A trajectory-conditional generative adversarial network model for missing vehicle trajectory imputation

Vehicle trajectory data plays a fundamental role in intelligent traffic management system. However, missing trajectory segments due to signal interruption or insufficient sampling frequency often fail to meet the requirements of high-precision applications. This paper proposes a trajectory-conditional generative adversarial network (T-CGAN) model to address the challenge of missing trajectory imputation. Firstly, in order to reduce the generation of redundant trajectories, we construct a directed graph based on discrete spatiotemporal grids, and propose a Shape-Based Missing Trajectory Generation (SBMTG) algorithm to mine conditional information. The SBMTG algorithm reformulates the trajectory filling task as a path optimization problem on a graph with predetermined source and target points, which uses the shape based distance as the optimization objective. Then the trajectories generated by the SBMTG serve as conditional input for the adversarial neural network. Gate Recurrent Unit for Imputation is applied to the adversarial neural network component, which takes missing intervals of trajectories into account. The proposed approach is validated using a local dataset from Xi’an, China. The results consistently demonstrate the algorithm's superior accuracy in infilling missing vehicle trajectories at either single intersections or across multiple consecutive intersections.