Car-following trajectory data imputation with adversarial convolutional neural network

Abstract

Missing values in the vehicle trajectory data undermine its application in traffic modelling and simulation. Traditional methods for missing data imputation rely on neighbor points of the missing/distorted data point and consequently can hardly handle trajectory with consecutive data loss. To fill in this research gap, external data with latent information should be considered to enhance the imputation. Hence, the primary objective of this study is to find an effective approach to impute consecutive missing data in the vehicle trajectory making use of leading vehicle trajectory. We proposed a novel imputation adversarial convolutional neural network (IACNN) by extending the CNN model with a multi-objective loss function and adversarial learning framework for vehicle trajectory data imputation. Performance of the model is evaluated on the commonly used trajectory dataset NGSIM with comparison to other baseline models. It turns out that the proposed IACNN outperforms other baseline models in most data loss scenarios, especially with consecutive data loss.