An intersection in Bellevue, Washington, outside Seattle, where Vision Analytics is being put to use.  Image by the author.

To really stop people from dying on city streets, identify and fix dangerous places before fatalities happen. New technology called "Vision Analytics" does just that. 

Vision Analytics is the result of a partnership between Microsoft, the University of Washington, and cities across North America to work toward towards Vision Zero, which seeks to eliminate traffic-based fatalities and serious injuries within ten years.

The traditional approach to Vision Zero is to make safety improvements where there are known issues. If a driver injures a cyclist at an intersection, someone from the department of transportation might go there and look into why the collision happened, and if the same collision happens repeatedly the city might fix the issue.

But it’s called Vision Zero, not Vision Close-To-Zero. To truly eliminate traffic-based fatalities and serious injuries, cities need to focus on known issues as well as areas where there are issues but no one has been hurt—yet.

Vision Analytics is a new application of big data and machine learning to identify and fix issues before someone gets hurts.

How does Vision Analytics work?

Vision Analytics takes frames of video collected by cameras at intersections and labels what is happening in each frame to track near-miss events. How many cars are there? Is there a car too close to a bicycle? Are the cars going too fast? Did a car collide with a pedestrian? Maybe a car passed a bike and only left two feet, maybe the pedestrian jumped out of the way just in time, or maybe there was a collision but no one reported it. These questions, and hundreds like them, are what Vision Analytics aims to answer.

The computer algorithm is able to answer all these questions with some help from people. After the computer collects video frames, someone needs to go through and help the computer categorize different objects as cars, buses, pedestrians, sidewalks, crosswalks, etc. Over time, the computer learns from these categorizations and learns from itself so that it can eventually answer these questions by itself.

Right now the computer is really good at knowing what is a car, but is still learning how to differentiate between pedestrians and bikes, and the project is asking for volunteers to help the computer learn.

Once the computer is done learning from humans it will keep learning from itself and can apply what it’s already learned to every new jurisdiction that joins the project. Anyone can help the Vision Analytics model learn better today by signing up to identify bikes and pedestrians at the Video Analytics website.

Eventually cities will be able to give Vision Analytics video from their intersections and have the computer determine the best places to make safety improvements. And over time, Vision Analytics can even help make those safety investments better because we will be able to track cyclist and pedestrian safety before and after different improvements.

We could do this in DC too

Already New York City, Los Angeles, Seattle, Calgary, and Hamilton have signed onto the pilot project led by Bellevue, Washington. Washington, DC and other jurisdictions in the region should join too.

To its credit, The District Department of Transportation knows where many collision hot spots already are. For the past two years, the agency has conducted site visits to a handful of problematic intersections across the District. Last year it visited eight sites and this year there are six sites—including one near my home.

This is a good start, but there are over 18,000 intersections in DC. At this rate it will take more than 2,000 years to survey and brainstorm safety solutions for each intersection. 

In DC, once we have data on the safety of our intersections, we can determine where we need safety improvements most and the type of improvements that would solve issues. Rather than making safety improvements off a list that is updated every so often, we could be making improvements from a database that updates every day and continually ranks the most dangerous intersections to those that are ok now. With limits on the amount of improvements we can make each day, it’s a way to get smarter about when, where, and how we make safety improvements.

Of course, this means we need to do two things right: ensure the Vision Analytics computer model works properly and build political will for making the changes needed to meet Vision Zero.

Why use Vision Analytics when we have people?

We all interact with our own built environment and know our neighborhoods and commutes. Why spend time building and training a complex computer algorithm when we all know where the real issues are?

The answer is that we cannot have people observing every intersection at all times. Even if we hired dozens of people to review videos, humans would not be able to classify each vehicle, pedestrian, and cyclist interaction over time. They might be able to see when there are issues, but not at every intersection and not at every moment. That is something only a computer can do.

Relying on collision reports can be problematic because not everyone reports collisions all the time, and often collision reports do not have full and accurate information to solve the issue before someone else gets hurt. Plus, collision reports are reactive and the goal of Vision Zero is to spot problems before people get hurt.

While Vision Analytics cannot and should not replace human judgement, it can be a very powerful tool in the Vision Zero toolbox and our region should be participating.