Li, ZhiHao (Key Laboratory of Road and Traffic Engineering, Ministry of Educ), Li, Ye (key laboratory of road and traffic engineering, ministry of educ), Gruyer, Dominique (Université Gustave Eiffel), TU, Meiting (Tongji University)

IWP-PIDL: Interpretable-Weight-Parameters Physics-Informed Deep Learning Model for Traffic State Estimation Based on Vehicle Trajectory

Scheduled for presentation during the Invited Session "Cooperative Driving Technology for Connected Automated Vehicles" (WeAT10), Wednesday, September 25, 2024, 12:10−12:30, Salon 18

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on October 8, 2024

Abstract

Traffic state estimation is crucial to improve the safety and efficiency of Connected-Automated Vehicle. In this task, physical information deep learning has been widely used by researchers due to its excellent performance. Existing studies have neglected the traffic shock waves generated due to inefficient road operation and the interpretation of weight parameters. Therefore, this paper proposes a framework considering various penetration rates that integrates physic information on traffic shock waves and establishes interpretable weighting parameters through the regional training priority metrics. Subsequently, the study employs next generation simulation dataset to randomly select trajectory data based on a specified ratio and evaluates the proposed model. The experiment shows that the proposed models all have more than 10% improvement in accuracy compared to the PIDL model with LWR and DL model. The design of weighting parameters proves that deep learning helps fast convergence in the early stage and physical information helps deep convergence in the later stage. The proposed model also provides the potential to identify "moving bottlenecks," which is a crucial link for enhancing the utilization of road resources.