HMM-based Tri-training algorithm in human activity recognition with smartphone

Abstract

With the popularity of smartphone, studies using sensors on smartphone have been investigated in recent years. Human activity recognition is one of the active research topics. User's context can be used for providing users the adaptive services and the advice about health based on a stream of activity data. In this paper, we introduce a HMM-based Tri-training algorithm. The Tri-training algorithm can automatically augment activity classifiers after they are deployed in a real environment. HMM model can use the relationship between previous and current states to help Tri-training algorithm chooses new samples for training set. This method can explicitly reduce the amount of noise introduction into classifier group and make the output state stream connect more smoothly.