Action recognition based on hybrid features

Human action recognition is a quite popular yet challenging problem in computer vision discipline, especially in automatical human motion understanding. This paper introduces a novel approach based on the second generation Curvelet transform to get the eigenvector for representing the human action in static images. As an exceptional multi-resolution feature extraction technique, the second Curvelet transform offers enhanced directional and edge representation that shows nice competitiveness. During feature descriptor extraction, the silhouettes and texture statistical information are extracted from the coefficients as the edge and the texture features. All the extracted features are aligned as the hybrid eigenvector of a frame. Experimental evaluation is performed on the benchmark Weizmann database and a comparison with the other counterparts is made. Results show that our method is rather competitive in quantitative index such as accuracy rate, which exhibits the descriptor developed from the second generation Curvelet to be a promising representation for such visual recognition tasks.