These Indian-origin researchers use satellite photos, AI to map crosswalks and fill gaps
As cities look to encourage people to switch from driving to greener forms of transportation, humble sidewalks and crosswalks play an important role. Walking, after all, is the original carbon-free transportation.
But in some cases, city officials lack money to invest in sidewalks or haven’t studied their pedestrian infrastructure to the same extent as streets and roads. Gathering data on sidewalks and crosswalks has historically been a labor-intensive job – taking months, in some cases – but transportation researchers are increasingly looking at ways for satellite imagery and artificial intelligence to speed the work.
Satish Ukkusuri, a professor of civil engineering at Purdue University, and Rajat Verma, a graduate researcher, recently developed a model to rapidly identify crosswalks in a city. They talked to The Washington Post about their research and how examining crosswalks could help urban leaders create more complete networks for pedestrians.
The Post: How well do cities understand their pedestrian infrastructure compared to highways?
Satish Ukkusuri: Typically, we have an overinvestment in highways, in terms of the driving infrastructure for cars and trucks, which is definitely the fuel of our economy. But cities – especially if we want to make cities livable and sustainable – we need very good sustainable transportation infrastructure, and that primarily comes from pedestrians. We have seen this in covid. People have really wanted to use green spaces more. People started biking more. But cities are constrained in terms of how much they actually can provide in terms of pedestrian infrastructure. So this project is really intended to fill in the gap by first identifying where are we in terms of looking at our pedestrian infrastructure in cities.
A city which has very good pedestrian infrastructure, as can be seen in both Asian cities and European cities, also leads to better health outcomes. It also leads to more economic opportunities for people to participate in and a better social environment for people to participate in a lot of activities. But U.S. cities are not very convenient for pedestrians. You can see this in the increasing number of pedestrian accidents that are currently happening, post-covid. While there are many reasons for why these type of accidents are happening, one key cause is the lack of connected pedestrian infrastructure that really complements what we have on the road networks.
The Post: What has traditionally been the approach to gathering data on sidewalks and crosswalks?
Rajat Verma: Traditionally, the MPOs, the metropolitan planning organizations, they either use existing inventories or more commonly what we’ve seen is this concept of walking audits, where individuals go to that target region and then they check the quality of existing infrastructure and locate the points that they find are suitable for development. Then they digitize all of these manually collected data, and then they go about identifying candidate segments or candidate places where they can make their investments. As you might imagine, this is a pretty laborious task. They did this in 59 European cities in 2021. They reported that it took almost 12 months for even trained analysts to go over each feature and code them. This is where we found the opportunity to exploit our methods and data.
The Post: You used a combination of satellite imagery and artificial intelligence. Walk me through what you did and how it works.
Verma: The study is broken down into two main parts. The first part is identifying crosswalks with the help of satellite images. And the second part is, once we have identified the crosswalks, how do we connect it to more traditional analysis that analysts do, typically in planning agencies. The first part is the heavy work, the AI-related work: how we can use images, how we can use different deep-learning methods.
Deep-learning is something that has been quite a buzzword. It’s actually proven to be one of the most very, very effective methods in computer vision. Deep-learning has almost completely revolutionized the field of computer vision because you can just pour humongous amounts of data – in this case, you can put thousands or millions of images – and the model will automatically identify the important pieces of information just from the data itself. You don’t even need to specify that a crosswalk has certain features.
If we look at Facebook auto-tagging of pictures, or if you put your picture on Google Photos, it’ll determine that it’s you or it’s your dog or a family member. This is known as “object detection,” and this has improved by leaps and bounds in the recent years.
The Post: How did you use data from the District to test your approach?
Verma: When we try to prove the effectiveness or efficacy of an algorithm, we also need to show that it’s effective on a true data set. In this case, we assumed that whatever the government of Washington, D.C., provided is correct. And on the basis of that, we showed that this is around 93 percent to 100 percent effective in counting the crosswalks.
Ukkusuri: One of the challenges with object detection and machine-vision algorithms, when you fine-tune them for these types of problems, is the scale. When you give a few images or you’re focused on just one roadway, it’s much simpler to identify crosswalks. But when you want to do this at the scale of an entire city, the problem just quickly becomes very intractable. For the Department of Transportation in Washington, D.C., they’re making investments at the network level, so they’re doing it for the entire city.
We need to address the problem of scale so that, one, we can look at the network aspect of the problem and we can identify the gaps that are there in pedestrian infrastructure. Two, we also need to come up with transferability of these solutions. We need to be able to develop it for a city, and then we should be able to take it to any other city, and we should be able to come up with solutions with minimum amounts of training. We also need to be careful that we are validating these methods with what’s really happening on the ground, which is why we need ground-truth information, which in this case is provided by the D.C. Department of Transportation.
The Post: You talked about other methods taking months potentially to gather this data. How long does it take you now?
Verma: It’s very interesting because once we had trained the model, it took us around 10 to 15 minutes to detect all the crosswalks in Los Angeles and D.C., combined.
The Post: You talk about pedestrian network completion as being an application here. What is that?
Ukkusuri: When we say the network level of connectivity, we want people to be able to walk from Point A to Point B without really having to be watching out for whether they’re going to get hit by a car. Or you’ll see long stretches where there are walking deserts. I mean, there’s really no infrastructure there. When cities make these types of decisions, they’re not really looking at that. I mean, they’re getting pressure from their local borough president, for example, and then they just build that infrastructure. We really have not seen whether this crosswalk is going to provide access to other parts of the city. So this global view, this network view of this pedestrian infrastructure, is something which is really important for us to do.
Sometimes the shortest path is just to use existing facilities that are there, and those facilities could be parks and green spaces that are there. So without having to necessarily have sidewalks or crosswalks everywhere, you can use existing infrastructure that’s already in place and make minor modifications to make them more pedestrian-friendly.
There are many planning organizations which do not have the wherewithal to invest in this pedestrian type of study. I think that’s where we can provide a value.