AI Case Study

"DataKind leveraged datasets from New York City’s Department of Transportation, NYC OpenData, New York State and other internal city data to examine the effectiveness of various street treatments to help inform the city’s future planning and investment of resources.

Lacking some of the data necessary to address the actual impact of existing street treatments, the team looked to answer other crucial questions regarding traffic safety that could help benefit the city.

Before they could answer these questions, they first needed to answer a more basic one — how many cars are on the road? Knowing the total volume of road users or “exposure” is necessary to understand the true rate of crashes, but most cities don’t have this data available. To overcome this, the team designed an innovative exposure model that can accurately estimate traffic volume in streets throughout the city.

The model has two main components. The first is an algorithm that propagates traffic counts on a single street segment to adjacent street segments. It assumes that traffic on one city block is very similar to traffic on adjacent blocks. This process can be run many times and allows one to widely propagate traffic count values along neighboring streets. However, some streets may not have any nearby traffic counts available, so the second component of the model is a machine learning model, with high predictive accuracy, that predicts traffic volumes on streets based on their characteristics.

The team also created a crash model for New York, allowing the city to examine individual locations and test how different street characteristics impacts the number of injuries. For example, the city may be able to look at a particular street and determine whether it is safer for the street to be a one- or two-way road.

The exposure model will prove to be invaluable to the City of New York, filling a crucial void in vehicle volume data that many cities face.

With it, the city can now perform initial safety project feasibility studies very quickly and provide context for a variety of other safety research work that requires an “exposure” rate. The model can also be altered to estimate other defined traffic volume measures, like peak hour traffic volumes. It can also help inform future work related to traffic congestion and citywide vehicle usage.

New York can also use the crash models to test the potential impact different engineering, land use and traffic scenarios would have on total injuries and fatalities in the city. They will continue to build upon the work started by DataKind, as the models developed set the stage for future research in crash prediction, congestion relief and city safety projects.

The team was able to leverage the work started in New York City to help develop and refine the approaches for two other cities, Seattle and New Orleans."