Intégration de la saillance visuelle dans la reconnaissance d’évènements rares

This paper presents a new method for the detection of rares events in video. It is based on the visual saliency and on the detection and local description of points of interest. The point-of-interest filtering is carried out using the saliency score, allowing only those with visual importance to be considered. A model of normal events is learned thanks to the probabilistic generative model "Latent Dirichlet Allocation" (LDA), known for its performance in textual data mining. The detection of an abnormal or rare event is carried out in a probabilistic way via the learned model. This paper proposes to combine a saliency based visual focalization and the use of automatic document classification technic in order to classify images from a video and to detect rare events.