Abstract
Recently, wristband-type acceleration sensors are used in the fields such as health care and sports. However, from previous research, the body posture estimation by movement's strength is poor accuracy. In this paper, we developed the highly accurate estimation method with extracted various features from 3-axis acceleration data. Especially, we discriminated the sedentary or no sedentary behaviors which are typical low-intensity activity. Firstly, various features for machine learning were calculated from time-series acceleration data. These features were meaning the exercise volume and inclination with an arm. Next, the most suitable feasures combination was selected from among these features. Finally, sedentary or no sedentary behaviors were discriminated by a random forest using selected features. In the result, we achieved higher accuracy (more than 90%) than previous methods. This result proves that our method can estimate the body posture by arm movement easily and correctly. We would try to apply this method to the posture estimation of the athlete playing games.
