Detecting spatio-temporally interest points using the shearlet transform

In this paper we address the problem of detecting spatio-temporal interest points in video sequences and we introduce a novel detection algorithm based on the three-dimensional shearlet transform. By evaluating our method on different application scenarios, we show we are able to extract meaningful spatio-temporal features from video sequences of human movements, including full body movements selected from benchmark datasets of human actions and human-machine interaction sequences where the goal is to segment drawing activities in smaller action primitives.