—In this paper, a localization system for a mobile robot, using a top-down multi-sensorial approach and a map of the environment, is proposed. Generally, the data sensors are associated with the map by a classical map-matching process. Popular methods try to optimize a global cost, to track multi-hypothesis or to reduce the problem by using multi-sensors. These approaches are bottom-up: each sensor data is analysed even if it is not relevant (like a GPS in indoor environment). The proposed approach is based on a Bayesian network, that is used in a top-down way to select the best feature to detect with the best sensor. This selection is done by taking into account the actual localization and the objectives of precision and integrity of the robot localization. The Bayesian network is also used to detect and fix association errors. This process makes possible to solve the kidnapped robot problem. Presented results show a real-time application of this method with a robot embedding several laser range-finders and a low-cost GPS. Both simulation and real data results are presented.