KMUTT RESEARCH & DEVELOPMENT JOURNAL

Volume 41, No. 04, Month OCTOBER, Year 2018, Pages 401 - 420

Feature selection for human activity classification from skeleton data using two-level selection technique

Duangpen Jetpipattanapong

Abstract Download PDF

various forms of vectors that represent the relation of the positions. The vectors were then used as features to train a classification model, which classifies 13 home-based human activities. A feature selection method using two-level selection techniques, which rank the priority of features based on the frequency of the selections of each feature via the use of the backward elimination method, was proposed and utilized. The features were grouped together by priority and the performance of features was tested by eliminating a group of features in order of priority and measuring the accuracy of the classification model using support vector machine. The results showed that, through the use of 48 features, the use of the two-level selection technique gave a mean accuracy of 92.78% and was able to select 29 major features with a mean accuracy of 92.25%.

Keywords

Activity Classification / Feature Selection / Two-Level Selection / Support Vector Machine

---