The paper presents a novel observer design algorithm for autonomous vehicles. The technique is based on the combination of a classical linear observer and the ultra-local model. The linear observer is easy to design and it requires only a linear model of the considered system. However, it performs poorly when the linear system cannot cover the system’s dynamics due to nonlinearities or unmodelled dynamics. The ultra-local model aims to compensate for the nonlinear effects and improve the overall performances of the observer. The proposed method is applied to a vehicle-oriented estimation problem: lateral velocity. The operation and the effectiveness of the presented algorithm is demonstrated through several test scenarios in CarSim and also using real-test measurements.