Cities across America will soon be able to easily map their pedestrian infrastructure using the power of artificial intelligence — and challenge them to have a long overdue conversation about why those networks have such an acceptable number of gaps.
This spring, an international coalition of urban studies and computer planning researchers has announced that they will debut Tile2Net, which they're calling "the first open-source scene classification model for pedestrian infrastructure from sub-meter resolution aerial tiles" — or, in layman's terms, the first free tool that uses machine learning technology to scan satellite imagery for the presence of sidewalks and crosswalks.
Named after the aerial "tiles" into which satellite maps like Google Maps are broken, the tool was successfully able to recognize more than 90 percent of those roadway features in cities like Boston and Cambridge, when compared against maps their governments had developed through conventional methods. Most notably, that included sidewalks and crosswalks that were obscured by tree cover, shadows from nearby buildings, and other barriers with which previous machine learning systems have struggled — and the tool can do it for virtually any city for which orthorectified aerial imagery is available.
As Tile2Net's hit rate improves, researchers hope it could revolutionize the way that municipalities document pedestrian infrastructure — because right now, many of them don't do it at all.
"Even in this day and age when so many things are digitalized and automated, many cities still rely on manual auditing and manual tracing [of their pedestrian infrastructure,]" said Maryam Hosseini, a post-doctoral researcher at the Massachusetts Institute of Technology and the lead author on the paper. "It’s time-consuming, it's labor intensive, and often, people just forgo it."
Hosseini added that by failing to even locate where their sidewalks and crosswalks are, cities are ill-equipped to take action to get ones that are broken fixed, never mind creating new pedestrian connections where none currently exist.
Because sidewalks are supposed to be maintained by property owners in most U.S. communities, cities often don't devote resources to inspecting them or scheduling regular, taxpayer-funded repairs. That leaves the burden on citizens to manually report when critical pavement is missing or falls into poor condition, followed by an often lengthy bureaucratic process to either fine the property owner into compliance or come up with the resources for transportation authorities to fix sidewalks themselves.
That inefficient process has particularly dire consequences for people with disabilities, many of whom are forced to painstakingly scrutinize Google Street View images for pixelated photos of wheelchair ramps and complete pavements before they visit unfamiliar neighborhoods — only to possibly find when they arrive that those photos were outdated, and the streets aren't actually accessible to them at all.
Maps of driving infrastructure, meanwhile, are meticulously maintained by both the public and private sectors, the latter of which has sunk billions of dollars into technology that keeps drivers apprised of up-to-the-minute road conditions that might affect their commute times. The autonomous vehicle industry, meanwhile, has devoted its own billions to helping cars recognize crosswalks — but not to helping cities recognize when crosswalks are missing.
"A lot of it is just how we prioritize our resources," Hosseini added. "We claim it’s easier to map road networks because you can just have a car with a camera on top of it driving around. But I believe that if we'd been actually looking for the solution, we could have found it much earlier than this."
Hosseni is careful to note that Tile2Net isn't a perfect replacement for conventional sidewalk mapping methods, even if it will be a big improvement to for cities who currently do no mapping at all.
Because it relies on aerial imagery, the algorithm can't yet identify "overground foot-bridges, under-ground pedestrian crossings, covered pathways, and public pathways inside buildings," nor can it see granular but critical details, like when a property owner hasn't plowed snow off their sidewalk in weeks, or when a tree root growing underneath the pavement is creating a trip hazard. It's also not designed to temporary obstructions, like large crowds of pedestrians exiting a sports game, and again, it's only 90 percent accurate when compared with on-the-ground models, and it will take more time and data to train it more fully.
Even so, Hosseini hopes that if nothing else, Tile2Net will help advocates rapidly highlight where their cities' walking networks are falling short and show their transportation leaders how those failures correlate with traffic violence, lost mobility, poor neighborhood health outcomes, and all the other negative impacts that follow when pedestrians are neglected.
"Nothing is perfect, nothing is 100 percent accurate," she adds. "There will be places that will have some uncertainties. But now, we’ve filtered it down to only a few points that [transportation officials] have to go out and manually check — and it’s much better to do it before people start having problems, before someone gets hurt."
