Title page for etd-0620118-143652


[Back to Results | New Search]

URN etd-0620118-143652
Author Yan-Bo Chiou
Author's Email Address No Public.
Statistics This thesis had been viewed 5343 times. Download 0 times.
Department Information Management
Year 2017
Semester 2
Degree Master
Type of Document
Language zh-TW.Big5 Chinese
Title A data-driven approach to recognize human daily living activity
Date of Defense 2018-06-25
Page Count 60
Keyword
  • machine learning
  • activity recognition
  • smart environments
  • random forest
  • deep learning
  • Abstract Recognizing activities of daily living is mainly used in medical care, record and track patients' daily living. It can substantially reduce burden of human resource. And recognizing daily living activity has become the key point of Smart home. However, compared with medical care, Smart home requires more privacy. Therefore, the main problem is how to improve accuracy without equipping with any invasive sensor.
      Using Knowledge-driven approach have been conducted well, but Knowledge-driven relies on subjective restrictions. And each individual resident has their own living style, subjective restrictions probably has reached the limitations and needs a lot of work to do in advance. Therefore, our research employees an approach that based on objective method to recognize human's daily activity.
      The research data was collected form WSU CASAS. Previous researches took advantage of camera or body-worn sensors to make rules of restriction, but we figure that body-worn sensor involve more privacy and convenience issue. Therefore, we only used data collected from non-invasive sensor. The coding method we designed features the time sequence in each individual datum, which allows us to randomly draw training data from the whole population. And it will solve the problem that the frequency of occurrence of activities vary from time to time.
      Our research employed three models, which are random forest, support vector machine and sequence to sequence model. We expect that machine learning won't require prior knowledge to have a better balance prediction result among all activities and have ability to predict new combination of activities, so do the accuracy, the precision, the recall and the F1-measure.
    Advisory Committee
  • Keng-Pei Lin - chair
  • Han-Wei Xiao - co-chair
  • Wei-Po Lee - advisor
  • Files
  • etd-0620118-143652.pdf
  • Indicate in-campus at 5 year and off-campus access at 5 year.
    Date of Submission 2018-07-20

    [Back to Results | New Search]


    Browse | Search All Available ETDs

    If you have more questions or technical problems, please contact eThesys