Clustering and classification of time series representing sign language words

Abstract

The paper considers time series with known class labels representing 101 words of Polish sign language (PSL) performed many times in front of a camera. Three clustering algorithms: K-means, K-medoids and Minimum Entropy Clustering (MEC) are compared. Preliminary partitioning of the data set is performed with help of immune based optimisation. Some time series representations and different clustering quality indices are considered. It is shown that clustering is able to reveal existing natural division. Moreover, it gives an opportunity to learn the issues of processing large number of multidimensional data and to identify potential problems which may occur in automatic classification of signed expressions. Results of ten-fold cross-validation tests for nearest neighbour classification are also given. © 2013 Springer-Verlag.