Validation of a Car Following Model within Vehicle Platoons on Highways using OBD Data

Abstract

In this study, we aim to evaluate a framework for microscopic car-following behavior models, considering the limitations of traffic flow prediction methods based on macroscopic traffic analysis. We conducted driving experiments using On-Board Diagnostics (OBD) on the Tomei and Shin-Tomei Expressways, where we processed the large amount of data obtained through image processing to create a database of the relative coordinates of vehicles in front of and behind the subject vehicle. In our analysis, we treated the positions of the vehicle groups as relative graphs and conducted a comparative study of Long-Short Term Memory (LSTM) and Graph Neural Network (GNN) models as methods for predicting phase transitions in time series data. Compared to conventional methods that assume a set of macroscopic physical laws for traffic prediction among sensors, detectors, and cameras, we found that it is challenging to reproduce speed reduction based on emergent mechanisms within vehicle groups using LSTM. In contrast, the reproducibility of the GNN model, which models the phase relationships within the vehicle group, was significantly improved.