Duke University has pipped companies such as Google and Microsoft to the post to develop an indoor GPS system that can locate users as accurately as 1.6 meters.
It could have a host of applications, says Romit Roy Choudhury, associate professor of computer engineering at Duke, for example allowing shoppers at stores to use their phones to learn more about the products in front of them, or helping parents to locate their children in malls. It could also help navigation in hospitals or smart homes.
The system, dubbed UnLoc, short for unsupervised indoor localization, is based on the way people use landmarks in outdoor environments. If Alice is in Bob’s neighborhood and calls him for directions, she can tell Bob she is standing in front of a fountain, and he can guide her from there.
“Our technique adopts the same intuition – it takes advantage of ‘invisible’ landmarks in indoor environments that a mobile phone can sense using its built-in sensors,” says Choudhury.
“Example landmarks could be distinct motion signatures created by elevators or stairwells, because the phone can detect motion, or certain dead spots where WiFi or 3G signals are absent.”
Once such an invisible landmark is sensed, the phone can work out its current location, and then track its forward path using motion sensors such as accelerometers, compasses and gyroscopes.
While the tracking can become inaccurate over time, the phone can continuously correct its location as it discovers other landmarks.
“The best part of the application is that it is recursive, which means that it starts with zero knowledge but ‘learns’ over time,” says He Wang, lead PhD student on the project.
“Therefore, it becomes more and more accurate the more it is used in a given building.”
Crucially, UnLoc doesn’t require any pre-deployment effort, or ‘wardriving’, whereby every location needs to be visited and calibrated to create a database of per-location fingerprints. This is expensive even as a one-off, and often needs to be repeated periodically.
It’s also less battery-hungry than GPS, thanks to the use of energy-efficient inertial sensors, available in almost all smartphones.
In tests at the Northgate shopping mall in Durham, NC, and on Duke’s campus, UnLoc achieved 1.6 meter accuracy on average.
“Further improvement is feasible by learning more landmarks and improving the tracking algorithms,” says Duke undergraduate Alex Mariakakis.
