Driver-Centric Predictive Risk Map Modeling via Deep Reinforcement Learning

In the field of autonomous vehicles, extensive research on risk assessment has been carried out to improve road safety. Nevertheless, in scenarios of assisted driving, particularly scenarios of human-machine co-driving, the establishment of driver-centric risk assessment models remains a challenge. The absence of consideration for human drivers may lead to a reduction in trust within the human-AI team and give rise to conflicts between humans and machines. This study presents a Driver-centric Predictive Risk Map (DPRM) model that combines the driver model with the AI risk assessment model through deep successor reinforcement learning. The proposed risk map not only takes into account future risks associated with driver behavior but also adapts to drivers with different driving styles.