Modeling Highway Traffic Volumes

The behavior of vehicular traffic is subject to many irregularities and stochastic fluctuations. Thus, our success in solving a variety of traffic management and optimizations tasks, such as accident detection or optimal vehicle routing, depends heavily on our ability to adequately model, reason about and predict such a behavior. The objective of this work is to develop a probabilistic model of traffic flows on highway networks that is realistic from the point of applications and at the same time supports efficient learning and inference. We study a number of multivariate probabilistic models, including the multivariate Gaussian model and the conditional autoregressive model and analyze their advantages. To address the efficiency concerns we develop a mixture of Gaussian trees model – and show its advantages in learning and inference with respect to alternatives. The performance of all models is analyzed on real-world traffic flow data obtained from the network of sensors installed on major highways in a metropolitan area.