Slashdot Mirror
New Navigation App 'Live Roads' Promises 1.5m-Accuracy With Standard Cellphone Hardware (arstechnica.com)
Jonathan M. Gitlin from Ars Technica reviews a new navigation app called Live Roads, which promises 1.5m-accuracy via your current smartphone without the need of any extra hardware. In a nutshell, the app provides more accurate mapping/navigation than what's currently available via Google Maps or Apple Maps, but it's still not quite as accurate as a true "HD map." HD maps are accurate to within a centimeter or two and are usually made by a combination of traditional surveying and lidar scanning. Here's an excerpt from the report: A few weeks after talking with the company, I was delivered a Samsung S7 loaded with Live Roads. I'll be honest: I'm not that familiar with Android, and this isn't really a review of the app. I used it enough to check that it does what it claims, but I didn't use it as my sole method of navigation. However, this brief bit of user-testing did let me check out the claims in that email. I don't think I'd equate the app with the HD maps that autonomous vehicles will need. For one thing it's readable by a human being; for another it's not quite that accurate. But the spatial resolution was indeed better than it should be on a consumer phone, and Live Roads was able to locate me down to a specific lane on a multi-lane road. Various navigation apps give you lane-specific instructions -- for instance, telling you to stay in the middle two lanes if you're approaching a complicated intersection. Where Live Roads differs is that it can also tell which lane you're actually in. Whether this is enough of a feature to build a business model around is an open question; I'm quite happy using Google Maps on iOS, with occasional forays into Waze (running in the background to warn of speed traps) and Apple Maps (if I'm driving something with CarPlay and the infotainment's built-in navigation sucks).
But it left me wondering: how does it work? Paul Konieczny, CEO of Live Roads, gave me an explanation -- up to a point. "Primarily it is based around sensor fusion and certain probabilistic models -- we call it the Black Box," he said. "The current release of the app that is available in the Play Store has an earlier revision of our Black Box. This initial version is missing some of the functionality of the full-fledged system and thus has a spatial resolution of ~2.5m. This compares favorably to standard GPS that has a resolution of 4.0 m+." By summer, Konieczny hopes that the system will be fully operational and that accuracy will be down to under 1.5m. Assuming a large enough user base, that should let it offer lane-specific traffic data, "as well as introducing an entire ecosystem of 3D objects that users will be able to interact with," he told me.
But it left me wondering: how does it work? Paul Konieczny, CEO of Live Roads, gave me an explanation -- up to a point. "Primarily it is based around sensor fusion and certain probabilistic models -- we call it the Black Box," he said. "The current release of the app that is available in the Play Store has an earlier revision of our Black Box. This initial version is missing some of the functionality of the full-fledged system and thus has a spatial resolution of ~2.5m. This compares favorably to standard GPS that has a resolution of 4.0 m+." By summer, Konieczny hopes that the system will be fully operational and that accuracy will be down to under 1.5m. Assuming a large enough user base, that should let it offer lane-specific traffic data, "as well as introducing an entire ecosystem of 3D objects that users will be able to interact with," he told me.
You can get a better GPS fix by having a base station at a surveyed point, but this isn't what they're doing, because it would be expensive. There are also corrections you can do based upon atmospheric conditions, but they say that wasn't good enough. They pretty obviously don't want to say how they're doing it, which makes sense for trying to corner the market. But I wonder how novel their methods are, and if they can really stay ahead for long.
They say it's by "sensor fusion" and probabilistic models. The sensors are presumably the accelerometer and maybe the barometer in a cell phone. They could have an idea of the nominal error of a sensor based upon the model of phone. The probabilistic models, I can only guess, come from having many users running the app at once so they can try to reduce the error using many measurements.
Once they narrow the precision of the fix to the size of a lane of traffic, they could start identifying the lane by the behavior of clients in that lane -- if everyone turns off the freeway at this point in this lane, it's probably an exit lane, that sort of thing. And those rules could become much more complex. Like, use a machine learning process and train it on a bunch of well-defined traffic situations.
But, two questions: first, can they compete with companies like Garmin and Google who have tons of money and clients, plus smart people who could try to reverse-engineer these improvements and roll them out basically for free? It could easily be worth it to crush competition. And, if the success of the probabilistic model is based upon having many clients, do they get this level of improvement everywhere, or only in higher-traffic areas like cities?