Inquiring the Next Location and Travel Time: A Deep-Learning-Based Temporal Point Process for Vehicle Trajectory Prediction

Trajectory prediction for individual vehicles has emerged as a vital component in Internet of Things (IoT)-based traffic management applications, inducing various control strategies for alleviating traffic congestion. This study focuses on a novel topic in this field, i.e., making joint predictions for the next location and travel time. Based on principles of vehicle mobility, we learn vehicle trajectories as discrete events in the spatiotemporal dimension and propose a neural temporal point process, named TrajTPP. This model employs two attention mechanisms to learn spatial and temporal dependencies, respectively, and a novel recurrent structure is proposed to integrate spatiotemporal features. Meanwhile, a gated residual attentive network (GRAN) is also designed to combine these learned dynamic features with static travel information. Then, the intensity-free learning strategy is employed to make probabilistic forecasting for the next travel times, and we develop a prior transition probability to involve historical travel behaviors in location predictions. Beyond the conventional prediction task, we design a sampling strategy to simulate vehicle mobilities by TrajTPP. Experiments from license plate recognition data in Changsha, China, demonstrate that our model outperforms advanced baselines, and sampling results provide evidence of its ability to accurately simulate vehicle mobilities. Moreover, its impressive accuracy on the latest next-location prediction benchmark is also listed in the Appendix.