An adaptive time window method for human activity recognition

This paper studies the problem of human activity recognition. Traditionally, the data collected by the accelerometer is preprocessed with a fixed time window, and features for human activity recognition model are extracted in this framework. However, some human activities are quasi-periodic, which means that classification accuracy can be improved if adaptive time window is adopted instead. As human activities can be divided into periodic and non-periodic class, in order to extract features more accurately for the classification, the adaptive time window is then designed specifically to cope with the two categories. Finally, experiment is conducted to show that the adaptive time window method improves the classification accuracy in the identification of six kinds of activities including sitting, walking, running, etc., compared with previous fixed time window method.