Real-Time Traffic Modeling and Estimation with Streaming Probe Data using Machine Learning

Traffic information systems play an important role in the world as numerous people rely on the road transportation network for their most important daily functions. This dissertation proposes a general system architecture for processing traffic data and for disseminating accurate, timely traffic information via the internet. It also specifically addresses the challenges with estimating arterial traffic conditions using only data from GPS probe vehicles. GPS probe data promises to be the most ubiquitous source of traffic data for years to come as transit agencies decrease their investment in traditional fixed-location sensing infrastructure. The dissertation introduces the architecture design and implementation of the Mobile Millennium system. A joint project between UC Berkeley, Nokia and Navteq, Mobile Millennium aggregates data from numerous sources, runs state of the art estimation and forecast algorithms, and provides timely traffic information to drivers and other targets. This system took over two years to build and the result is a robust framework for any traffic estimation researcher to access vast stores of data quickly and easily as well as test any number of estimation algorithms. For estimating arterial traffic conditions, this dissertation proposes a hybrid approach leveraging advances in the fields of machine learning and traffic theory (based on hydrodynamic theory). This approach provides a foundation for any arterial traffic estimation model. A variety of model/algorithm approaches are presented, with one ultimately proving to be superior to the rest and the one that should be carried forward as research in this area continues.