A Machine Learning Method for Real-Time Traffic State Estimation from Probe Vehicle Data

Reliable Traffic State Estimation (TSE) is an important precursor to developing sophisticated traffic controls for intelligent transportation systems (ITS). Historically, TSE is calculated using stationary sensors with occasional vehicle probe data as supplementary data. However, even with recent developments that apply machine learning to TSE calculations, the literature reports having to fuse probe data with stationary data or focus solely on freeways where the penetration is greater. This work proposes and analyzes an Ordinal Regression model developed using XGBoost to compute TSE exclusively from probe data that can be used for real-time model predictive control on signalized corridors. Our results show our model to have an mean absolute error of less than half a class and show promising preliminary results in a real-world control experiment.