Car-following models are essential parts of any traffic simulation package. In the past, researchers have successfully explained car-following behaviour mathematically. The recent availability of accurate datasets from drone videography has facilitated the development of data-driven car-following models. This thesis investigates if data-driven models are more precise than mathematical models, despite their inferior interpretability. To this aim, the highway drone dataset (highD) from Cologne, Germany is clustered into two traffic phases based on Gaussian mixture models clustering. Subsequently, car-following models with two leaders are developed in each phase using multiple linear regression, including penalisation techniques, random forest and gradient boosting machines regression. Gipps’ and Krauss’ models are calibrated with the same dataset and then compared with the trained data-driven models. Further, a validation procedure using unseen data with varying reaction times confirms that double data-driven methods can surpass the performance of conventional models. It is the first time that a sensitivity analysis of Krauss’ car-following model takes place and the calibrated model is put to comparison with Gipps’ and other models. The results of this study are meant to improve the results of microscopic traffic simulations.     «

Car-following models are essential parts of any traffic simulation package. In the past, researchers have successfully explained car-following behaviour mathematically. The recent availability of accurate datasets from drone videography has facilitated the development of data-driven car-following models. This thesis investigates if data-driven models are more precise than mathematical models, despite their inferior interpretability. To this aim, the highway drone dataset (highD) from Cologne, Ge...     »