To guide a vehicle, the localization system must provide an accurate and reliable estimation. Generally, the estimation of the vehicle's state is dealt with a Bayesian approach like a Kalman filter. However, if this technique is a good mean to merge information of different sensors, it gives any idea of the result's reliability. We propose here to include a confidence level on the estimated vehicle's pose. This confidence level is updated after each landmark detection. This update is function of characteristics of the perception system. Thus, we propose also a method to characterise feature detection algorithms in order to obtain the most realistic confidence level. We demonstrate the practicality of this approach by guiding an experimental vehicle in real outdoor environment.