Data-Driven Modelling of Car-Following Behavior in the Approach of Signalized Urban Intersections

The increasing focus on virtual testing and development of automated driving systems implies high standards to the accuracy of a virtual testing environment. Especially traffic participants surrounding a vehicle under test must perform realistically in order to compare simulated test results to reality for validation purpose. In this paper, we therefore combine extended floating car data with traffic light signal data and propose a data-driven CNN-LSTM based model to replicate car-following behavior in approaches towards traffic light actuated intersections. The model considers human characteristics like memory effects as well as a reaction delay. The performance of the proposed model is compared to the existing fixed-form models IDM and an extension of the FVD model regarding approaches to signalized urban intersections. The results of the analysis indicate that the developed model outperforms the fixed-form models in replicating car-following trajectory data, especially in situations in which the driver is forced to stop by a red light.