A Hierarchical Intention Inference Model for Connected and Automated Vehicles

Lane-changing is one of the typical maneuvers with high risk, which can cause uncomfortable brake and dangerous crash. An experienced human driver keeps an eye on surrounding vehicles, catches lane-changing hints and conducts deceleration in advance. To make the connected and automated vehicle (CAV) more like a skilled driver, it is significant to earlier identify potential lane-changing intentions. Motivated by this, we design a hierarchical model to realize accurate and early inference on driving intentions. A novel module based on parallel recurrent neural networks (PRNN) is proposed to generate intention probabilities and a Gaussian hidden Markov model (GHMM) is suggested to replace the classical ratio comparison to determine intention output. Comparative results on public datasets verify that our model outperforms other homologous methods. Under 95% specificity, our model can accurately identify more than 93% of lane-changing maneuvers 5.10 seconds before the insertion. In addition, driver-in-the-loop (DIL) experiments on a driving simulator are carried out to show the effectiveness of our method. More importantly, we integrate our hierarchical intention inference scheme with a finite state machine (FSM)-based controller. By doing so, the actual significance of our module on the controller is investigated in a real driving task. The comparison of different strategies illustrates that with the help of intention inference, a softer deceleration and a safer distance headway can be ensured by the CAV in a cruise task.