Optimal sensor placement and density estimation in large-scale traffic networks - Thesis defense

Thesis defense of Martin Rodriguez-Vega on April 9, 2021.

Title: Optimal sensor placement and density estimation in large-scale traffic networks

Abstract:

The contributions of the PhD work are mainly related to the monitoring of large-scale traffic network states. We deal with three main problems: the location of sensors under budget constraints, the estimation of traffic density using heterogeneous data sources, and the estimation of an aggregated state for a region of an urban traffic network.

First, the optimal sensor placement problem is considered. Our contribution consists in the analysis of two sensor technologies, one that measures the absolute flow in a road, and another that estimates intersection parameters called turning ratios, which are a relative measure of route choice behavior of drivers. We propose an approach to choose the locations for each type of technology such that a minimal number of total sensors are required, and that traffic flow can be estimated for each road of the network.

The second problem considers the estimation of flow and density using heterogeneous sources of information. In addition to the fixed flow sensors used in the first problem, another available data source is the so called Floating Car Data (FCD) which provides the trajectories of individual vehicles, albeit at an unknown penetration rate and generally aggregated due to privacy regulations. We analyze how to fuse these data sources to be able to estimate the density and flow of every road in the network, for the static and dynamical cases depending on the amount of available information.

For the third problem, we consider the estimation of the aggregated density of an urban network. This is of interest when the density of every individual road of a zone is not required to be known, or when computational power is limited. The convergence of an estimator for the average density of the zone was analyzed, but it was found that in general, the estimator does not converge. To solve this problem, we propose a method to calculate a virtual representation of the same underlying physical network where each road is divided into a number of cells, such that the estimator for the virtual system converges. Under certain conditions, we show that the difference between the real and virtual averages is small.
The effectiveness of our contributions were tested using simulated and real data. In the first case, the simulation consists of an application of the well known microscopic traffic software Aimsun, where the dynamics of individual vehicles are calculated in a modeled real network. In the second case, real data is obtained from sensors located in the downtown area of the city of Grenoble and collected using the Grenoble Traffic Lab (GTL).

[juanraa]

Very proud of the great work Martin has done!