Talk

Multi-fidelity traffic flow estimation
Alexandre Chasse (IFPEN)

Rapid increase of bicycle mobility requires a precise traffic flow monitoring. A clear vision of the mobility demand is required for optimal transformation and maintenance of the cycling infrastructure. Different kind of traffic measurements are commonly available: GPS traces and fixed counting stations.

- The counting stations provide precise counts of cyclists passing by the detectors. These measurements are expected to be unbiased and non-noisy. However, the sensors are generally sparsely spread over metropolitan areas and only provide a limited picture of traffic flows.

- A dataset of GPS traces from a large community of cyclists is a valuable source of information to estimate route preferences. However, even large cyclist communities are expected to remain a biased sampling of the overall cycling population. Moreover, the trips made by the members of the community can be lost in the data treatment processing pipeline. The map-matching operation (the projection of the GPS traces on the road network) is subject to errors. This rich source of information is then inherently biased and noisy.

In this talk, we will describe a multi-fidelity data reconciliation learning algorithm exploiting information for the road network structure, a large database of bicycle GPS traces and fixed counting station measurements. It provides traffic flow estimation on each road. The method is applied over the capital city of Paris. The estimation offers confidence intervals and has limited biases and noises.