Using SVD for Segmentation and Classification of Human Hand Actions

An automated strategy for decomposing time series into small, elementary subsequences is proposed. This is accomplished in two steps: first the time series must be decomposed into simpler sub-series (segmentation), next each sub series has to be suitably modeled or uniquely characterized (classification). In this paper, an approximation employing the first right singular vector of the data matrix is considered, and two new criteria for segmenting data are proposed and compared. The effectiveness of the proposed strategy is shown on a time series resulting from sensory data on a data-glove when a human picks a tin can. The strategy proves to be simple and reliable, and can be used as a basic ingredient for real-time detection and interpretation of human gestures.