Human Actions Modelling and Recognition in Low-Dimensional Feature Space

Abstract

The aim of this paper is to prove that it is possible to model the set of multiple human actions in low-dimensional feature space so they can be nearly unambiguously distinguish with the help of classification algorithm. The most important condition that has to be satisfied to solve this problem is selection of proper features sets that fit to a particular actions group. We evaluate our methodology on the dataset consisted of 16 different Oyama Karate techniques performed by two professional sport (black belt) instructors and masters of Oyama Karate. The dataset consisted of 640 actions samples. As a classification algorithm we have used Gesture Description Language. We have used four different angle-based features sets. With that selection we have made actions descriptions that transformed initial motion capture dataset to 6 or 8 dimensional features space with maximally three keyframes. Beside only one class the recognition rate was at level of 88% to even 100%.