Construction of Personalized Driver Models Based on LSTM Using Driving Simulator

Many automated driving technologies have been developed and are continuing to be implemented for practical use. Among them a driver model is used in automated driving and driver assistance systems to control the longitudinal and lateral directions of the vehicles that reflect the characteristics of individual drivers. To this end, personalized driver models are constructed in this study using long short-term memory (LSTM). The driver models include individual driving characteristics and adapt system control to help minimize discomfort and nuisance to drivers. LSTM is used to construct the driver model, which includes time-series data processing. LSTM models have been used to investigate pedestrian behaviors and develop driver behavior models in previous studies. We measure the driving operation data of the driver using a driving simulator (DS). The road geometry of an actual section of the Tomei Expressway, which comprises straight and curved roads, between Tokyo and Nagoya in Japan was simulated in the DS. Personalized driver models were constructed using LSTM based on the data of driving maneuvers on the expressway. Simulation results indicate that model accuracy decreases for the entire experimental road compared to that for each curved road; the model accuracy of each curved road was improved. In order to improve the accuracy, it is effective to build a model for each curve or section, and the accuracy is lower at the exit than at the entrance of the curve, and highest at the middle. And then it is necessary to consider both the time required to improve accuracy and the change in curvature when considering the construction of personalized models on curved roads.