indoor localization using fm signals

Indoor Localization Using FM SignalsYin Chen, Dimitrios Lymberopoulos, Jie Liu, and Bodhi Priyantha

Abstract—The major challenge for accurate fingerprint-based indoor localization is the design of robust and discriminative wireless

signatures. Even though WiFi received signal strength indicator (RSSI) signatures are widely available indoors, they vary significantly

over time and are susceptible to human presence, multipath, and fading due to the high operating frequency. To overcome these

limitations, we propose to use FM broadcast radio signals for robust indoor fingerprinting. Because of the lower frequency, FM signals

are less susceptible to human presence, multipath, and fading, they exhibit exceptional indoor penetration, and according to our

experimental study they vary less over time when compared to WiFi signals. In this paper, we demonstrate through a detailed

experimental study in three different buildings across the US, that FM radio signal RSSI values can be used to achieve room-level

indoor localization with similar or better accuracy to the one achieved by WiFi signals. Furthermore, we propose to use additional signal

quality indicators at the physical layer (i.e., SNR, multipath, etc.) to augment the wireless signature, and show that localization

accuracy can be further improved by more than 5 percent. More importantly, we experimentally demonstrate that the localization errors

of FM and WiFi signals are independent. When FM and WiFi signals are combined to generate wireless fingerprints, the localization

accuracy increases as much as 83 percent (when accounting for wireless signal temporal variations) compared to when WiFi RSSI

only is used as a signature.

Index Terms—FM, indoor localization, mobile systems, fingerprinting, wireless

Ç

1 INTRODUCTION

ACCURATE indoor positioning information has thepotential to revolutionize the way people search,

locate, and navigate to points of interest inside buildingsin a similar way that GPS revolutionized the way peoplenavigate outdoors. For instance, a user in a mall couldleverage his mobile device, equipped with accurate indoorpositioning technology, to instantly search, locate, andnavigate with real-time turn-by-turn directions to any storein the mall. When entering a store, the user’s mobile devicecould automatically provide directions to the exact aisle orsection where the desired product is located. At the sametime, businesses and advertisers could push couponsand offers to the user in real time based on his currentposition within the mall or the store, maximizing customertargeting effectiveness.

Enabling these scenarios has been challenging mainlydue to the unavailability of GPS signals in indoor environ-ments. In the absence of GPS, fingerprint-based indoorlocalization techniques have been the most accurateapproach to indoor localization [1], [2], [3]. The majorchallenge for fingerprint-based approaches is the design ofrobust and discriminative signatures. The most popularapproach, that does not require any hardware deployment,has been to leverage already available wireless signals(e.g., WiFi, cellular) to profile a location, usually in the formof received signal strength indicator (RSSI) values [1], [4].In previous work, RSSI values of WiFi signals have been

primarily used for this purpose, as WiFi access points arewidely deployed indoors, and every mobile device isequipped with a WiFi receiver.

Even though this approach has been successful inlocalizing people at a coarser grain (e.g., at the buildinglevel [5]), it exhibits several limitations when consideringindoor environments where a person needs to be localizedat the room level. First, the operating frequency range ofWiFi signals makes them susceptible to human presenceand orientation as well as to the presence of small objects ina room. This introduces variability in the recorded finger-prints that can lead to localization errors. Second, several ofthe deployed WiFi access points are commercial in natureand employ optimizations, such as frequency hopping, toimprove network’s throughput. These optimizations canresult in variations in the observed received signal strength(i.e., RSSI values change across WiFi channels), and there-fore in the localization process. Third, WiFi RSSI valuesexhibit high variation over time that, as we show in thispaper, can adversely impact localization accuracy. Fourth,the area of coverage of a WiFi access point is significantlyreduced in indoor environments due to the presence ofwalls and metallic objects, easily creating blind spots (i.e.,basement, parking lots, corner rooms in a building, etc.).

To address these limitations, we study the feasibility ofleveraging alternative wireless signals to augment or evenreplace WiFi signals for fingerprinting. In particular, wepropose to use FM broadcast radio signals for fingerprint-ing indoor environments. FM signals operate at thefrequency range of 88-108 MHz in the US, which makesthem less susceptible to the presence and orientation ofhumans and small objects [6], [7]. Furthermore, FM signalsare significantly stronger than WiFi signals in the sense thatthey can easily cover areas of hundreds of kilometers, whileachieving good indoors penetration (see Table 1). From theinfrastructure point of view, there are thousands ofcommercial and amateur FM signals being broadcasted

1502 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 12, NO. 8, AUGUST 2013

. Y. Chen is with Qualcomm Research Silicon Valley, Apt. 02, 601 AlmaridaDrive, Campbell, CA 95008. E-mail: [email protected].

. D. Lymberopoulos, J. Liu, and B. Priyantha are with Microsoft Research,1 Microsoft Way, Redmond, WA 98052.E-mail: {dlymper, liuj, bodhip}@microsoft.com.

Manuscript received 14 Sept. 2012; revised 5 Feb. 2013; accepted 6 Apr. 2013;published online 5 May 2013.For information on obtaining reprints of this article, please send e-mail to:[email protected] and reference IEEECS Log Number TMCSI-2012-09-0464.Digital Object Identifier no. 10.1109/TMC.2013.58.

1536-1233/13/$31.00 � 2013 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS

continuously across the world, eliminating the need fordeploying any custom infrastructure. Also, most mobiledevices, even the lower end ones, are equipped with FMradio receivers that are lower power and less costlycompared to the WiFi receivers (see Table 1).

However, in the case of FM radio signals, the accesspoints (FM towers) are located up to several hundred ofkilometers far away from the user and transmit signals at avery high power. As a result, the recorded FM RSSIsignatures might not exhibit significant variation acrossnearby locations, and therefore fine grain localization mightnot be feasible. Previous work that has already examinedthe use of FM radio signals for localization in outdoorenvironments has verified this intuition by demonstratingcoarse-grained localization accuracies (e.g., zip code level[8] or tens of meters [9]).

In this paper, we demonstrate through a detailedexperimental study that FM broadcast radio signals can beused to achieve room-level indoor localization with similaror better accuracy to the one achieved by WiFi signals. Eventhough FM radio reception may not vary significantlyacross nearby outdoor locations, in the case of indoorenvironments the internal structure of the building cansignificantly affect the propagation of FM radio signals,providing enough resolution in the FM signal signatures toaccurately localize mobile devices.

This paper makes the following contributions:

1. We demonstrate through detailed experiments inthree representative buildings across the US (aresidential building, an office building, and ashopping mall) that FM radio signals can achieve

similar room-level accuracy in indoor environmentswhen compared to WiFi signals.

2. We propose to exploit additional information at thephysical layer, such as multipath or frequency offsetinformation, to create more reliable fingerprinting ofindoor spaces, and demonstrate through real-worldexperiments, that this approach can improve theaccuracy of FM-radio-based indoor localization bymore than 5 percent when compared to the accuracyachieved by FM or WiFi RSSI-only signatures.

3. We study in detail the effect of wireless signaltemporal variation and demonstrate that WiFi RSSIvalues exhibit significantly higher variation over timecompared to FM RSSI values. This enables FM-basedindoor localization to achieve approximately 57 per-cent higher room-level localization accuracy whenconsidering temporal variations of wireless signals.

4. We experimentally demonstrate that FM and WiFisignals are complementary in the sense that theirlocalization errors are independent. Our experimen-tal results indicate that when FM and WiFi signalsare combined to generate fingerprints, the localiza-tion accuracy increases by 11 percent (withoutaccounting for temporal variation) or up to 83 per-cent (when accounting for wireless signal temporalvariation) compared to when WiFi RSSI only is usedas a signature.

2 ARCHITECTURE OVERVIEW

Fig. 1 provides an overview of the proposed indoorlocalization approach. As in most fingerprinting ap-proaches, there is a training and a positioning stage.The training stage is responsible for collecting location-annotated wireless signal fingerprints that form thefingerprint database. The fingerprint database can beautomatically crowdsourced from real mobile users as theycheck-in to different businesses or it can be manuallycreated through detailed profiling. Every time a businesscheck-in takes place, the wireless fingerprint is recorded on

CHEN ET AL.: INDOOR LOCALIZATION USING FM SIGNALS 1503

TABLE 1Basic Properties of WiFi and FM Broadcast Signals

Fig. 1. (a) Training and positioning stages for indoor localization. (b) Signal fingerprinting using FM and WiFi.

the mobile device and the business location is retrievedfrom freely available web services. The recorded wirelessfingerprint is properly annotated with the business’ locationinformation and stored in the database. At the positioningstage, the mobile device records its wireless signalfingerprint and compares it against the available finger-prints in the database. The location associated with thefingerprint in the database that is the closest to thefingerprint recorded on the mobile device, in terms of adistance metric, such as euclidean distance, is assumed tobe the current location of the device.

The most challenging task in fingerprint-based localiza-tion is the engineering of the fingerprint itself. To enableaccurate localization, fingerprints need to be carefullyengineered so that even nearby locations have sufficientlydifferent fingerprints. Most previous approaches haveadopted the received signal strength (RSSI) of nearby WiFiaccess points as the wireless signal fingerprint. In thispaper, we extend this approach in two fundamental ways.First, we augment the wireless fingerprint to include theRSSI information obtained by FM radio signals. As Fig. 1bshows, mobile devices record RSSI information from severalFM radio station signals that are broadcasted from one ormore radio towers. The RSSI value for each FM frequencycan be used along with the WiFi RSSI values to form thewireless fingerprint. By combining WiFi RSSI with RSSIvalues from another wireless signal that is less susceptibleto human presence and orientation, small objects, andmultipath and fading due to its lower wavelength, wemanage to encode a more robust profile of the location intothe wireless signal fingerprint that, as it will be shown later,can lead to better localization accuracy.

Second, to enable unique fingerprints even for nearbylocations, we propose to extract more detailed information,that goes beyond RSSI, at the physical layer (see Fig. 1b).Even though RSSI has been proven to be a good high-levelsignal indicator, it does not provide the necessary granu-larity to enable robust fine grain localization. For instance,RSSI values at different rooms inside a building might beidentical due to different reasons such as human presenceand multipath. However, lower level information at thephysical layer, such as signal-to-noise ratio (SNR), and

multipath indicators, can provide enough insight on how anRSSI value was generated. For instance, the way wirelesssignals are reflected in a room is unique and it depends onthe room’s setup and location in the building. As a result,multipath indicators might be different across rooms eventhough the RSSI values for these rooms might be identical.By augmenting wireless signal fingerprints with additionallow-level signal indicators, we enable fingerprints tocapture more robust information about the wireless signaltransmission and the way it is affected from the current room’s

structure. This information could eventually be used todifferentiate two rooms with the same RSSI values.

Note that this information could be leveraged for anywireless signal (e.g., WiFi) as long as it is exposed throughthe software driver. In fact, most recently Sen et al. [12],[13] exploit 802.11n PHY layer (OFDM) impulse responsesand report notable localization accuracy gains. In thispaper, we focus on FM radio signals and thereforeleverage the additional signal indicators for FM radiosignals only. Nevertheless, we believe that in addition toWiFi and FM, getting additional signal indicators from thePHY layer can improve the localization accuracy of otherwireless signals too.

3 EXPERIMENTAL SETUP

To evaluate the proposed approach we conducted detailedexperiments in three typical building environments (seeFig. 2): an office building in a major corporate campus, amall consisting of various restaurants and retail shops, andan apartment in a typical US residential building. The officeand mall buildings are part of a major corporate campuslocated in the west coast of the US. Both buildings have asteel skeleton, and their perimeter is covered by largewindows. The office building consisted of three differentfloors, with each floor containing approximately 40 roomsof 9 ft � 9 ft size each. The mall building consisted ofa single floor with a total number of 13 large rooms(i.e., 100 ft � 30 ft) of varying size and shape (see Fig. 2b).The residential building is located in a major city on theeast coast of the US, and is built using steel-reinforced


Fig. 2. Maps and dimensions of the different buildings used in our experimental study.

concrete. The particular apartment profiled in this studyconsisted of five different rooms as shown in Fig. 2c.

All buildings had exceptional WiFi and FM signalcoverage. During data collection we recorded 434, 379,and 117 unique WiFi access points at the office, shoppingmall, and residential buildings, respectively. In each roomof the office building, we could listen, on average, 32 WiFiaccess points. In every room of all three buildings, the FMreceiver was able to tune to more than 32 FM radio stations.

3.1 Hardware

FM radio signatures were collected using the SI-4735 FMradio receiver from Silicon Labs [11] (see Fig. 3). Theparticular receiver was chosen for two reasons. First, SiliconLabs’ FM radio receivers are very popular among a widevariety of consumer products such as cars, cell phones,portable media players and more. This enabled us toexperiment with one of the most widely used FM receiversin the market today. Second, the particular receiver isamong the few ones that expose low-level reception signalinformation to the application layer. In particular,besides reporting RSSI values, it provides three additionalindicators1 of signal reception: signal-to-noise ratio, multi-path (MULTIPATH), and frequency offset (FREQOFF). SNRtakes values between 0 and 128 db, and indicates howstrong the received signal is compared to noise (i.e.,interference, signal reflections, etc.). MULTIPATH takesvalues between 0 and 100, and indicates the severity of themultipath effect (i.e., the number and power of wirelesssignal reflections) in the current signal reception. FREQOFFmostly takes values between �10 and 10, and quantifies thedifference in carrier wave frequency between the actualreceived signal and the nominal frequency (e.g., 88.1 MHz).Even though the major source for carrier wave frequencyvariations are the differences in reference crystal frequen-cies between the transmitter and the receiver, other sources,such as Doppler spread, also contribute to FREQOFF. Inparticular, multipath components of the transmitted signalthat arrive at the receiver over different paths of varyinglength, caused due to either static or moving nearby objects,contribute to the Doppler spread effect, increasing thesignal bandwidth, and affecting FREQOFF values. Thecombination of RSSI, SNR, MULTIPATH, and FREQOFFindicators represents the FM signature collected at eachlocation during data collection (see Fig. 1b).

For embedded designs, such as cell phones, where spaceand size is important, the FM receiver can be connected to apatch antenna that can be directly implemented on the mainPCB of the device. This would enable the device to acquirereliable signal readings, eliminating the unpredictability ofloose earphone wires that are currently used as the FMantenna on most cell phones. Unfortunately, we were notable to find any patch antennas that meet the design criteriaof Silicon Labs for the current FM receiver. As a result, weopted to use the typical FM antenna provided in theevaluation board. To approximate the behavior of a patchantenna, we folded the antenna as much as possible asshown in Fig. 3. In our measurement setup the length of theantenna has twice the length of a typical smartphone.

WiFi signatures were collected using an 802.11a/b/g/ncompatible WiFi Link 5300 card from Intel. All WiFisignatures consisted of RSSI values only. Both WiFi andFM receivers were connected to a Lenovo T61p laptop thatsimultaneously recorded WiFi and FM wireless signalfingerprints for a given location (see Fig. 3).

3.2 Data Collection and Evaluation

The goal of our experiments was to achieve accurate room-level localization.2 In other words, given a wireless signalfingerprint from a mobile device, provide the roomnumber where the device is located (“Position” in Fig. 1is actually a room number). To achieve this, we profiledeach room in the building using both WiFi and FMsignatures as described in Fig. 1b. For every room, wecollect the WiFi and FM signatures for three random pointsinside the room. To properly take into account thestochastic fluctuations of radio signals, we record eachsignal quality indicator nine consecutive times for eachprofiled location, and then average all nine values to createthe fingerprint for that location. We chose to only profile asmall number of points within the room so that we canevaluate the ability of FM and WiFi signals to achieve highlocalization accuracy with sparse profiling, i.e., a smallfingerprint database. We explore the benefit of havinglarger databases in Section 5.2. Depending on the experi-ment, multiple data collections were performed for eachroom over different time windows.

Specifically, at each location we record the FM signaturesfor 323 FM broadcast stations and scan the signal strengthsof all available Wi-Fi access points, as described inSection 3.1. We denote ri; si;mi; fi 2 R32 as the RSSI, SNR,MULTIPATH, and FREQOFF values for the 32 FM broad-cast stations at the ith location, and similarly wi 2 RM as theRSSI values for the Wi-Fi access points. Here, M is the totalnumber of Wi-Fi access points in the building, and i is theprofiled location. We concatenate the signature vectors ofthe ith location and denote as ai 2 R128þM . In total, for everyprofiled location i within a room there are 128þM valuescorresponding to four values for every one of the 32 FM


Fig. 3. Data collection setup based on the SI-4735 FM radio receiverfrom Silicon Labs and the Intel WiFi Link 5300 wireless card connectedto a Lenovo T61p laptop.

1. The exact algorithm used to compute these indicators is sensitive, andis not provided by the manufacturer.

2. We focus on room-level resolution as this is the resolution at whichdata can be crowdsourced in a robust way from business check-in events. InSection 7, we evaluate FM’s capability to achieve fine-grain indoorlocalization.

3. The seek/tune time for the SI4735 FM receiver is 60 ms per channel,and the power-up time is 110 ms. It takes approximately 2 s to scan 32 FMradio stations.

channels and M values, one for every WiFi access point. Asa result, the whole data set for a given building can bewritten as A ¼ fai : i ¼ 1; 2; . . . ; 3�Rg, where R is thenumber of rooms.

A typical fingerprinting-based localization scheme in-volves an offline phase to construct the database and anonline phase for positioning unknown locations. Toevaluate the localization performance, we emulate thistwo-phase process by separating the signatures intotraining and test sets. We first partition the whole data setinto three complimentary subsets of equal size: A1, A2, andA3, where each set Aj contains one and only one locationfrom each room. Next, we group two subsets together as thetraining set, i.e., the fingerprint database, and use the thirdsubset as the test set. We repeat this process using each ofthe three subsets as the test set and correspondingly theother two subsets as the fingerprint database, and reportthe average localization accuracy across all combinations.For each test location in the test set, we compare itssignature vectors against the fingerprint database andreturn the location of the nearest neighbor in signal spaceas the localization result.

In all experiments, we report the localization accuracywhen euclidean and Manhattan distance metrics are used tocompute the distance between wireless signal signatures.Even though more distance metrics have been evaluated, weopted to show only these two, as they consistently providedthe highest localization accuracy across all experiments.

4 FM-BASED INDOOR LOCALIZATION

In this section, we focus on the office building environmentconsisting of three different floors and 119 rooms in total.We first investigate the performance of using RSSI valuesalone for both FM (i.e., ri) and Wi-Fi signals (i.e., wi). Next,we use more signal quality indicators for FM (i.e., si, mi, fi)to see whether extracting more information from thephysical layer can improve the localization accuracy. Last,we combine the FM and Wi-Fi vectors to investigate theeffect on the localization accuracy and perform sensitivityanalysis on the number of FM radio stations and WiFiaccess points used for fingerprinting.

4.1 RSSI-Based Indoor Localization

Table 2 lists the room-level localization accuracy resultswhen signature vectors consist of FM or WiFi RSSI valuesonly. It is clear that FM and WiFi RSSI values achievesimilarly high accuracies that are close to 90 percent. Of thetwo distance metrics, Manhattan distance (i.e., the L1 norm)yields slightly higher accuracy than euclidean distance(i.e., the L2 norm).

Fig. 4 shows the distribution of the localization errors interms of physical distance when using FM and WiFi RSSIsignatures. Although both signals exhibit similar room-levelaccuracies (see Table 2), the localization errors are lower inthe case of WiFi. In other words, when WiFi localizationerroneously predicts rooms, those rooms are closer toground truth compared to FM-based localization. This isexpected given that there are orders of magnitude differ-ence between WiFi and FM signals in terms of bothdeployment density and communication range. In general,a WiFi access point is only visible in a subset of the rooms inthe building. This significantly limits the search space, andtherefore the localization error that is generally lower than30 ft. Conversely, in the case of FM signals, there are only ahandful of radio towers at a given region that might be tensor even hundreds of kilometers away from the building.These FM signals can be received throughout the wholebuilding, making every room in the building a possiblecandidate location.

This effect is better illustrated in Figs. 5a and 5b, wherethe Manhattan distance between every pair of profiledlocations is shown when FM RSSI and WiFi RSSI signaturesare used, respectively. In the case of WiFi signatures (seeFig. 5b), errors are usually constrained within the vicinity ofthe diagonal (three squares along the diagonal, where eachsquare corresponds to one of the three floors profiled)mainly due to the communication range of access points.Very rarely the distance between WiFi RSSI vectors is lowfor distant locations in the building. On the other hand, theerror profile of FM RSSI signature is the exact opposite(see Fig. 5a). The effect of the three squares shown in Fig. 5bhas disappeared, but now distant locations in the buildingcan generate low distance values. As a result, even thoughFM and WiFi achieve similar localization accuracy overall,the localization error of FM signals is higher in terms ofabsolute physical distance.

4.2 Robust Fingerprinting by Exploiting thePhysical Layer

To further increase localization accuracy and to constrainterrors, in this section, we leverage additional information atthe physical layer to generate more robust signatures. TheSI4735 FM receiver provides three additional signal qualityindicators (SNR, MULTIPATH, and FREQOFF) as de-scribed in Section 3. Each of these signal indicators couldbe used as an individual signature, or they could all becombined with RSSI to form a single more detailed


TABLE 2Room-Level Localization Accuracy for 119 Rooms on Three

Floors Using FM and WiFi RSSI Values as Signatures

Fig. 4. Distribution of the localization errors for FM and Wi-Fi RSSI.The nearest neighbors in signal space are determined using theManhattan distance.

signature. The additional signal indicators can enhance the

resolution of FM signatures by providing more insight

about signal reception. Signal-to-noise ratio, multipath, and

frequency offset, all capture detailed information about the

wireless signal reception and the way it is affected from the

current room’s structure and position in the building.When combining multiple signal indicators into the same

signature, calculating the distance between signatures

becomes more challenging. Different signal indicators have

different value ranges that could result into biasing the

distance calculation (i.e., higher value/range indicators

become more important). For example, the multipath value

range is between 0 and 100 whereas the frequency offset

value is usually in the range of �10 to 10. Therefore,

we normalize the value of each signal indicator using the

standard deviation of the signal indicator’s values in

the fingerprint database. For example, we can compute the

standard deviation for the RSSI signatures as

�r ¼1

N jDj � 1

Xi2D

XNj¼1

rij � r� �" #1

2

; ð1Þ

where D represents the set of location indices that are in the

fingerprint database, and rij is the RSSI value of the jth FM

broadcast station at the ith location. r is the average RSSI

value in the database, jDj is the cardinality of the set D, and

N ¼ 32 is the number of FM broadcast stations. Using thestandard deviation, we can normalize the RSSI signatures as

ri ¼ri�r; 8i 2 D [ T; ð2Þ

where T represents the set of location indices that are in thetest set. All signal indicators’ values are normalized in thesame way, enabling us to compute unbiased distance valuesbetween signatures.

Table 3 lists the room-level localization accuracy wheneach signal indicator is used as a single signature, andwhen all signal indicators are combined together into asingle signature. The “FM All” signature corresponds tocombining all raw signal indicator values into a single


TABLE 3Room-Level Localization Accuracy for 119 Rooms onThree Floors Using Additional FM Signal Signatures

Fig. 5. The Manhattan distance of signature vectors between all pairs of profiled locations in the office building and for four different signature types.The distances are normalized by the maximum pairwise distance in each figure to have the same range of ½0; 1� across all figures. Diagonal valuesare 0 as they are the distances between identical vectors that correspond to the same locations.

signature. The “FM All Normalized” signature correspondsinto combining all normalized signal indicator values into asingle signature.

It is clear that among all individual signal indicators,RSSI achieves the best accuracy. On the other hand,multipath and frequency offset indicators seem to not beable to provide the necessary resolution to achieve accuratelocalization on their own. However, combining all signalindicators into a single signature achieves higher accuracythan any individual signal indicator. This highlights thebenefit of extracting more information from the physicallayer and reflects the intuition that each type of signalindicator can capture a unique set of interplays between thepropagating radio wave and its surrounding environment.

The impact of the additional signal indicators on thelocalization accuracy of FM signature is better illustrated inFigs. 5a and 5c, where the Manhattan distance betweenevery pair of profiled locations is shown when the FMsignature ignores or takes into account the additional signalindicators. By comparing Figs. 5a and 5c, it is obvious thatwhen all signal indicators are leveraged in the FMsignature, the distance matrix appears to be significantlyless noisy, in the sense that the distances between non-neighboring locations in the matrix are significantly highercompared to the FM RSSI matrix. As a result, higherlocalization accuracy is achieved when all FM signalindicators are combined into a single signature.

Furthermore, Table 3 shows that normalizing thesignatures, as described above, can further improvelocalization accuracy. When compared to Table 2, FM-based localization achieves 5.7 percent higher accuracycompared to the WiFI RSSI signatures.

Normalization not only improves accuracy, but alsoconstraints the error when wrong predictions are made.Fig. 6 shows the distribution of the localization errors forFM-based fingerprinting. Normalization increases the per-centage of correctly identified locations, and also reducesthe errors for the incorrectly identified locations.

4.3 Combining FM and Wi-Fi

In this section, we investigate whether the FM and WiFisignal indicators could be combined into a single signatureto further improve indoor localization accuracy.

Table 4a lists the room-level localization accuracy whenWiFi and all FM signal indicators are combined into asingle signature. The combination of WiFi and FM signalscan eliminate almost all localization errors, achieving98 percent accuracy; an 11.3 percent increase compared to

WiFi RSSI fingerprinting (see Table 2). This suggests that

the localization errors generated by the FM signatures are

not correlated with the errors generated by WiFi signatures.

To further investigate the correlations, we collect and count

the number of locations misidentified by each signature

type, as shown in Fig. 7. E1 and E2 are the set of locations

misidentified by FM RSSI and WiFi RSSI, respectively. The

fact that jE1 \ E2j � jE2j indicates that the set of locations

misidentified by FM RSSI rarely overlap with those

locations misidentified by WiFi RSSI. As a comparison,

we also collect the locations misidentified by all normalized

signatures of FM and denote as set E3. One can see that

jE1 \ E3j � jE3j, and therefore the locations misidentified

by the “FM All Normalized” signature highly overlap with

those misidentified by the FM RSSI signature. Overall, one

can see that the FM localization errors are not correlated

with the WiFi errors. On the other hand, embedding

additional information from the physical layer in the FM

signatures removes many of the localization errors intro-

duce by the FM RSSI signature.


Fig. 6. Distribution of the localization errors when all available signalindicators from the FM signals are combined, and Manhattan distanceis used.

Fig. 7. E1 and E2 represent the set of locations that are misidentified byFM RSSI only and WiFi RSSI only, respectively. E1 \ E2 are the set oflocations that misidentified by both FM and WiFi RSSI signatures.jE1 \ E2j � jE2j, suggesting that FM and WiFi positioning errors are notcorrelated. As a comparison, E3 denotes the set of misidentifiedlocations using the FM All normalized signature.

TABLE 4Room-Level Localization Accuracy

for 119 Rooms on Three Floors

(a) The localization accuracy achieved across different signature types.(b) A complete list for the room and floor numbers of the incorrectlyidentified locations.

The complementary nature of FM and WiFi signals isclearly illustrated in Fig. 5. Initially, FM RSSI signatures(see Fig. 5a) provide high localization accuracy, whichfurther increases when the additional signal indicators areleveraged (cleaner distance matrix in Fig. 5c). However,errors are still distributed throughout the building, in thesense that signature distances even for distant locations inthe building are low. On the other hand, localization errorsin WiFi RSSI signatures are mostly constrained to onlynearby locations as demonstrated by the three dark squaresin Fig. 5b. When FM and WiFi signal indicators arecombined into a single signature (see Fig. 5d), the benefitsof both FM and WiFi signals show up in the resultingdistance matrix. FM signatures significantly reduce the darksquare effect which is the main source of errors in the caseof WiFi signals (see Fig. 5b). At the same time, WiFi signalsreduce the number of cases where distant locations havelow distance values; the major source for errors in the caseof FM signals. As a result, a mobile device can leverage itsability to receive both signals to effectively enhancelocalization accuracy with marginal overhead.

Table 4b lists the floor and room numbers of the locationsthat are identified incorrectly when combining the normal-ized FM and WiFi signatures (i.e., the last row of Table 4a)with the Manhattan distance. A total of seven test locations(2 percent of the 357 locations) are misidentified. However,all the erroneously predicted rooms are on the same floorand nearby the true rooms.

4.4 Sensitivity Analysis on the Number ofFM Stations and WiFi Access Points

So far we have been using the signatures from all 32 FMbroadcast stations and 434 visible WiFi access points in theoffice building for indoor localization. However, it isunclear how many FM radio stations and WiFi accesspoints are actually needed to provide accurate localization.To answer this question, we perform a sensitivity analysiswhere we withhold a certain percentage of FM radiostations and WiFi access points, and rerun the nearestneighbor-based localization algorithms. Specifically, we sortthe FM stations and WiFi access points in descending orderof their RSSI values averaged over all locations. At eachstep, we incrementally add one station/access point at atime, and rerun the localization algorithm. As a result,when we evaluate the localization performance withn stations, we use the n strongest stations in terms of theiraverage RSSI values.

Fig. 8 shows the localization accuracy achieved whendifferent number of FM radio and WiFi access points areused. It is clear that for both signals, additional infra-structure leads to higher accuracy. To achieve the maximumlocalization accuracy (i.e., accuracy when all radio stationsor access points are used), 30 FM radio stations andapproximately 50 WiFi access points are required. In thecase of FM, the accuracy increases very fast till around eightstations, but it does not saturate afterwards. Instead it keepsrising, but at a lower pace, indicating that a large number ofradio stations is required to achieve high localizationaccuracy. In the case of WiFi, given the high density ofaccess points and their limited communication range, the

number of useful access points saturates relatively fast ataround 50 access points.

When we start by using the 50 strongest WiFi accesspoints as the base line, the localization accuracy increases aswe are incrementally adding FM stations, and seems tosaturate at the point where 25 radio stations are used(bottom graph of Fig. 8).

5 TEMPORAL VARIATIONS

The results in the previous section are derived withoutconsidering the temporal variations of FM and WiFi signals.However, it is known that signal signatures are likely tochange overtime. For example, Haeberlen et al. [2] achieve95 percent room-level localization accuracy using WiFi RSSIsignatures when the test and training data are collected inclose time proximity. With data from different time andday, however, the localization accuracy drops to 70 percent,as pointed out in [14].

In this section, we explore the temporal variations of thebroadcasted FM signals and the impact on localizationaccuracy. First, we continuously monitor the FM signals forten days at a fixed location in one room to gain intuition onhow signatures vary over time. To quantify the impact oftemporal variations on localization accuracy, we collectfingerprints for the 40 rooms on the second floor ondifferent days and run the localization algorithm againstfingerprint databases that were recorded at differentpoints in time.

5.1 Continuous Monitoring of FM Signalsover 10 Days

Fig. 9 shows the raw RSSI values of the 32 FM stations at afixed location in one room over the course of 10 days. This


Fig. 8. The sensitivity of localization accuracy on the number of FMbroadcast stations and WiFi access points. Stations and access pointsare added in descending order of their average signal strength. In thelast graph, we employ the 50 strongest WiFi access points and only varythe number of FM radio stations.

room is a regular office and therefore its door and furniturecould change states due to the presence of humans. Theroom is at the perimeter of the office building and has awindow that faces a busy street. We note that there wererainy, cloudy, and also sunny days during the experiment.We configured the receiver to record all signal indicatorsfor the 32 FM broadcast stations once per 3 minutes.Therefore, Fig. 9 includes more than 4,000 rows of data.

Fig. 9 shows that, overall, RSSI values at a givenfrequency do not change drastically over time. However,all FM stations seem to exhibit fluctuations in RSSIvalues, but to a different extent. For instance, FM station27 seems to exhibit way larger fluctuations in RSSI valueswhen compared to FM station 21. We believe that this isdue to the fact that different radio stations are broad-casted from different radio towers and at differenttransmission power levels.

To quantify whether these fluctuations would impactthis room’s localization result, we use the first three rows ofdata in conjunction with the fingerprints collected inSection 4 as the fingerprint database, and use the rest ofthe rows in Fig. 9 as the test set. Table 5 lists the percentageof rows that are identified correctly. The localizationaccuracies are consistently high for all signature typesexcept frequency offset. This suggests that as long as alocation has been profiled before, the temporal variationsshould not cause this location to be mislabeled in the future.

5.2 Collecting Fingerprints on Different Days

In this section, we extend the temporal variation analysis tomultiple locations. Specifically, we collect four additional

sets of fingerprint measurements for the 40 rooms on thesecond floor in exactly the same way as before, but ondifferent days. We chose to study the second floor becausethis floor exhibits most localization errors as shown inTable 4b.

We first study the pairwise localization performancebetween two data sets, where one data set is chosen as thetest set and a different data set is chosen as the fingerprintdatabase. Note that this way the size of the test set is thesame as that of the fingerprint database. We run the nearestneighbor localization algorithm on all combinations of pairsof data sets (i.e., 20 pairs across five data sets) and presentthe average room-level localization accuracy in Table 6a.Compared to the results in Table 4, it is obvious thattemporal variations of the signal signatures can lead tonoticeable degradation of localization accuracy. WiFi RSSIsignatures seem to be affected the most by temporalvariations as the localization accuracy decreases by 44 per-cent (from 88 to 49 percent) in the presence of temporalvariations. On the other hand, FM signatures seem to be lesssusceptible to temporal variations, as the localizationaccuracy decreases by only 13 percent (from 93 to81 percent), and the achieved accuracy remains above80 percent. In general, because of the differences infrequency and wavelengths between FM and WiFi signals,WiFi signals are more susceptible to human presence/orientation, and to the presence of even small objects in theroom. However, both human presence and setup of objectsin a room changes over time, significantly lowering thelocalization accuracy of WiFi RSSI-based fingerprinting.

The signature that is affected the least by temporalvariations is the one that combines WiFi RSSI and the fourFM signal indicators. In particular, localization accuracydecreases by 8 percent, and absolute accuracy is 90 percent.These observations are consistent with the observations


Fig. 9. Raw RSSI values in dB�V of the 32 FM broadcast stations overthe course of 10 days.

TABLE 5Localization Accuracy Using the Data Collected at

a Fixed Location over the Course of 10 Days as theTest Set, against the Database Collected in Section 4

The first three measurements taken at the fixed location were insertedinto the fingerprint database.

TABLE 6Room-Level Localization Accuracyfor 40 Rooms on the Second Floor

(a) The localization accuracy where one data set is chosen as the testset and a different data set as the fingerprint database; and (b) includesfour data sets in the fingerprint database.

in Section 4.3, indicating that WiFi and FM errors areuncorrelated.

As more data is crowdourced or manually collected overtime, the quality of the fingerprint database improves. Therationale is that more data sets that are collected acrossdifferent days can potentially capture more patterns of thesignal signatures in the temporal domain. Table 6bquantifies the impact of the size of the fingerprint databaseon the localization accuracy. It lists the results of using fourdata sets as the fingerprint database and one data set as thetest set. It loops through the five different combinations andreports the average numbers. Clearly, adding more data setsinto the database can lead to notable gains in the localizationaccuracy, indicating that a bigger fingerprint database canbetter cope with temporal variations. However, even with alarger fingerprint database, only FM-based signatures seemto be able to fully recover in terms of localization accuracy.In the case WiFi RSSI signatures, accuracy increases from 49to 61 percent when a larger data set is used as the database,but it is nowhere close to the initial 90 percent accuracyachieved under no temporal variations.

We note that during the course of our experiments, mostof the changes in the environment were the movements ofpeople, chairs, doors, and other smaller objects. Ourexperimental results indicated that both WiFi and FMsignals are susceptible to these changes, however, FMsignals are affected significantly less compared to WiFi.More dramatic changes such as moving big metal shelvesshould affect the signatures to a larger extent for both WiFiand FM signals, but have not been studied in this work.

6 DIFFERENT TYPES OF BUILDINGS

In this section, we investigate whether the results obtainedin office environments can apply to other types of buildingsand geographic regions in the US.

6.1 Shopping Mall

Shopping malls are different from office buildings inmany aspects. The ceilings are taller and the rooms aresparser and bigger, which makes malls resemble outdoorenvironments more than office buildings do. Given thatFM-based indoor localization depends on the internalstructure of the building to achieve high localizationaccuracy, it is unclear whether FM signatures can be usedin this type of environments.

We opted to collect fingerprint measurements at the firstfloor of a two-story mall building that hosts variousrestaurants and retail shops (see Fig. 2b). We collected fivefingerprint data sets on three different days. For each dataset, we take measurements in 13 rooms and at three randomlocations for each room. Therefore, every data set includes atotal of 39 locations. The first four data sets are collectedduring a weekend, and the fifth data set is collected on aWednesday afternoon. Table 7a shows the average pairwiselocalization accuracy across all possible combinations ofdatabase and test data sets. Interestingly, FM signaturesperform slightly worse compared to the office building (cf.,Table 6), but WiFi signatures perform significantly better.The degradation of FM signatures’ accuracy can beattributed to the fact that mall buildings resemble more ofoutdoor environments as the impact of the internalstructure of the building on signal propagation is lowercompared to office buildings. On the other hand, WiFisignatures can more reliably distinguish the 13 roomsbecause of the large size and clear spatial separation ofthese rooms, to the extent that some of them are covered bycompletely different sets of access points. For instance,Fig. 10 shows the recorded RSSI values from all visible WiFiaccess points across all profiled locations. It is clear thatthere are sets of locations that are covered by completelydifferent sets of access points due to the fact that the size ofa lot of the rooms in the mall is almost equal to thecommunication range of WiFi access points. Nevertheless,combining WiFi and FM signatures still provides thehighest positioning accuracy.

Furthermore, according to Table 7a, combining all FMsignal indicators into a single fingerprint decreases theoverall localization accuracy. This, again, is an artifact of themall space we profiled. The mall consisted of huge spaces(open areas with very tall ceilings, that are sparsely


Fig. 10. The RSSI values recorded for all WiFi access points across allprofiled locations in the shopping mall.

TABLE 7(a) Lists the Localization Accuracy Where One

Data Set Is Chosen as the Test Set and a DifferentData Set as the Fingerprint Database; and (b) Includes

Four Data Sets in the Fingerprint Database

finished) where individual stores were in many cases onlyseparated virtually, in the sense that there was no concretestructure for every store. The minimal internal structure ofthis type of buildings, has minimal impact on the broad-casted FM radio signals, resulting into low resolution in therecorded signal quality indicators. This is the reason whyeven FM RSSI fingerprints achieve lower localizationaccuracy compared to the office building. However, whenall the signal indicators are combined together into a singlefingerprint, the individual errors from the low resolution ofeach signal indicator add up, resulting into even lowerlocalization accuracy. This follows the trends we wouldexpect to see in outdoor environments.

Table 7b shows the localization accuracy when usingonly one data set as the test set and the remaining four datasets altogether as the database. Similar to the case of theoffice building environment, having more fingerprints inthe database increases localization accuracy.

6.2 Residential Building

Residential buildings differ from office and mall buildingsin size, shape, structure, and often building materials. Inthis section, we study the performance of FM-based indoorlocalization in an apartment unit that has multiple rooms ofvarious sizes (see Fig. 2b). This apartment is located on theeast coast of the US, and therefore the list of FM stations arecompletely different from the ones used in the office andmall buildings which are on the west coast. Nevertheless,we still chose 32 audible stations such that the number ofstations remains consistent.

We collected two data sets on two different days. Eachdata set gathered measurements in five rooms and at threerandom locations per room. We chose each of the two datasets in turn as the test set and fingerprint database toevaluate the positioning accuracy. Table 8 shows theaverage room-level localization accuracy for FM and WiFisignatures. Overall, both FM and WiFi signatures demon-strate above 90 percent accuracy in this environment, withFM signatures achieving perfect room-level localizationaccuracy. These results are comparable to the results at theoffice and mall buildings, suggesting that: 1) the achievedlocalization accuracies are independent of the building type,and 2) the FM-based indoor localization approach isapplicable to other geographic regions with different FMbroadcast infrastructure.

7 FINE-GRAIN LOCALIZATION AND DEVICE

VARIATIONS

So far, we have been focusing on room-level localization asthis is the resolution at which data can be crowdsourced

reliably from current business check-in events. However, itis still feasible to collect more fine-grained, locationannotated indoor fingerprints through detailed profiling.To study the feasibility of using FM radio signals to performfine-grain location estimation, we performed experimentsalong the hallway on the second floor of the office building.In addition, we also investigated the effect of devicevariations on localization performance by using a differentFM receiver of the same model and manufacturer inhallway experiments. Finally, we placed the two FMreceivers inside an office room to study the devicevariations in room environments.

7.1 Hallway Experiments

In the hallway experiments, we gathered FM and WiFisignatures at 100 locations that formed a straight line alongthe hallway, with the distance between every two adjacentlocations being approximately one foot (�0:06 ft). Wecollected a total of three data sets in different days tocapture the temporal variation of signal signatures. Theexact same locations were profiled in all three data sets. Thecoordinates of each location were measured accuratelyusing a laser distance meter.

7.1.1 Leave One Out Evaluation

For each one of the three data sets, we perform the leave-one-out evaluation. In particular, we remove one and onlyone location at a time from the data set and compare itssignature against the other 99 signatures in the data set. Theposition of the closest signature is assumed to be theposition of the test location. In such an evaluation, a robustsignature scheme should always return one of the twoneighboring locations to the test location.

Fig. 12 shows the distribution of the localization errorsfor FM and WiFi RSSI signatures across all three data sets.FM RSSI signatures provide highly accurate positioningwith errors around 1 ft. Recall that the 100 locations formeda line with the distance between two neighboring positionsset to one foot (�0:06 ft). This indicates that each location isidentified as one of its two neighbors on the line. On theother hand, WiFi RSSI signatures exhibit significantly largererrors, with the 90 percentile error being 10 ft.

Fig. 11 visualizes the spatial resolution of the FM andWiFi RSSI signatures by plotting the Manhattan distancesbetween all pairs of locations for one of the three data setscollected. The distances are linearly mapped to the ½0; 1�range for each signature such that the graphs are directlycomparable. The diagonals are always zero because thevectors are identical for the same location. By comparingFigs. 11a and 11b, it becomes apparent that not only FMRSSI signatures have the necessary spatial resolution foraccurate fingerprinting, but they provide significantly finergrain resolution compared to WiFI RSSI signatures. In thecase of FM signals, the RSSI distances are low only for thesame or neighboring locations. In most other cases, the FMRSSI signatures differ significantly. Conversely, in the caseof WiFi signals, RSSI distances are low even for locationsthat can be more than 15 ft far away from each other,leading to high localization errors as shown in Fig. 12.

These results indicate that FM radio signals are notfundamentally constrained to room-level localization accu-racy. Even though limited to a 2D deployment, our


TABLE 8Localization Accuracy for Five Rooms in a ResidentialBuilding When the Manhattan Distance Metric Is Used

experimental setup shows that fine-grain fingerprinting hasthe potential to enable even higher localization accuracies.

7.1.2 Temporal and Device Variation

All three data sets were collected at different days, and thethird data set was collected using a different FM receiverchip of the same model and manufacturer. This allowed usto study how temporal and device variations affect theaccuracy of fine-grain localization.

First, we examine temporal variations using data sets 1and 2. We let one of them be the test set and the other be thefingerprint database, and then flip their roles to computethe average localization error. The 50th and 90th percentileerrors are shown in Table 9. Both FM and WiFi RSSIsignatures show higher localization errors compared toFig. 12, due to temporal variations. Nevertheless, FM stilloutperforms WiFi significantly.

To study the effect of device variations, we choose dataset 3 as the test set (fingerprint database) and one of theother two data sets in turn as the fingerprint database (testset), and compute the average CDF of localization errors.The results are summarized in the last row of Table 9. Notethat since data set 3 was also collected in a different day, theresults are subject to both temporal and device variations.Even under device variations, though, the localization errordoes not increase significantly as compared to temporal-only variations. In fact, the localization error seems to beslightly lower with device variations, which, at a first glance,can be counter-intuitive. This decrease in error is due to thetemporal variations in radio signals. Distinguishing betweentemporal and device variations is a very hard task, and

when these two sources are combined, this type of counter-intuitive results can be produced. As an analogy, it is likeadding two random variables (corresponding to device andtemporal variations) that both follow normal distributionwith different variance (e.g., a ¼ Nð1; 1Þ and b ¼ Nð1; 10Þ).The sum of those variables could occasionally be smallerthan b, although the expected sum is greater than b. Inpractice, repeating the same experiment multiple times andaveraging the errors would compensate for these variationsand lead to slightly higher error in row 3 of Table 9 whencompared to the first row.

7.2 Device Variation in Room Environments

We have seen in the previous section that device variationshave marginal effects on the localization performance. Inwhat follows we explore the device variations inside anoffice room. Specifically, we gathered FM signatures at16 locations that form a 3 ft � 3 ft grid inside the room. Wecollected one data set on the first day using the first FMreceiver, and the second and third data sets on the secondday using the first and second FM receivers, respectively.Every data set contains 16 FM signatures, each of whichcorresponds to the RSSI values for the 32 FM stations.

We derive the pairwise variations among the three datasets and list the results in Table 10. Specifically, for eachpair of data sets, we compute the differences betweenevery pair of the two vectors that were recorded at thesame point on the 3 ft � 3 ft grid and thus get 16 vectors ofthe differences. Next, we take the average L-1 norm of the


Fig. 12. Distribution of the leave-one-out localization errors for FM andWiFi RSSI signals when the Manhattan distance is used.

TABLE 9Localization Errors Using FM and WiFi RSSI

Signatures in the Hallway Using the Manhattan Distanceto Determine the Nearest Neighbor in Signal Space

All data sets were taken at different days. Data sets 1 and 2 were takenusing the same FM receiver, while data set 3 leveraged a different FMreceiver of the exact same type.

Fig. 11. The Manhattan distance between the RSSI vectors measured at 100 evenly spaced (1 ft) locations in the hallway of an office building. Thedistances are normalized by the maximum pairwise distance in each figure to have the same range [0,1] across the two figures. Diagonal values are0 as the RSSI values for the same location are identical.

16 difference vectors to quantify the variations between thepair of data sets, as listed on the last column of Table 10.Data sets 1 and 2 were both recorded using the first FMreceiver but on different days, therefore the first rowreflects the temporal variations of FM signals. Data sets 2and 3 were recorded using two different FM receivers buton the same day, so the second row reflects the devicevariations of FM signals. Accordingly, the third row reflectsthe combination of the temporal and device variations.Conceptually, the lower the variation shown in the lastcolumn of Table 10 is, the lower the noise introduced intothe localization process is going to be. The fluctuationsacross the different variation sources shown in Table 10 isquite informative. Device variations are significantly lowercompared to temporal variations, an observation consistentwith the results in the previous section.

8 LOCALIZATION USING SHORT-RANGE FMTRANSMITTERS

So far we have been relying on FM broadcast stations toderive FM signatures. In this section, we investigate thefeasibility of deploying low-power FM transmitters to assistor even replace broadcast FM radio stations in the indoorlocalization process. Individual FM transmitters could beused to either assist the localization process by providing aset of extra short-range radio stations that the receiver canrecord, or even replace the broadcast radio stations.However, similarly to any approach that requires thedeployment of custom transmitters, for such an approachto be scalable and feasible, only a small number oftransmitters should be able to reliably cover the entirebuilding. Due to FM radio signal’s excellent indoorpenetration, FM transmitters are excellent candidates forachieving high indoor area coverage in low quantities.

To evaluate this approach, we placed five SI-4713 FMtransmitters from Silicon Labs [15] in five rooms on thesecond floor of the office building as Fig. 13 illustrates.Four of those rooms are at the corners, and the fifth roomis in the central area. We configured the FM transmitters tocontinuously broadcast at five empty channels and at themaximum possible transmission power level. We thenused the FM receiver in 39 rooms on the second floor torecord the FM signatures of both the 32 FM broadcaststations and the five FM transmitters. Note that the areaused in this experiment (see Fig. 13) corresponds toapproximately 15 percent of the area of a single floor inthe four-floor office building. As a result, even at thisdeployment density, several hundreds of sensors would berequired to cover the whole building.

Table 11a lists the room-level localization accuracy whenonly the FM signatures for the five transmitters are used. Itis clear that custom FM transmitters provide lower localiza-tion accuracy compared to broadcast FM radio stations.A closer inspection of the collected data revealed that themajor reason for the reduced accuracy is the limitedcoverage of the FM transmitters. Note that in all experi-ments, all deployed transmitters were set to transmit at themaximum possible transmission power level that is prop-erly set by the hardware manufacturers to comply with theFCC rules. We believe that this observation represents afundamental limitation in deploying and using customtransmitters for indoor localization, since a deploymentdenser than WiFi access points would be required to achievesimilar localization accuracy to WiFi or FM radio stations.


TABLE 10Temporal and Device Variations for 16 Locations Inside a Room

Data sets 1 and 2 were recorded using the same FM receiver but ondifferent days, while data sets 2 and 3 were recorded using two differentFM receivers but on the same day.

Fig. 13. Location of the five FM transmitters.

TABLE 11Localization Accuracy When (a) Only the Five Transmitters

Are Used, (b) Only the 32 FM Stations Are Used, and(c) When Both the FM Transmitters and Stations Are Used

As a comparison, Table 11b lists the localization perfor-mance when only the 32 FM stations are used. The betterresults reflect the strong coverage of FM broadcast radiostations in indoor environments. Nevertheless, Table 11csuggests that adding custom FM transmitters can effectivelyincrease the localization accuracy, and is therefore a feasibleoption when necessary. In essence, the short-range customFM transmitters can help eliminate some of the ambiguityintroduced by the long-range broadcast FM radio signals,and therefore can improve the overall accuracy even withoutdense deployment requirements.

9 RELATED WORK

Previous approaches to fingerprint-based indoor localizationcan be roughly classified into two categories: infrastructure-based and infrastructure-less approaches. Infrastructure-basedapproaches rely on the deployment of customized RF-beacons, such as RFID [16], infrared [17], ultrasound [18],Bluetooth [19], and short-range FM transmitters [20]. Theadvantage of these approaches lies on the fact that thedeployed beacons can be carefully engineered/optimizedfor indoor localization, and of course they can be deployed atthe necessary density to provide accurate indoor positioning.However, the high overhead of deploying custom hardwareusually prohibits the feasibility of infrastructure-basedapproaches.

Infrastructure-less approaches, on the other hand, do notrequire any hardware to be deployed, as they leveragealready available wireless signals to profile a location,usually in the form of RSSI values. The state-of-the-artapproach to signal fingerprinting relies on WiFi signals asWiFi access points are widely deployed indoors, and everymobile device is equipped with a WiFi receiver. The earlyRADAR [1] indoor localization system demonstrated theeffectiveness of WiFi fingerprinting by achieving localiza-tion accuracies in the range of 2 m. More recent work [2], [3]reported higher accuracy by statistically modeling thesignal strength as Gaussian distributions. WiFi signals,however, operate at the 2.4 GHz or 5 GHz range that makesthem particularly susceptible to multipath, fading, smallobjects and most importantly to human presence. Inparticular, human body and its orientation can drasticallyimpact WiFi RSSI values. As a result, profiling of locationsbecomes extremely tedious as for every location signaturesneed to be recorded for different body orientations [1]. Also,in agreement with our findings, the temporal variations ofWiFi RSSI values tend to be high due to the sensitivity of thesignal to the presence of humans and small objects. Mostrecently, Sen et al. [12], [13] exploit 802.11n PHY layer(OFDM) impulse responses and report more robustlocalization performance.

Varshavsky et al. [21], [4] developed GSM fingerprint-based indoor localization systems that can achieve slightlyworse accuracy to that of WiFi-based systems. The keyconcept of their system is to record fingerprints from notonly the six surrounding GSM stations, but also otherstations whose signal can still be heard by the phones.

Recently, Tarzia et al. [14] explored the possibility of usingacoustic background spectrum for room-level localization.Their localization system takes advantage of the observation

that each room tends to have its own unique backgroundnoise. Using typical smartphones and their embeddedmicrophones, they demonstrate 70 percent room-levelaccuracy from an experiment involving 33 rooms.

Chung et al. [22] investigated indoor localization withgeo-magnetic sensors. Their system is based on the ob-servation that the steel and concrete skeletons of buildingscan distort the geomagnetic field, such that fingerprinting isfeasible. The localization error of this system is within 1 m88 percent of the time. However, the values of thegeomagnetic sensors change drastically even for nearbylocations, requiring enormous profiling. Furthermore, theproposed system uses custom geomagnetic sensors that arenot integrated into current phones.

In this paper, we consider FM broadcast radio signals forfingerprint-based indoor localization. Because of the lowerfrequency, FM signals are less susceptible to humanpresence, multipath and fading, they exhibit exceptionalindoor penetration, and according to our experimentalstudy they vary less over time when compared to WiFisignals. From the infrastructure point of view, there arethousands of commercial and amateur FM signals beingbroadcasted continuously across the world, eliminating theneed for deploying any custom infrastructure. Also, manymobile devices are equipped with FM radio receivers.

FM radio-based localization systems have been studiedbefore in the context of outdoor localization. Krumm et al.[8], [23] measured the signal strength from a set of FMbroadcasting stations in outdoor environments and usedthe strength rankings to distinguish six suburbs. Also inoutdoor environments, Fang et al. [9] compared theperformance of FM radio signals and GSM signals using aspectrum analyzer. They report that both signals achievesimilar accuracy in the order of tens of meters using thefingerprinting approach. In this paper, we consider FMradio signals for indoor localization, and we experimentallydemonstrate that higher accuracies can be achieved indoors.This is due to the impact that the internal structure of thebuilding has on the FM radio signal propagation. Theinternal walls, floors, and ceilings affect FM radio signalpropagation enabling higher spatial resolution in therecorded RSSI signatures.

In parallel with our work, Popleteev et al. [24], [25] aswell as Moghtadaiee et al. [26], [27], [28] explored FMbroadcast signals for indoor localization. Popleteev et al.[24] reported median localization error of 0.91 m whenleveraging FM RSSI signatures. However, this is prelimin-ary work that only included profiling a single room in abuilding. In a more recent effort [25], Popleteev et al.validated FM-based indoor localization at a greater scaleincluding several rooms at different floors of an officebuilding. More recently, Moghtadaiee et al. [26] measuredRSSI values of FM broadcast radios in several rooms withUSRP2, and reported a mean localization error of 2.96 musing fingerprinting. Also, in parallel with our work,Moghtadaiee et al. [27] explored the concept of combiningFM and WiFi signals into a single fingerprint to improveindoor localization accuracy.

Our work differs from these approaches in five funda-mental ways. First, to the best of our knowledge, this is thefirst large scale study of using FM radio signals for thepurpose of indoor localization. We collected measurements


in more than 100 rooms on multiple floors, buildings andregions in the US. The volume of experimental data enabledus to investigate room level as well as fine-grain localizationperformance in the 2D, as well as in the 3D space. Second,we go beyond RSSI-based fingerprinting. We propose andevaluate the use of additional signal strength indicators atthe physical layer to create more robust and discriminativesignatures. Third, we study in detail the impact of temporaland device variations on the localization accuracy for bothWiFi and FM signals. Fourth, we evaluate the effect that thesize of the fingerprint database has on the localizationaccuracy when FM and WiFi fingerprinting are leveragedindividually or combined. Fifth, we experimentally demon-strate that WiFi and FM errors are independent, and thatthe two signals can be combined to generate signatures thatcan achieve up to 83 percent higher localization accuracywhen accounting for wireless signal temporal variations.Even though the work in [27] has demonstrated, in parallelwith this work, the complementary nature of WiFi and FMsignals, this paper has done a much deeper analysis. Ourwork has evaluated this approach in different buildingenvironments, including more than 100 rooms spanningthree different floors. In contrast, the evaluation in [27]included seven rooms located on the same floor of abuilding. Furthermore, our work addresses in detail theeffect of radio signal temporal variations, fingerprintdatabase size, and the use of additional signal indicatorsin the localization process.

10 DISCUSSION

Even though the concept of combining multiple radio signalson a mobile device to achieve accurate indoor localization isusually associated with notable increase in the mobiledevice’s power consumption, this is not the case with FMradio signals. As Table 1 shows, the power consumption ofFM radio receivers is an order of magnitude lower than thatof WiFi receivers (40 mW versus 800 mW). In addition, thetime it takes a typical FM receiver (like the one used in thiswork) to lock to a given frequency is about 60 ms, whichresults into a total time of approximately 1.6 s for scanning27 different FM radio channels. As a result, the poweroverhead of activating an FM radio receiver is negligiblecompared a WiFi receiver.

Even though power overhead is not a major bottleneck inleveraging FM radio receivers for indoor localization, thereare still obstacles that need to be overcome in the design ofcommercially available devices to enable reliable usage ofbroadcasted FM radio signals. FM infrastructure is alreadywidely deployed across the world, and many of the mobiledevices include FM receivers, as most of the SoCs on thesedevices already include an FM radio receiver. However,individual phone manufacturers or wireless carriers usuallyopt to not include the FM drivers or expose the FM radioAPIs to the application layer. In addition, several FMradio chips do not expose rich signal quality indicators thatgo beyond RSSI information.

In addition, current mobile devices lack a dedicated,reliable FM antenna. In most devices, FM radio chips canonly be used when earphones are plugged in to the device.The reason is that the earphones act as the actual FM

antenna. An embedded dedicated antenna is required forreliable indoor localization. This could be a patch antennaor even a custom antenna layout design embedded in theprinted circuit board of the mobile device. We believe thataccurate indoor localization can justify and motivate chipvendors and phone manufacturers to included dedicatedFM antennas and to expose low level FM signal qualityindicators to the application layer.

The fact that FM signals can achieve accurate indoorlocalization individually or when combined with WiFisignals, creates a unique opportunity for other broadcastedsignals in nearby frequency spectrums, such as TV andAM signals. TV signals are transmitted over the VHF lowand high bands, as well as over the UHF band. They arehigher frequency signals than FM (in the range of severalhundreds of MHz), but still low frequency compared toWiFi signals. Given the wide availability of transmissioninfrastructure all over the world, TV signals are idealcandidates for indoor localization. In addition, TV signalsare becoming increasingly important as the traditional TVfrequency spectrum is slowly being repurposed to become amore traditional networking medium. The recent rise ofWhite Space networking [29] points to this direction. As thisfrequency spectrum is transitioned to a networking med-ium, it becomes a natural candidate for indoor localization.

In addition, AM signals that are broadcasted in evenlower frequencies than FM signals (KHz to low MHzfrequency range) could potentially be exploited for indoorlocalization. However, because of the different modulationand propagation characteristics of AM signals, a morecareful study is required to understand the ability of thesesignals to achieve accurate indoor localization. In general,AM signals use amplitude modulation, and are susceptibleto atmospheric and electrical interference. This sensitivitycould potentially affect the stability of fingerprints, andtherefore the localization accuracy.

11 CONCLUSIONS

We have presented and evaluated a new approach tofingerprint-based indoor localization that leverages FMbroadcast radio signals. Our experimental results show thatwhen FM RSSI values are combined with additionalinformation about signal reception at the physical layer toform the wireless signal signature, localization is 5.7 percentmore accurate than WiFi-based techniques. Furthermore,we experimentally demonstrated that due to differences inoperating frequency and wavelength, FM signals are morerobust to temporal variations when compared to WiFisignals. More importantly, we have shown that FM andWiFi signals exhibit uncorrelated errors. The combination ofFM and WiFi signal indicators provides up to 83 percenthigher accuracy than when WiFi only signals are used.

REFERENCES

[1] P. Bahl and V.N. Padmanabhan, “RADAR: An In-Building RF-Based User Location and Tracking System,” Proc. IEEE INFO-COM, 2000.

[2] A. Haeberlen, E. Flannery, A.M. Ladd, A. Rudys, D.S. Wallach,and L.E. Kavraki, “Practical Robust Localization over Large-Scale802.11 Wireless Networks,” Proc. ACM MobiCom, 2004.


[3] M. Youssef and A. Agrawala, “The Horus WLAN LocationDetermination System,” Proc. ACM MobiSys, 2005.

[4] A. Varshavsky, E. de Lara, J. Hightower, A. LaMarca, and V.Otsason, “GSM Indoor Localization,” Pervasive and Mobile Comput-ing, vol. 3, pp. 698-720, Dec. 2007.

[5] Skyhook, Inc., http://www.skyhookwireless.com, 2013.[6] A. Alomainy, Y. Hao, Y. Yuan, and Y. Liu, “Modelling and

Characterisation of Radio Propagation from Wireless Implants atDifferent Frequencies,” Proc. Ninth European Conf. WirelessTechnology, 2006.

[7] M.A.A. Karunarathna and I.J. Dayawansa, “Energy Absorption bythe Human Body from RF and Microwave Emissions in Sri Lanka,”Sri Lankan J. Physics, vol. 7, pp. 35-47, 2006.

[8] J. Krumm, G. Cermak, and E. Horvitz, “RightSPOT: A NovelSense of Location for a Smart Personal Object,” Proc. Int’l Conf.Ubiquitous Computing (UBICOMP), 2003.

[9] S.-H. Fang, J.-C. Chen, H.-R. Huang, and T.-N. Lin, “Metropolitan-Scale Location Estimation Using FM Radio with Analysis ofMeasurements,” Proc. Int’l Wireless Comm. and Mobile ComputingConf. (IWCMC), 2008.

[10] D. Halperin, B. Greenstein, A. Sheth, and D. Wetherall, “Demys-tifying 802.11n Power Consumption,” Proc. Int’l Conf. Power AwareComputing and Systems (HotPower), 2010.

[11] Silicon Laboratories, “Si4735 Broadcast AM/FM/SW/LW RadioReceiver IC Rev D60,” technical specification, 2011.

[12] S. Sen, B. Radunovic, R.R. Choudhury, and T. Minka, “PreciseIndoor Localization Using PHY Layer Information,” Proc. Work-shop Hot Topics in Networks (HotNets), 2011.

[13] S. Sen, B. Radunovic, R.R. Choudhury, and T. Minka, “You AreFacing the Mona Lisa: Spot Localization Using PHY LayerInformation,” Proc. ACM MobiSys, 2012.

[14] S.P. Tarzia, P.A. Dinda, R.P. Dick, and G. Memik, “IndoorLocalization without Infrastructure Using the Acoustic Back-ground Spectrum,” Proc. ACM MobiSys, 2011.

[15] Silicon Laboratories, “Si4713 FM Radio Transmitter IC,” technicalspecification, 2011.

[16] L.M. Ni, Y. Liu, Y.C. Lau, and A.P. Patil, “LANDMARC: IndoorLocation Sensing Using Active RFID,” Wireless Network, vol. 10,pp. 701-710, Nov. 2004.

[17] R. Want, A. Hopper, V. Falcao, and J. Gibbons, “The Active BadgeLocation System,” ACM Trans. Information Systems, vol. 10, pp. 91-102, Jan. 1992.

[18] N.B. Priyantha, A. Chakraborty, and H. Balakrishnan, “TheCricket Location-Support System,” Proc. ACM MobiCom, 2000.

[19] R. Bruno and F. Delmastro, “Design and Analysis of a Bluetooth-Based Indoor Localization System,” Proc. Int’l Conf. PersonalWireless Comm., 2003.

[20] A. Matic, A. Popleteev, V. Osmani, and O. Mayora-Ibarra, “FMRadio for Indoor Localization with Spontaneous Recalibration,”Pervasive and Mobile Computing, vol. 6, pp. 642-656, Dec. 2010.

[21] V. Otsason, A. Varshavsky, A.L. Marca, and E. de Lara, “AccurateGSM Indoor Localization,” Proc. Seventh Int’l Conf. UbiquitousComputing (UbiComp), 2005.

[22] J. Chung, M. Donahoe, C. Schmandt, I.-J. Kim, P. Razavai, and M.Wiseman, “Indoor Location Sensing Using Geo-Magnetism,” Proc.ACM MobiSys, 2011.

[23] A. Youssef, J. Krumm, G. Cermak, and E. Horvitz, “ComputingLocation from Ambient FM Radio Signals Commercial RadioStation Signals,” Proc. IEEE Wireless Comm. and Networking Conf.(WCNC), 2005.

[24] A. Popleteev, “Indoor Positioning Using FM Radio Signals,” PhDdissertation, Univ. of Trento, Apr. 2011.

[25] A. Popleteev, V. Osmani, and O. Mayora, “Investigation of IndoorLocalization with Ambient FM Radio Stations,” Proc. IEEE Int’lConf. Pervasive Computing and Comm. (PerCom), 2012.

[26] V. Moghtadaiee, A.G. Dempster, and S. Lim, “Indoor LocalizationUsing FM Radio Signals: A Fingerprinting Approach,” Proc. Int’lConf. Indoor Positioning and Indoor Navigation (IPIN), 2011.

[27] V. Moghtadaiee, A.G. Dempster, and S. Lim, “Indoor PositioningBased on FM Signals and Wi-Fi Signals,” Proc. Int’l GlobalNavigation Satellite Systems Soc. Symp. (IGNSS), 2011.

[28] V. Moghtadaiee, A.G. Dempster, and B. Li, “Accuracy Indicatorfor Fingerprinting Localization Systems,” Proc. IEEE/ION PositionLocation and Navigation Symp. (PLANS), 2012.

[29] P. Bahl, R. Chandra, T. Moscibroda, R. Murty, and M. Welsh,“White Space Networking with Wi-Fi Like Connectivity,” Proc.ACM SIGCOMM, 2009.

Yin Chen received the BE degree in automationfrom Tsinghua University in 2007 and the PhDdegree in computer science from Johns HopkinsUniversity in 2012. His research interests in-clude mobile sensing and computing, indoorpositioning, and networked embedded sensingsystems. He is a member of Qualcomm Re-search Silicon Valley.

Dimitrios Lymberopoulos received the PhDdegree in electrical engineering from YaleUniversity. He is a researcher in the Sensingand Energy Research Group at MicrosoftResearch, Redmond, Washington. His researchinterests include mobile sensing, context-awaremobile services, and networked embeddedsensing systems.

Jie Liu received the PhD degree in electricalengineering and computer sciences, Universityof California, Berkeley, in 2001. He is aprincipal researcher at Microsoft Research,Redmond, Washington, and the manager ofits Sensing and Energy Research Group. Hisresearch interests include understanding andmanaging the physical properties of computing,such as timing, energy, and the awareness ofand impact on the physical world. He is an

associate editor for the ACM Transactions on Sensor Networks andthe IEEE Transactions on Mobile Computing.

Bodhi Priyantha received the PhD degree inelectrical engineering and computer sciencesfrom the Massachusetts Institute of Technology.He is a researcher at Microsoft Research,Redmond, Washington. His research interestsinclude low-power systems design, wirelesscommunication, and networked sensing.

. For more information on this or any other computing topic,please visit our Digital Library at www.computer.org/publications/dlib.


indoor localization using fm signals

Documents