Deep neural networks have been adopted to recognize human car-following behaviors under the assumption that data would be all that was needed. These attempts, however, are inefficient because the knowledge accumulated by previous theory-based car-following studies is not utilized. In order to combine both approaches, we investigated the potential for using coefficients in a theory-based car-following model to introduce stochasticity to car-following behavior. To achieve this, we developed a probabilistic graphical model (PGM) that generates an ego vehicle's car-following response and the trajectories of the ego and surrounding vehicles. The proposed modeling framework integrates a theory-based car-following model with two variational autoencoders (VAEs) to embed the trajectories of the ego vehicle and surrounding vehicles into the hidden driving regimes and the corresponding random coefficients of the car-following model. The reaction time embedding was also incorporated into the modeling framework. The PGM was estimated using the variational inference (VI) within a Bayesian framework. As a result, the proposed car-following model outperformed other benchmark models in reproducing real driver responses.

深度神经网络已经被用来识别人类的跟车行为，假设数据就是所需要的。然而，这些尝试是低效的，因为没有利用以前基于理论的跟车研究积累的知识。为了结合这两种方法，我们研究了在基于理论的跟车模型中使用系数将随机性引入跟车行为的可能性。为实现这一目标，我们开发了一种概率图形模型 (PGM)，可生成自我车辆的跟车响应以及自我和周围车辆的轨迹。所提出的建模框架将基于理论的跟车模型与两个变分自动编码器 (VAE) 集成在一起，以将本车和周围车辆的轨迹嵌入到隐藏的驾驶状态和跟车模型的相应随机系数中。反应时间嵌入也被纳入建模框架。PGM 是在贝叶斯框架内使用变分推理 (VI) 估算的。因此，所提出的跟车模型在再现真实驾驶员反应方面优于其他基准模型。