Spatiotemporal information fusion for human action recognition in videos

Many applications are concerned by human action recognition notably in multimedia and more particularly for video retrieval and archival. Usual approaches focus on probabilistic methods and assume a still camera. In this paper, a method based on the Transferable Belief Model fusion process and considering a moving camera is proposed. In this framework, the affine camera motion estimation and temporal variations of three major human points are combined. The method is tested on videos of athletics meetings in which running, jumping and falling actions have to be recognized. Results show the validity of the method for action recognition.