Modeling pedestrians' near-accident events at signalized intersections using gated recurrent unit (GRU)

Pedestrian safety plays an important role in the transportation system. Intersections are dangerous locations for pedestrians with mixed traffic. This paper aims to predict the near-accident events between pedestrians and vehicles at signalized intersections using PET (Post Encroachment Time) and TTC (Time to Collision). With automated computer vision techniques, mobility features of pedestrians and vehicles are generated. Extreme Value Theory (EVT) is used to model PET and minimum TTC values to select the most appropriate threshold values to label pedestrians' near-accident events. A Gated Recurrent Unit (GRU) neural network is further used to predict these events. The established model reaches an AUC (Area Under the Curve) value of 0.865 on the test data set. Moreover, the proposed model can also be applied to develop collision warning systems under the Connected Vehicle environment.