A CNN-based generative model for vehicle trajectory reconstruction in mixed traffic flow

Jianghan Hu, Song Lian, Simon Hu, Cristoforo Demartino, Gaoang Wang, Xin Liu, Yongfu Li, Claudio Roncoli, Der-Horng Lee

Abstract: With the breakthrough of connected and autonomous vehicles (CAVs) technology, vehicle trajectories can be collected by various sensors installed on CAVs continuously and sent to the traffic control center for operation and management. However, the trajectory data collected by CAVs may contain incomplete part owing to sensor limitation, thus hindering the data availability. To address this issue, we propose a CNN-based generative model for reconstructing multiple vehicle trajectories in multi-lane traffic scenarios using hybrid detection data from both CAVs and fixed sensors. Specifically, we generate missing trajectories using a Generative Adversarial Network (GAN) architecture with spatio-temporal features extracted by Convolutional Neural Networks (CNNs). The performance of the method is examined on a simulated arterial by assessing the mean absolute error (MAE) of the reconstructed data. The results indicate our method is robust even at a low CAV penetration rate.