Data-Driven Prediction for Red-Light-Running at a TJunction

Running red lights is a serious road safety problem worldwide, which often leads to severe injuries and fatalities. Most recent works focus on identifying red-light-running behavior through surveillance cameras for punishment of violations. A few works predict the red-light-running behavior of drivers at intersections with Support Vector Machines (SVM) method. But they pay little attention to non-motor vehicles and the accuracy needs to be further improved. To address this problem, we conduct an observational experiment and construct a trajectory dataset (RedRun dataset) with the software Petrack. We also propose an Environment-Aware Red-light-running and Trajectory prediction Network (EA-RTN). It predicts the trajectories and red-light-running behavior of individuals (i.e. pedestrians, bicycles, electric vehicles, tricycles and cars) at T-junctions to help road users judge others’ movement in advance. Specifically, EA-RTN consists of two modules: one is a fully connected neural network (FCNet), which uses two hidden layers to predict whether a road user will run a red light. The other is a two-layer long short-term memory neural network. It predicts the trajectories of road users in the next 2 seconds and then assists drivers to plan ahead. The losses of these two tasks are combined to update the weights for realizing the multi-task learning. To evaluate our model, experiments are conducted on RedRun dataset. The results show that our approach predicts red-light-running behavior of road users more accurately. The accuracy is about 10% higher than SVM method.