Résumé en anglais

This paper introduces a generic method for the accurate analysis of junctions, relying on a statistical modeling of normalized image gradients. We analyze junctions as local visual events that do not happen by chance under a background model derived from the a-contrario methodology. The method not only provides thresholds for the detection of junctions, but also enables their accurate characterization, including a precise computation of their type, localization, scale and geometrical configuration. The efficiency of the method is evaluated through various experiments.