Fusion de données pour l'estimation des temps de parcours via la théorie de l'évidence

Abstract

The emergence of new information technologies and the transformation that has occurred in traffic management have both increased drivers' already considerable need for road traffic information. The travel time is one of the forms in which this information is presented, and a number of systems are based on its dissemination. In this context, this indicator is used as a measure of the impedance (or cost) of routes on the network and/or a congestion indicator. This raises the problem of estimating travel times with an acceptable degree of accuracy, which is a particularly difficult task in urban areas as a result of difficultes of a theoretical, technical and methodological nature. Thus, in order to find out the traffic conditions that prevail on an urban road, the traffic sensors that are usually used to measure traffic conditions are ineffective under certain circumstances. New measurement devices (cameras, GPS or cellphone tracking, etc.) mean that other sources of data are increasingly used in order to supplement the information provided by conventional measurement techniques and improve the accuracy of travel) time estimates. As a result, travel time estimation becomes a typical data fusion problem. This study deals with a multisource estimate of journey times and attempts to provide a comprehensive framework for the utilization of multiple data and demonstrate the feasibility of a travel time estimation system based on the fusion of data of several different types. In this case two types of data are involved, data from conventional induction loop sensors (essentially flow and occupancy measurements) and data from probe vehicles. The selected modelling framework is the Dempster-Shafer Evidence Theory, which has the advantage of being able to take account of both the imprecision and uncertainty of the data. The implementation of this methodology has demonstrated that, in each case, better results are achieved with fusion than with methods based on a single source of data and that the quality of the information, as measured by correctly classified rates, improves as the degree of precision required of the estimate is increased.