Abstract:

Smartphones and other vehicular sensors equipped with GPS
and wireless networking capabilities, are becoming ubiqui-
tous in transportation systems. They provide us with op-
portunities to gather timely information about road traffic
conditions, fuse (assimilate) it with traffic 
ow models to improve upon the accuracy of these models, and 
hence supply valuable information for real-time transportation deci-
sion making. Macroscopic traffic flow models are described
by systems of partial differential equations (PDEs), which
are usually only solved numerically. Adaptive moving mesh
methods have shown promise in handling high variability of
the spatio-temporal features (e.g. shocks and discontinu-
ities) in model's solutions.

We propose a novel low-overhead strategy to adaptively
select observation sites in real time, by relying on informa-
tion from the adaptive moving mesh of the numerical solver
of the underlying PDEs. The idea is to place more of the
limited observational resources to locations of higher vari-
ability in the numerical solution. We incorporate our strat-
egy into a particle filter based data assimilation framework,
and compare it with the strategy of gathering and assimilat-
ing measurements from evenly spaced observation sites. We
experimentally show that our strategy reduces the relative
error by up to 53% in estimating vehicle density on a road
during phantom jams and traffic jams due to bottlenecks.

Keywords
traffic prediction, adaptive observation, Helbing's model, da-
ta assimilation