Mixed traffic microscopic interactions modelling in shared space using machine learning

Modelling the dynamics of the movement of different types of road users, as well as the interactions among them is a very challenging task, especially in shared space environments where road users share the same priority and space. In this paper, a set of Machine Learning models are developed that can predict vehicles’ decisions, pedestrians’ movement, as well as the priority between them in cases where vehicle and pedestrians interact. In order to feed the algorithms a feature engineering step is preceded where the traffic scene is reconstructed in a variety of parameters that describe users’ positions, movement dynamics and their interactions. Out of six well-established ML models, the Random Forest classifier seems to outperform the rest when modelling car or pedestrian movements. For the priority model, the Gradient Boosting algorithm achieves the higher accuracy and further investigation of the results, through the estimation of SHapley Additive exPlanations (SHAP) values, revealed that the distance between the subject pedestrian and the vehicle, the number of existing pedestrians in the influential area of the vehicle and the acceleration of the pedestrian are the most critical factors. Results and conclusions drawn in this work can be used in other complex environments to model multiple interactions and can be incorporated into simulation applications to define priority between two interacting road users.