Construction of Driver Models for Cut-in of Other Vehicles in Car-Following Situations

The purpose of this study was to construct driver models using long short-term memory (LSTM) in car-following situations, where other vehicles change lanes and cut in front of the ego vehicle (EGV). The development of autonomous vehicle systems (AVSs) using personalized driver models based on the individual driving characteristics of drivers is expected to reduce their discomfort with vehicle control systems. The driving characteristics of human drivers must be considered in such AVSs. In this study, we experimentally measured data from the EGV and other vehicles using a driving simulator consisting of a six-axis motion device and turntable. The experimental scenario simulated a traffic congestion scenario on a straight section of a highway, where a cut-in vehicle (CIV) changed lanes from an adjacent lane and entered in between the EGV and preceding vehicle (PRV). To construct a highly accurate model, we analyzed critical variables as input information affecting the output of the LSTM model using a random forest (RF) model. The results showed the high importance of the EGV velocity, THW, and relative velocity as information related to the traveling lane, in addition to the CIV velocity as information related to the CIV. The CIV data obtained after the lane change were used for the PRV in this analysis. Based on the variables analyzed in the RF model, we constructed personalized driver models using LSTM, and the mean coefficient of determination was greater than 0.95, indicating that this system is more accurate than the conventional car-following models. The driver models constructed in this study are expected to improve the usability of AVSs employing the driver model.