Implementation of Human-Like Driver Model Based on Recurrent Neural Networks

Driver model is the most basic and important model for moving direction control of autonomous vehicles and it has been extensively studied from the perspective of precision and robustness of control, driving safety, and riding comfort. However, human-like driver model is a rarely mentioned research issue. In this paper, we first establish three preview-based driver models in PreScan+Simulink and collect human drivers' steering data on four two-way and two-lane free curved roads with 20 experienced drivers and one experimental vehicle under four specified speeds. Then, the similarities between steering wheel angles of preview-based models and those of human drivers are compared through dynamic time warping (DTW). From the calculation results of DTW and analysis of human drivers' gaze positions, it shows that the preview-based models are hard to reflect the characteristics of human drivers' maneuver. To this end, we propose a human-like driver model based on the continuity of human drivers' steering wheel angles. The experienced drivers' steering wheel angels are modeled with three different kinds of multivariate multi-step recurrent neural networks (RNNs) and the inputs of models are historical speeds, historical road curvatures, future road curvatures, and historical steering wheel angles, as well as the outputs are future steering wheel angles. By comparing the three RNN-based driver models with different configuration structures and historical steps, it is found that the long short-term memory (LSTM) model has the best prediction performance in validation and testing period. In this way, a data-driven human-like driver model is developed to generate human-like steering wheel angles on curved roads.