A Reinforcement Learning-Based Adaptive Forward Collision Warning System by Considering Drivers’ Reaction Time in Real Time

This study proposes a reinforcement learning-based adaptive forward collision warning system (RL-AFCWS) with individualized real-time driver reaction time (DRT). Specifically, the curve matching algorithm is presented to obtain the real-time DRT by matching the trajectories of preceding-following vehicle pairs. A deep Q network (DQN) model is constructed to predict the activation state of the collision warning by using trajectory data and DRT as inputs. To improve the accuracy of the RL-AFCWS, real driving data set NGSIM from two different periods is selected for samples. The effectiveness of the proposed study is verified by numerical experiments. The result indicates that the proposed RL-AFCWS can effectively improve the accuracy rate and true positive rate to 94% and 49%, respectively, while greatly reducing the alarm interference, compared to the TTC-based classical model and adaptive model. It also has realized adaptive early warning for drivers with different driving styles.