Personalized forward collision warning model with learning from human preferences

The Forward Collision Warning (FCW) system has been widely equipped on vehicles to reduce rear-end crashes, which are considered the most common type of crash. However, existing FCW systems have the problem of low response rates, which restrict their safety improvement effects. This study aims to address this issue by building personalized FCW models based on human risk preferences. First, a warning feedback index ranks the gaps between drivers' risk perceptions and FCW models. Then, reward models are developed to characterize the risk perception preferences of each individual driver. After that, the reward models serve as guidelines to fine-tune the benchmark FCW model using the Proximal Policy Optimization (PPO) algorithm. In the empirical analyses, a total of 95,814 warning fragments collected from 74 drivers are used, and the proposed method generates pseudo warning results. By comparing the pseudo and historical warnings, it shows that the precision of pseudo warning results increases from 53.5% to 78.2%. Furthermore, the average differences between the moment of warning and the moment of braking behavior decrease from 2.4 s to 1.6 s. This demonstrates a higher synchronization level in the timing of risk perception between the personalized FCW models and individual drivers, which enhances the driver's trust in the warning system.