3 ms·

they're fingerprinting locations based on rssi from devices in the signal environment. meaning your phone tries to connect, gets a set of rssis back and that th

they're fingerprinting locations based on rssi from devices in the signal environment. meaning your phone tries to connect, gets a set of rssis back and that then finger prints where you're standing with the phone (by the collection of rssis). the "space" it's locating in isn't physical space but "rssi tuple" space.

konschubert 9y ago
So, how does it get mapped to physical space then? As long as I am in "rssi tuple" space, I maybe know if I am in a place where I have been before, and I can learn the topology of the place in terms of "which places are adjacent". But I can't put my location on a map. How do they do the mapping?
- mlevental 9y ago
  >So, how does it get mapped to physical space then? don't quote me but i don't think it does. you basically go around and set up hotspots in rssi space that the server knows about. when other people enter that hotspot things happen. this isn't as silly as it sounds because for a lot of applications you don't need to know where you are in physical space but just what you're near (i.e. i put my tv somewhere, hotspot it, then when my phone gets near that hotspot it's near the tv).
  - coupdejarnac 9y ago
    Yeah, that makes sense. But that's not going to be very accurate. That will be accurate at the tuple space, but not down to the nearest centimeter in X Y coordinates or anything like that. Plus, if there is any mass between the marker and the tracked device, that will throw off readings. RSSI measurements are really only good for open air measurements or if you have characterized all the materials the rf passes through.
    - mlevental 9y ago
      >lus, if there is any mass between the marker and the tracked device there are no markers so there is no attenuation happening in that way.
    - nonis 9y ago
      Yeah, BLE beacons aren't particularly accurate for an IPS. I worked on this problem a few years ago at FB, for the purposes of helping employees find themselves + meeting rooms on a mobile app. We really only needed 2-3 meter accuracy, so people could figure out approximately where they are. BLE signals are crazy noisy, so you need to tack on a lot of filtering (kalman filter, etc) and it really helps if you have multiple data sources (wifi signals, gyroscope, RF, etc). We could also kinda cheat by adding a track to the center of each hallway on the map, then magnet users towards the track.
      - qrv3w 9y ago
        That sounds very similar to this! Did FB ever open-source it?
        nonis 9y ago
        Unfortunately no. Last I worked on it it was pretty heavily integrated with their mobile maps application (https://code.facebook.com/posts/504406169737284/instrumenting-meetings-at-facebook/ https://code.facebook.com/posts/504406169737284/instrumentin...) and there wasn't any plan on open sourcing.
      - ghaff 9y ago
        There was a lot of interest in BLE beacons a couple of years back for retail displays and the like. But as you say beacons don’t really work all that well, people had to install apps, etc. and the whole space seems to have sort of fizzled. The wireless beacons also don’t really have good enough battery life for a lot of applications. I played around with some of this stuff but it wasn’t good enough for anything practical that I was interested in. I wish there was a way to do better phone tracking but it’s hard to come up with a system that would be fast and reliable if you left your phone down somewhere while not driving you crazy with false alarms on a day to day basis.
    - qrv3w 9y ago
      > That will be accurate at the tuple space, but not down to the nearest centimeter Yes, you're right. Sorry, I didn't mean to imply centimeter-level precision anywhere. I said "high precision" because I compare it to GPS and not RTT (which was not common when I started this project). Unlike FIND, any nearest-centimeter precision requires specialized hardware (to be available on Android P in the future! [1]). The FIND system fills a niche where you might need room-level or sub-room level precision without having to install anything except an app on your phone or computer. > any mass between the marker and the tracked device will throw off readings Its true that individual readings will be affected by varying obstacles (doors/people), and this will throw off location classification if you have very few signal generating devices in the vicinity. However, generally places will see Bluetooth/WiFi coming from their neighbors in all directions so it would be hard to attenuate all signals simultaneously. The machine learning is pretty robust too, so if one of the readings get thrown off because there are still several others that can compensate. [1]: https://developer.android.com/preview/features.html#rtt https://developer.android.com/preview/features.html#rtt
      - michaelt 9y ago
        nearest-centimeter precision requires specialized hardware (to be available on Android P in the future! [1]). Sadly, that link says "The result is typically accurate within 1 to 2 meters" so it's nearest meter rather than nearest centimetre.
- MrQuincle 9y ago
  This is kind of the resolution you can get with fingerprinting: https://youtu.be/ZtnYH53A4zE https://youtu.be/ZtnYH53A4zE But it all depends on the density of nodes. If you are also allowed to use odometry on the phone you can suddenly go up in resolution quite a bit. In that case you'll have a conventional SLAM problem from robotics, which can be considered more or less solved.

konschubert 9y ago

So, how does it get mapped to physical space then? As long as I am in "rssi tuple" space, I maybe know if I am in a place where I have been before, and I can learn the topology of the place in terms of "which places are adjacent". But I can't put my location on a map. How do they do the mapping?

mlevental 9y ago
>So, how does it get mapped to physical space then? don't quote me but i don't think it does. you basically go around and set up hotspots in rssi space that the server knows about. when other people enter that hotspot things happen. this isn't as silly as it sounds because for a lot of applications you don't need to know where you are in physical space but just what you're near (i.e. i put my tv somewhere, hotspot it, then when my phone gets near that hotspot it's near the tv).
- coupdejarnac 9y ago
  Yeah, that makes sense. But that's not going to be very accurate. That will be accurate at the tuple space, but not down to the nearest centimeter in X Y coordinates or anything like that. Plus, if there is any mass between the marker and the tracked device, that will throw off readings. RSSI measurements are really only good for open air measurements or if you have characterized all the materials the rf passes through.
  - mlevental 9y ago
    >lus, if there is any mass between the marker and the tracked device there are no markers so there is no attenuation happening in that way.
  - nonis 9y ago
    Yeah, BLE beacons aren't particularly accurate for an IPS. I worked on this problem a few years ago at FB, for the purposes of helping employees find themselves + meeting rooms on a mobile app. We really only needed 2-3 meter accuracy, so people could figure out approximately where they are. BLE signals are crazy noisy, so you need to tack on a lot of filtering (kalman filter, etc) and it really helps if you have multiple data sources (wifi signals, gyroscope, RF, etc). We could also kinda cheat by adding a track to the center of each hallway on the map, then magnet users towards the track.
    - qrv3w 9y ago
      That sounds very similar to this! Did FB ever open-source it?
      - nonis 9y ago
        Unfortunately no. Last I worked on it it was pretty heavily integrated with their mobile maps application (https://code.facebook.com/posts/504406169737284/instrumenting-meetings-at-facebook/ https://code.facebook.com/posts/504406169737284/instrumentin...) and there wasn't any plan on open sourcing.
    - ghaff 9y ago
      There was a lot of interest in BLE beacons a couple of years back for retail displays and the like. But as you say beacons don’t really work all that well, people had to install apps, etc. and the whole space seems to have sort of fizzled. The wireless beacons also don’t really have good enough battery life for a lot of applications. I played around with some of this stuff but it wasn’t good enough for anything practical that I was interested in. I wish there was a way to do better phone tracking but it’s hard to come up with a system that would be fast and reliable if you left your phone down somewhere while not driving you crazy with false alarms on a day to day basis.
  - qrv3w 9y ago
    > That will be accurate at the tuple space, but not down to the nearest centimeter Yes, you're right. Sorry, I didn't mean to imply centimeter-level precision anywhere. I said "high precision" because I compare it to GPS and not RTT (which was not common when I started this project). Unlike FIND, any nearest-centimeter precision requires specialized hardware (to be available on Android P in the future! [1]). The FIND system fills a niche where you might need room-level or sub-room level precision without having to install anything except an app on your phone or computer. > any mass between the marker and the tracked device will throw off readings Its true that individual readings will be affected by varying obstacles (doors/people), and this will throw off location classification if you have very few signal generating devices in the vicinity. However, generally places will see Bluetooth/WiFi coming from their neighbors in all directions so it would be hard to attenuate all signals simultaneously. The machine learning is pretty robust too, so if one of the readings get thrown off because there are still several others that can compensate. [1]: https://developer.android.com/preview/features.html#rtt https://developer.android.com/preview/features.html#rtt
    - michaelt 9y ago
      nearest-centimeter precision requires specialized hardware (to be available on Android P in the future! [1]). Sadly, that link says "The result is typically accurate within 1 to 2 meters" so it's nearest meter rather than nearest centimetre.
MrQuincle 9y ago
This is kind of the resolution you can get with fingerprinting: https://youtu.be/ZtnYH53A4zE https://youtu.be/ZtnYH53A4zE But it all depends on the density of nodes. If you are also allowed to use odometry on the phone you can suddenly go up in resolution quite a bit. In that case you'll have a conventional SLAM problem from robotics, which can be considered more or less solved.