Author:

Keywords:

CIB_TRAFFIC

Abstract:

Demand estimation problems based on traffic counts have been investigated for decades. These are traditionally solved by an optimisation problem, where some distance between measured and simulated link flows is minimised, in order to find the most likely origin-destination (OD) flows. To partly limit the effects of solution under-determinedness, typical of these problem types, a-priori assumptions on the OD matrix structure are adopted. This structure is provided by either adding additional components in the objective function to be minimised, namely by adding the deviation between estimated OD flows and an available seed matrix as additional performance metric, or by using an opportune initial matrix as starting point for the chosen search algorithm, hence looking for the closest local optimum. In this paper we study the impact of sensor locations on the quality of OD estimations. We show that linear correlations between link flow data may negatively affect the estimation reliability. By contrast, efficient sensor location models, able to identify sensor positions that collect linearly independent link flows allow improving the reliability of the estimation process, especially when no good prior information is available on the OD matrix structure. Hence, we study the gain offered by efficient sensor positioning, which is threefold: 1) it reduces the overall solution space were to look for the true OD flows, 2) it allows better estimation of the gradient of the objective function, and 3) it allows enhancing the objective function by including information from links where sensors are not installed, whose flow is however fully determined by the set of sensors available in the network. In addition, we show that by selecting sensor locations according to partial observability metrics introduced by the authors in recent research, the contribution of the additional information in the objective function also improves with respect to alternative metrics introduced in the literature. Tests are done on a small and a mid-sized network example, showing the improvements in terms of a better estimation of the gradient of the objective function and the overall improvement in the objective function values up to 10%.