Continuously Improving Model of Road User Movement Patterns using Recurrent Neural Networks at Intersections with Connected Sensors

Abstract

Intersections with connected infrastructure and vehicle sensors allow observing vulnerable road users (VRU) longer and with less occlusion than from a moving vehicle. Furthermore, the connected sensors are providing continuous measurements of VRUs at the intersection. Thus, we propose a data-driven prediction model, which benefits of the continuous, local measurements. While most approaches in literature use the most probable path to predict road users, it does not represent the uncertainty in prediction and multiple maneuver options. We propose the use of Recurrent Neural Networks fed with measured trajectories and a variety of contextual information to output the prediction in a local occupancy grid map in polar coordinates. By using polar coordinates, a reliable movement model is learned as base model being insensitive against blind spots in the data. The model is further improved by considering input features containing information about the static and dynamic environment as well as local movement statistics. The model successfully predicts multiple movement options represented in a polar grid map. Besides, the model can continuously improve the prediction accuracy without re-training by updating local movement statistics. Finally, the trained model is providing reliable predictions if applied on a different intersection without data from this intersection.