A novel reduced feature set and a hierarchical system of classifiers for human action recognition based on the natural domain knowledge of the human figure

In this paper, a method with a significantly reduced feature set is described which recognizes human actions from video sequences. The proposed method is based on the natural domain knowledge of the human figure such as proportions of the human body. The method is evaluated on two of the publicly available databases in the literature of activity recognition, the Weizmann and the UIUC databases. Results show that the efficiency of our method is higher than that of other state-of-the-art methods on both datasets with a perfect correct classification rate (CCR) of 100% when the LOOCV is used. The experimental results are promising with much fewer features (between 27 and 170 times less features), compared with other methods, and the possibility of processing more than 500 fps.