Can car-following models simulate vehicle-bicycle interactions? A benchmarking study

Understanding vehicle-bicycle traffic dynamics has become important in urban mixed traffic environments, whereby their accurate modelling and simulation are paramount for infrastructure design and safety evaluation of mixed traffic. However, the literature is devoid of concrete evidence on whether existing car-following models can reproduce longitudinal interactions of vehicle-bicycle traffic. To this end, this study benchmarks three widely used car-following (Gipps, Krauss, and Wiedemann 99) models to simulate vehicle-bicycle following using six naturalistic trajectory datasets of vehicle-bicycle interactions and performs empirical analyses and model calibration/validation. Comparing the numerical and behavioural soundness of the models, the Gipps model outperforms its competitors in both calibration and validation, whereas the Wiedemann model performs poorly. Further, given that these models are originally developed for vehicle-vehicle following, a comparison of vehicle-vehicle and vehicle-bicycle interactions reveals two main findings. First, significant behavioural differences, like large standstill distance and shorter reaction time for vehicle-following-bicycle, reflect cautious driving when following bicycles. Second, models show higher calibration and validation errors for vehicle-bicycle interactions, suggesting that these models may not adequately capture behavioural complexities inherent in vehicle-bicycle interactions, like stimulus perception, multi-vehicle anticipation, and intra-driver heterogeneity. Finally, this study also assesses the suitability of short trajectories (less than 10 s), and results confirm the feasibility of using such trajectories for calibration and validation. Our benchmarking findings highlight several gaps in current vehicle-bicycle modelling capabilities and suggest the need for developing new, realistic, and behaviourally sound models for accurately reproducing the dynamics of vehicle-bicycle interactions.