Hybrid-learning-based driver steering intention prediction using neuromuscular dynamics

The emerging automated driving technology poses a new challenge on driver-automation collaboration. In this study, oriented by human-machine mutual understanding, a driver steering intention prediction method is proposed to better understand human driver's expectation during driver-vehicle interaction. The steering intention is predicted based on a novel hybrid-learning-based time-series model with deep learning networks. Two different driving modes, namely, both hands and single right-hand driving modes, are studied. Different electromyography (EMG) signals from the upper limb muscles are collected and used for the steering intention prediction. The relationship between the neuromuscular dynamics and the steering torque is analyzed first. Then, the hybrid-learning-based model is developed to predict both the continuous and discrete steering intentions. The two intention prediction networks share the same temporal pattern exaction layer, which is built with the Bi-directional Recurrent Neural Network (RNN) and Long short-term memory (LSTM) cells. The model prediction performance is evaluated with a varied history and prediction horizon to exploit the model capability further. The experimental data are collected from 21 participants of varied ages and driving experience. The results show that the proposed method can achieve a prediction accuracy of around 95% steering under the two driving modes.