Learning a Discriminative Feature Descriptor with Sparse Coding for Action Recognition

In this paper, we propose a novel algorithm for action recognition. The contribution of our work is three-fold. First, modified Weber local descriptor (IWLD) is proposed to capture the form cues of the action video sequences. Through introducing novel Weber magnitude and orientation components, our proposed IWLD can represent local patterns more effectively and accurately than existing Weber local descriptor (WLD). Second, to describe the form feature, histogram of improved Weber orientation Magnitude (HIOWM) is constructed. Considering motion and context cues also have discriminative power, we further propose a scheme that fuses HIOWM with motion and context cues to generate motion context HIOWM (MCHIOWM) descriptor to represent action video sequences. Third, for the sake of the more discriminative feature, we adopt sparse coding method to further refine the selected MCHIOWM. We present experiments to validate that the proposed framework obtains the competitive performance compared with the state-of-the-art methods.