-Modern automotive technologies try to predict the driver's intention in order to control the vehicle effectively. However mathematical models describing the driver's steering behavior with sufficient accuracy are not available. The difficulties arise from the time-varying properties of the driver's behavior under rapidly changing traffic conditions. In this paper, a timevarying system identification method using maximum a posteriori estimation is proposed. An efficient iterative procedure is presented for maximizing the posterior probability of the parameters conditioning on observed data. Then it is applied to the experimental driving data, and the driver's time-varying steering models are identified and analyzed. The results indicate that the time-varying model reduces the output estimation errors significantly. Moreover, changes of driving strategies are observed from the identified models after drivers drive for a period of time.
I. IntroductionThe recent development of advanced automotive technologies is aimed at enhancing driving safety when the driver is involved in the control loop. For example, CWS (collision warning system) issues alarms to draw the driver's attention whenever the vehicle is too close to the preceding one; ESP (electronic stability program) compares the driver's intention with the vehicle's response, correcting the vehicle's trajectory such that the driver maintains control of the vehicle. Because of the close interaction between the driver and the active or passive driving assistant systems, it is desirable to derive a mathematical model which describes and predicts the driver's behavior with sufficient accuracy.The study of driver's steering model can be dated back to 1960's [1]. It has been shown that almost all manually controlled systems can be characterized by the "crossover model", i.e. the loop transfer function consisting of the driver and the vehicle behaves as an integrator around the gain crossover frequency [2]. As the advances in control theories, more structured driver models were proposed which took into account the driver's driving strategies, reaction time delay, and responses of neuromuscular systems [3] [4].Due to the complexity of human behavior, it is very difficult to derive the driver model from physical laws. On the other hand, system identification techniques allow researchers to establish models from simulated or experimental data. Chen, Pilutti, and Ulsoy used ARX and ARMAX models to represent the driver's steering behavior. System parameters and the range of model uncertainties were identified [5][6]; Kuge et al proposed an HMM-based framework to detect