Mapping the roads manually across the globe can take almost forever by just taking aerial images, even for companies with vast resources like Google. That is why existing mapping apps such as Google Maps have yet to completely trace out the Earth's more than 20 million miles of roads.
Thankfully, researchers from the Massachusetts Institute of Technology's Computer Science and Artificial Intelligence Laboratory (CSAIL) aim to fill the gap in maps by creating RoadTracer, an AI-based system for mapping the roads with 45% more accuracy than existing methods.
It's important to note that current approaches to building roadmaps rely on image segmentation, a method of training neural networks to analyze and label individual pixels on aerial images as either “road” or “not road.” However, this approach is prone to errors due to visual obstructions like trees, buildings, and shadows that could get in the way of accurately identifying the starting and end points of the roads.
To minimize the level of imprecision, the segmentation approach uses some post-processing methods to link two road segments, though this is limited to the assumption that two roads are connected just because of their close proximity.
The automated system developed by MIT still uses aerial photos, but instead of partitioning a digital image into multiple sets of pixels, RoadTracer identifies a known starting point on the road network and analyzes the surrounding area using a neural network to pinpoint the most likely next part on the road. The automated system then factors in the previous points it created and performs the entire process all over again in order to create the road network.
Fayven Bastani, a graduate student at MIT and a co-author of the study, explains:
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. This is in many ways actually a lot closer to how we as humans construct mental models of the world around us.
The team consists of researchers from CSAIL and the Qatar Computing Research Institute. The group worked to train RoadTracer using the aerial images of 25 cities across six countries in North America and Europe before evaluating its abilities to map out the roads across 15 other cities.
Bastani adds that RoadTracer's 45-percent lower rate of error makes automatic mapping systems more practical for companies like Google that develop mapping solutions. The team also claims the AI-based approach to mapping is also more cost-effective than existing methods.
MIT professor Mohammad Alizadeh, one of the study's co-authors, says the automatic mapping system can also significantly help smaller organizations with limited resources remove errors from maps. He adds:
RoadTracer is well-suited to map areas of the world where maps are frequently out of date, which includes both places with lower population and areas where there’s frequent construction. For example, existing maps for remote areas like rural Thailand are missing many roads. RoadTracer could help make them more accurate.
That means humans aren't completely removed from the mapping process. Instead, the team envisions the system recommending roadmaps for a specific region and then letting human supervisors correct the errors more easily. The goal, according to Alizadeh, is to reduce the load of work human experts need to perform in curating and refining the maps.
The team will present their research paper in June at the Conference on Computer Vision and Pattern Recognition in Salt Lake City, Utah.
Source: MIT News