A Hybrid Physics-Based Car-Following Model with Bayesian Convolutional Neural Network

Physics-based car-following models (CFMs) are widely used for microsimulation and predictive control purposes. However, human driving often deviates from the machine-like driving described by such models due to intra- and inter-personal heterogeneity in driving behaviours. Various data-driven and hybrid approaches have been proposed in the literature to address the limitations of physics-based models, leveraging advancements in machine learning and the availability of high-quality traffic data. Hybrid models, in particular, integrate the merits of physics and deep neural networks to capture such heterogeneity and achieve more accurate trajectory estimates under a broad range of traffic conditions. In this study, we propose a novel hybrid CFM in which a Bayesian Convolutional Neural Network (B-CNN) is employed as an additional computational layer on top of a physics-based model, such that the B-CNN captures (i) the systematic errors inherent to the physics-based model and (ii) variations in driving patterns. The proposed model is trained and tested on both simulation and real-world datasets. Experimental results indicate that our hybrid model effectively captures systematic errors and heterogeneity in driving behaviours, thereby producing significantly more accurate trajectory (speed and acceleration) predictions than a well-calibrated Intelligent Driving Model (IDM). The developed model has strong potential for a variety of traffic management applications, such as anticipatory traffic signal control.