Wang, Chengming (Xi'an Jiaotong-Liverpool University), Jia, Dongyao (Xi'an Jiaotong-Liverpool University), Zheng, Zuduo (The University of Queensland), Wang, Wei (Xi'an Jiaotong-Liverpool University), Wang, Shangbo (University of Sussex)

A Novel Feature-Sharing Auto-Regressive Neural Network for Enhanced Car-Following Model Calibration

Scheduled for presentation during the Regular Session "Driving behavior models" (ThBT6), Thursday, September 26, 2024, 16:10−16:30, Salon 14

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 December 26, 2024

Abstract

Accurate calibration and validation are crucial for physics-based car-following models (CFMs) to effectively capture longitudinal human driving behaviors. It may also facilitate the development of hybrid CFMs that integrate physics-based and data-driven models. However, existing calibration methods often produce inaccurate parameters, especially with incomplete trajectories, where certain driving regimes are missing. Additionally, most research only uses individual trajectory data for calibration, neglecting common patterns across trajectories. This research proposes to address the above issues by introducing a feature-sharing approach using an auto-regressive neural network for parameter calibration. This approach allows parameters to generalize to missing driving regimes by leveraging shared information through common features, such as lane information, among different trajectories. We validated the effectiveness of our approach through parameter estimation accuracy with simulated data and trajectory simulation accuracy with real-world traffic data, showing it outperforms existing calibration methods. Furthermore, we evaluated our approach in the recent promising paradigm of physics-informed deep learning (PIDL). Experiments show significant performance improvements of PIDL upon integration with a more accurate CFM acting as a physics informer.