Data-Driven Approach for Modeling the Nonlane-Based Mixed Traffic Conditions

Abstract

The diverse nature of vehicle categories and the resultant lane discipline in mixed (heterogeneous) traffic cause complex spatial interactions. As a result, the driving behavior process in mixed traffic conditions is meaningfully different, where both longitudinal and lateral movements of the vehicles continuously occur. Under prevailing homogeneous traffic conditions in developed countries, driving behavior is partially discrete, where following longitudinal behavior and outboard lane-change models can model traffic behavior. However, the established car-following and lane-change models cannot be directly used in shaping mixed traffic conditions. Such conditions also warrant the use of high-quality microlevel vehicular trajectory data. Accordingly, realizing this need, vehicular trajectory data for different traffic flow conditions were developed. The data were used to extract the parameters required for modeling the vehicles' positions using machine learning algorithms. Three established supervised machine learning algorithms (k-NN, random forest, and regression tree) and deep learning are selected to model mixed traffic conditions. The parameters which influence longitudinal and lateral movements are identified using Spearman correlation analysis. Furthermore, simulation runs are performed using the python language. The performance of the algorithms is evaluated both at the microscopic and macroscopic levels using relevant traffic indicators. The results show that a deep learning model and k-NN tend to replicate better-mixed traffic conditions than random forest and regression trees.