In this article, we present a parsimonious high level multi-sensor fusion architecture for robot localization using several types of localization techniques. Operating in a Top-Down mode, the parsimonious localization system based on the use of an existing absolute environment map, selects the most adequate modality in an economical and efficient way. Using a bayesian network associated to an Extended Kalman Filter and referring to contextual information, our proposed method aims to ensure a good localization level by selecting the best technique which responds the best to a fixed objective. Presented results show the parsimonious aspect of our application deploying a robot embedding several sensors (laser range-finder, UWB beacons and a low-cost GPS).