Traffic Flow Estimation Using Higher-Order Speed Statistics

In this article, we consider the problem of estimating traffic flow on a multi-lane road using a set of point speeds, either crowd-sourced or collected from the fixed infrastructure. We specifically investigate the relation between higher-order speed moments and the expected value of traffic flow. The algorithm proposed is based on the selection of optimal covariates constructed as speed moments, for a class of conditional mean predictors. The second contribution of this article consists in the analysis of specific components of the speed moments with significant correlation with flow values. In particular, we show that for more than 75\% of the fixed sensing devices considered, the correlation coefficient between the inter-lanes speed variance and the aggregate flow is more than 0.75. Additionally, for more than 70\% of these fixed sensing devices the lane speed variance increases with flow. The third contribution of this article consists of identifying the explanatory features for the high correlation between speed moments and flow values. The algorithms presented in this article are trained and tested on a large dataset from the Mobile Millennium system, collected in the Bay Area from August 2009 to October 2009.