计算机应用 ›› 2015, Vol. 35 ›› Issue (11): 3157-3160.DOI: 10.11772/j.issn.1001-9081.2015.11.3157

• 2015年全国开放式分布与并行计算学术年会(DPCS 2015)论文 • 上一篇    下一篇

基于随机森林的跌倒检测算法

罗丹1, 罗海勇2   

  1. 1. 北京邮电大学 国际学院, 北京 100083;
    2. 中国科学院计算技术研究所 普适计算研究中心, 北京 100190
  • 收稿日期:2015-06-17 修回日期:2015-07-09 发布日期:2015-11-13
  • 通讯作者: 罗丹(1994-),女,北京人,主要研究方向:物联网工程、移动计算.
  • 作者简介:罗海勇(1967-),男,湖北麻城人,副研究员,博士, CCF会员,主要研究方向:移动计算、普适计算.
  • 基金资助:
    国家自然科学基金资助项目(61374214);深圳市战略性新兴产业发展专项(深发改[2014]1787号).

Fall detection algorithm based on random forest

LUO Dan1, LUO Haiyong2   

  1. 1. International School, Beijing University of Posts and Telecommunications, Beijing 100083, China;
    2. Research Centre of Pervasive Computing, Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China
  • Received:2015-06-17 Revised:2015-07-09 Published:2015-11-13

摘要: 针对现有跌倒检测算法由于缺乏真实老人跌倒样本以及使用年轻人仿真跌倒样本规模较小导致的过拟合和适应性不足等问题,提出了基于随机森林的跌倒检测算法.该算法采用滑动窗口机制,对窗口内的加速度数据进行时间域和变换域处理,提取时间域和变换域特征参数后,在所有样本集中进行有放回的Bootstrap随机抽样和属性随机选择,构建多个基于最佳属性分割的支持向量机(SVM)基本分类器.在线跌倒检测阶段,对多个SVM基本分类器的分类结果采用少数服从多数的原则,给出最终判定结果.实验表明,随机森林跌倒检测算法可获得95.2%的准确率、90.6%的敏感度和93.5%的特异性,明显优于基于SVM和反向传播(BP)神经网络跌倒检测算法,反映出随机森林跌倒检测算法能更准确地检测跌倒行为,具有较强的泛化能力和鲁棒性.

关键词: 跌倒检测, 随机森林, 支持向量机, 反向传播神经网络, 属性

Abstract: To handle the over fitting and inadaptability problem of current fall detection algorithms caused by lack of real fall samples of elderly people and the use of small size of fall samples collected by young people, a fall detection algorithm based on random forest was proposed. By adopting sliding window mechanism, the sequentially collected acceleration data within the window were firstly processed to extract feature parameters of time domain and frequency domain, and then the Bootstrap approach was employed to randomly select partial samples with the same number from the whole training sample collection, after that random selection of features was performed to construct a collection of basic SVM (Support Vector Machine) classifiers with best feature partition. On the online fall detection stage, the final classification result was obtained with vote of results by multiple basic SVM classifiers according to the majority criteria. The experimental results demonstrate that the proposed algorithm outperforms the SVM and BP (Back Propagation) neural network algorithms with 95.2% accuracy, 90.6% sensitivity and 93.5% specificity, and reflects that the fall detection algorithm based on random forest can accurately recognize the fall activity, and has strong generalization ability and robustness.

Key words: fall detection, random forest, Support Vector Machine (SVM), Back Propagation (BP) neural network, feature

中图分类号: