HERE Chicago Hackathon: Third Place Team - Solution to automatically detect roundabouts from Probe data

HERE Chicago Hackathon Winners: 3rd place

Background

In early October, HERE Technologies hosted its first on-site hackathon for local university students. 32 students attended the two-day event, where they developed, proposed and presented solutions to a geospatial/big data problem: Automatically detecting roundabouts from HERE probe data.

Today, we are highlighting the third-place team comprised of Shriya Kethireddy (Data Science), Han Zhang (Computer Science) and Sadia Haidari (Computer Science & UX Design) all pursuing Bachelor’s degrees from the University of Michigan.

Approaches and Challenges to Problem Solving

Their main approach was to:

Filter the data to reduce its size and improve the quality of the data.

Cluster the data by latitude and longitude to create potential candidates for roundabouts.

Implement a mathematical model that utilizes the principles of directionality in a roundabout.

They also experimented with neural networks and logistic regression but due to the complex relationships between parameters, they faced challenges with convergence in our logistic regression model. Additionally, differentiating between U-turns, left turns, and roundabouts proved to be a challenge due to their similar geometries. However, they were able to overcome these and come up with a great final solution that impressed our panel of judges.

Participating in the HERE hackathon was an incredible experience where I could apply my GIS skills to real-world data. It was amazing to see how geographic insights can drive innovative solutions and make a tangible impact!

Shriya Kethireddy

Final Solution

Their final solution was part of a 6-step approach, which included quite a bit of cleaning and filtering the data to start. Data from peak traffic hours, or those that had time inconsistencies and location jumps were removed. Additionally, a list of common roundabout attributes was created and used to filter out unnecessary data points. They then employed DBSCAN to group the remaining points spatially followed by Kernel Density Estimation (KDE) clustering to identify high traffic areas by analyzing the density of data points. They were able to identify those that exhibited circular patterns from there.

Data following the first round of clustering and after the KDE analysis

Those that did not meet this requirement were filtered out. The final step was to create and apply a mathematical model to the data points within each cluster to identify the most likely roundabout candidates. For each point in the cluster, a +1 score was assigned if it adhered to the directionality rules within each quadrant. They classified a cluster as a roundabout if at least 50% of the points conformed to these rules.

Results

With this approach, they were able to achieve a true positive rate of nearly 80%, with a false positive rate of around 40%.

Scaling?

One of the criteria for judging the hackathon results was the ability to scale the process globally. Give that this team’s solution is not dependent on specific latitudes or longitudes for the data points, they were confident that the model could be applied globally and efficiently.