MatMap: An OpenSource Indoor Localization System

Transcription

1 MatMap: An OpenSource Indoor Localization System Richard Ižip and Marek Šuppa Faculty of Mathematics, Physics and Informatics, Comenius University, Bratislava, Slovakia Abstract. In the past few decades a huge amount of work has been done in the area of Indoor Localization Techniques and Systems. In this paper we present MatMap: A simple OpenSource Indoor Localization System which can be used on any device that is capable of scanning WiFi access points in its proximity, does not require addition of any specialized hardware in the building and only needs a simple pre-mapping of the desired locations. The system is based on the same design principles as the Redpin[2] system that is to the author s best knowledge the first open source indoor localization system that used the intensity of the received signal (received signal strength RSS). Unlike the Redpin system, MatMap uses a Naive Bayes Classifier whose performance has been shown to exceed that of the Redpin system. Keywords: indoor localization, mobile computing 1 Introduction Precise location of a mobile device has become a crucial information for mobile devices application in recent years. While it is currently possible to obtain position estimates that are satisfactory for vast amount of applications in outdoor settings by using GPS, indoor localization still remains to be an open problem for many applications. For instance, if meter-level accuracy is desired the state of the art approach is to install additional hardware which then greatly improves the accuracy of such a system. However, installation of additional hardware is undesirable in many applications. In such setting the existing infractructure of WiFi access points needs to be utilized and the current state of the art methods provide room-level accuracy of position estimates. In our work we focus on the design of a system that requires no additional hardware to be installed, is simple to implement and suitable to be used as part of a smartphone application while providing room-level accuracy. 1.1 Motivation The campus of Faculty of Mathematics, Physics and Informatics of Comenius University in Bratislava can be a very confusing place, especially for a stranger.

2 Given the vast amount of rooms and halls that look all alike, one can get lost fairly easily (most of the time this is the case of freshman students). Our goal is to provide a framework for official applications of the Faculty which would be capable of detecting user s position and then navigating them around the Faculty. 1.2 Related work In the past few decades a huge body of research has been done in the area of indoor localization. The systems it describes can be categorized by numerous factors into classes such as continuous (position is expressed as a coordinate in some sort of 2D or 3D coordinate system)[3], discrete (where the locations are represented in form of distinct labels, such as rooms)[5], those that rely on existing infrastructure (such as WiFi access points)[2] and those that rely on specific hardware that needs to be installed as part of their setup procedure (these solutions, such as [1], are usually based on ultrasound). The system proposed in this paper could be categorized as discrete, since locations are distinct labes of places and would also belong to the category of those systems that rely on existing infrastructure. It shares many principles with the Redpin system [2] but uses a localization method described in [6] that yields more accurate results in therms of localization precision. 2 Model 2.1 Naive Bayes As mentioned above, the system described in this work uses a naive Bayes classifier. All possible locations need to be known beforehand. We denote them L 1, L 2,..., L N where N is the number of locations. These locations are estimated by a set of features F 1 ( ), F 2 ( ),..., F M ( ). These features ought to depend on the location only 1. Under this assumption the probability that certain variables f 1, f 2,..., f M were observed can be denoted as P ( ) = P (f 1 ) P (f 2 )... P (f M ) P ( ) P (f 1 ) P (f 2 )... P (f M ) We then estimate the location where these variables were measured as the most probable location, that is P (f 1 ) P (f 2 )... P (f M ) P ( ) L prob = argmax P (f 1 ) P (f 2 )... P (f M ) We can note that the dominator would be the same for all, thus we can simplify the formula to 1 Hence the name naive, since this assumption often times does not hold.

3 L prob = argmax P (f 1 ) P (f 2 )... P (f M ) P ( ) In our definition of the problem we assume that all locations are equally likely to be observed. That means that P ( ) does not contribute to the overall result and thus we can remove it from the equation. L prob = argmax P (f 1 ) P (f 2 )... P (f M ) In many applications of naive Bayes classification it is customary to take the logarithm of the product of probabilities and then express it as the sum of logarithms of probabilities. This is done in order to simplify the computation since probabilities are often very small numbers and taking their product might be troublesome to implement 2. As we describe below this is not necessary in our use case. 2.2 Naive probability estimation In some works that use naive Bayesian classifier the strength of the WiFi access point signal at a given location is completely ignored and thus the features depend on existence of a given access point. This doesn t yield particularly good results, as described in [4]. We try to avoid this problem by taking the signal strength of an access point into consideration. The features f 1,..., f M are in our case access points that were observed at the current scan. They can be uniquely described as a tuple (ssid, signal) which we denote as (fi ssid, f signal i ). In the same way we can express the recorded features for locations as (L ssid i, L signal i ). It is necessary to note that since the features were recorded more than once in all locations they are grouped by L ssid and the final L signal is produced by taking the mean of all values L signal that belong to the same L ssid. The probability P (f i ) can be then expressed as P (F i = f i ) = S max L signal i f signal i where L ssid i = fi ssid and at the same time L signal i, f signal i are real values ranging from 0 to S max and S max is the value of maximal signal that is obtainable 3. The probability is therefore also a number ranging from 0 to S max and thus it is not necessary to use the logarithm trick described above to simplify the implementation of probability estimation. 2 Often times the biggest problems are with precision of floating point values, such as under-runs. 3 In many systems S max = 100

4 2.3 Other approaches to probability estimation In some works other approaches to estimating P (F i = f i ) are used, most notably modeling the data by normal distribution. The assumption is that the manually tagged locations are distributed normally. It is further supported by arguing that for WiFi singals this seems to be a sensible assumption since the variation in their levels can be attributed to to random noises from multiple sources which leads to Gaussian distribution. [6]. In order to estimate parameters of this distribution, mean values µ L1 f 1,..., µ L1 f M,..., µ L N f M and standard deviations σ L1 f 1,..., σf L1 M,..., σf LN M are computed. The probability P (F i = f i ) can be then estimated as P (F i = f i ) = 1 e σf Li i 2π (f i µ ) 2 f i 2(σ Li ) f 2 i as The formula for estimating the most probable location can then be rewritten L prob = argmax M m=1 1 e σf Li m 2π (fm µ fm )2 2(σ Li fm )2 2.4 Choosing probability estimation method In order to choose from the two methods described above we evaluated them on the same dataset which consisted of labeled locations and respective scans of access points in these locations. In order to avoid overfitting we did multiple rounds of evaluations. In each round the ordering of the data points in the data set was randomized and 10-fold cross validation was used. The results in Table. 1 show that the naive method has preformed better overall in terms of accuracy while maintaining a similar standard deviation than the method which used Gaussian distribution to model the data. round mean sd naive gaussian Table 1. Results of 10 evaluation rounds. Based on this evaluation we have decided to implement the naive method. Apart from being more accurate it is also simpler to implement and requires less computational resources.

5 3 Technical details of the localization application Application contains main activity from which another activities are called. Main activity is now blank, just with action bar and two icons in it. Also it contains two hidden elements called History and Setting. These two are available for user when he presses mobile button usually used for settings or info. When user presses History, Settings or action bar icon called New location nothing impressive happens just a new activity is created and a simple text message is written. When user presses search button in action bar, also a new Activity is created and it starts scanning for WiFi access points. The WiFi info is displayed in a text field which will later be used for users actual location. The historical information about scans is maintained for further use. 3.1 Volatile wifi into problem Instance of a class WifiManager is created when a user presses search button and creates a new activity. The problem is that there is no period in which new data are scanned. It might be that there is but the application does not react in this way. Sometimes it so happens than it takes some time while new data is delivered. There is also a problem when user is already connected to some WiFi access point. In this case data is delivered just when user creates an activity. We will try to solve this problem by using another tread or doing more research on this problem. 4 Goals 4.1 Short term goals localize the user and display his actual position use database to maintain location data on the device (SQLite) solve the problem with volatile wifi info Long term plans Use a graph and a breath-first search to find required path Make visual map to display real rooms and paths References 1. Filonenko, Viacheslav, Charlie Cullen, and James D. Carswell. Indoor positioning for smartphones using asynchronous ultrasound trilateration. ISPRS International Journal of Geo-Information 2.3 (2013):

6 2. Bolliger, Philipp. Redpin-adaptive, zero-configuration indoor localization through user collaboration. Proceedings of the first ACM international workshop on Mobile entity localization and tracking in GPS-less environments. ACM, Kessel, Moritz, and Martin Werner. SMARTPOS: Accurate and precise indoor positioning on mobile phones. MOBILITY 2011, The First International Conference on Mobile Services, Resources, and Users Lin, Hsiuping, et al. Enhanced indoor locationing in a congested Wi-Fi environment. (2009) &context=silicon_valley 5. Luo, Yan, Orland Hoeber, and Yuanzhu Chen. Enhancing Wi-Fi fingerprinting for indoor positioning using human-centric collaborative feedback. Human-centric Computing and Information Sciences 3.1 (2013): com/content/3/1/2 6. Kovalev, Maxim. Indoor Positioning of Mobile Devices by Combined Wi-Fi and GPS Signals.

7 Fig. 1. Activity scanning information about nearby WiFi access points. Fig. 2. Main application window.