Modeling of traffic dynamics and evolution is difficult due to complex spatial and temporal aspects of traffic across a network. For example, consider how downstream congestion from a bottleneck can spillback on a freeway for miles in some cases. When one further considers how traffic crashes and special events (among many other factors) can lead to irregularities in demand, it is clear to see how complicated traffic modeling at a network scale can become. Recently, the STAR Lab took a variety of research projects studying traffic modeling on large transportation networks that are rooted in the application of multiple sources of traffic sensing data.
A critical first step in this research is to store and maintain these various data streams. The STAR Lab thus developed and maintains an online platform, called the Digital Roadway Interactive Visualization and Evaluation Network (DRIVE Net), for exactly this purpose. This platform facilitates large-scale online data storage, sharing, visualization, modeling, and analysis.
A second step in this effort involves using this platform to collect and analyze data for a given area of analysis. Over the years, the STAR Lab developed numerous state-of-the-art machine learning and statistical-based algorithms to model and forecast traffic states for freeway networks (such as that in the Seattle area).
A third step involves sharing our knowledge and experience with the community. Recently, we built an online platform called AI Net, where we have created workspaces for transportation network model evaluation and related data sharing. On the platform, we provide datasets and implement models and demonstrate their applications such as network-level speed prediction. Visitors of the platform are also welcome to upload their own models for training online and compare to our models and other benchmarks.
Data storage and sharing platforms
Machine Learning for traffic prediction