Application of hidden markov models and gesture description language classifiers to Oyama karate techniques recognition

Abstract

The karate movements classification is extremely challenging task due to the speed of body movements. From the other hand movements patterns are highly repetitive because they are practiced for many years by skilled martial artists. Those two facts make karate techniques classification tasks reliable tests of classifiers potential. Also, nowadays there is a growing interest on commercial market for solutions that are capable to be used in computer entertainment and coaching systems. Those factors motivated us to evaluate our Gesture Description Language (GDL) classifier trained with unsupervised reversed-GDL (R-GDL) method on karate techniques dataset and to compare it with state-of-the-art approach namely multivariate continuous hidden Markov model classifier with Gaussian distribution. The evaluation of capability of R-GDL methodology to karate techniques classification is main novelty of this paper. We have achieved very promising results. Only one class of actions has average recognition rate on the level of 88% while other where between 90% and 100%. GDL has also important advantages over state-of-the-art HMM classier that we will discuss in this paper.