In this paper an ATV (All-Terrain Vehicle) rollover prevention system is proposed. It is based on the online estimation and prediction of the Lateral Load Transfer (LLT), allowing the evaluation of dynamic instabilities. Using a vehicle model based on two 2D representations, the LLT can be estimated and predicted. As we consider off road vehicle, grip conditions must be encountered and are here estimated thanks to observation theory. Nevertheless, two main behaviours (over/under-steering) may be encountered pending on grip, and vehicle configuration. Because of the low cost sensor, these two opposite dynamics cannot be explicitly discriminated. As a result, two observers are used according to the vehicle behaviour. Based on a bicycle model and a low cost perception system, they estimate on-line the terrain properties (grip conditions, global sideslip angle and bank angle). A "supervisor" selects on-line the right observer. Associated to a predictive control algorithm, based on the extrapolation of rider's action and the selected estimated dynamical state, the risk can be anticipated, enabling to warn the pilot and to consider the implementation of active actions. Simulations and full-scale experimentations are presented to discuss about the efficiency of the proposed solution.