Dictionary of gray-level 3D patches for action recognition

This paper deals with action recognition in the context of video analysis based on the use of a sparse dictionary defined by the 3D spatio-temporal representation of the actions. A 3D volume can be seen as a set of gray-level 3D patches comprising 2D patches taken in successive frames in order to capture a motion pattern. The goal of our proposal is to recognize human actions within these 3D volumes whose 3D patches are described with the dictionary atoms. To that end, we compute a motion signature by building a histogram based on the use of the atoms of the dictionary. Paired with a SVM, we show that these signatures can be exploited in the context of action recognition. This method has been tested on the KTH database with good results.