Probabilistic model for robust traffic state identification in urban networks

Efficient estimation of local traffic states from fundamental diagram at each detection site in urban and freeway networks is crucial for many real-time traffic management applications. Usually, these traffic states are inferred from the bivariate relationship between traffic flow and density using a deterministic approach. However, due to traffic congestion and position of detection sites especially in urban networks, this relation is highly scattered making these methods not suitable to handle the associated uncertainty in the process. We propose a probabilistic model that allows the inclusion of prior knowledge on traffic states and part of their relative parametrization according to the expert user's judgment. The model is formulated in a Bayesian framework where we also introduce several constraints as per the fundamental diagram shape to solve the common problem of identifiability in this kind of generative models used to estimate latent variables. Derived probability distributions can be efficiently updated in real-time with new data observations. The model performance has been evaluated in three networks: M4 Western Motorway in Sydney, and urban city centers of Santander and Leicester. Results demonstrate the robustness of our approach to infer traffic states even with low data availability in some parts of the fundamental diagram.