Volunteers will “teach” the program to better identify people walking and biking
Despite signing onto Vision Zero, a campaign that aims to eliminate traffic-related deaths and serious injuries, Washington, D.C. and other large cities have struggled to make progress in reducing deaths on their streets.
While traffic deaths in D.C. have held steady in the high 20s over the past few years, traffic fatalities nationwide have increased by 6 percent, just from 2015 to 2016. Traffic safety data, especially, can be difficult to quantify and analyze in support of VZ plans. However, a new tool is in development that will provide planners with insight to better understand where they should target their safety efforts.
Franz Loewenherz, a transportation planner in Bellevue, Washington, has, in partnership with Microsoft and a number of city departments of transportation, developed a video analytics program to evaluate the safety of any intersection with a traffic camera. The program analyzes various road users – cars, bicyclists, pedestrians – in short clips of about 10 seconds, and follows their behaviors. This way, the tool can identify dangerous situations – like near-misses between people on bikes and drivers – and give planners the opportunity to make proactive changes that improve an intersection’s safety before people are hurt.
But before this program can officially generate recommendations, it needs to learn more. Its developers are currently crowdsourcing involvement from volunteers to teach the computer how to better recognize relevant subjects. According to Loewenherz, the program is more than 90 percent accurate in identifying cars, but still struggles with people walking and biking. As participants manually categorize the various objects passing through their frame, they build a baseline knowledge for the computer to automatically determine what it sees more accurately across multiple contexts.
So far, a number of local departments of transportation throughout the U.S. have contributed about 20 intersections’ worth of footage, creating hours of material for participants to catalogue. The platform will display the same clips to multiple people, using the “wisdom of the crowd,” as Loewenherz describes it, to rectify inaccuracies and fine-tune identification. The more people who contribute, the more the program learns, and the better the data planners have to work with.
A sample view of several pedestrians identified within a downtown Seattle intersection.
Gathering proactive data
In analyzing everyday street movements, the program captures important information about vehicle conflicts that is not necessarily tracked by traditional incident reporting. Loewenherz explains that most agencies “wait for people to become statistics … and generate enough data points” to determine dangerous spots and take reactive safety measures.
On the other hand, Loewenherz says the machine-learning approach should “create a system that provides advance knowledge of how effective changes are” in improving intersections. Most evaluations of safety campaigns, through engineering or enforcement, rely on putting observers at the problem spot to manually track how people on the street act before and after implementation, and only at select times. Not surprisingly, the results don’t always paint a complete picture.
With video analytics, agencies can see immediately what effect their street redesigns and outreach changes have as they’re carried out. In addition, Loewenherz explains that “while [agencies] implement a treatment, they see what approach gives them more bang for the buck” in reducing the number of traffic conflicts, such as choosing between engineering changes or enforcement tactics.
Using protected bike lanes as an example, Loewenherz explained that anecdotal feedback from riders could suggest positive results, but there can be scant data to support that. Creating a program to constantly count how users and behaviors shift after a street design has been changed provides evidence-based insight into the lane’s effects, from which planners can learn.
Loewenherz expects this program will develop a wealth of collective knowledge for participating agencies by creating scalable, accessible, and actionable data from multiple jurisdictions. Ultimately, the data should help cities better learn from each other and better understand how they can eliminate dangerous traffic situations.
Part of the broader solution
“We can’t achieve Vision Zero without this,” Loewenherz claims. Without a proactive model for analyzing their efforts, cities can’t get ahead of dangerous traffic situations, and therefore can only address problems after it’s too late for some people.
“We have a shared responsibility to address this,” Loewenherz explains. Governments, businesses, and citizens are coming together in a concerted effort to prevent deaths and injuries before they happen, and new insights from the video analysis tool could be a crucial step toward achieving Vision Zero.
To try your hand at teaching the analytics program, click here.
Photo, top: A mockup of data generated through the program (courtesy of Video Analytics Towards Vision Zero).