Modeling Mixed Traffic in Shared Space Using LSTM with Probability Density Mapping

Abstract

In shared spaces, vulnerable road users (i.e., pedestrians) are encouraged to directly interact with other road users (i.e., vehicles). They self-organize to give or take right-of-way without being regulated by explicit traffic rules. The safety based on their behavior patterns in such areas is, hence, critical and needs to be fully investigated. In this paper, we first carry out a collision probability method based on safety distance to quantitatively study behavior patterns in mixed traffic. Then we propose a Long Short-Term Memories recurrent neural networks model that takes 2D trajectory data at discrete time steps and incorporates collision probability that captures user behavior patterns as a density mapping function for trajectory prediction. The model handles collisions explicitly based on the relative positions of the neighboring users in an ego user's vicinity with considering the impact of personal space and vehicle geometry. It also provides a friend flock detection mechanism to allow close interactions between friends. After training by real-world trajectories, the model outputs comparative results to the state-of-the-art methods for predicting three-second trajectories in complicated situations in a shared space. It can be applied for intent detection and on-board alarming system for autonomous driving when interacting with multimodal road users in shared spaces.