Prediction of global traffic states in road networks, and estimation of traffic state by aggregation of floating car data

Stage « Master Recherche » ou ingénieur

Télécharger le sujet au format PDF

Lieu du stage :

Centre de Robotique (CAOR), Mines ParisTech 60 boulevard Saint Michel 75006 Paris + partiellement (si stage long) à BERKELEY, Californie, USA


In Intelligent Transportation Systems, short-term prediction of traffic status is necessary for automatic route selection and prediction of traveling time along pre-specified roads. Furthermore, planning and management of urban traffic need to optimally regulate traffic states in different kinds of roads. To achieve this goal, prediction, especially short-term forecasting of traffic states in the whole road network is essential. In this field, most of current research focus on performing prediction on separately on each link of the road network. However, traffic states of one road usually correlates with those of its geometrical neighbours and affected by regular behaviour patterns of drivers shared by different roads. Therefore, constructing traffic state prediction with respect to each link independently ignores the correlation and the common factors, which leads to under-estimation of real traffic dynamics due to constraints of model complexity. More important, we usually need the networked traffic state information, not just traffic state estimation of individual links, in order to regulate traffic flows over the whole network. Therefore, it is necessary to figure out how to predict the global traffic states in a networked way. However, the high dimensionality of the global traffic state observations makes it a difficult task to perform prediction directly. We propose a novel state-space model to evaluate and predict evolution of global traffic states in a road network based on compressive sensing and regression models. To construct the state-observation module, we firstly assume that the vector of global traffic states can be decomposed to (or approximated by) a linear combination of orthonormal basis. After that, we perform a random projection matrix of the high-dimensional global traffic state vectors to obtain low-dimensional measurements as latent states in the proposed state-space model. To construct the transition function between the latent states, we plan to adopt regression models, such as neural network or vector-output SVM (Support Vector Machine), to approximate transition relations between the multi-dimensional latent states of neighbouring time sample points. During temporal evolution of the state-space model, currently observed global traffic state is projected to a low-dimensional latent state variable. Based on the regression based transition function, we could predict the latent state at the successive time. The predictive global traffic state is then reconstructed using the estimated latent state based on compressive sensing theory. Following this scheme, the evolution of the proposed state-space model can be shown in the above. An important practical aspect of traffic data analysis is the way it is obtained. Until recently most information used to come from magnetic detection loops buried into some main roads. Nowadays, due to wide usage of GPS navigation, more and more data are collected from fleet of probe vehicles (aka floating car data). The individual travel times of one single probe vehicle being potentially a very noisy estimator of the traffic state, these are usually aggregated over time, space and vehicles. However the optimal sampling and data aggregation strategy may depend on the application requirements. An interesting open research topic is therefore testing the quality/fidelity/characteristics of obtained traffic data depending on sampling and aggregation strategy. Depending on the starting time, duration, and on prior knowledge and preference of the intern, the internship work may focus on one, or several, of the following tasks of our global workplan: 1. Verification of the model assumption: Verify the existence of a dictionary of spanning orthonormal basis, based on which an observed global traffic state vector can be decomposed to a sparse combination of the basis in the dictionary. 2. Test the proposed algorithm on two different databases, which contain simulations of traffic evolution in road network. 3. Develop a toolbox which uses the proposed algorithm for global traffic prediction in road networks of different scales. 4. Develop and evaluate sampling and data aggregation strategies for processing floating car data to provide high-fidelity inputs to new urban traffic signal control strategies that are being designed based on the assumed availability of this richer set of input data (compared to existing conventional point detector data). Traffic simulations based on a commercially available simulation platform (VISSIM) will be combined with models of the data sampling and aggregation strategies and the new signal control strategies for individual intersections, arterial corridors and possibly area-wide control as well. [This part of internship is relatively independent from the rest, and would be conducted in BERKELEY UNIVERSITY (CALIFORNIA) UNDER SUPERVISION BY Pr SHLADOVER, WITHIN A FRANCO-CALIFORNIAN COLLABORATION PARTNERSHIP]


1. Fluency in English absolutely necessary 2. Some knowledge of statistical learning and optimization. 3. Programming skills in matlab and C++ is required. 4. Knowledge in linear algebra; some notions of sparse coding / compressive sensing would be a plus.

Contexte de travail

Durée de 3 à 6 mois (à définir). Indemnités de stage selon profil (+ frais déplacement si séjour à Berkeley)


Yufei Han, post-doctoral researcher (tél :, Fabien Moutarde, enseignant-chercheur (tél :,

caor/positions/2011_internship_roadtraffic.txt · Last modified: 2011/11/18 12:44 by MOUTARDE Fabien
Recent changes · Show pagesource · Login