Multiple Model Particle Filter for Traffic Estimation and Incident Detection

This paper poses the joint traffic state estimation and incident detection problem as a hybrid state estimation problem, in which a continuous variable denotes the traffic state and a discrete model variable identifies the location and severity of an incident. A multiple model particle smoother is proposed to solve the hybrid estimation problem, in which the multiple model particle filter is used to accommodate the nonlinearity and switching dynamics of the traffic incident model, and the smoothing algorithm is applied to improve the accuracy of the estimate when data are limited. The proposed algorithms are evaluated through numerical experiments using CORSIM as the true model. The proposed algorithm is also compared with a standard macroscopic traffic estimator via particle filtering and the California incident detection algorithm. The results show that jointly estimating the state and incidents in one algorithm is better than two dedicated algorithms working independently.