Histogram of Multi-Directional Gabor Filter Bank for Motion Trajectory Feature Extraction

Abstract: In this paper, we decompose motion flow using multidirectional Gabor filter bank to create robust descriptors for action recognition problem. The dense trajectory is utilized to allocate the Region of Interest (ROI) for extracting Gabor descriptors. In addition, the feature distribution is adopted instead of the conventional sum of energy. Our experiments are conducted in an open data set (UCF11) and our self-constructed one (CNU). The results indicate that the proposed feature outperforms all other ones in both individually and collectively.