MIT improves method of making road maps based on satellite photos

MIT scientists have developed the RoadTracer system, using neural networks to step by step to map them. This would be considerably more accurate than the existing methods.

RoadTracer takes a location on a satellite photo that is certain to be a road and uses neural networks to analyze the environment and thus determine which point is most likely to next part of the road. This is different from the existing way, in which neural networks are also trained, but in order to determine from the analysis of the pixels whether segments are a road or not.
Incorrect assumptions can be strengthened and parts where the roads have been removed from the eye, as by trees, are difficult to recognize as such. “RoadTracer focuses on the simpler problem of figuring out which direction to follow from a starting point that we know is a road, adding a point and repeating it gradually, step by step, establishing the road network”, Fayven Bastani, who is a student at MIT’s Computer Science and Artificial Intelligence Laboratory and will present his team’s research in June.
Bastani and his team trained RoadTracer on satellite images of twenty-five cities in six countries in North America and Europe. The operation was then applied for evaluation to images of fifteen other cities, where it was not trained. The system was able to accurately map 44 percent of the crossings in New York City, while the method based on segmentation allowed for only 19 percent. Overall, RoadTracer’s margin of error is 45 percent lower than that of existing systems.
According to Bastani, RoadTracer makes it easier to automate company floor plans, although a human check still needs to be done. According to him, their work could also be more efficient with the system, because they can restart the tracing after correction from the point of the error.

 

Loading...