FUSION OF DATA FROM INERTIAL SENSORS, RASTER MAPS AND GPS FOR ESTIMATION OF PEDESTRIAN GEOGRAPHIC LOCATION IN URBAN TERRAIN

Transcription

1 Metrol. Meas. Syst., Vol. XVIII (0), No., pp METROLOGY AND MEASUREMENT SYSTEMS Inex , ISSN FUSION OF DATA FROM INERTIAL SENSORS, RASTER MAPS AND GPS FOR ESTIMATION OF PEDESTRIAN GEOGRAPHIC LOCATION IN URBAN TERRAIN Przemyslaw Baransk, Macej Polanczyk, Pawel Strumllo Techncal Unversty of Loz, Insttute of Electroncs, /5 Wolczanska, Loz, Polan ) Abstract An electronc system an an algorthm for estmatng peestran geographc locaton n urban terran s reporte n the paper. Dfferent sources of knematc an postonng ata are acqure (.e.: accelerometer, gyroscope, GPS recever, raster maps of terran) an jontly processe by a Monte-Carlo smulaton algorthm base on the partcle flterng scheme. These ata are processe an fuse to estmate the most probable geographcal locaton of the user. A prototype system was esgne, bult an teste wth a vew to ang bln peestrans. It was shown n the conucte fel trals that the metho yels superor results to sole GPS reaouts. Moreover, the estmate locaton of the user can be effectvely sustane when GPS fxes are not avalable (e.g. tunnels). Keywors: peestran navgaton, ata fuson, GPS, gtal maps, partcle flterng, electronc travel as. 0 Polsh Acaemy of Scences. All rghts reserve. Introucton Relable an accurate estmaton of outoor personal postonng s an mportant problem n many professonal an lesure actvtes. Applcatons are numerous, e.g.: trackng of solers or fremen n acton [], ang travel for the vsually mpare [], locatng the whereabouts of chlren an the elerly [3], or navgatng toursts aroun ponts of nterest [4]. In open outoor spaces, an obvous soluton s to use the Global Postonng System (GPS) that, together wth ata erve from gtal maps, offers a powerful postonng an navgaton tool. Car navgaton evces mostly mplement such solutons. The avalable software s capable of mprovng postonng accuracy by usng a pror nformaton concernng the layout of roa networks. However, a peestran navgaton task poses a more complex challenge [5, 6, 7]. Frstly, the path the peestran walks s not confne to a pror gven routes. Seconly, there are few maps avalable that prove etale nformaton about pavement layout n urban terran (e.g. unergroun pathways, large squares, parks etc.). Thrly, the moblty pattern of a peestran s more complex than that of the vehcles movng along roa lanes. Hence, the routng algorthms use n car navgaton proceures cannot be rectly apple n peestran navgaton. Fnally, n the urban envronments, peestrans walk alongse bulngs walls makng the GPS reaouts even less accurate (the postonng errors can reach several ozens of meters). The ncate problems of peestran navgaton tasks call for atonal sources of postonng ata that coul be fuse wth GPS reaouts an enhance the navgaton performance of a peestran. There have been a number of reporte systems ang GPS base peestran navgaton that can be groupe nto two major types: Artcle hstory: receve on Nov. 6, 00; receve n revse form on Feb. 0, 0; accepte on Feb. 8, 0; avalable onlne on Mar., 0.

2 P. Barańsk et al.: FUSION OF DATA FROM INERTIAL SENSORS, RASTER MAPS AND GPS FOR ESTIMATION OF PEDESTRIAN smart envronment systems: n whch a ecate nfrastructure s embee nto the urban terran or alternatvely an alreay exstng telecommuncaton system (GSM base statons) are use for postonng, wearable sensors that are carre by the navgate user. In the smart envronment systems, a network of electronc tags (rao, ultrasoun or nfrare beacons) s embee nto the envronment. Example solutons employ Rao Frequency Ientfcaton Devces (RFID) [8] or nfrare base systems lke the Talkng-Sgns [9] esgne for navgatng the vsually mpare. The peestran poston s tracke by the beacons whose lanmark s known. Unfortunately, such systems are costly n mplementaton. Large number of tags nee to be mounte an servce, although the unt cost of such evces s consstently roppng. In smart envronment systems, a peestran s poston can be etermne by multlateraton algorthms, smlar to the ones use n the GPS [0]. Wearable sensors are becomng an attractve moble soluton for ang peestrans. Currently, such evces as accelerometers, electronc gyroscopes an compasses, gtal cameras an barometers are of small sze an consume low power. Some of the evces can be ntegrate nto cloths. The wearable sensors allow estmaton of a peestran s knematcal parameters, hs carnal rectons or local postonng aganst the surrounng objects. However, ue to lmte accuracy of the sensors an an teratve manner of moton estmaton, specal ea-reckonng algorthms nee to be employe. Dea-reckonng can be efne as the task of etermnng own (e.g. peestran) poston on the bass of a prevous fx an current knematcal parameters (e.g. spee, movement angle). Ae navgaton can only be successfully mplemente f varous sources of postonng or knematc ata can be sutably fuse an contnuously upate for effcent poston trackng of a movng object [0]. The paper proposes an teratve Monte-Carlo metho (also known as partcle flterng) that fuses an processes the ata from an accelerometer, gyroscope, GPS recever an raster maps. The accelerometer serves as a step meter for calculatng the number of steps as well as estmatng ther length []. The gyroscope proves the angular velocty whch s use to calculate a relatve rotaton of the traveller. Shoul the GPS fxes verge, the user s poston s precte n a ea-reckonng manner from the step meter an gyroscope reaouts. The raster map of the terran s converte nto a probablty map, n whch fferent probablty values are assgne to the fferent regons the user passes (e.g. who s more lkely to walk along a pavement than traverse walls, fences or pons). In a number of fel trals, t was shown that the bult prototype system an the propose ata fuson algorthm sgnfcantly mprove the estmaton of a peestran s geographc locaton n urban terran.. System components The system block agram s shown n Fg.. A small portable laptop serves the purpose of a computatonal unt. The system can estmate locaton on-lne as well as store ata from the sensors for off-lne analyss. All the sensors (accelerometer, gyroscope an GPS) are encompasse n a small PCB boar ( Sensor boar see Fg. ) that connects to the PCB va an USB nterface. The sensors boar can work nepenently storng ata on a Mcro-SD car. The boar can be then connecte to the PC an all recore ata can be rea off. Fg. shows a pcture of the sensors PCB. The PC recors on-lne the ata stream from the sensors boar an processes them. The accelerometer reaouts are gtal-fltere to estmate the number of steps an ther length. The angular velocty, prove by the gyroscope, s ntegrate an a relatve rotaton angle s 46

3 Metrol. Meas. Syst., Vol. XVIII (0), No., pp calculate. The step lengths an rotatons thus etermne a relatve user splacement, whch s use to prect the geographcal locaton shoul the GPS fxes be naccurate. Fg.. System block agram, ACC an GYR stans for accelerometer an gyroscope respectvely an uc enotes a mcrocontroller. Fg.. The sensors boar (65mm 48mm 9mm). The large cube s a gtal 6DOF sensor (Analog Devces ADIS6355) that houses 3-axs accelerometer an 3-axs gyroscope. The smaller cube s a GPS recever (wth an ntegrate antenna). The bottom layer comprses a Mcro-SD slot an a battery connector for the nepenentwork moe. When connecte to a PC, the power s rawn from the USB nterface... The GPS recever The FGPMMOPA6B GPS recever comes wth an ntegrate ceramc antenna of compact measurements (6mm x 6mm). Apart from the locaton estmates, every GPS recever wthn the NMEA protocol proves a parameter that reflects the qualty of a fx.e. HDOP (Horzontal Dluton of Precson) [0]. Larger values of ths parameter correspon to less accurate estmates. Ths parameter s crucal wth a vew of the whole algorthm that s to take ecson whether to beleve the GPS recever or prect the locaton from the step meter, gyroscope an raster maps. We categorze the value of HDOP parameter nto three ranges: HDOP <.0 accuracy better than 5m HDOP <.0 meocre accuracy, better than 5m HDOP.0 poor accuracy, worse than 5m. 47

4 P. Barańsk et al.: FUSION OF DATA FROM INERTIAL SENSORS, RASTER MAPS AND GPS FOR ESTIMATION OF PEDESTRIAN These values ffer greatly from recever to recever. We teste three GPS recevers: HACOM HI-04III, Pentagram Pathfner P305 an FGPMMOPA6B whch returne competently fferent values n the same locatons at the same tme. It was notce that the HDOP parameter, n many cases, oes not reflect the actual precson, whch ntally complcate the mplementaton of the algorthm. To account for such events, after a pero of poor qualty estmates, we ntrouce the so calle guar tme,.e. the value of HDOP s fltere so that t oes not abruptly change from hgh to low values. The followng smple rule was apple for HDOP flterng: IF (current_hdop < prevous_hdop) THEN current_hdop := prevous_hdop + A. (prevous_hdop current_hdop) Factor A ranges from 0.05 to 0. an epens on the GPS recever moel. Lower values of ths parameter correspon to a longer guar pero. In Fg. 3 an example of flterng the HDOP value by usng the propose rule s shown. Fg. 3. Flterng the HDOP parameter: thn lne correspons to HDOP values obtane from the GPS recever an thck lne enotes the flterng result. The mplementaton of ths non-lnear flterng metho has mprove overall system performance. However, unque stuatons take place n whch naccurate values of HDOP confuse the system... The accelerometer The ata from the accelerometer s use to estmate the length an number of steps of a person. The sample rate s 50Hz. The resoluton of the reaouts s 4-bts per sample. As a person walks the boy moves up an own n accor wth the steps. The step etecton algorthm examnes the acceleraton of a boy n the vertcal recton only (.e. parallel to the gravty axs). The algorthm estmates step lengths wth.5% accuracy when the user walks on a flat groun. The step etecton algorthm s shown n Fg. 4. A etale explanaton of the algorthm together wth the tests performe on a number of peestrans follow n the authors paper []. The peometer s senstve to orentaton, as t measures the acceleraton n a gven axs to estmate the step length. Any evaton ø from the recton of the gravty wll scale the accelereometer s reaouts by cos(ø ), as t s clarfe by Fg. 5. Accorng to the step estmaton algorthm, the lengths of steps wll be scale by cos 0.5 (ø ). The estmate step length s then gven by formula: 48

5 Metrol. Meas. Syst., Vol. XVIII (0), No., pp = K ( a a ) = K ( a a ) cos ( ϕ ), () m max m mn where: a mmax an a mmn enote the measure peaks n the acceleraton (cf. Fg. 4), a gmax an a gmn stan for the peaks of acceleraton n the gravty axs. g max g mn Fg. 4. Step etecton algorthm. The acceleraton samples are fltere to reject unesre consttuents lke boy swngs an so on. A maxmum an mnmum of acceleratons are etecte an multple by a scalng factor K, whch s nvual for a gven person. Then the result s rase to an emprcal power of 0.5. a g φ a m Fg. 5. A sensor msalgnment causes that the sensors measures the acceleraton a m whch s equal a g cos(φ ). The error ue to the sensor msalgnment s gven by the followng formula: ϕ K ( ag = cos max a 0.75 g mn ( ϕ ) ) 0.5 sn( ϕ ). () For small reasonable angles ø, the algorthm accuracy s not sgnfcantly compromse. For example, for ø =0 o, the step length s scale by It was notce urng tests that trppng over ha an nfluence on the algorthm performance. These events cause the sensor to regster to 4 spurous steps. Moreover, bln testers were notce to make steps n a place when they were not sure were to go. The latter error can be elmnate by measurng the acceleraton n the walkng recton. Ths soluton has yet to be mplemente. The accelerometer reaouts are perturbe by whte nose []. As the stance estmaton s base on the prevous one an the current step length, the whole process s escrbe by a ranom walk, whose varance grows unbounely over tme. Ths wll be explane n the next chapter whch covers the gyroscope accuracy. The accelerometer bas nstablty oes not play an mportant role, because the step lengths are calculate as the fference between subsequent extrema. Calculatng the splacement of the user by ntegratng twce the acceleraton wll quckly lea to enormous errors. Any bas that may appear ue to e.g. sensor msalgnment or temperature change, results n the splacement error growng quaratcally wth tme. 49

6 P. Barańsk et al.: FUSION OF DATA FROM INERTIAL SENSORS, RASTER MAPS AND GPS FOR ESTIMATION OF PEDESTRIAN.3. The gyroscope The ata from the gyroscope s ntegrate to calculate a relatve rotaton angle φ, accorng to the followng formula: t 0 ( ω + ) t φ = ω, (3) where: t s the ntegraton tme,.e. the samplng nterval, ω s the angular velocty of a boy, ω rft s the angular rft velocty. Every gyroscope exhbts a so calle rft,.e. the sensor shows an angular velocty espte no movement. Ths epens on the temperature, sensor algnment an crcut voltage supply. The manufacturer clams that the gyroscope s temperature compensate. Nonetheless the rft was approx. 0.º/s. Thus n the tme pero of 5mnutes the gyroscope can stray 80º from the true recton. The systematc bas can be easly measure an subtracte. Smlarly to the accelerometer, the gyroscope measurements ω(t) are flawe by whte nose n ω (t). The estmate angle φ(t) s then corrupte by the error expresse by the formula: t n 0 = ( n t) ) t rft ω ( Nω ( ) δt, (4) where: δt s the samplng pero, N ω () are samples of angular spee ω(t), s the measurement number, whereby t=n δt. The varance of the aforementone error (of zero mean) can be calculate accorng to the followng formula: Var n = N n ω ( ) δt = E ( Nω ( ) δt) = n δt E{ Nω ( ) } = n δt σ ω = t δt σ ω =, (5) where: σ ω s the varance of the angular velocty measurements. The varance of the estmate angle σ φ grows bounlessly wth the tme. The manufacturer of the sensor [] states that the evaton σ φ equals 4. 0 / h, where h s gven n hour tme unts. Hence, after one hour the evaton equals σ φ =4. o an after two hours reaches σ φ =5.8 o etc. σ φ grows proportonally wth tme ue to bas nstablty at the level of 54 o /h []. There are other factors, however, of much lesser mportance, that egrae the gyroscope s performance lke: change of gyroscope senstvty ue to temperature, axs nonorthogonalty, nonlnearty, voltage senstvty etc. What shoul be also taken nto account s the sensor msalgnment whch causes the reaouts to be scale, smlarly as wth the accelerometer. Laser gyroscopes mprove by far on MEMS gyroscopes. For example, GG30AN Dgtal Laser Gyro s a sngle axs gyroscope wth a bas nstablty of o /h an an angular ranom walk of / h. The mensons rather squalfy them from moble applcatons 8cm 8cm 4.5cm. By comparson, the ADIS6355 measures cm cm cm. The cost of the laser gyroscope s at least several tmes hgher than the MEMS (Mcro Electro-Mechancal Systems) base evce ADIS6355. The authors performe also tests wth an electronc compass whch proves absolute orentaton. The compass was very susceptble to any magnetc fel stortons cause by cars, power lnes, trams etc. For example, a tram, 30 meters away, renere the compass reaouts useless for aroun 30s []. 50

7 Metrol. Meas. Syst., Vol. XVIII (0), No., pp The raster map The raster map s ve nto three stnct regons wth fferent weghts assgne as follows: forben areas bulngs, pons, walls, fences etc., w map =0., probable areas lawns, fels etc., w map =0.5, preferre areas pavements, streets, squares, alleys etc., w map =.0. Each weght correspons to the probablty of the user beng n a gven regon. The probablty map mtgates the problem of gyroscope rfts an step meter naccuraces. The weght value for the forben regons (w map =0.) was chosen to prove some flexblty for the ata fuson algorthm. Intally we set w=0 for the forben regons. However, cases occurre, ue to naequate HDOP parameter reaouts, that the estmate locatons (partcles) enter the wall an e out ue to the zero weght value assgne. Then all the partcles ha to be ranomly rentalze. Hence, the algorthm requre extra tme to converge to an optmal soluton, n whch case t worsene the overall computatonal performance. The values of the map s weghts were assume emprcally on the bass of fferent trals. The envronment s subject to change, e.g. a new pavement, roa an a gtal map s not always up to ate. The weghts shoul prove some margn for that. A small value for probable areas woul cause that the partcles woul e out quckly e.g. when a user crosses a we street an wants to travel on the other se. The algorthm woul then estmate that the peestran s on the wrong se. Fg. 6 shows an example of a raster probablty map. Fg. 6. A raster map of the Unversty campus. Whte, grey, black correspon to preferre, probable an forben regons respectvely. 3. The partcle flterng algorthm apple to postonng estmaton 3.. Introucton Partcle flterng s an avance moel estmaton technque erve from a sequental verson of Monte Carlo methos [3, 4]. Ths approach uses a smulaton technque that generates a large number of canate solutons (partcles) n search for the best soluton to 5

8 P. Barańsk et al.: FUSION OF DATA FROM INERTIAL SENSORS, RASTER MAPS AND GPS FOR ESTIMATION OF PEDESTRIAN a complex problem. An estmate system state s represente by a large number of partcles. Each partcle unergoes certan rules an then t s assesse how accurately t approxmates the soluton (measurement upate). Therefore, every partcle has an assocate weght (mportance). The fnal system state s a weghte average of partcle states (precton). Wth successve teratons of the smulaton, some partcles assume neglgble weghts an o not effectvely take part n the smulaton. These partcles are replace wth the ones that more accurately converge to the optmal soluton ths s the so-calle resamplng technque [3]. Kalman flters are commonly use for ea reckonng applcatons. The Kalman flter s a recursve algorthm, whereby the current state epens only on the prevous one an the rvng nputs [5]. The flter s much less computaton expensve than a partcle flter. When the object moel s nonlnear, then the Extene Kalman Flter (EKF) s apple. The flter s base on lnearzaton aroun the current state by the Taylor approxmaton. The algorthm performs well when the system exhbts moerate nonlnearty an moerate non-gaussan measurement nose. Due to lnearzaton, the covarance matrx s propagate wth an error whch grows wth every teraton. The algorthm, calle the Unscente Kalman Flter (UKF), eals wth sgnfcant non-lneartes by samplng the probablty strbuton aroun the mean n so-calle sgma ponts [5]. The above scusse technques can effectvely be use when the posteror strbuton s unmoal [6]. Often, there are stuatons when two solutons are possble, for example, a pavement can bfurcate nto two paths of a smlar orentaton. The traveller mght be on ether path but not between them, as a Kalman algorthm woul estmate. One path can be ponte as more probable but at a later tme. For these kns of problems, the ensemble Kalman flter, also referre to EnKF, s employe [7]. The system state s represente by a large collecton of ensemble members. Each member mplements the ea of the Kalman flter an t s nvualze by ang a small perturbaton to ts state. The EnKF works uner the assumpton that all probablty strbutons are Gaussan [8]. As a matter of fact, the EnKF s closely relate to the partcle flter, whch can assume any probablty strbuton, albet at a heavy computatonal cost. The man avantage of a partcle flter s quck mplementaton an flexblty when ang new ata sources. 3.. Implementaton of the partcle flterng algorthm In our smulaton, every partcle c ( s partcle number) that takes part n the smulaton s a hypothetc user locaton, thus t can be escrbe by the followng vector enote by formula 6. c = [ x, y, ϕ, w ], (6) where: x (t), y (t) a partcle locaton n Cartesan coornates, φ (t) an angle assocate wth the partcle,.e. angle versus the coornates, w (t) a partcle weght. In the scusse applcaton the number of partcles must be at least 500 to obtan goo smulaton accuraces. For the algorthm outset, all the partcles are ntalze wth ranomly strbute angles φ (t), an the coornates x (t), y (t) are sprea aroun GPS estmates converte nto Cartesan coornates x GPS (t), y GPS (t). Ths algorthm step s calle ntalzaton. As the user walks, the step meter an gyroscope prove ata about relatve user splacement (t) an φ(t) respectvely. Every partcle s upate wth ths ata. Moreover, the algorthm smulates sensor naccuraces by ntroucng measurement errors. The algorthm generates (ranomly) measurements errors nvually for every partcle. A Gaussan nose moel was use both for the step meter an the gyroscope. A partcle transton from the current to the next state s explane n Fg. 7. 5

9 ( x ( t ), y ( t ) ) Metrol. Meas. Syst., Vol. XVIII (0), No., pp ( x ( t + ), y( t + ) ) ( t + ) + ε ( t +) ( x, y ) ϕ ( t + ) = ϕ ( t ) + ϕ ( t ) + ε ( t ) + ε ϕ Fg. 7. Partcle transton ntrouce by the step meter an gyroscope. (t) s a step length, φ(t) s the relatve rotaton. ε (t) an ε φ (t) are smulate measurement errors for the step meter an gyroscope respectvely. The algorthm generates these values ranomly for every upate partcle. Followng Fg. 7, a partcle transton to a new state s gven by equaton (7). x ( t + ) x + + = + cos( ϕ ϕ( t) ε ) ϕ ( ( t) + ε ( t)). (7) y ( t + ) y sn( ϕ + ϕ( t) + ε ) ϕ Then every partcle s assesse,.e. how accurate t approxmates the user locaton. The assessment s base on GPS measurements an the map. If a partcle encounters a forben regon an n the long run oes not conce wth GPS measurements, t means that the partcle has a wrong orentaton φ (t) or ts locaton (x (t),y (t)) s erroneous. The weght n (t) of ths partcle s accorngly lowere. The weghts of partcles are upate on the bass of GPS measurements accorng to equaton (8). w ( t + ) = w PDFGPS ( δ ( t + )), (8) where PDF GPS (δ (t)) s a probablty ensty functon for the e (t) error. We use a normal error strbuton, expresse by equaton (9). PDF GPS e = ( δ ) exp. (9) σ GPS π σ GPS σ GPS s stanar evaton of GPS measurements. Emprcally we assume that σ GPS s proportonal to the square HDOP parameter. e (t) s the stance between a gven partcle locaton (x (t),y (t)) an the GPS measurement (x GPS (t),y GPS (t)), thus t can be calculate from equaton (0). e t = ( x xgps ) + ( y ygps ( )). (0) The raster map s use also to upate partcle weghts equaton (). w ( t + ) = w w ( x, y ), () w map (x (t),y (t)) s an area weght for gven coornates (x (t),y (t)) (see secton.4 The raster map). The weghts of the partcles are normalze accorng to formula. map ϕ 53

10 P. Barańsk et al.: FUSION OF DATA FROM INERTIAL SENSORS, RASTER MAPS AND GPS FOR ESTIMATION OF PEDESTRIAN w w : = L, () = w where L s the total number of partcles. After a number of algorthm teratons, all but a few partcles have neglgble weghts an therefore o not partcpate n the smulaton effectvely. Ths stuaton s etecte by calculatng the so calle egeneraton ncator gven by equaton (3). = g = L. (3) L w If g(t) falls below a gven threshol of e.g. 0.6, then a process calle resamplng s ncorporate nto the algorthm n whch a new set of partcles s create. The probablty of copyng a partcle to the new set s proportonal to ts weght w (t). Therefore partcles that poorly approxmate the system state are replace by more accurate partcles. Then all partcles are assgne the same weght L -. The estmate system state,.e. the user locaton (x u (t),y u (t)), s a weghte average of all partcle states: 4. Results L L ( x t), y ) x w, y w u ( u =. (4) = = Fg. 8 shows an example path followe by a peestran n urban terran. The route was ca..9km long. The hstogram of geographc estmaton errors for the path s shown n Fg. 9. Fg. 8. A path followe by a peestran: Sol lne true path, crosses estmate path, trangles GPS reaouts. Frst, the absolute errors between the true locaton (rea from a map) an the locaton estmate by means of the apple algorthms are calculate. Next, the errors are sorte wth respect to ther values that were expresse n meter unts. The number of errors fallng nto each hstogram bn s ve by the number of measurements an multple by a factor of 00. From the hstogram shown n Fg. 9, we can conclue that the estmaton algorthm (3) s the most precse, snce more that 70% of locaton estmatons are wthn the range of 0 3m. In other wors, the propose algorthm brngs large errors to smaller range. Hence, the number of smaller errors ncreases. It woul be eal f 00% of the errors fell n the range of 0 3m. 54

11 Metrol. Meas. Syst., Vol. XVIII (0), No., pp Fg. 9. The error hstogram for the path. () sole GPS reaouts, () GPS + nertal sensors + partcle flterng, (3) GPS + nertal sensors + map + partcle flterng. The hstogram clearly shows that the propose metho combnng ata from nertal sensors an raster maps mproves on GPS reaouts. The errors were cut own to the range of up to 0 meters. Ths s a sgnfcant mprovement, but t s stll not enough to bul a evce that woul robustly gue the user on a certan se of the pavement, n whch case the accuracy must be wthn meters. As shown n Fg. 9, mplementaton of nertal sensors mproves on sole GPS reaouts, by elmnatng large errors (ca. 75m cf. Fg. 0). Snce the gyroscope proves only recton change an a new locaton s base on the prevously estmate one, the precaton strays away over tme, as the errors quckly accumulate, e.g. when the recton s estmate wth 0º accuracy, after traversng a 00m stance the precte locaton s ca. 0m from the correct locaton. Moreover, GPS errors n many cases ten to be systematcally base. Introucton of new ata moalty taken from the raster maps gves a synergetc beneft. As the partcles attempt to rft away from the correct recton, they are trmme by the bulng regons cf. Fg. 0. In consequence, the recton s self-algne an the precton precson mproves. Fg. 0. A screenshot of the partcle flterng smulaton: large roun ot GPS locaton reaout, large square estmate user locaton, small squares partcles. Note that the algorthm elmnates the partcles that woul enter forben areas (.e. bulngs, fences, etc.). 55

12 P. Barańsk et al.: FUSION OF DATA FROM INERTIAL SENSORS, RASTER MAPS AND GPS FOR ESTIMATION OF PEDESTRIAN 5. Conclusons an future work Applcaton of nertal sensors an raster maps gves a measurable mprovement n locaton estmaton n urban terran (Fg. 9). The errors were reuce several tmes from ca. 70m to 0m. The presente example path (Fg. 8) was taken from one of the best results that were obtane. On several occasons, however, the partcles entere an ajacent street or a square an encountere a ea-en n a form of connecte bulngs or walls. In these cases the algorthm was nferor to the GPS recever. A relable recovery algorthm nees to be worke out. It was notce urng the tests that the algorthm lacks parallelsm. When the partcles met several smlar recton paths, they accorngly branche, but over tme one path not always the correct one monopolze the partcles. Ths le n many cases to the aforementone ealock. A larger number of partcles allevate the problem, but the computaton buren was prohbtve for a real-tme smulaton. An atonal source of ata s currently consere to further mprove postonng precson. Namely, the stereovson technque s beng teste an compare aganst 3D maps of the terran. The frst successful fel trals were conucte on the Unversty Campus. Acknowlegements Ths work has been supporte by the Natonal Centre for Research an Development of Polan grant no. N R n years References [] Fscher, C., Gellersen, H. (Jan. Mar. 00). Locaton an Navgaton Support for Emergency Responers: A Survey. Pervasve Computng, [] Hersh, M.A., Johnson, M.A. (008). Assstve Technology for Vsually Impare an Bln People. Sprnger- Verlag. Lonon Lmte [3] Helal, S., Mokhtar, M., Abulrazak, B. (Es). (008). The Engneerng Hanbook of Smart Technology for Agng. Dsablty an Inepenence. John Wley & Sons. [4] Arkawa, M., Konom, S. (Jul.-Sept. 007). Navtme: supportng peestran navgaton n the real worl. Pervasve Computng, 9. [5] Retscher, G., Thenelt, M. (004). NAVIO a navgaton an guance servce for peestrans. Journal of Global Postonng Systems, 3( ), [6] Mezentsev, O., Lachapelle, G., Colln J. (005). Peestran ea reckonng a soluton to navgaton n GPS sgnal egrae areas. Geomatca, 59(), [7] Fang, L., Antsakls, P.J., Montestruque, L., McMckell, M.B., Lemmon, M., Sun, Y., Fang, H., Koutrouls, I., Haengg, M., Xe, M., Xe, X. (005). Desgn of a Wreless Assste Peestran Dea Reckonng System The NavMote Experence. IEEE Trans. on Instrumentaton an Measurement, 54(6), [8] Kourog, M., Sakata, N., Okuma, T., Kurata, T. (006). Inoor/Outoor Peestran Navgaton wth an Embee GPS/RFID/Self-contane Sensor System. Proc. Int. Conf. Artfcal Realty an Telexstence CAT006,30-3. [9] The Smth-Kettlewell Eye Research Insttute. (00 November) [0] Farell, J. A. (008). Ae Navgaton. McGraw Hll. New York. [] Baransk, P., Strumllo, P. (00). An electronc moule mplementng ea-reckonng algorthm for mprovng geographc locaton accuracy from GPS reaouts n urban terran. VIII-th Natonal Electroncs Conference June 00. (n Polsh). [] The webste of Analog Devces. (0 February). 56

13 Metrol. Meas. Syst., Vol. XVIII (0), No., pp [3] Arulampalam, M.S., Maskell, S., Goron, N., an Clapp, T. (00). A tutoral on partcle flters for onlne nonlnear/non-gaussan Bayesan trackng. IEEE Trans. Sgnal Processng, 50, [4] Ceranka, S., Neźweck, M. (003). Applcaton of Partcle Flterng n Navgaton System for the Bln. Proceengs of the IEEE Seventh Internatonal Symposum on Sgnal Processng an ts Applcatons. [5] Brown R. (983). Introucton to Ranom Sgnal Analyss An Kalman Flterng. John Wley & Sohns. [6] Cappe O., Gosll S., an Moulnes E. (007). An overvew of exstng methos an recent avances n sequental Monte Carlo. Proceeng of the IEEE, 95, [7] Manel J. (007). A Bref Tutoral on the Ensemble Kalman Flter. Unversty of Colorao at Denver an Health Scences Center. [8] Mackenze D. (October 003). Ensemble Kalman Flters Brng Weather Moels Up to Date. SIAM News, 36(8),