Evaluating the Effect of Local Variations in Visually-Similar Motions on the Clustering of Body Sensor Features

Abstract

Body Sensor Network-related applications such as assistive-living environment, orthopedic, physical medicines, and rehabilitation use wearable body sensors like motion trackers to track joint movements and electromyogram sensors to track muscular activities. These sensors provide information in the form of multidimensional time series data. Generally, for these applications, classification or similarity retrieval of human motions is performed by traditional clustering of dimensionally-reduced feature vectors based on joint movements and/or muscular activities. However, local variations in visually-similar sets of human motions cause them to group in different clusters resulting to a lower recall during retrieval. Hence, it is important to evaluate the effect of local variations on the given clustering of feature vectors. In this paper, we represent the local variations in the form of quantitative attributes that are measured from sensors' time series data. And further, we propose a multivariate analysis of variance technique for evaluating the effect of quantitative attributes on the clustering results that are based on different configurations of feature vectors representing joint movements and muscular activities.