In this chapter, we have presented a method to improve localization systems based to data association with GNSS receiver. This method increases the precision and the reliability of localization based on an Kalman filter. It consists to take care the characteristics of GNSS error. This error is an unpredictive stochastic process and it drifts the estimated position which is calculated by a Kalman Filter. The developed idea is to establish a prediction model of GNSS bias and to insert it in the localization system so as to modify the observation error from low-cost GNSS receiver to zero-mean, white, Gaussian noise. We have seen a possible model of GNSS error is Autoregressive process. We have determined its parameters and its order. Then, we have shown how this model is inserted in the Kalman Filter. However, the bias estimation needs to have sometimes absolute data (position of landmark of the environment) coming from exteroceptive sensors. To do that we propose to use a multi-sensor system (Tessier, Debain, Chapuis & Chausse, 2007) in which landmarks detection is given by autonomous entities called perceptive agents. The main weakness of this multi-agent fusion system is about the focusing process and the measure of the accuracy of the estimated vehicle's pose. Thanks to numerous experiments we noticed a strong robustness and a good accuracy of the guidance process allowing using it at high speed even in an environment with lots of elements like trees or buildings.