Mixed Traffic Trajectory Prediction Using LSTM–Based Models in Shared Space

Abstract

Real–world behaviors of human road users in a non-regulated space (shared space) are complex. Firstly, there is no explicit regulation in such an area. Users self-organize to share the space. They are more likely to use as little energy as possible to reach their destinations in the shortest possible way, and try to avoid any potential collision. Secondly, different types of users (pedestrians, cyclists, and vehicles) behave differently. For example, pedestrians are more flexible to change their speed and trajectory, while cyclists and vehicles are more or less limited by their travel device—abrupt changes might lead to danger. While there are established models to describe the behavior of individual humans (e.g. Social Force model), due to the heterogeneity of transport modes and diversity of environments, hand-crafted models have difficulties in handling complicated interactions in mixed traffic. To this end, this paper proposes using a Long Short–Term Memory (LSTM) recurrent neural networks based deep learning approach to model user behaviors. It encodes user position coordinates, sight of view, and interactions between different types of neighboring users as spatio–temporal features to predict future trajectories with collision avoidance. The real–world data–driven method can be trained with pre-defined neural networks to circumvent complex manual design and calibration. The results show that ViewType-LSTM, which mimics how a human sees and reacts to different transport modes can well predict mixed traffic trajectories in a shared space at least in the next 3 s, and is also robust in complicated situations.