Traffic sensor location using Wardrop equilibrium

Abstract

This paper proposes a strategy for optimal traffic sensor placement that does not require previous traffic measurements. Our approach could be used to determine how many sensors are needed and where to place them in order to obtain an estimation of the network traffic state. We first generate a traffic-flow dataset based on the transport network and some transportation demands. Specifically, the traffic flow is obtained by calculating the Wardrop equilibrium associated with each demand. Then, a neural network autoencoder with a l_1 regularization is trained with that dataset. Two initialization strategies were used and their performances were compared and validated. The final neural network weights indicate where sensors should be placed and also give the traffic flow reconstruction from those measurements. This approach was tested on several well-known traffic networks present in the literature, including a real large-scale network, with promising results.