Unsupervised Discovery of Repetitive Activity Patterns in Acceleration Time Series Data

Abstract

In order to automatically analyze human motion data, it is important to extract basic actions such as “stand up” and “walking”. The most conventional methods are several issues: (1) they need labeled data, (2) they only classify data but not extract the structure of data, (3) they require a window size of time series in advance. This paper proposes a novel unsupervised pattern discovery method which extracts a series of repetitive actions by recognizing basic actions in time series. The proposed method consists of a basic action enumeration process and a grammatical inference process. The former discovers the subsequence of basic actions using the Motif discovery method. The latter discovers grammatical patterns constructed with the sequence of the basic actions. Furthermore, we evaluate the accuracy of extracted basic actions and the duration time of a series of actions for packing operations and screw-tightening operations.