Improving Urban Travel Time Estimation Using Gaussian Mixture Models

In this paper a methodology to improve the accuracy of the estimation of path travel times in urban areas is proposed, as they play an important role in advanced management applications. Heterogeneous travel time measurements provided by different types of monitoring technologies (e.g. Floating Car and Bluetooth Antennas) have been first used to derive current and historical information on traffic states based on Gaussian Mixture Model (GMM) and then fused according to a Bayesian approach. Accuracy of the resulting estimate has been then evaluated by comparison with ground truth. The proposed methodology has been implemented in a full laboratory test relative to the real size network of a district in Rome: starting from traffic counts collected by fixed detectors in a period of 40 days, the 960 hourly Origin - Destination traffic demand matrices, split in 20 minutes time slices, have been estimated. The ground-truth information has been assumed as vehicles trajectories for any considered time slice obtained by a Dynameq microscopic simulation based on within-day user equilibrium principles. Measurements from different sensors have been simulated adopting different values of technologies penetration rates, errors bias (mean value) and dispersion (variance and distribution). Then, identification of the traffic state of simulated current measurements according to the historical GMM has allowed fusing different heterogeneous data. According to promising results that have been obtained, future research will deal with travel time forecasts.