A lane keeping system using adaptive model predictive control with linear time-variant prediction model is proposed in this paper. First, real-time on-line system identification using recursive least square method is employed to obtain the estimated tire cornering stiffness of the bicycle model. The vehicle velocity within the prediction horizon is predicted using the longitudinal acceleration to obtain the linear time-variant bicycle model. A cost function which consists of the errors between the target trajectory and predicted trajectory, and the steering angles within the prediction horizon is minimized to generate the optimal steering angle command to perform the lane keeping control. For curved road tests with different road frictions and non-constant speed profiles, simulation results show that the proposed control can effectively reduce the lateral displacement error and achieve better lane keeping performance than the conventional model predictive control and the adaptive model predictive control with linear time invariant system.