Human Action Recognition Using GLAC Features on Multi-view Binary Coded Images

This paper presents a human action recognition framework by focusing on the auto-correlation features extracted on the binary coded motion and static information images of depth action video clips. At first, the action video clips are passed to the 3D Motion Trail Model (3DMTM) in order to generate the 2D motion and static information images. Clearly, for a depth video clip, the 3DMTM yields a set of three motion information images and a set of three static information images about the front, side and top projections of all the action video frames. Next, those images are transformed into the binary coded images by the Local Binary Pattern (LBP) operator. Finally, the Gradient Local Auto-Correlation (GLAC) algorithm is employed on those binary coded images for representing depth actions through the auto-correlation features. To classify the multiple actions with the gained features, the Extreme Learning Machine (ELM) is adopted here with the kernel trick. The introduced approach is extensively validated on the Microsoft Research Action3D (MSRAction3D) database. The experimental result demonstrates, the classification outcome of our system is better in comparison with the state-of-the-art systems.