Geonformatcs 2004 Proc. 12th Int. Conf. on Geonformatcs Geospatal Informaton Research: Brdgng the Pacfc and Atlantc Unversty of Gävle, Sweden, 7-9 June 2004 A TEST-BED SIMULATOR FOR GPS AND GIS INTEGRATED NAVIGATION AND POSITIONING RESEARCH: - BUS POSITIONING, USING GPS OBSERVATIONS, ODOMETER READINGS AND MAP MATCHING Jng L 1, George Taylor 2, Chrs Brunsdon 3, Andrew Olden 4, Dörte Steup 5 and Maryln Wnter 6 GIS Research Centre, School of Computng, Unversty of Glamorgan, CF37 1DL, Wales 1 jl@glam.ac.uk, 2 getaylor@glam.ac.uk, 3 cbrunsdo@glam.ac.uk, 4 aolden@glam.ac.uk, 5 dsteup@glam.ac.uk, 6 mwnter@glam.ac.uk Abstract A test-bed applcaton, called Map Matched GPS (MMGPS) processes raw GPS output data, from RINEX fles, or GPS derved coordnates. Ths developed method uses absolute GPS postonng, map matched, to locate the vehcle on a road centre-lne, when GPS s known to be suffcently accurate. MMGPS software has now been adapted to ncorporate postonng based on odometer derved dstances (OMMGPS), when GPS postons are not avalable. Relatve GPS postons are used to calbrate the odometer observatons. In OMMGPS, GPS pseudorange observatons are combned wth DTM heght nformaton and odometer postons to provde a vehcle poston at one second epochs. Generally, odometer postonng s used much more often, to poston the vehcle, than GPS. Typcally the rato s 7:3 odometer postons to GPS postons. In total, over 15,000 vehcle postons were computed usng OMMGPS. The descrbed experment used GPS observatons taken on a bus on a predefned route, hence the correct road s always known. Therefore, map matchng technques are used to mprove the GPS postonng accuracy, and to dentfy grossly naccurate GPS postons. Calbrated odometer correctons are made usng odometer count at the current epoch and relatve GPS dstance travelled. If a GPS poston s detected to be naccurate, t s not used for postonng the bus, or for calbratng the odometer correcton factor. In general the poston qualty provded by GPS alone was extremely poor, due to multpath effects caused by the urban canyons of central London. In the case of one partcular trp, OMMGPS provdes a mean error of poston of 8.8 metres compared wth 53.7 metres for raw GPS alone. INTRODUCTION Global navgaton satellte systems (GNSS), such as the Global Postonng System (GPS) have been ncreasngly used n real tme trackng of vehcles. Especally, when GPS s ntegrated wth ever powerful geographc nformaton system (GIS) technologes, the accuracy and relablty of low cost standalone GPS recevers can be sgnfcantly mproved to meet the techncal requrements of varous transportaton applcatons of GPS such as vehcle navgaton, fleet management, route trackng, vehcle arrval/schedule nformaton systems (bus/tran) and on demand travel nformaton. 31
G. Taylor, C. Brunsdon, J. L, A. Olden, D. Steup and M. Wnter There have been many attempts to mprove the relablty of vehcle postonng through the fuson of observatons obtaned by the ntegraton of varous postonng and navgaton nstruments. The vast majorty of such systems use a Global Navgaton Satellte System (GNSS), for absolute postonng, and a varety of other sensors to provde relatve postonng. The usual model beng to use Global Postonng System (GPS) to poston a vehcle whenever possble, and use some form of nertal navgaton system (INS), or dead reckonng (DR) system; odometer, gyro, compass, to determne a vehcle s poston relatve to an ntal poston. Kealy et al. (1999) and Ramjattan (now Kealy) and Cross (1995) descrbe a typcal soluton, ntegratng GNSS and DR usng a Kalman Flterng technque. In ths experment a test route for the system was establshed n the centre of Perth, Western Australa. The results of ths work found that DGPS/DR soluton starts to degrade from 1m to errors as much as 35m by the end of a 10 mnute perod. (Kealy et al., 1999). Kalman flterng technques do have an nherent problem, for vehcle navgaton, on road networks, n terms of stablty, computatonal load, mmunty from nose effects and observablty (Chang et al., 2002). The performance of the flter s heavly dependent on the models used. The model used s a compromse between a statstcal predctve dynamc model and the measurement (observaton) model. If too much weght s gven to the dynamc model, an overly smooth track s the result,.e. rapd changes of drecton are not recognsed quckly enough. If too much weght s gven to the measurement model, errors would be construed as sharp changes n drecton. Devsng the correct model s very dffcult, and wthout a very good model a Kalman flter wll delver the wrong result. Other accounts of usng a Kalman flter for mult-sensor vehcle navgaton are gven by Stephen and Lachapelle (2000), usng GPS and low cost gyro, Petrovello et al. (2003) provdes an nformatve dscusson on levels of ntegraton, also see Mezentsev et al. (2002). Hales (1999) use Kalman flterng wth map-matchng. Fe et al. (2000) descrbe fuzzy logc technques as an alternatve to Kalman flterng for GPS/INS ntegraton. Furthermore, Mayhew and Kachroo (1998) compare solutons usng varous confguratons of GPS, steerng poston, odometer, gyroscope, forward accelerometer and map-matchng, wth sensor fuson methods Kalman flterng, rule based and fuzzy logc. Chang et al. (2002) have developed a GPS/INS mult-sensor navgaton system that utlses an artfcal neural network (ANN) as another alternatve to Kalman flterng. Over the past four years a group of researchers from the GIS Research Centre, School of Computng, Unversty of Glamorgan, have desgned, development and mplemented a software applcaton package for researchng algorthms and technques to mprove GPS based map matchng for navgaton and trackng. Ths test-bed applcaton, called Map Matched GPS (MMGPS) processes raw GPS output data, from RINEX fles, or GPS derved coordnates. It provdes lnkage to a GIS for access and analyss of approprate spatal and related attrbute data (prmarly road and heght nformaton). MMGPS dentfes the correct road on whch a vehcle s travellng on, and snaps the vehcle poston onto that road. Furthermore, MMGPS corrects the derved poston usng ts own computed correcton parameters; Correcton Dluton of Precson (CDOP), usng hstory of prevous poston estmates and road geometry (Blewtt and Taylor, 2002). Varous research experments utlsng MMGPS have been conducted and results have been fully descrbed n Taylor et al. (2001). 32
A test-bed smulator for GPS and GIS ntegrated navgaton and postonng research Snce the man objectves of ths work are to determne both the accuracy and relablty of poston, of a publc transport bus, that can be provded usng Global Postonng System (GPS), odometer and map matchng technques, a new algorthm has been developed that ntegrates odometer observatons wth exstng Global Postonng System (GPS) mapmatchng software, called MMGPS. Now called OMMGPS wth odometer observatons. In OMMGPS, heght nformaton obtaned from dgtal terran models (DTM) are used to obtan 3D GPS pont postons, when only three GPS satelltes are vsble to the recever. More mportantly, heght adng mproves the accuracy of GPS pont postons wth poor satellte geometry (hgh PDOP), and when severe sgnal multpathng occurs (multple reflected GPS satellte sgnals). The developed method uses absolute GPS postonng, map matched, to locate the vehcle on a road centre-lne, when GPS s known to be suffcently accurate. When ths s not the case, odometer readngs are used to locate the vehcle on a road centre-lne. Relatve GPS postons are used to calbrate the odometer observatons. The accuracy of ths method s a functon of the frequency of accurate GPS ponts, relable map-matchng and correct odometer calbraton. Standard Ordnance Survey (OS) dgtal plan and heght map products were used for road map matchng and heght adng. A number of trps along a bus route were made to test the method. A typcal result of map matched GPS postonng s shown n Fgure 1. Fgure 1: Map Matched GPS postons (MMGPS). METHODOLOGY The general approach that was taken, was to use GPS to poston the vehcle, and also to calbrate the odometer readngs, but only when GPS was avalable and of suffcent accuracy. At all other tmes odometer readngs were used to poston the bus. Odometer postonng was acheved by tracng the dstance measured by the odometer along the bus route road centre-lne, actually a 5m offset centre lne was used, left of drecton of travel, see Fgure 1. Map matchng technques are used to mprove the GPS postonng accuracy, and to dentfy grossly naccurate GPS postons. The prevous 10 GPS/odometer postons are used for map matchng calculatons. 33
G. Taylor, C. Brunsdon, J. L, A. Olden, D. Steup and M. Wnter Map matchng The exstng MMGPS software has been adapted to ncorporate postonng based on odometer derved dstances, when GPS postons are not avalable. The man map matchng crtera for a snappng a Raw GPS poston to a road centre-lne Ref GPS poston, are that each of the followng s below a set maxmum value: Dstance error (absolute value of the dfference between Raw dstance and Ref dstance, between the current and prevous epochs). Bearng error (absolute value of the dfference between Raw bearng and Ref bearng, between the current and prevous epochs). Resduals of CDOP Maxmum dstance of Raw GPS poston from the road centre-lne Also, that the number of satelltes vsble to the recever s the same for current and prevous epochs. If a GPS Ref poston passes the check, t s used for postonng the bus, and for calbratng the odometer correcton factor. Otherwse the poston of the bus s derved from the calbrated odometer dstance, and the odometer calbraton correcton factor s not updated. The values used for map matchng, are obvously open to adjustment, or tunng, for dfferent road geometres and for dfferent envronmental scenaros. Dstance correcton factor The prevous secton assumes that dstances obtaned from odometer readngs are multpled by a correcton factor C, so that the dstance suppled to OMMGPS s actually Cd rather than just d. It s not reasonable to assume C s fxed, as dfferent roads, and ndeed dfferent road condtons on the same road wll all nfluence C. For ths reason, when GPS sgnals and odometer sgnals are both avalable, we wll calbrate C over a tme wndow by comparng dstances travelled based on odometer readngs wth those estmated from the GPS (.e. relatve dstances between two GPS ponts). If the GPS goes offlne, the value of C obtaned just before the GPS sgnal s lost (or regarded as unrelable) s used together wth the odometer method to estmate locaton (.e. calculate the current odometer poston based on the prevous GPS or odometer poston and the current odometer readng whch s multpled by a correcton factor C). Estmatng C C can be regarded as a correcton factor between odometer-based dstance estmates as used above, and those obtaned from the GPS-based method. At each second t, we obtan an odometer dstance, d t, and GPS-based coordnate estmates 1 (X t,y t,z t ). Here we assume d t s a cumulatve varable, so that the dstance travelled between t 1 and t s d t d t 1. Call ths d. Also, from the GPS measurements, we can compute the cumulatve dstance travelled usng OMMGPS. Call these dstances D t. Smlar to the odometer dstances, defne D as D t D t 1. Thus, a model of the relatonshp between d and D s D = C d + error We can calbrate ths model by estmatng C usng least squares technques - that s, choose C to mnmse the expresson 2 ( D C d ) 34
A test-bed smulator for GPS and GIS ntegrated navgaton and postonng research It may be verfed that n ths case, the estmate for C s D d C = (1) 2 d However, ths assumes that C s a constant correcton factor. In realty, C s lkely to change, dependng on traffc condtons, road shape, weather and so on. A more realstc model allows C to vary wth tme, so that at each tme t we have a dstnct C t. One approach n ths stuaton s to estmate C accordng to the same model, but to use a `movng wndow' least squares estmate. At each tme t we consder only the values for D and d n a tme wndow of k seconds - so only data from tmes t k,t k+1,..., t s consdered. At tme t+1 the data from tme t k s dropped, and that for tme t+1 s added. Also, a weghtng scheme s used n the least squares method, so that the squared errors for data close to t have a hgher weghtng. Ths gves an estmaton method whch places more emphass on mnmsng errors close to tme t. In ths case, the least squares expresson to be mnmsed s w = 0... k ( D C d ) 2 t where w s the weght placed on the error at lag seconds before tme t. In ths case, we have w D d t t = 0... k C = (2) t 2 w d t = 0... k Weghtng scheme for w Some thought should be gven to the weghtng scheme for the w 's. Obvously, we wsh for a tme-decay effect, so that w 0 > w 1 > > w k. One possblty s to choose an exponental fall-off up to w k, so we could choose w k = γ k, provded γ < 1. Note that we may fx w 0 = 1 wthout loss of generalty. We may experment wth dfferent values of k and γ n order to obtan the best performance of the trackng algorthm as a whole. Thus, the estmate for C t may be wrtten as C t = γ = 0... k γ = 0... k D Heght adng In OMMGPS heght adng s used throughout to add an extra equaton n the least squares approxmaton computaton of GPS poston. Heght nformaton obtaned, usng blnear nterpolaton, from a dgtal terran model (DTM) are used to obtan 3D GPS pont postons, when only three GPS satelltes are vsble to the recever. More mportantly, heght adng mproves the accuracy of GPS pont postons wth poor satellte geometry (hgh PDOD), and when severe sgnal multpathng occurs (multple reflected GPS satellte sgnals). The number of GPS postons computed s nearly always ncreased by usng addtonal heght nformaton, dsplayed n Table 1. Trp 3.4 s the excepton. t t d d t t 2 t (3) 35
G. Taylor, C. Brunsdon, J. L, A. Olden, D. Steup and M. Wnter Table 1: Number of GPS postons computed. Trp no. GPS only GPS +Heght Adng 3.4 3834 3834 4 3468 3635 6 3320 3991 9.2 3713 4014 IMPLEMENTATION OMMGPS conssts of a Dynamc Lnk Lbrary (DLL), wrtten n C++, together wth a GUI for use n ESRI s GIS products ArcVew or ArcGIS. The GUI was orgnally wrtten n ArcVew s Avenue and has recently been translated to Vsual Basc for use n ArcGIS (Steup and Taylor, 2003). The GIS s used to vsualse the results graphcally, usng background OS mappng. DATA PROCESSING AND RESULTS A number of separate trps along the the bus route were made to test the method. Durng each of these trps, GPS and odometer observatons were taken at each second, on the bus. Also, on each trp, the bus recorded the tme when t detected a beacon at the beacon s known locaton, actually to a normal ntersect wth a 5 metre offset centre-lne, see Fgure 2. The postons of the bus at these tmes were used as the true poston of the bus. There are altogether 13 beacons on route 2 whch are used for determnng OMMGPS accuracy. Beacon 15º 15º Bus Poston Intersecton of offset busroute and normal from beacon Fgure 2: Beacon Detecton. Usng ths data the exact equvalent OMMGPS poston used to calculate dstances were obtaned by nterpolaton, usng the OMMGPS postons at the nearest second before and after the beacon detecton tme. Ths s smple lnear nterpolaton. A bus poston was avalable at each second, ether computed by GPS or by odometer observatons. For the whole route, odometer postons are used much more than GPS postons: 70% Odometer postons, 30% GPS postons. For the calculaton of OMMGPS poston at beacon detecton tme for the 13 beacons used; 4 postons were calculated usng only GPS and 9 postons calculated usng only odometer. The results obtaned for trp 4 are presented n Table 2. Ths bus trp s the one on whch the algorthm was developed, and tuned, and shows the excellent potental of the method. The results of the other three trps, that were processed,.e. trp3.4, 9.2 and 6 are shown n Table 3. There s a substantal mprovement n the accuracy of bus poston usng OMMGPS nstead of only raw GPS; In the case of trp4, OMMGPS provdes a mean error of 8.8 metres compared wth 53.7 metres for raw GPS wthout odometer. 36
A test-bed smulator for GPS and GIS ntegrated navgaton and postonng research Table 2: OMMGPS statstcs for all beacons- trp 4 and for 95% - trp 4. Error for 100% Error for 95% Mean 8.8m 5.2m Standard Devaton 6.6m 3.6m Range 18.7m 11.2m Mnmum 0.9m 0.9m Maxmum 19.6m 12.2m Table 3: Average errors for all trps on route2. Trp no GPS only OMMGPS 95% cut off OMMGPS 3.4 27.9m 11.3m 6.6m 4 53.7m 8.8m 5.2m 6 >100m 28.3m 14.1m 9.2 40.7m 22.8m 14.4m CONCLUSIONS A new algorthm that ntegrates odometer observatons wth the exstng MMGPS mapmatchng software was developed and successfully mplemented. Ths new algorthm s called OMMGPS. In OMMGPS, GPS pseudorange observatons are combned wth DTM heght nformaton and odometer postons to provde a vehcle poston at one second epochs. Generally, odometer postonng s used much more often, to poston the vehcle, than GPS. Typcally the rato s 7:3 odometer postons to GPS postons. Ths predomnant use of odometer postonng s due ether to GPS not beng avalable, or GPS postons beng consdered to be too naccurate to use. Ths lack of GPS postons s due to satellte maskng by buldngs (urban canyon effect) or the result of severe GPS sgnal multpath. Four bus trps along the same bus route were used to test OMMGPS. The results obtaned from these four trps are most encouragng. In total, over 15,000 vehcle postons were computed usng OMMGPS. The postons provded by OMMGPS at the tme of beacon detecton can be consdered to be a random sample of the accuracy provded by OMMGPS, compared to the accuracy provded by GPS alone. That s, f a GPS poston was avalable at all, on or near the beacon detecton tme. The average error of OMMGPS postons, over all vehcle postons, usng the random sample of beacon detecton tmes, s 17.8m overall, and 10.1m for a 95% cut off. Ths compares wth an average error for GPS alone of at least 55.6m overall. Moreover, the effectveness of OMMGPS s entrely dependent of the frequency and accuracy of GPS derved postons. The GPS data provded for testng compared very poorly wth smlar raw L1 pseudorange GPS data collected ndependently along the same bus routes, albet wth a much more expensve recever and antenna. Smlarly, recever coordnates collected usng another low cost L1 GPS recever also provded mproved postons, although ths was most probably due to smoothng provded by ths partcular recever s own navgaton flter. In concluson, the technque developed n OMMGPS works well, and can be further mproved wth more superor low cost GPS recever technology or a more careful attenton to ts operatonal applcaton. 37
G. Taylor, C. Brunsdon, J. L, A. Olden, D. Steup and M. Wnter ACKNOWLEDGEMENTS Ths work was undertaken as part of a project commssoned by PA Consultng Group and Transport for London. Partcular thanks must be gven to Dr Phl Whte and Kenny Steele, for ther help wth the work, and problem solvng and deas. REFERENCES Blewtt, G. and Taylor, G., 2002: Mappng dluton of precson (MDOP) and map matched GPS, Internatonal Journal of Geographcal Informaton Scence, ISSN 1365-8816, 16(1), pp 55-67. Chang, K., Noureldn, A. and El-Shemy, N., 2002: Mult-sensor ntegraton usng neuron computng for land-vehcle navgaton. GPS Solutons 6(4), 209 218. Fe, P., Qshan, Z. and Zhongkan, L,. 2000: The applcaton of map matchng method n GPS/INS ntegrated navgaton system. Internatonal Telemeterng Conference, USA Instrument Socety of Amerca 36(2), 728 736. Hales, T.A., 1999: Integratng technologes: DGPS Dead Reckonng and map matchng. Internatonal Archves of Photogrammetry and Remote Sensng 32(2W1), 1.5.1 1.5.8. L, J., Taylor, G. and Kdner, D., 2003: Accuracy and relablty of map matched GPS coordnates: dependence on terran model resoluton and nterpolaton algorthm. AGILE Conference on Geographc Informaton Scence, Lyon, France. Kealy, N, Tsakr, M. and Stewart, M., 1999: Land vehcle navgaton n the urban canyon - ~A Kalman flter soluton usng ntegrated GPS, GLONASS and Dead Reckonng. Proceedngs of ION GPS 99 Conference, pp. 509 518. Mayhew, D. and Kachroo, P., 1998: Mult-rate sensor fuson for GPS usng Kalman flterng, fuzzy methods and map-matchng. Proceedngs of SPIE Conference on Sensng and Controls wth Intellgent Transportaton Systems, Boston, Massachusetts, USA, November 1998, Vol. 3525, pp 440 449. Mezentsev, O., Lu, Y, Lachapelle, G. and Klukas, R., 2002: Vehcle navgaton n urban canyons usng a hgh senstvty GPS recever augmented wth a low cost rate gyro. Proceedngs of ION GPS 2000 Conference, Portland, OR, USA, September 2002. Petrovello, M.G., Cannon, M.E. and Lachapelle, G., 2003: Quantfyng mprovements from the ntegraton of GPS and a tactle grade INS n hgh accuracy navgaton systems. Proceedngs of ION NTM Conference, Anahem, CA, USA January 2003. Ramjattan, A and Cross, P.A., 1995: A Kalman flter model for an ntegrated land vehcle navgaton system. Journal of Navgaton, Cambrdge Unversty Press 49(2), 293-302. Stephen, J and Lachapelle, G., 2000: Development of a GNSS-based mult-sensor vehcle navgaton system. Proceedngs of ION NTM Conference, Anahem, CA, USA January 2000. Steup, D. and Taylor, G., 2003: Portng GIS applcatons between software envronments. Proceedng of the GIS Research UK 2003 11th Annual Conference, London, pp. 166-171. Taylor, G. and Blewtt, G., 1999: Vrtual dfferental GPS & road reducton flterng by map matchng. In: Proceedngs of ION'99, Twelfth Internatonal Techncal Meetng of the Satellte Dvson of the Insttute of Navgaton, Nashvlle, USA, pp. 1675-1684. Taylor, G., and Blewtt, G., 2000: Road reducton flterng usng GPS. Proceedngs of 3rd AGILE Conference on Geographc Informaton Scence, Helsnk, Fnland, pp. 114-120. Taylor, G., Blewtt, G., Steup, D., Corbett, S. and Car, A., 2001: Road reducton flterng for GPS-GIS navgaton. Transactons n GIS 5(3), 193-207. 38