Human action recognition in stereoscopic videos based on bag of features and disparity pyramids

In this paper, we propose a method for human action recognition in unconstrained environments based on stereoscopic videos. We describe a video representation scheme that exploits the enriched visual and disparity information that is available for such data. Each stereoscopic video is represented by multiple vectors, evaluated on video locations corresponding to different disparity zones. By using these vectors, multiple action descriptions can be determined that either correspond to specific disparity zones, or combine information appearing in different disparity zones in the classification phase. Experimental results denote that the proposed approach enhances action classification performance, when compared to the standard approach, and achieves state-of-the-art performance on the Hollywood 3D database designed for the recognition of complex actions in unconstrained environments.