Deep learning-based queue forecasting model for signalized intersection using sampled vehicle trajectory data

Abstract

Traffic congestion is a common problem in urban areas, and accurate forecasting of queue lengths is necessary to alleviate congestion and manage roads efficiently. Previous queue forecasting studies have not sufficiently considered the characteristics of traffic theory, resulting in a lack of interpretability of the methodology. Previous studies that utilized vehicle data had the limitation of low performance in the case of insufficient vehicles. To overcome these limitations, we propose a new methodology based on deep learning to predict queue length by processing sampled vehicle trajectories based on the theory of shock waves in traffic engineering. The methodology consists of three steps: vehicle feature extraction, queue estimation, and forecasting. In the vehicle attribute extraction phase, vehicle trajectories are analyzed to extract a new concept of data that reflects the shock wave theory. In the queue estimation step, the queue lengths of upstream, target stream, and downstream are estimated based on the extracted data. Finally, the estimation of queues takes into account the data of three adjacent links to analyze the influence of individual characteristics of each link on the queue length of the target stream. The mean absolute error (MAE) is as low as about 3 vehicles, which demonstrates the potential of the proposed integrated structure for queue length forecasting. In summary, this study develops a deep learning model applying shock wave theory to propose an accurate and efficient method for queue length forecasting under diverse traffic demand conditions at signalized intersections based on sampled vehicle data.