On-line calibration of traffic prediction models

A methodology for the on-line calibration of the speed-density relationship is formulated as a flexible state-space model. Applicable solution approaches are discussed and three of them (extended Kalman filter (EKF), iterated EKF, and unscented Kalman filter (UKF) are selected and presented in detail. An application of the methodology with freeway sensor data from two networks in Europe and the U.S. is presented. The improvement in the estimation and prediction of speeds due to on-line calibration (compared with the speeds obtained from the off-line calibrated relationship) is demonstrated. The EKF provides the most straightforward solution to this problem, and indeed achieves considerable improvements in estimation and prediction accuracy. The benefits obtained from the -more computationally expensive-iterated EKF algorithm are shown. An innovative solution technique (the UKF) is also presented. The UKF has a number of unique qualities and advantages over the EKF, including no assumption of analytical representation of the model and no need for explicit computation of derivatives. Empirical results suggest that the UKF outperforms the other two solution techniques in prediction accuracy.