测绘通报 ›› 2017, Vol. 0 ›› Issue (7): 29-33,44.doi: 10.13474/j.cnki.11-2246.2017.0218

• 学术研究 • 上一篇    下一篇

基于随机森林算法的多维情境特征活动识别

刘克强1,2, 汪云甲1, 陈锐志2, 褚天行3   

  1. 1. 中国矿业大学环境与测绘学院, 江苏 徐州 221116;
    2. 武汉大学测绘遥感信息工程国家重点实验室, 湖北 武汉 430079;
    3. 德州农工大学科普斯分校地理空间计算实验室, 美国 德州 科普斯 TX78412
  • 收稿日期:2016-11-28 出版日期:2017-07-25 发布日期:2017-08-07
  • 作者简介:刘克强(1988-),男,博士生,主要研究方向为无缝定位与人类活动识别。E-mail:cumtlkq@163.com
  • 基金资助:
    国家重点研发计划(2016YFB0502102);国家高技术研究发展计划(863计划)(2013AA12A201);现代工程测量国家测绘地理信息局重点实验室经费资助(TJES1302);2014江苏省普通高校研究生科研创新计划(KYLX_1394)

Research on Human Activity Recognition with Multiple Contexts by Using Random Forest

LIU Keqiang1,2, WANG Yunjia1, CHEN Ruizhi2, CHU Tianxing3   

  1. 1. School of Environment Science and Spatial Informatics, China University of Mining and Technology, Xuzhou 221116, China;
    2. State Key Labortatory of Information Engnineering in Surveying, Mapping, and Remote Sensing, Wuhan University, Wuhan 430079, China;
    3. Conrad Blucher Institute for Surveying & Science, Texas A & M University Corpus Christi, Corpus Christi, TX 78412 USA
  • Received:2016-11-28 Online:2017-07-25 Published:2017-08-07

摘要: 利用智能手机传感器可感知时间、空间、时空和用户等多维情境的特征,可识别用户活动,但原框架模型中仅利用了单一分类器中的朴素贝叶斯算法,存在分类精度效果受限的问题。本文利用集成分类器中的随机森林算法对原有框架中的单一分类器进行了改进。在获取的3个数据集上的十倍交叉验证结果表明,加权平均F1量测值均有较大提高,表明利用随机森林算法在分类精度效果上有所提升;但由于集成算法结构相对复杂,其学习效率相对较低。此外,随机森林算法的分类混淆矩阵表明,导致识别误差的因素主要为活动的定义与室内定位精度。

关键词: 活动识别, 情境感知, 机器学习, 随机森林, 手机传感器

Abstract: Previous research has verified the effectiveness of a method which utilizes multiple contexts including temporal, spatial, spatiotemporal and user from smartphone sensors to recognize human activities. But it has limitation in classification accuracy performance when recognizing human activity on smartphones by using naïve Bayes. This paper presents an improving research work on recognizing activity by utilizing a random forest algorithm in ensemble learning instead of original naïve Bayes under the same framework. The result of a ten-fold cross-validation experiment on three subjects' datasets shows that the weighted average F1 measures are improved by using random forest. However, the training time efficiency measures are worse than naive Bayes because of the complex structure of random forest. Meanwhile, the confusion matrix shows that the main factors for classification error are activity definition and indoor positioning error.

Key words: activity recognition, context awareness, machine learning, random forest, smartphone sensors

中图分类号: