Indoor Location Detection using Wifi | Marko Tisler | WLPC EU Budapest 2016

[Music] [Music] so my name is Marco - Claire I work for aero hive networks and today I'd like to talk about indoor location detection using Wi-Fi more particularly about the accuracy the expectations and what actually gets delivered because there's a lot of misconceptions about the accuracy both in time and in locate in terms of location so what I'm talking about is the customer who wants to deploy Wi-Fi location detection it expects GPS like experience without you know a scale of a city or a country but a scale of a floor or maybe a large hole and so basically there are expecting hey you're here and you're here now so what's actually happening is we're saying you're sort of somewhere here in the building that's the actual experience so you're sort of somewhere here in the in the room and we're not going to put our hand in the in the fire and say hey we guarantee you that you're actually here right so that's how Wi-Fi location works work with a certain accuracy we're sort of certain probability we're saying you're probably here and so once you start this honest discussion with a customer it gets tricky so the way it works is you go to your prospect so your customer let's say if it's a retailer you are there to present your location analytics solution and they ask you ok so if I understand correctly this is what you can do if my customer stands in front of a shelf with coffee you'll be able to tell me whether they are looking at coffee or chocolate on this side and you say no no no it's not what we do you explain it's it cannot detect which way the customer is facing etc they say ok that's fine it's fine I know I was too optimistic so but you what you can do is here's one shelf there not a shelf behind it so you'll be able to tell me whether the customer is looking at the coffee or at milk at the other health behind it and that's probably like two meters apart or me or a meter apart and you say not quite because you know there's a tricky thing about Wi-Fi it's actually not that accurate and then you get to a point where the customer or the prospect says okay so what you can do again you say yours we can tell you that the customers are mostly sort of here at this shelf not at the other shelf and it then it gets hard to sell the next question you know you have is okay so let's say you sort of positioned the accuracy bit then you start talking to the developers and they say okay so where is this information available when can I actually see where a certain device or a certain user is is it real time is it not real time is it sort of real time and your answer is going to be depends it depends how accurate you want your positions to be it depends whether your client is associated or not and because most developers don't have any Wi-Fi concept covered they say what and then you start saying ok so if you hear the thing if your client is not associated will we're talking about 30 to 60 maybe 2 minutes 2 minutes delay so 4 into 60 seconds or 2 minutes delay before you get your information because here's the thing the client has to probe and we have to catch that probe in order in order to actually do any triangulation okay then developer says so it's not real time yes well it's sort of real time and you get into this awkward conversation again then you get into the conversation okay how about history well I can it depends right are you storing the history are you processing the history so and from the perspective of a developer what I need to do one API call I need my location for client X and now okay so again is this misconception and this expectation versus reality paradox so when you come in this is what your customer things are doing you are the expert for location analytics you are very accurate everything is really straight to the point so like architecture what you're actually doing is this you're navigating a stormy sea trying to find that client as soon as possible RF is messy we all know Wi-Fi is messy as well and here's going back to something that Keith said is it's hard to make Wi-Fi work really good it's even harder to do Wi-Fi location really good it's very easy to do Wi-Fi location very bad yeah so for example this is from a personal experience I had a project where it was in an emergency room of the hospital and they said okay so we have this operating rooms and we want to know exactly in real-time when somebody is in a certain room and the rooms are basically like five meters across one next to each other well you can do that with Wi-Fi Kenya so we had to include some other technologies to be able to do that but their expectations was hey Wi-Fi can do that no sir and especially in an ER people died there so you don't really want to mess with hey we're sort of doing this right now it has to be it's either right or it's wrong the reason behind us is we are still relying on our SS based methods so we've seen something similar to this graph yesterday and this is the relationship between the RSSI in DBM versus distance in meters and this is a theoretical model it only works in a desert free of any noise and only in perfect perfect perfect conditions and you can actually produce this relationship for any frequency you want and there's the equation so we are able to calculate the distance from the transmitter so between the transmitter and the receiver based on that equation and we are able to evaluate the distance if we do that for at least three devices we get the possibility of doing triangulation so we're able to say okay so if I have three distances I'll be able to get X&Y only in two dimensions X and Y for that specific device cool and it turns out you have to solve this system of equations which is pretty simple once you have the measurements and this is what we call two alliteration problem number one the further away you get from an AP and this is what we saw yesterday I think Carl presented this the further away you go from the AP the same drop in RSSI means a larger change in distance so when your artists eye changes by 3 DB here that means let's say I don't know 6 meters when it changes by 3 DB here that is 10 or 20 meters which means your error grows the further away you go from the AP which means this sort of procedure works sort of okay if you're near all the APS and it starts degrading as soon as you move away it's actually even worse than that so we all know that RSSI in real life doesn't behave in a linear fashion it jumps around like crazy so instead of having this smooth smooth hockey stick we have this jagged weird approximate even approximation we have this quasi random behavior of RSS RSS I so in reality what happens is so at first it behaves okay for a couple of meters then it starts jumping up sharply presenting you to be closer than you actually are then it drops sharply making it look like you're further away than you actually are and something interesting happens once you get really far away so let's say 15 20 meters away artists I sometimes jumps up making you again appear close to the AP now try to explain that to your customers so what happens in reality is the theory said your position somewhere around here so behind this wall in that office according to that RSSI model because of several factors first of all absorption of wolves and other materials in the room the artists eye is lower than according to that model so in reality you're actually closer to the AP how do you how do you work with that and one of my experiences was I was helping a startup that was sort of pitching retails in inside retail analytics to customers make sense out of all of this messy messy stuff the startup felts on here and I am able to talk about it first thing that we can do is we can do our fingerprinting our fingerprinting means okay so let's let's throw the theoretical mother away because it's obvious not good and let's try to make real measurements of actual environment so we go around the floorplan taking measurements of our SSI from different devices and put that in a some sort of database and then once we take actual measurements from environment we compare that to the measurements we've made and it's called what are a fingerprinting it takes a lot of time one of my colleagues called this the Wi-Fi dance because what you need to do is you need to take one of your clients you walk around and take RSSI measurements now the problem is different clients will make different measurements you are doing the measurements with your MacBook but the people coming in to the store are actually using Android smartphones HD HTC Samsung whatever so but you are sort of accounting for non-line of sight situations material observe absorption etc it does sort of improve location detection however there there is a lot of effort involved especially if this is a large venue and even if you do this we all know that RF is and will always remain a dynamic environment so even if you go through all this effort two things are going to change so one thing is sooner or later somebody's going to change the layout or an AP is gonna stop working or it's just the position of the device so when you did the measurements you are holding the device horizontally when somebody is using the Wi-Fi he's using a device like this and the are the RSSI just goes like this so this doesn't really help you that much and it comes at the cost of doing a lot of work okay so we live in 21st century we have this cool thing called data science let's get some big data scientists or data engineers and let's take this messy data and try to filter it out see if that will give us a better accuracy so we take this large amount of messy data throw away the lowers and our SSI measurements because we saw really correspond to our model that well and then maybe we take into account the fourth or the fifth strongest signal to some sort of multi literation but you soon find out it this is very costly in terms of resources and it's still an assumption so you do a lot of post-processing you pay the price in terms of CPU memory probably data storage as well because this sort of things have thousands and thousands of clients right this is not a deployment with 50 clients for example if you look a at a deployment like Dubai Mall or something like that this is thousands and thousands of clients all the time you're taking measurements all the time and you're putting that if you're smart in a non-sequel database if you're not so smart you're putting it in a sequel database and see what happens one month later and then you do post-processing yeah and now we come back to the question when has the data available well once we post process it when is that going to be in an hour two hours okay so real the real time stuff just goes down so it's really problematic there are other things we can do we can cheat we can say okay we know enough about the actual environment so that we are pretty sure that some locations are invalid for clients so we can say okay if a location is invalid snap it to a valid to the nearest valid location now the problem is you start with an error how do you know you're stepping the client took your application maybe you're just introducing more error or a popular popular method seems to be snapped to line you draw lines across your floor mat floor plan and you try to snap a client to that path now what if the client is outside your location and they're being snapped to your path contributing to the statistics that you want or your customer wants and making it false again so basically what you're doing any sort of alteration to that data just skews it further there's just more and more fallacy in that data making it more and more unusable obviously it look is good to the customer right so suddenly have this nice lines the customer says ok now I know where my clients are moving I know where my customers are moving I'm happy because this is what you wanted to sell the first time but if you actually went in and check every single data set that you have you see that a lot of it is clients coming out from the outside of the environment just invalid measurements probably even clients that were snapped from a valid invalid location to a path or things like that so ok this doesn't work so there is another approach let's do let's do instead of counting on x and y-coordinates let's try to do something like a path analysis so let's forget about x and y let's just try to guess on average where do people move or where well do people dwell so on average that may be correct it may be not correct it's really hard to prove whether hey whether it's really 90 percent of the people that are going across that path or maybe there's something we're not doing correctly and it's hard to evaluate those methods you don't have access to the data set it it's troublesome to do the actual measurements and compare them to what the application is giving you so the problem now becomes what's the credibility of this stuff how can I trust it because the proposition is hey mr. customer you have some business processes and we can give you insight to help you with those business processes and now you're trying to change your business processes to the data that may or may not be reliable oops so RSS has some strengths the data is available right now any Wi-Fi network can provide that data to you it's very opportunistic it's very cheap to start with it's expensive one just once a scale remember the sequel database stuff and it supports unassociated clients because they are as soon as the Wi-Fi is turned on they're probing and you are able to evaluate their position accuracy somewhere between five to ten meters you normally need to do some sort of post-processing which makes it costly to scale and it's even more problematic in terms of accuracy when our clients are on the move when they're moving around then your accuracy just goes down so is there any other way there are some alternatives we do know about time based methods and we have two different approaches here one is let's measure round-trip time for a frame because if we know the distance between different nodes we obviously know the speed at which the signal moves we are able to calculate the distance or let's take several measurements and compare time of let's say reception of a frame between different nodes and then we can say okay from this nut from the station to this node it took this much time and to the other two it took this much time which means we can calculate the distance and do our tool iteration again so the first method is called round-trip time and you have two choices here you can measure the amount of time which is known for a frame to be transmitted to the client so from the AP to the client and until the time you receive an acknowledgment these are known values and based on that you can estimate okay this is the distance between the client and the AP or you could use RTS CTS if you're using that so because this is a known time and because the signals travels in known speed you can calculate the distance cool the problem is the client needs to be connected so forget about using this for the unassociated now back to the question how do we trust these methods it also goes back to something that's been mentioned here today and yesterday what are the values whether something is good or bad in Wi-Fi or what are they well how do we say whether something is good in bad in location detection well there's something called what we call a confidence factor and it's not often given out if you look at the probability distribution of your errors or what's called the cumulative distribution factor CDF and you compare our SSI to round-trip time on the bottom you have position error and on the y-axis you have the probability that the error is lower than that so if we check our society we have somewhere like let's say 90% probability that the error is lower than eight or nine meters that's our side round-trip time less than 90% of the time the error is going to be less than two meters so okay that's better that's better and this is actually something that you can use to weigh or evaluate the methods that you're using right nobody's giving this out nobody's saying hey we're doing location accuracy with this sort of confidence factor how is it obtained well through experimentation or simulation you can't really say for sure this is the accuracy we're saying we're 80% sure that the accuracy is going to stay within these limits oops the other methods which is already known from outdoor location positioning time difference of arrival basically you work under the assumption that there is going there's going to be several nodes that will be detecting the frame being sent out from the client to the AP because Wi-Fi is a polite protocol and somebody talks everybody else listen so when you listen you have time to do stuff so you listen for when you receive a frame and you compare those timestamps to evaluate how long did the frame require to travel over the air and again you can get the distance so T do a what does it give us in terms of our confidence factor and this is a this is an actual experiment done at the University we are between one and two meters of accuracy the problem there sub several problems first of all we need line of sight from at least three stations all the station is need to listen on the same channel how often do you get that in your in your high-density Wi-Fi it requires several samples but that can be achieved more importantly the clock needs to be synchronized so the only way to these measurements make any sense if you have synchronized clock to something called a global clock and we're not talking about NTP here we're talking about microseconds yeah so NTP is not going to do this for you you need a specific clock a dedicated equipment it's going to provide the clocking for you now there's good news to this you can do that without the clock if there's so it's been proved and I'm going to go into the details but it's been proved that there exists a synchronous method to doing this if you are able to have 20 megahertz or larger clock frequency on your receiver so if you're able to do 20 Hertz clock ticks you are able to derive this information the same information without having a synchronized clock which is good because now you can use cheap equipment to do this the problem still remains all the stations needs to be on the same channel you need several sources several samples of data and more importantly you need line of sight at least three stations needs to need to see this in of sight which in some cases so if you imagine a warehouse or if you imagine a store which shelves that can actually be achievable so we're getting closer so in terms of weaknesses and strengths round-trip time it's pretty accurate in terms of distance measurement especially when you're compare it to our society the problem is only works for connected clients so yes we often need to combine it with something else t do a again pretty accurate can work for both connected and unassociated clients and it's been also proven that it works pretty well for the client that are on the move and best of all it's pretty real-time you don't have to do any special post-processing and things like that it can give you that data pretty quickly so in terms of resources that you're using it is pretty efficient especially if you go for an as a synchronous mode now there are some problems the ap's need to be on the same channel and you need to have cells frame samples which may not be available when your client is probing right it depends how many frames it's sending out the last method angle of arrival the reason why there is a fighter jet over here is I've been told this same technology is used in fighter planes basically you have an antenna array doesn't have to be larger it could be four elements what you're able to do is based on the phase you're able to say hey the signal came from this direction so based on your antenna position you can say okay the client is over there and based on the sub based on your signal strength or round-trip time or whatever can say okay that's the distance so you're getting more accurate with this the problem is there's three versions of this implementation one is cheap you just need several elements several antenna elements need to derive the face do a cross correlation for up a signal and then you'll be able to discern at what angle did the signal come from the problem is this method the cheap method is very very it doesn't tolerate noise very well it's sensitive to signal-to-noise ratio so as long as your signal-to-noise ratio is above 15 or maybe even more 15 dB it's gonna work as soon as it drops below that forget about it so how do you control that there are two more expensive methods but then you're spending money on on your hardware so angle of arrival accurate positioning can work with a single ap by the way which is pretty cool it can cost a lot especially if you want to go for the signal-to-noise ratio or noise resistant method or there's something else so you either invest in hardware or tune the environment which means position your ApS and then you have to do some measurements to measure the signal response and then you're able to do the cross correlation to get the angle of the signal which is time-consuming and well nobody wants to do it there's more there have been several hybrid methods that have been implemented and have been proven to work pretty well one is okay so we have several sources of truth where the client is let's say we have the RSS the received signal strength we have the round-trip time so let's combine those two sources of truth to use RSS RSS worst case so when the clients are unassociated and then combine it with our dt1 once the client is actually connected this gives us so if the client is connected this gives us sub 2 meter accuracy if the client is unassociated this gives us 5 to 10 meter accuracy which is still better than missing the client or maybe pair RSS with one of the video a types so if enough samples are gathered we're confident that our tdo is going to work well so you see do a if we don't have enough samples use RSS or the for example we can actually use client-side information especially if that client is on the move so most of smartphones today have accelerometers and if you have an accelerometer you know how your device is oriented what the position of it is and where you're moving so and how fast you're moving if you use that data from the device to augment what you're measuring you can actually get pretty decent location the problem is you need an app on your phone who likes apps on their phones what's the what's the retention rate of an app on your phone supposedly up to three weeks so any let's say there are several retailers using this are you going to really install apps for all the retailers you're visiting I certainly wouldn't but yes it's been proven that this can work pretty well now once you figure out how to get that data and once you've convinced your customer hey this is the confidence factor I am 90% sure I'm giving you the information that is within 2 meters 90% of the time now how to get the data out well you need an API for that and usually you don't or can't do all the post-processing on your own because when you're do post-processing on your own maybe you're discarding information that the customer wants maybe you are interpreting that information differently then it should be interpreted in a specific environment so you don't know that is the business processes and you don't know what the data is actually going to be used for then there's the question how regal is real-time when is that application is that data available so if I request it now will I get data from one second ago one hour ago one week ago what am I getting am I getting the confidence factor are you telling me how sure you are this data is actually correct what about historical data so how long do you keep it is it available and especially in Europe what are the privacy concerns of storing and exchanging that data so this is this is sort of customer information right and I was consulting one of the legislative bodies in the EU and they said that your max oh sorry your phone's MAC address is actually personal information they treat it as personal information so it like your passport your passport personal information your phone MAC address personal information so okay what do we do about it do we encrypt if we encrypt what does that mean for the data interpretation do we do a one-way hash what do we do what do we have to do and most importantly how does it all scale so I've been through several implementations of this and trust me the only thing that the only way this scales is for the cloud it just doesn't scale any other way and when I say scale I mean thousands and thousands of clients hundreds or thousands of APs that's what you want and don't use traditional relational databases you're going to get burned it's going to burn that because they are not equipped to handle this sort of this amount of data getting in especially when you do long term trend analysis and things like that okay so that's pretty much it from me thanks for your attention any questions [Applause] question oh oh in the back you see then he handed to Scott when you're done first of all thank you so much very very good and informative presentation and you know being in the artists industry for 15 years myself as well you've clearly done your homework so you know put a lot of thought on the drawbacks of the systems and the benefits and I agreed with nearly all of them except maybe the RSI Icarus II can be tuned to be slightly higher than 5 meters to my experience and the second thing was that clients on the move at least with the the system we used to have we sold our artists business but armed the system we used to have it was actually more accurate when the clients were in the move cause we were frequently getting pulse from either from the infrastructure or the client RSI voice so so you know when the client was moving the the historical trends that we followed made it actually much more accurate than when it was stationary and it started to hover because it started to get more and more unsure of where it is just just an observation last thing if I get an OK from my management will you be willing to come and talk about these things to our webinar as well I'm sure it could have to talk with my manager but I'm sure I could and that would be nice thank you so one thing about the RSSI bit 5 meter accuracy is on average across the different verticals so you don't necessarily you know different verticals different environments means different things and with the the thing where you're saying about the moving client is it depends on so if you're doing it like this it works but what people are actually doing is going like this going like this doing the Wi-Fi dance and there are side just goes so I've been doing this in a hospital right and I I assumed it's going to be like what you said so when you have our tag it works like you said when you have a smartphone a device that user has they push it against their heads they I don't know they sit on whatever it just dances around yeah that's why it works pretty well in hospital basements I get the feeling that RTLS maybe isn't dying technology but particularly with the analytics we've got ble and other technologies that are more accurate is our TLS in a traditional healthcare was obviously the main market is that still you know it a customer still is that still being deployed is that it's our Allah vendors now looking at the these more accurate and the technologies we're looking at hybrids to be honest so that's a very good observation the RTL has a traditional archaea RTLS it's still there for health care but for other verticals like retail logistics and things like that it actually turns out so really is a good example and by the way these techniques don't apply to Wi-Fi radios only they can be used across different frequencies across different radios it's it's not something that's Wi-Fi specific and ideally so I think what the future is using several radio technologies like some information from Wi-Fi complement that with ble you have several sources of truth and then you try to approximate what the truth actually is so I think that is the future and maybe maybe Google or Apple will expose some sort of API to their accelerometer and you can tap into that information as well but we'll see where that goes yeah angle of arrival so from what I know Cisco has this halo antenna the question is what is that how does that correlate to what was presented its angle of arrival several antenna elements I think pretty too or something like that so it from my experience you don't actually need that many but yeah it's on steroids so basically makes you invest in more Hardware you're investing in more hardware to get more accuracy basically that's the idea that's the big problem and so most retailers what they don't want to force you to go to a captive web portal register or they don't want you to use an app because that proved to be inefficient you know the home Wi-Fi monetization schemes just fell apart five years ago we started talking about it and then just didn't stick so now they say okay we want the passive detection right and passive detection has some limitations in with all these methods

Channel: Wireless LAN Professionals

Views: 37,294

Keywords:

Id: vtfnlgTj_-A

Length: 38min 51sec (2331 seconds)

Published: Mon Jun 05 2017