 与跌倒和接近跌倒行为相关的内在关联提供了基础；接着，给出了用于跌倒检测的RNN训练算法以及基于RNN的 跌倒检测算法，将跌倒检测转换为输人序列的分类问题：最后，在前期实现的基于分布式神经元大规模RNN系统的 基础上，在Spark平台上实现了基于RNN的跌倒检测系统，使用Fall adl data数据集进行了测试与分析，验证了其 能有效提高跌倒检测的准确率和召回率，F值相比现有跌倒检测系统提高12%和7%，同时能有效检测出接近跌倒 的行为.有助于及时采取保护措施减少伤害。 关键词：跌倒检测：接近跌倒检测：传感器数据：递归神经网络：大数据：跌倒检测算法：训练算法：RNNFD 中图分类号：TP391文献标志码：A文章编号：1673-4785(2018)03-0380-08 中文引用格式：牛德姣，刘亚文，蔡涛，等.基于递归神经网络的跌倒检测系统.智能系统学报，2018,133)：380-387 英文引用格式：NIU Dejiao,LIU Yawen,,CAI Tao,etal.Fall detection system based on recurrent neural networkJl.CAAI transac- tions on intelligent systems,2018,13(3):380-387. Fall detection system based on recurrent neural network NIU Dejiao,LIU Yawen,CAI Tao,PENG Changsheng,ZHAN Yongzhao,LIANG Jun (School of Computer Science and Communication Engineering,Jiangsu University,Zhenjiang 212001,China) Abstract:The existing methods of fall detection have poor adaptability and limited functions.In this paper,a recurrent neural network based fall detection system is introduced to improve the performance of fall detection and to make it able to identify more dangerous near-falls by exploring the relationship of the position sensor data.Firstly,a serialization rep- resentation method on position sensor data,training and test data is designed as the basis for intrinsic relationship ex- ploration.Then,the training algorithm for RNN based fall detection is proposed,where the fall detection is transformed into a classification problem of the input sequence.Finally,using the large-scale RNN system based on distributed neur- ons,the fall detection system is implemented on the Spark platform.Evaluations are carried out on Fall_adl_data.The experimental results prove that the proposed system can improve the precision and recall of fall detection effectively. Compared with the existing fall detection systems,F-measure has improved by 12%and 7%,respectively.Moreover, the system is also able to detect the near-fall behavior effectively which helps provide timely protective measures to re- duce the damage caused by falls. Keywords:fall detection;near fall detection;sensor data;recurrent neural network;big data;fall detection algorithm; training algorithm;RNNFD 跌倒行为会对人体造成伤害，特别是对老人、 收稿日期：2017-10-17.网络出版日期：2018-04-09. 基金项目：江苏省科技厅重点研发计划产业前瞻与共性关键技术 小孩和病人等群体造成的伤害尤其严重。如何对可 项目(BE2015137):江苏省自然科学基金项目 能出现的跌倒行为进行及时的预警，从而减少和避 (BK20140570):中国博土后基金项目(2016M601737) 通信作者：牛德蛟.E-mail:dniu@ujs.edu.cn. 免其带来的伤害就显得非常重要。当前国内外相关DOI: 10.11992/tis.201710013 网络出版地址: http://kns.cnki.net/kcms/detail/23.1538.TP.20180408.1522.018.html 基于递归神经网络的跌倒检测系统 牛德姣，刘亚文，蔡涛，彭长生，詹永照，梁军 （江苏大学 计算机科学与通信工程学院, 江苏 镇江 212013） 摘 要：针对现有跌倒检测方法存在适应性差和功能较单一等问题，引入递归神经网络，通过发掘位置传感器数据之 间的内在联系提高检测跌倒行为的效果。首先，设计了传感器、训练与检测输入数据的序列化表示方法，为发掘其中 与跌倒和接近跌倒行为相关的内在关联提供了基础；接着，给出了用于跌倒检测的 RNN 训练算法以及基于 RNN 的 跌倒检测算法，将跌倒检测转换为输入序列的分类问题；最后，在前期实现的基于分布式神经元大规模 RNN 系统的 基础上，在 Spark 平台上实现了基于 RNN 的跌倒检测系统，使用 Fall_adl_data 数据集进行了测试与分析，验证了其 能有效提高跌倒检测的准确率和召回率，F 值相比现有跌倒检测系统提高 12% 和 7%，同时能有效检测出接近跌倒 的行为，有助于及时采取保护措施减少伤害。 关键词：跌倒检测；接近跌倒检测；传感器数据；递归神经网络；大数据；跌倒检测算法；训练算法；RNNFD 中图分类号：TP391 文献标志码：A 文章编号：1673−4785(2018)03−0380−08 中文引用格式：牛德姣, 刘亚文, 蔡涛, 等. 基于递归神经网络的跌倒检测系统[J]. 智能系统学报, 2018, 13(3): 380–387. 英文引用格式：NIU Dejiao, LIU Yawen, CAI Tao, et al. Fall detection system based on recurrent neural network[J]. CAAI transac￾tions on intelligent systems, 2018, 13(3): 380–387. Fall detection system based on recurrent neural network NIU Dejiao，LIU Yawen，CAI Tao，PENG Changsheng，ZHAN Yongzhao，LIANG Jun (School of Computer Science and Communication Engineering, Jiangsu University, Zhenjiang 212001, China) Abstract: The existing methods of fall detection have poor adaptability and limited functions. In this paper, a recurrent neural network based fall detection system is introduced to improve the performance of fall detection and to make it able to identify more dangerous near-falls by exploring the relationship of the position sensor data. Firstly, a serialization rep￾resentation method on position sensor data, training and test data is designed as the basis for intrinsic relationship ex￾ploration. Then, the training algorithm for RNN based fall detection is proposed, where the fall detection is transformed into a classification problem of the input sequence. Finally, using the large-scale RNN system based on distributed neur￾ons, the fall detection system is implemented on the Spark platform. Evaluations are carried out on Fall_adl_data. The experimental results prove that the proposed system can improve the precision and recall of fall detection effectively. Compared with the existing fall detection systems, F-measure has improved by 12% and 7%, respectively. Moreover, the system is also able to detect the near-fall behavior effectively which helps provide timely protective measures to re￾duce the damage caused by falls. Keywords: fall detection; near fall detection; sensor data; recurrent neural network; big data; fall detection algorithm; training algorithm; RNNFD 跌倒行为会对人体造成伤害，特别是对老人、 小孩和病人等群体造成的伤害尤其严重。如何对可 能出现的跌倒行为进行及时的预警，从而减少和避 免其带来的伤害就显得非常重要。当前国内外相关 收稿日期：2017−10−17. 网络出版日期：2018−04−09. 基金项目：江苏省科技厅重点研发计划产业前瞻与共性关键技术 项 目 (BE2015137) ；江苏省自然科学基金项 目 (BK20140570)；中国博士后基金项目 (2016M601737). 通信作者：牛德姣. E-mail：djniu@ujs.edu.cn. 第 13 卷第 3 期 智 能 系 统 学 报 Vol.13 No.3 2018 年 6 月 CAAI Transactions on Intelligent Systems Jun. 2018