Better-performing mapping system searches the streets
When it comes to making city maps based on aerial photos, manually tracing all the roads can be quite the hassle. As a result, we're now seeing computer programs that do so automatically. Scientists at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed a program of their own, that is promised to be even better at the job.
Ordinarily, mapping programs use neural networks that analyze an entire aerial photo all at once, deciding which pixels represent roads and which ones do not. It's not a highly accurate system, however, as it can be thrown off by obstructions such as trees and shadows. This can result in gaps in the map, that have to be manually corrected in a post-processing step.
MIT's RoadTracer system, however, works differently.
It begins at a known location on the road network within a photo, then uses a neural network to gradually work its way out from there, determining which point is most likely to be the next one in the road. That point gets added to the map, and then the system continues to search out from it looking for the next point, building the map up in a step-by-step process.
RoadTracer was trained on aerial images of 25 cities across six countries in North America and Europe, and was then trialled on a further 15 cities that it hadn't trained on. In the case of New York City, it was able to correctly map 44 percent of the road junctions, which was over twice as good as the 19 percent managed by traditional computer systems.
"Rather than making thousands of different decisions at once about whether various pixels represent parts of a road, RoadTracer focuses on the simpler problem of figuring out which direction to follow when starting from a particular spot that we know is a road," says grad student Fayven Bastani. "This is in many ways actually a lot closer to how we as humans construct mental models of the world around us."
The technology can be seen in use, in the following video.