Improving action recognition by selection of features

Abstract

In this paper we are interested in knowing, which features provide useful information for recognizing a gesture or an action, and how the set of selected characteristics impact the accuracy of detection. Then we define a large set of possible features, which are angles calculated from the joints of the skeleton provided by the kinect device. Our contribution is to propose an algorithm: Reduction of Feature Dimensions based on Standard Deviation (RFD-SD), to select the relevant set of features. In some cases we are evaluating if using a Genetic Algorithm (GA) contributes to improve the recognition. The experiments are carried out with two public datasets, to analyze the effect of using the complete set of proposed features, and the set of selected features. The obtained results suggest that action recognition improves by using the selected features.