In this paper, a rollover prevention system dedicated to ATV (All-Terrain Vehicle) in off road context is presented. It is based on a multi-model observer associated to a bicycle model allowing the estimation of the preponderant variables influencing the LLT evolution. Dynamic instability evaluation is then based on the on-line estimation and prediction of the Lateral Load Transfer (LLT) from a vehicle 2D multimodel. More precisely, a predictive control algorithm, relying on the extrapolation of rider’s action and the estimated data from the observer, allows the rollover risk to be anticipated. However as the driver behaviour has been neglected, it is here proposed to study his influence on the different parameters and consequently to test the robustness of the system. It is realized thanks to full-scale experimentations.