Nonlinear Temporal Modeling for Motion-based Video Overviewing

This paper presents a method for video collection overviewing based the dynamic content of the scenes. In an unsupervised context, our approach relies on the nonlinear temporal modeling of wavelet-based motion features directly estimated from the image sequence. Based on SVM-regression, the nonlinear model is able to learn the behavior of the motion descriptors along the temporal dimension and to catch useful informations of the dynamic content. A similarity measure associated to the temporal model is then defined. It allows to compare video segments according to motion descriptors and thus defines a high-dimensional feature space where the video sequences under investigation are projected. The Curvilinear Component Analysis algorithm is finally used to map the feature space onto a 2D space. This operation enables us to display the video collection and gives an overview of the content according to motion features.