LANDMARC: Indoor Location Sensing Using Active RFID*

Transcription

1 LANDMARC: Indoor Locaton Sensng Usng Actve ID* Lonel M. N,2, Yunhao Lu, Yu Cho Lau and Abhshek P. Patl Department of Computer Scence & Engneerng Mchgan State Unversty East Lansng, Mchgan, USA 2 Department of Computer Scence Hong Kong Unversty of Scence and Technology Kowloon, Hong Kong, Chna Abstract Growng convergence among moble computng devces and embedded technology sparks the development and deployment of context-aware applcatons, where locaton s the most essental context. In ths paper we present LANDMARC, a locaton sensng prototype system that uses Rado Frequency Identfcaton (ID) technology for locatng objects nsde buldngs. The major advantage of LANDMARC s that t mproves the overall accuracy of locatng objects by utlzng the concept of reference tags. Based on expermental analyss, we demonstrate that actve ID s a vable and costeffectve canddate for ndoor locaton sensng. Although ID s not desgned for ndoor locaton sensng, we pont out three major features that should be added to make ID technologes compettve n ths new and growng market. Keywords: Locaton-aware computng; Sensng network; ID; Sgnal strength; Wreless communcaton. I. INTRODUCTION The prolferaton of wreless technologes, moble computng devces, and the Internet has fostered a growng nterest n locaton-aware systems and servces. Many applcatons need to know the physcal locaton of objects. Over the years, many systems have addressed the problem of automatc locaton- sensng. Trangulaton, scene analyss, and proxmty are the three prncpal technques for automatc locaton-sensng []. One of the most well known locaton-based systems s the Global Postonng System (GPS), a satellte-based navgaton system made up of a network of 24 satelltes placed nto orbt [2]. GPS s wdely used to track movng objects located outdoors. However, GPS, as t s satellte dependent, has an nherent problem of accurately determnng the locaton of objects located nsde buldngs. Dfferent approaches have been proposed and tested for ther effectveness and utltes n order to acheve the ablty to locate objects wthn buldngs. The objectve of our research s to develop an ndoor locaton-sensng system for varous moble commerce applcatons. Our goal s to mplement a prototype ndoor locaton-sensng system usng easly accessble wreless devces so that we can make use of exstng nfrastructures. At present, there are several types of locaton-sensng systems, each havng ther own strengths as well as lmtatons. Infrared, 802., ultrasonc, and ID are some examples of these systems. Secton 2 wll gve a comparatve overvew of these technologes and some related work. We are nterested n usng commodty off-the-shelf products. The results of our comparatve studes reveal that there are several advantages of the ID technology. The no contact and non-lne-of-sght nature of ths technology are sgnfcant advantages common among all types of ID systems. All tags can be read despte extreme envronmental factors, such as snow, fog, ce, pant, and other vsually and envronmentally challengng condtons. They can also work at remarkable speeds. In some cases, tags can be read n less than a 00 mllseconds. The other advantages are ther promsng transmsson range and costeffectveness. Secton 3 wll gve an overvew of the ID technology. Secton 4 ntroduces LANDMARC, a locaton-sensng prototype system based on ID technologes. Snce ID s not desgned for locaton sensng, the purpose of our prototype ndoor locaton-sensng system s to nvestgate whether the ID technology s sutable for locatng objects wth accuracy and cost-effectveness. In Secton 5, we present the expermental results of the LANDMARC system. Based on the analyss of ths study, suggestons are gven for manufacturers of ID products to use these products n alternatve and vable ways. Secton 6 concludes the paper and descrbes future research.

2 II. RELATED WORK A number of wreless technologes have been used for ndoor locaton sensng. Infrared. Actve Badge, developed at Olvett Research Laboratory (now AT&T Cambrdge), used dffuse nfrared technology [4] to realze ndoor locaton postonng. The lne-of-sght requrement and short-range sgnal transmsson are two major lmtatons that suggest t to be less than effectve n practce for ndoor locaton sensng. IEEE RADAR s an based system for locatng and trackng users nsde buldngs [3], usng a standard 802. network adapter to measure sgnal strengths at multple base statons postoned to provde overlappng coverage n a gven area. Ths system combnes emprcal measurements and sgnal propagaton modelng n order to determne user locaton thereby enablng locaton-aware servces and applcatons. The major strengths of ths system are that t s easy to set up, requres few base statons, and uses the same nfrastructure that provdes general wreless networkng n the buldng. The dffculty s that the object beng tracked must be supported by a Wave LAN NIC, whch may be mpractcal on small or power constraned devces. In most cases to date, the overall accuracy of the systems, usng 802. technologes, s not as optmal as desred. For example, RADAR s mplementaton can place objects to wthn about 3 meters of ther actual poston wth 50 percent probablty, whle the sgnal strength lateraton mplementaton has 4.3-meter accuracy at the same probablty level []. Ultrasonc. The Crcket Locaton Support System [0] and Actve Bat locaton system [2] are two prmary examples that uses the ultrasonc technology. Normally, these systems use an ultrasound tme-of-flght measurement technque to provde locaton nformaton. Most of them share a sgnfcant advantage, whch s the overall accuracy. Crcket for example can accurately delneate 4x4 square-feet regons wthn a room whle Actve Bat can locate Bats to wthn 9cm of ther true poston for 95 percent of the measurements. However, the use of ultrasonc ths way requres a great deal of nfrastructure n order to be hghly effectve and accurate, yet the cost s so exorbtant that t s naccessble to most users. ID. One well-known locaton sensng systems usng the ID technology s SpotON [3]. SpotON uses an aggregaton algorthm for three dmensonal locaton sensng based on rado sgnal strength analyss. SpotON researchers have desgned and bult hardware that wll serve as object locaton tags. In the SpotON approach, objects are located by homogenous sensor nodes wthout central control. SpotOn tags use receved rado sgnal strength nformaton as a sensor measurement for estmatng nter-tag dstance. However, a complete system has not been made avalable as of yet. The above are popular technologes for ndoor locaton sensng. Some other technologes, such as ultrawdeband [5], are also beng nvestgated. The choce of technque and technology sgnfcantly affects the granularty and accuracy of the locaton nformaton. There are some other projects usng the above technologes. Due to the lack of avalablty of costeffectve ndoor locaton sensng products, we have tred both nfrared and 802.b products. Nether was satsfactory for the above reasons. We do not ntend to buld our own devces due to cost constrant. We selected commercally avalable ID devces as our prototypng technology, whch s descrbed below. III. ID TECHNOLOGY AND OUR FIRST ATTEMPT ID ( Identfcaton) s a means of storng and retrevng data through electromagnetc transmsson to an compatble ntegrated crcut and s now beng seen as a radcal means of enhancng data handlng processes [4]. An ID system has several basc components ncludng a number of ID readers, ID tags, and the communcaton between them (Fgure ). Fgure : ID System Components The ID reader can read data emtted from ID tags. ID readers and tags use a defned rado frequency and protocol to transmt and receve data. ID tags are categorzed as ether passve or actve. Passve ID tags operate wthout a battery. They reflect the sgnal transmtted to them from a reader and add nformaton by modulatng the reflected sgnal. Passve tags are manly used to replace the tradtonal barcode technology and are much lghter and less expensve than actve tags, offerng a vrtually unlmted operatonal lfetme. However, ther read ranges are very lmted. Actve tags contan both a rado transcever and a button-cell battery to power the transcever. Snce there s an onboard rado on the tag, actve tags have more range than passve tags. Actve tags are deally suted for the

3 dentfcaton of hgh-unt-value products movng through a tough assembly process. They also offer the durablty essental for permanent dentfcaton of captve product carrers. After lookng nto the specfcatons of dfferent avalable systems, we have chosen the Spder System manufactured by Code [8] to mplement the prototype framework. Ther actve tags have a read range of 50 feet. If necessary, ths range can be ncreased to 000 feet wth the addton of a specal antenna. Fgure 2 shows the ID readers and tags used n our system and ther relatve sze compared wth a US quarter. The range that can be acheved n an ID system s essentally determned by [7]: The power avalable at the reader/nterrogator to communcate wth the tag(s). The power avalable wthn the tag to respond. The envronmental condtons and structures (the former beng more sgnfcant at hgher frequences ncludng sgnal to nose rato). Our frst attempt s to nstall a number of readers as shown n Fg. 3. Each reader has a pre-determned power level, thus defnng a certan range n whch t can detect ID tags. By properly placng the readers n known locatons, the whole regon can be dvded nto a number of sub-regons, where each sub-regon can be unquely dentfed by the subset of readers that cover that subregon. Gven an ID tag, based on the subset of readers that can detect t, we should be able to assocate that tag wth a known sub-regon. The accuracy of ths approach s then determned by the number of readers requred, the placement of these readers, and the power level of each reader. Such a ncely formulated optmzaton problem turns out to be useless because the range n whch a reader can detect a tagged object s not just due to the power level (smlar to sgnal strength). There are many factors that wll affect the range ncludng both statc obstructons and dynamc human movement. Due to these dynamc nterferences, even a statc object could be reported n dfferent sub-regons from tme to tme. Ths s the same reason why approaches based on the sgnal strength of 802.b are not very useful. Fgure 2: The ID reader and tag used n our prototype system. The feld or wave delvered from an antenna extends nto the space surroundng t and ts strength dmnshes wth respect to dstance. The antenna desgn wll determne the shape of the feld or propagaton wave delvered, so that range wll also be nfluenced by the angle subtended between the tag and antenna. In space free of any obstructons or absorpton mechansms, the strength of the feld reduces n nverse proporton to the square of the dstance. In our system, the ID Reader s operatng frequency s 308 MHz. It also has an 802.b nterface to communcate wth other machnes. The detecton range s 50 feet. The reader provdes dgtal control of read range va provdng confguraton software and API wth 8 ncremental read ranges. Each reader can detect up to 500 tags n 7.5 seconds. Each ID tag s pre-programmed wth a unque 7-character ID for dentfcaton by readers. Its battery lfe s 3-5 years. Tags send ther unque ID sgnal n random wth an average of 7.5 seconds. Note that the ID reader has 8 dfferent power levels. Based on the sgnal strength receved by the ID reader, the reader wll report or gnore the receved ID, where power level has the shortest range and level 8 has the longest range. Fgure 3: Placement of 9 readers wth two dfferent ranges and the sub-regons. IV. LANDMARC APPROACH In order to ncrease accuracy wthout placng more readers, the LANDARC (Locaton Identfcaton based on Dynamc Actve ID Calbraton) system employs the dea of havng extra fxed locaton reference tags to help locaton calbraton. These reference tags serve as reference ponts n the system (lke landmarks n our daly lfe). The proposed approach has three major advantages. Frst, there s no need for a large number of expensve ID readers. Instead we use extra, cheaper ID tags. Second, the envronmental dynamcs can easly be accommodated. Our approach helps offset many envronmental factors that contrbute to the varatons n

4 detected range because the reference tags are subject to the same effect n the envronment as the tags to be located. Thus, we can dynamcally update the reference nformaton for lookup based on the detected range from the reference tags n real-tme. Thrd, the locaton nformaton s more accurate and relable. The LANDMARC approach s more flexble and dynamc and can acheve much more accurate and close to real-tme locaton sensng. Obvously, the placement of readers and reference tags s very mportant to the overall accuracy of the system. The LANDMARC approach does requre sgnal strength nformaton from each tag to readers, f t s wthn the detectable range. However, the current ID system does not provde the sgnal strength of tags drectly to readers. Readers only report the power level ( to 8 n our system) of the tag detected. We mght do a prelmnary measurement to know whch power level corresponds to what dstance. However, ths may work only n free space. As ndcated earler, the power level dstrbuton s dynamc n a complcated ndoor envronment. Thus, the physcal dstance cannot be computed accurately by usng power levels drectly. We have to develop an algorthm to reflect the relatons of sgnal strengths by power levels. A. System Setup The prototype envronment conssts of a sensng network that helps the locaton trackng of moble users/objects wthn certan granularty and accuracy, and a wreless network that enables the communcaton between moble devces and the Internet. The sensng network prmarly ncludes the readers and Tags as mentoned earler. The other major part of the nfrastructure s the wreless network that allows wreless communcaton between moble devces lke PDAs and the Internet. In addton, t also acts as a brdge between the sensng network and the other part of the system. As the reader s equpped wth the capablty of communcatng wrelessly usng IEEE 802.b wreless network, all the tag nformaton gathered from readers s sent over to the suppled API sttng on a specfc server (the locaton server). Ths feature does not have the problem of havng a wre-connecton to the readers, thus reducng the possble restrctons of where the readers could be placed. In addton, the wreless network wll serve as the fundamental framework of all the communcatons n the nfrastructure. To be able to track an object s locaton, the locaton nformaton receved from the Readers has to be processed before beng useful. The followng s a bref explanaton of some of the major confguraton values n the API software: Devce ( Readers) Setup: Used for confgurng the IP addresses of the Readers. Range: Used for specfyng what range for tags s to be scanned. Mode (Excepton versus Contnuous): () Excepton mode: The reader wll report the tag when t s nsde the detected range whle t wll not report agan untl the reader realzes the tag has gone out of range. (2) Contnuous mode: The reader wll contnuously report the tag ID as long as t was n the confgured range Tme/tag lmt per log fle: Used to confgure how long and how much tag events recorded before the API wll start a new log fle. Ths s n fact somewhat crtcal to the confguraton n the sense of ts effect on effcency. After the sgnal s receved by the readers, the readers then report the nformaton to TagTracker Concentrator LI (a software program/api provded by Code, Inc.) va a wred or wreless network. Moreover, the software also acts as a central confguraton nterface for the readers. For example, t can be used to adjust the detecton range and rate of the readers. After the nformaton from the readers s processed by the TagTracker Concentrator LI, the processed locaton nformaton can be buffered locally as a fle on the same machne or transmtted va a network socket (confgurable n the API). B. Methodology Suppose we have n readers along wth m tags as reference tags and u trackng tags as objects beng tracked. The readers are all confgured wth contnuous mode (contnuously reportng the tags that are wthn the specfed range) and a detecton-rang of -8 (meanng the reader wll scan from range to 8 and keep repeatng the cycle wth a rate of 30 seconds per range). We defne the Sgnal Strength Vector of a trackng/movng tag as r S = ( S, S2,..., Sn ) where S denotes the sgnal strength of the trackng tag perceved on reader, where (, n). For the reference tags, we denote the correspondng Sgnal Strength vector as v θ = ( θ, θ 2... θ n ) where θ denotes the sgnal strength. We ntroduce the Eucldan dstance n sgnal strengths. For each ndvdual trackng tag p, n 2 where p (, u), we defne: E j = (θ S ) = where j (, m), as the Eucldan dstance n sgnal strength between a trackng tag and a reference tag r j. Let E denotes the locaton relatonshp between the reference tags and the trackng tag,.e., the nearer reference tag to the trackng tag s supposed to have a

5 smaller E value. When there are m reference tags, a v trackng tag has ts E vector as E = ( E, E2,... Em ) Ths algorthm s to fnd the unknown trackng tags nearest neghbors by comparng dfferent E values. Snce these E values are only used to reflect the relatons of the tags, we use the reported value of the power level to take the place of the value of sgnal strength n the equaton. There are three key ssues we examne through the process of locatng the unknown trackng tags. The frst ssue s the placement of the reference tags. Snce the unknown tag s ultmately located n a cell surrounded by some reference tags, the layout of reference tags may sgnfcantly affect the locaton accuracy of an algorthm. The second ssue s to determne the number of reference tags n a reference cell that are used n obtanng the most accurate approxmate coordnate for each unknown trackng tag. For example, the smplest way to fnd the nearest reference tag to the trackng tag s to use the coordnate of the reference tag wth the smallest E value as the unknown tag s coordnate. We call ths as -nearest neghbor algorthm. Or, we can choose a trackng tag s two nearest neghbors and call t 2-nearest neghbor algorthm. When we use k nearest reference tags coordnates to locate one unknown tag, we call t k-nearest neghbor algorthm. The unknown trackng tag coordnate (x, y) s obtaned by: ( x, y) = k = w ( x, y ) where w s the weghtng factor to the -th neghborng reference tag. The choce of these weghtng factors s another desgn parameter. Gvng all k nearest neghbors wth the same weght (.e., w = /k) would make a lot of errors. Thus, the thrd ssue s to determne the weghts assgned to dfferent neghbors. Intutvely, w should depend on the E value of each reference tag n the cell,.e., w s a functon of the E values of k-nearest neghbors. Emprcally, n LANDMARC, weght s gven by: E w j = k 2 = E Ths means the reference tag wth the smallest E value has the largest weght. Note that our approach can be easly extended to a 3-dmensonal coordnate. 2 whle the other 8 tags (u=8) as objects beng tracked, as llustrated n Fgure 4a. Wth the setup, the data are collected va the socket from the TagTracker Concentrator LI n groups of a onehour perod and the system wll compute the coordnates of the trackng tags based on each group of data. To quantfy how well the LANDMARC system performs, the error dstance s used as the bass for the accuracy of the system. We defne the locaton estmaton error, e, to be the lnear dstance between the trackng tag s real coordnates ( x0, y0 ) and the computed coordnates ( x, y), gven by e = ( x x y ) + ( y 0 ) Wth the placement of the reference tags and the trackng tags shown n Fgure 4 for over 48 hours, we keep collectng data of the power levels from 4 readers contnuously. Thus we obtan 48 groups of one-hour data. For each of 8 trackng tags per hour, the system computes the coordnates of ths tag by usng the algorthm dscussed n Secton 4. We then compute the locaton error e for each trackng tag. Thus, we have 48 groups of 8 e values. We may examne the locaton accuracy by analyzng the dstrbuton of these e values under dfferent condtons. Note that we have repeated the experments many tmes to avod statstcal errors. A. Effect of the Number of Nearest Neghbors One of the key ssues s to fnd a best k value n the algorthm. We choose dfferent k values as k=, 2, 3, 4, Reference Tag Trackng Tag Reader m m 4 m 3 m 2 m m 9 m 8 m 7 m 6 m 5 m 4 m 3 m 2 m m (0,0) m m Reference Tag Trackng Tag Reader 4 m 3 m 2 m m 9 m 8 m 7 m 6 m 5 m 4 m 3 m 2 m m (0,0) V. EXPERIMENTAL RESULTS AND PEORMANCE EVALUATION We conduct a seres of experments to evaluate performance of the postonng of the LANDMARC System. In the standard setup, we place 4 readers (n=4) n our lab and 6 tags (m=6) as reference tags Fgure 4a: Placement of readers and tags (standard placement) Fgure 4b: Placement of readers and tags ( placement confguraton 2 )

6 cumulatve % e (meters) k=2,av e=.47, Worst=2.68 k=3, Av e=.3, Worst=.98 k=4, Av e=.09, Worst=.8 k=5, Av e=.3, Worst=.99 F gure 5: Cumulatve percentle of error dstance for k from 2 to 5. and 5 and compute the coordnates of the trackng tags, respectvely. Fgure 5 shows the results of usng dfferent k values n the formula. As shown n Fg. 5, k=4 works the best and the postonng accuracy does not mprove as the k value further ncreases. Keepng the same placement, we repeat the process for another 48 hours. Though the postonng error dstrbuton changes, k=4 stll gves the best locaton nformaton. In fact, n all the later experments except on a few occasons that k=3 and k=5 worked better, n most cases k=4 s the best choce Hence, we set 4 as the value of k n our formula n the followng experments. Based on the statstcs, t can be seen that the 50 percentle has an error dstance of around meter whle the maxmum error dstances are less than 2 meters. Ths s very promsng because the 50 percentle of the RADAR project s around 2.37m-2.65m and ts 90 percentle s around 5.93m-5.97m [9]. B. Influence of the Envronmental Factors In order to see how well the LANDMARC approach works n dfferent envronments, we collect 0 groups of data from mdnght to early mornng (durng whch tme there s lttle movement) and another 0 groups of data from 0 AM to 3:00 PM (at whch tme varyng level of actvtes that would result n varatons n transmsson of the tags). Fgure 6 shows the comparson. We know that durng the daytme, the lab s very busy wth many people so there s more nterference than at nght. From the results, we do not see much dfference n the overall accuracy. Ths shows that our reference tag approach can successfully offset the dynamcs of nterference. As the postons of trackng tags n the real world would be unpredctable, we change the placements of trackng tags randomly and expect the dstrbuton of e could be changed but the accuracy of the system should be at the same level. We change the placement of the trackng tags as shown n Fg. 4b wth the reference tags placement unchanged and repeat the process. cumulatve % cumulatve % 0 5 Fgure 6: Cumulatve percentle of error dstance n the daytme and at nght e(meters) e(meters) Daytme, Worst=.956 Nght, Worst=.783 orgnal setup, Worst=.8 change TrkTag, Worst=.82 Fgure 7: Cumulatve percentle of error dstance between two trackng tag placement confguratons n Fg 4. Fgure 7 s the comparson of the results between the two trackng tag placements shown n Fg. 4. As we expected, the dstrbuton s changed but the overall accuracy s at the same level. Fgures 6 and 7 show that the approach of usng reference tags effectvely helps offset some of the envronmental factors that contrbute to the varatons n a detected range. Snce the reference tags are subject to the same effect n the envronment as the tags to be located, we can dynamcally update the reference nformaton for lookup, based on the detected range from the reference tags n real-tme. C. Effect of the Number of Readers One of the problems of usng to locate objects s the nconsstency of the sgnal strength recepton. Ths can prmarly be due to the envronment and the devce tself. In most cases, the envronmental factors always have the most mpact on the accuracy and maxmum detectable range. These nclude ssues lke furnture placement, people s movement, and so on. Besdes, non-lne of sght (NLOS) s another source of reducng the locaton sensng accuracy. Even NLOS does not prohbt transmsson as that of nfrared, t does create the mult-path problem, meanng the sgnal can possbly take dfferent paths to reach the recever and result n nterference among the receved sgnals.

7 cumulatve % e (meters) 4 readers data, Worst=.8 3 readers data, Worst=2.59 Fgure 8: Cumulatve percentle of error dstance for 3 and 4 readers. To better deal wth the problem, we can use more readers to mprove the accuracy. Wth more readers, a better decson can be made for locaton sensng because more data can be gathered by havng extra readers to do the sensng as shown n Fg. 8. However, the readers are usually qute expensve so placng more readers means extra costs for the whole system. Due to budget constrants, we have only four readers. Addng more readers may not necessary sgnfcantly ncrease the accuracy. It does ncrease the processng overhead. D. Effect of Placement of Reference Tags Intutvely, placement of reference tags should have an effect on the measurement accuracy. Consder the two confguratons shown n Fg. 9. In the case of Fg 9a, where there s not any obstacle, t s probable that the system can easly fnd the trackng tag s four nearest neghbors whch are tags e, f, h, by comparng the reported sgnal strengths and E f could be the smallest n ths trackng tag s E vector. Thus, the trackng tag could be located among the four reference tags. However, sometmes the envronment could be more complcated. Suppose there s a partton (or sometmes even a person standng lke the partton) as shown n Fg. 9b. Under these crcumstances, t s possble that the recepton of the sgnal strength from the reference tag f s nfluenced by the partton (or the unexpected people). Consequently, the readers wll report a weaker sgnal strength from tag f so that tag f could fal to be ncluded n the four neghbors of the trackng tag. Instead, tag k may become one of the four reported nearest neghbors to the trackng tag as shown n Fg. 9b. Usng e, h,, k as the four reported neghbors, the poston of the trackng tag s lkely to be computed as ndcated n Fg. 9b. Thus, more error occurs. Thngs wll change f we place more reference tags as llustrated n Fgure 0. Around the trackng tag, now we have placed more reference tags (the green ones n the Fgure). Together wth the tag, tags m, n, o could be ncluded n the four reported nearest neghbors of the trackng tag. Thus, better locaton nformaton wll be provded. In our next experment, we place all of the reference tags wth a hgher densty as shown n Fgure. In the frst 48 hours we keep the orgnal postons of all the trackng tags unchanged (case Near n Fg. ). In the next 48 hours we move the postons of the trackng tags as ndcated n the case of Near 2 n Fg.. a b c Reader a b c Reader Partton P a b c Reader d e f Four Nearest trackng tag d e f real poston Four nearest d e f m n g h g h computed poston g h o Reader2 j k l (a) wthout a physcal partton Reader2 j k l (b) wth a physcal partton Reader2 j k l Orgnal Reference Tags New Reference Tags Trackng Tag Fgure 9: A physcal partton to separate reference tags c and f from others Fgure 0: More reference tags are used.

8 m m m m m m m m Reference Tag Trackng Tag Reader Reference Tag Trackng Tag Reader near near 2 Fgure : Two hgher densty, comparng wth those n Fg. 4, placements of reference tags far far 2 Fgure 3: Two lower densty, comparng wth those n Fg. 4, placements of reference tags cumulatve % e(meters) Orgnal, Worst=.8 near, Worst=.76 near 2, Worst=.69 cumulatve % e (meters) Orgnal, Worst=.8 far,worst=2.59 far 2,Worst=3.7 Fgure 2: Cumulatve percentle of error dstance wth a hgher reference tag densty It can be seen n Fg. 2 that the accuracy of the LANDMARC System s mproved wth a hgher reference tag densty, as we have dscussed. However, the mprovement s not as great as we expected. We wll dscuss ths pont later. Fgure 3 shows two confguratons of a lower reference tag densty. The correspondng dstrbuton of error dstance s shown n Fg. 4. As expected, the accuracy has dropped qute sgnfcantly. It s obvous that the accuracy of the LANDMARC approach decreases greatly wth a lower densty of reference tags. Thus, there s a tradeoff between the accuracy and the number of reference tags. Concevably, we can mprove the accuracy of the LANDMARC System by placng as many reference tags as we can, for example, even n every cubc square centmeter of the space where we want to locate objects. Ths does not make much sense due to the ncreased complexty, overheads, nherent devce error, and measurement error. The expermental results have ndcated that the LANDMARC approach works well. Usng 4 readers n our lab, we roughly need one reference tag per square meter to accurately locate the objects wthn the error dstance such that the worst error s 2 meters and the average s about meter. Fgure 4: Cumulatve percentle of error dstance wth a lower reference tag densty We beleve that the accuracy can be greatly mproved f ID vendors can make some desgn changes to be dscussed n the next secton. VI. CONCLUSIONS AND FUTURE RESEARCH Ths paper has presented a prototype ndoor locaton sensng system usng actve ID. Although ID s not desgned for ndoor locaton sensng, the proposed LANDMARC approach does show that actve ID s a vable cost-effectve canddate for ndoor locaton sensng. However, there are three problems that ID vendors have to overcome n order to compete n a new and growng market. The frst problem s that none of the currently avalable ID products provdes the sgnal strength of tags drectly. Instead, the reader reports detectable or not detectable n a gven range. Ths forces LANDMARC to spend approxmately one mnute each tme to scan the 8 dscrete power levels and to estmate the sgnal strength of tags. By sendng the sgnal strength nformaton drectly from readers, t wll not only elmnate unnecessary processng tme, but also reduce errors. Ths feature can easly be added as readers do have the sgnal strength nformaton from tags.

9 The second problem s the long latency between a trackng tag beng physcally placed to ts locaton beng computed by the locaton server. There are two factors contrbutng to ths long latency. One s the scannng tme of dfferent power levels as ndcated above. Ths factor can be elmnated by sendng the sgnal strength drectly. The second factor s the tme nterval of emttng two consecutve IDs from an actve tag. Our product has an average nterval of 7.5 seconds n order to avod sgnal collson for handlng up to 500 tags. Dependng on the total number of tags expected n a detectable area, ths tme nterval can be further reduced. ID vendors should provde a mechansm to allow users to reconfgure the tme nterval. The thrd problem s the varaton of the behavor of tags. When employng the LANDMARC approach, the basc assumpton s that all tags have roughly the same sgnal strength n emttng the sgnal. In our experments we found that the power level detected by the same reader from two tags n an dentcal locaton may be dfferent. A possble explanaton for the dfference may be due to the varaton of the chps and crcuts, as well as batteres. In fact, before our experments, we had conducted a seres of pre-tests to classfy tags based on ther sgnal strength. Repeated experments are needed to classfy these tags due to the potental sgnal collson causng a detectable tag not detectable. Ths s another factor decreasng the accuracy of the system. If all the above problems can be overcome, the accuracy and latency wll be greatly mproved. However, the error due to the dynamcs of the envronment can hardly be allevated. A major advantage of LANDMARC s to help offset some of the envronmental factors that contrbute to the varaton of accuracy of locatng objects. However, the dynamc envronment s stll one of the man reasons for ncreasng measurement errors, as we can never guarantee all the nearest neghbors and the trackng tag tself are always nfluenced equally. Sometmes a person standng n front of a trackng tag may greatly ncrease the error dstance of locatng ths tag and such an error s often unpredctable. In our experments all reference tags are organzed n a grd array. Ths may explan the reason of usng 4 nearest neghbors. The nfluence of havng other shapes of reference tags to the selecton of the number of nearest neghbors wll be nvestgated. Our methodology can easly be appled to 3D coordnates. The accuracy of the 3D case wll also be studed. We are also nvestgatng the use of Bluetooth for locaton sensng based on the same methodology. Each Bluetooth devce emts a 48-bt unque ID, but wth a shorter range of up to 0 meters. As Bluetooth, lke ID, s not desgned for locaton sensng, t also has some nherent problems. The mplementaton of a locaton-sensng system based on Bluetooth s currently takng place n our lab. REFERENCES [] Jeffrey Hghtower and Gaetano Borrello, A Survey and Taxonomy of Locaton Sensng Systems for Ubqutous Computng, CSE , Unversty of Washngton, Department of Computer Scence and Engneerng, Seattle, WA, Aug 200, hghtower200survey/hghtower200survey.pdf. [2] Garmn Corporaton. About GPS. Webste,200, [3] P. Bahl and V. N. Padmanabhan, RADAR: An Inbuldng -based User Locaton and Trackng System, Proceedngs of IEEE INFOCOM 2000, Tel- Avv, Israel (March 2000), mcrosoft.com/~padmanab/papers/nfocom2000.pdf. [4] R. Want et al., The Actve Badge Locaton System, ACM Trans. Informaton Systems, Jan. 992, pp [5] The Bat Ultrasonc Locaton System. [6] Rado Frequency Identfcaton (ID) home page. [7] AIM (Automatc Identfcaton Manufacturers), the trade assocaton for the Automatc Identfcaton and Data Collecton (AIDC) ndustry, Gettng Started n ID - A Step approach, 2002, s/papers/rfd_bascs_prmer.htm [8] Code, Inc., [9] P. Bahl, V. N. Padmanabhan, and A. Balachandran, Enhancements to the RADAR User Locaton and Trackng System, Mcrosoft Research Techncal Report, February 2000, mcrosoft.com/~bahl/papers/pdf/msr-tr pdf. [0] Nssanka B. Pryantha, Ant Chakraborty, and Har Balakrshnan. The crcket locaton-support system. In Proceedngs of MOBICOM 2000, pages32-43, Boston, MA, August ACM, ACM Press. [] Jeffrey Hghtower, Roy Want, and Gaetano Borrello, SpotON: An Indoor 3D Locaton Sensng Technology Based on Sgnal Strength, UW CSE , Unversty of Washngton, Department of Computer Scence and Engneerng, Seattle, WA, Feb 2000, htower2000ndoor/hghtower2000ndoor.pdf. [2] Andy Harter, Andy Hopper, Pete Steggles, Any Ward, and Paul Webster. The anatomy of a contextaware applcaton. In Proceedngs of the 5th Annual

10 ACM/IEEE Internatonal Conference on Moble Computng and Networkng (Mobcom 999), pages 59-68, Seattle, WA, August 999, ACM Press. [3] Jeffrey Hghtower, Chrs Vakl, Caetano Borrello, and Roy Want, Desgn and Calbraton of the SpotON AD-Hoc Locaton Sensng System, UW CSE , Unversty of Washngton, Department of Computer Scence and Engneerng, Seattle, homes/jeffro/pubs/hghtower200desgn/hghtower2 00desgn.pdf [4] Maro Chesa, Ryan Genz, Franzska Heubler, Km Mngo, Chrs Noessel, Natasha Sopeva, Dave Slocombe, Jason Tester, ID, March 04, 2002, c.noessel/id/research.html [5] AEtherwre. [6] Jeffrey Hghtower, Roy Want, and Gaetano Borrello, "SpotON: An Indoor 3D Locaton Sensng Technology Based on Sgnal Strength," UW CSE , Unversty of Washngton, Seattle, WA, Feb