ACC, one of the ADAS, helps to reduce driver fatigue. However, vehicle control by ACC is uniformly applied to all drivers. Therefore, it differs from normal driving and may cause drivers a sense of anxiety or discomfort. To improve the system, it is necessary to develop a system that reflects the driver's driving characteristics and adapts to the personal driving behavior. To achieve the system, driver models that simulate driving behavior are required. The purpose of this research is to construct a personalized driver model using LSTM (Long Short-Term Memory) for the merging behavior of drivers at highway merging sections. LSTM models were constructed for a scenario of merging into the main line at a highway merging section. Driving simulators were used in the experiment, and seven participants were asked to drive as they normally do, and their driving data was measured. The LSTM model was constructed using the measured data, and the models were objectively evaluated using unknown data that was not used for training the models. We also verified the influence of the model input values on the model accuracy. The accuracy of the constructed models was evaluated by RMSE (root mean square error). As a result, the RMSE was low for all drivers, and a highly accurate models were constructed. The results of this study were similar to the RMSE of a previous study in which a driver models were constructed for a following scenario, and therefore, models with a highly adapted to individual drivers could be constructed. It was also found that the input values of the models related to the merging section affect the model accuracy, and that there are individual differences in the influence of the model accuracy. In this study, other vehicles drive on main line at 100 km/h and keep sufficient distance between vehicles without interfering with ego vehicle's merging. However, in actual situations, the position, velocity, and distance of other vehicles may vary. In merging sections, it is considered that the ego vehicle is strongly affected by other vehicles, so it is necessary to verify the generality of the driver models. In addition, since the participants in this study were young people, it is necessary to investigate the effects of different age groups and driving experience on the accuracy of the models. Several studies have been conducted to model personal driving behavior using LSTM models. These models have been constructed for each scenario such as following, overtaking, and interrupting, which are assumed to occur on highways, and showed a highly adapted to the individual. However, few studies on driver models have focused on merging behaviors, and especially there are few studies that focus on individual merging behaviors. In this study, driver models were constructed for merging behavior using the LSTM model. The accuracy of the driver models was quantitatively evaluated by comparing the output of the constructed models with actual driving data of experimental participants measured by a driving simulator. The experimental results showed that models with highly adapted to the individual were constructed. It was also found that the input values related to the merging section affected the model accuracy, and that there were individual differences in the influence of the model accuracy.
Mr. Sota Sakai, student, Shibaura Institute of Technology