An approach based on statistical features for extracting human actions from multivariate time series

Abstract

Extracting system behavior from the multivariate time series collected by sensors is more challenging. If regarded people as system, the various physical actions can be considered as different states of the system. Our intention was to produce a simple, flexible and accurate method for clustering multivariate time series to obtain the system state. Previously, we proposed a distribution-based state extraction method (DBSE) that can effectively extract motion from the multivar-iate time series of human leg activities. In this paper, we propose an Auto-segmented state extraction method (AS-DBSE) by improving the size of the segmented window to adapt to more complex situations. The method uses vectors composed of statistical feature parameters to represent time series. It uses distance as the similarity measure and clusters these vectors to get the state of the system. We have proved the effectiveness and simplicity of this method through research on human arm activity data. Finally, compared with clustering based on Toeplitz inverse covariance (TICC) and symmetric non-negative matrix factorization (SymNMF), the results of the new method are more satisfactory.