How to identify dangerous overtaking manoeuvres based on real-time sensor data?

Abstract

Overtaking manoeuvre which involves both lateral and longitudinal control is considered as one of the most dangerous and complex manoeuvres that a driver can perform. In all types of road accidents related to overtaking manoeuvres, there is a risk of rear-end accidents that the overtaking vehicle no longer maintains the safe distance from the car ahead in preparation for overtaking. The research is an effort to contribute to the development of an Advanced Driver Assistance System which helps predict dangerous overtaking manoeuvres with respect to rear-end accidents before the headway between the driven and preceding vehicles reaching its critical threshold and alert the driver about these possible dangers, giving him enough time to react. A sensory-fusion deep learning architecture based on Recurrent Neural Networks (RNNs) with Long Short-Term Memory (LSTM) units is proposed to monitor vehicle dynamics and driving context and signal predictions. The prediction performance of regular LSTM-RNN and bidirectional LSTM-RNN are also compared with other models based on Feedforward neural network (FFNN). In terms of experiment settings, the model is trained in simulation with driving scenarios on two-lane rural roads but is tested in natural freeway and city driving.