You are on page 1of 5
‘There are thre kinds of es: es, damn lies, and statistics.” So reportedly sai Benjamin Disraeli, prime minister of Great Britain from 1874 101880, And just asthe notoriously wily ‘Statesman noted, the science of analyzing deta, or statistics, sometimes yields results that one can interpret ina variety of ways, depending (on politics or interests, Likewise, we in the satelite navigation field interpret results depending othe information ve wish to produce: Using various statistical methods, we can create many different GPS and GLONASS position accuracy measures. Itcan seem confusing, even misleading, but as we'll seein ‘this month's colwnn, there's some rhyme to ‘our reason, We'll examine some ofthe most ‘commonly used accuracy measures, reveal their relationships to one another, and correct several common misconceptions about ‘accuracy. Our author is Frank van Diggelen of Ashuech, Inc, In Sunnyvale, California. Van Diggelen is the OEM (original equipment ‘manufacturer) and navigation products ‘marketing manager. “Innovation” isa regular column featuring discussions about recent advances in GPS technology and its applications as well as the fundamentals of GPS positioning. The column is coordinated by Richard Langley ofthe Depariment of Geodesy and Geomatics Engineering at the University of New Brunswick, who appreciates receiving your comments aswell as topic suggestions for uaure columns, To contact hin, se the "Colunniss” section on page 4 ofthis issue, GPS Accuracy: lies, Damn lies, and Statistics Frank van Diggelen Ashtech, Ine Root mean square (rms), twice the distance root mean square (2drms), circular error probable (CEP), spherical error probable (SEP), so on, and so forth — Why do we have so many different position accuracy measures? The answer lies in the fact that the errors of position coordinates determined using a GPS or GLONASS unit are nt con- stant — they vary statistically, If you observe the reported position of stationary receiving system over ime, you will notice it wanders. Graphing these moving points yields a “scat- ter plot"; how you analyze the scatter depends on the information you want to obtain To complicate matters ihe poston is fundamentally three dimensional, but not everyone is interested in obtaining three- ‘mensional accuracy. One user might care aout horizontal accuracy, another might ‘want vertical. Thus, clearly, we must con- sider different accuracy measures ‘Before we take a ook at some ofthe com- ‘mon aceuracy evaluations and their rlation- Ships t one another, we must say a word about the meaning of accuracy itself. To ascertain how accurately a system has dete Table 1. Position accuracy measures TNNOVATION mined a point’s coordinates, you must know the point's true coordinates. Typically, this comes from measurements made using 2 sys- tem with an inherently higher accuracy than the one being tested. Simply averaging a sys tems reported position will provide an indi- cation of system precision or repeatability, ‘but the measurements might contain a bias thatcould affect the results. So, when we talk bot system accuracy, we must consider the possibility of such a mean error. POPULAR ACCURACY MEASURES Table 1 lists the most commonly used GPS positon accuracy measures and their defini- tions. Note that the first two methods are explained in terms of average squared error, and the last three are defined diretly from the position errr distribution (the scatter). ‘Thus, we can immediately associate these last three with error probabilities. If we assume thatthe error distribution along any axis (east, north, or up) i “normal” or Gauss- jan, then we can also derive probabilities associated with the rms and 2dmms accuracy measures. [Te normal or Gaussian distribu- tion isthe ome to which the dispersion of the sum of a very large number of very small errors always converges. The famous Ger- ‘man polymath Carl Friedrich Gauss used this distribution to develop his error theory inthe early nineteenth century, To honor the importance of this and the scientst’s other ‘accomplishments, Germany features a por- trait of Gauss and his probability distribution ‘omits 10-mark bank note. —R.B.L.] ‘Table 2 shows how the accuracy measures reused and what probabilities ean be associ- ated with them. Note that the probability associated with ms depends on whether one js using rms in one, two, or three dimensions (L-D, 2D, or 3-D). The later “Common Mis- conceptions” section discusses this further. ‘sotaing Acaracr. A ranle, Now, suppose {you are comparing the specifications of two ‘Measure ‘Abbreviation Definition ‘eotmean square ms the square root of he average ofthe squared enor ‘los dstance ms 2am twice the ims ofthe horizontal errs ‘cuter ener probable CEP a orci’ radius, contared atthe tue nina poston, cortaning 50 percent ‘ofthe panei tho hoon seston sot hodzortal 95 percent ROS 1 roe’ radius, certrod at the rue securacy lana poston, contarning 86 percent of points inthe Porzontal seater plot spherical ror sep 1 shares radius, centered at the we antenna probable peatton, conaiifg 50 poroont of the piri. Fine tree smensiona seaton at ime 198 OPS WORD a INNOVATION ‘Table 2. Common GPS and GLONASS accuracy measures ‘Accuracy ge Dimensions ‘measure (dimension) 1 we oe vertcal 2 cer 50 horizontal 2 ios 03.8 horizontal 2 Fs 95 hotaontal 2 2s 95-98 hotizontl 3 co er-68 30 3 sep 0 30 Table 3. Theoretical equivalent accuracies ROS rms rms | (horizontal rms (vertical) |CEP | (horizontal) |95 percent) 2drms |(3-D) | SEP 1 Joss] ose os | ut | at | ose 1 12 24 24 [25 | 20 7 17 2 [aa | a7 me (horizontal) 1 12 | 12 | 096 [R05 horizontal '95 percent) 1 [at [oss | adem ce) 1 ‘SEP, Meters Figure 1. Measured and theoretical horizontal error distribution. The vertical axis indicates the relative frequency of errors occurring in each error interval, Horizontal axis values are rounded, a PS Wo ny 98 positioning systems. One unit has a quoted accuracy of 3 meters (3-D rms) and another has a quoted accuracy of 2 meters (CEP) ‘Which system is more accurate? By making three assumptions about the ratio of east, north, and up errors, we can relate different accuracy determinations 10 each other, as shown in Table 3 (entitled “Theoretical Equivalent Accuracies"). To tse that table, idemtify the desired measure in the top row and the orginal measure in the right hand column, Take the number in the cell at which the row and column intersect and multiply it by the original measure value to yield the desired number. The three assumptions, from which the conversion values derive, are true on aver age, First assumption, the error distribution is Gaussian. Second, the ratios of the postion Ailton of precision (PDOP) tothe horizon- tal DOP (HDOP) and the vertical DOP (VDOP) to the HDOP are 2.1:1 and 1.9:1, respectively. Third, the horizontal ero dis- tribution is circular. These supposition. are based on simulations performed over a grid covering the entire globe between latinides 16 degrees south and 66 degrees north. In general, horizontal distributions are elliptic cal, the ellipses are often very close to cireu- lar, and the circular Ganssian distribution ‘model is very good at estimating the true dis- tribution, as shown in Figure I To answer the question posed earlier (hich is more aceurate — 2 2-meter [CEP] ‘ameter [-D ms] system? follow these four steps 8 Go down the “rms -D)" column to the “CEP” row. ‘The entry inthis cel is 2.5, ® According to the table, ems (3-D) 25x CEP, 1 So, CEP = rms (3-D)/2.5 = 3/2.5 = 1.2 meters Thus, a system with 3. rms of 3 meters will have a CEP of 1.2 meters and i, there- fore, more accurate than a system with a CEP of 2 meters, For specific details about how we created Table 3, see the “Deriving the Equivalent ‘Accuracies Table” sidebar onthe last page of this article, Making Valid Assumptions. To understand the table in more general ers, one must realize that the three assumptions and the table are valid for the average measurement. This ‘means that if someone takes measurements all day, then, on average, the different accu- racies are related by the nurbers inthe table, At any instant, however, the satelite geome try may produce a different relationship between various accuracies (for example, between vertical and horizontal). But the best ‘way to make a comparison, seemingly, is to use average relationships. Stating a Sal et. In any specific example, the table entries are apparently good to within #20 percent. Tha is, ifthe table says 2drms = 1.2 X horizontal 95 percent, then particular experiment may show 2drms to be anywhere from 0.96 to 1.44 X horizontal 95 percent ‘To evaluate Table 3's efficacy. we used ata from more than $50 hours (2 million data points) of differential GPS positions (DGPS), obiained with a US. Coast Guard reference station providing the differential corrections. Our results were 42 centimeters CEP, 91 centimeters horizontal 95 percent, and 104 centimeters 2drms, Table 4 shows hhow these results compare with Table 3°s theoretical values. Cinig the Cele. The circular Gaussian dis tribution model for horizontal errors is su prisingly good. Figure | portrays a histogram {in green) generated from the DGPS data ‘These data have a horizontal rms value of (0.52 mete. Overlaid on the green histogram is-a bar graph, showing the theoretical his- togram that would be obtained from data that truly were cireulary distributed and Gauss- jan and that had the same horizontal rms a the measured data, As the figure shows, the ‘measured and theoretical distributions agree extremely well. We can obtain similarly good fits for vertical error distributions modeled as -D Gaussian ‘COMMON MISCONCEPTIONS By now. one may fee that accuracy measures are rather simple to understand. Still, general discussions about GPS and GLONASS posi- tion accuracies frequently contain several misconceptions. Here’s our attempt to set the record straight “Misconception Number I —rms precisely le 4. Sample dat TNNOVATION equals one sigma (Ia or 1 standard devia- tion), Well, actually this true, as long asthe ‘mean error is zero. With most GPS or GPS/GLONASS systems, the mean errors (over a sufficiently long time interval) are zero, o close to ero, and so rms may be con sidered essentially equivalent to one sigma. Misconception Number 2 — 2drms means “1wo-dimensional rms.” Infact, 2érms usu- ally stands for “twice distance rms,” in which the “distance” is measured ina 2-D space, the horizontal plane. Thus, 2dims is a very con- fusing abbreviation: I i a two-dimensional measure, but the “2d” usually stands for twice distance. (Some publications about navigation accuracies, notably those issued by the North Adantic Treaty Organization, sults compared with theory Ratio Theoretical theoretical values Barmamherental@S percent 1.14 12 5 2anmsiCeP 248 2a a Horizontal 95 percentiCEP 2a7 24 3 GPS Antennas from yokouwo Suppress GPS Interference with ERI’s ISU ‘The world's largest supplier of GPS antennas {or car navigation. Already the OEM of choice for major automobile manufactures. Don't let commercial communication systems fr intentional jammers interfere with your GPS reception. ERI's Interference Suppression Unit {ISU} blocks in-band interfering signals to give ‘you maximum navigation, position, and time performance ‘Yokowo has a volume, in-house, ceramic manufacturing capability and can custom manufacture large quantities of GPS antennas for almost any application, Contact us to learn more about the ISU's €Rl Circle 37 Yokowo America Corporation 415 West Golf Road, Suite 44 Arlington Heights, IL 60005 : 847-228-0454 » Fax: 847-228-0469 E-mail: yokowoa@aol.com Electro-Radiation Inc. 39 Plymouth Street, Fairfield, NJ 07004, USA Tel. 973-808-9033 Fax. 873-808-9557 Circle 36 Jr 8 PS WUD « INNOVATION use the alternative meaning of “24.” Thus, their “2dems" is exactly one-half of the usual measure.) Misconception Number 3 — 2drms is ‘exactly equivalent to a 95 percent probability level This untrue belief stems from the fact ‘that, for a 1-D Gaussian distribution, 95 per- ‘cen of it ies inside an interval from ~2a to +420. However, 2dmms is a measure for a2-D distribution. The percentage of scatter lying within a circle with radius equal to 2drms depends on the distribution shape. For acir- cular distribution, the percentage of scatter inside a 2drms circle is 98 percent. The “Deriving the Equivalent Accuraces Table” sidebar shows this. As the scatter becomes ‘more elliptical (with different eror distibu- tions for the two horizontal coordinates), it also becomes more one-dimensional, causing the percentage of elliptical distribution val- tues inside a 2drms circle to tend toward 95 percent. For GPS units, when the whole sky is vis- ible above a 10-degree mask angle, seater is approximately circular. Typically, distribu- tions become very elliptical when HDOP. gets large (much greater than 1). Thus, for any GPS receiver in any environment, the le with a radius equal o 2dems coniains between 95 and 98 percent ofthe scatter. ‘When HDOP is low, the percentage is closer 1098 percent; when HOP is high itis closer 1095 percent. Misconception Number 4 — rms is per= Jeedly comparable with a 68 percent proba- bility level. This is true for only -D Gaussian Aistibutions. For 2-D or3-D Gaussian dsi- butions, the percentage ofthe values disuib- uted inside a circle (or sphere), with a radius equal tothe rms value, depends on distribu- tion shape Misconception Number 5 —The errors tribution really is Gaussian. We use the assumption that the error distribution is Gaussian for analytical purposes, and over time, one can show that a eteular Gaussian distibtion can model the errors very well (Gee Figure 1), However, certain errors may not have a Gaussian distribution: 1 Stand-alone GPS errors are dominated by selective availabilty. Because this isan atificial err source, the erat contributes are no always Gaussian. Stand-alone GPSIGLONASS rors a TY I SPIRIT RTK technologies offer: ‘1 centimeter accuracy at 0.95 rela ‘* 1-2 second latency time SPIRIT develops algotis and sofware fr igh ‘pion TK Posting weg both GPS & GLONASS 1 phe meairemeis We ploy GLONAS expe CAN world is now turning to To license the future's technology today, turn to SPIRIT. ASS RTK When you've ready to advance to (GPS/GLONASS RTK, come to where the experts are. Come tous, SPIRIT CORP <9 “he Rustin SotvareSerce forte avemen of ie eg ae {GGLONAS tong to tome tse wens WW.spiritcorp.com telig iS compari Compan Bc NIG xh a gM) ETD saaSineg ne eat ee! SAR tegen nearer “a GPS WORD Jamu 1998 Circle 38 show distributions that match Gaussian Aistrbutions quite well (o about 10 percent) vera time period of, say, several hours, 1 Differential errors over a long time display distributions that match Gaussian pattems to within a few percent. Tis is true for both code differential and carier-phase differential (commonly referred to as real- time kinematic, or RTK). Differential errors ‘over a short time produce scater dominated ‘by multipath, which is fairly constant over a ‘ew minutes, and, hence, the distribution is distinctly non-Gaussian. IN CONCLUSION ‘As Disraeli also noted, “An investment in ‘knowledge pays the best interest.” We hope that this brief note has proven to be a warth- ‘hile “investment” to readers, shedding ight ‘on the sometimes murky subject of accuracy ‘measures used in GPS and GLONASS posi- tioning. With the simple information pro- vided, you should be able to compute the positioning accuracy ofa system in a variety fof measures and also, contrary to the old adage, be able to compare “apple A” with “orange B." Ml LET EL Ti ‘The function inichsq jf.) computes the square of sci’ radus euch hat the eum of equares ‘ftw random varabes, ech wth m3 = = 4, hasa probably pof‘aling insde the cece. ‘This function allow uses to relate isto probability or a two-dimensional cular ‘stration ‘Comments are ehown in eury brackets “." “able 3 labout equivalent accuracy contains ‘nies with eschion of only two dg. tis impossible, n general to provide more pecine rates because th mao kal assumptions ae ‘averages over the whole word, and. thus, are ‘goed ony o within 2 few percent of ere in any parieular rego. ims (erica) = 19 x ms fhorzorta using YDOPIHOOP = 1.9} (CEP = 50 porcent cle frst solv for CEP =x 2% re (hoiontal, then use that resu fo derive other ratios} fms (horizorta) = v2 rms near (assuming & a}erky= 088-12 PROOF THAT COME IN SMA ME GPS RECEIVER MODULE omar EIT ‘Actual size [77 (OEP =x x ms Werical) =x x ms porzontal) x 19= 085 » ms (hoizoria) => x= 089/19 = 04s (2jenty=0.44 R95 = 05 percent el st save for ROS =x x ‘ms fotizaota, then use that rsut to dove ‘the ratios rchisg (095.2) = 1.78% 2 => FOS =1.73 2X rms near) = 1.78 ‘ms hoizontal (Ga) enny = 1.7 ROS =x x rms (vccal = x x ms hier) = 179g 5 = 1.78 ms fhorizona = 1.73 cepa gs (24) envy = + 79/063 =2:1 cams = 2% rms (crizonta fb defriton) 2 2owms =X X CEP =x x 0.88 me leona = 086 « 2drmel2 => x= 2/088 =24 28) enty=24 Beene = x ROS = 1.73 X ms orzo) D173 x 2am? => x=2/.73= 1.16 as enty 12. ms (0) =2.1 x rms horizontal (sing POOPMDOP =2.1)=2.1 x 2erms/2 = 2.1 x me verioay 8 (3. entry=24 4 (gen =297 mme(0-0)=1 x CEP =x x 089 < ss (hozontal = 0.88% rms @-DY2. x=21/088 2259 Ce enty=25 rms 8-0) =x x ROS = 1.73 x ts (heczotall =x > 1.73 rms @-0Y2.1 Wi73= 121 Gemy=12 ‘SEP based on simulation: ms froxizota ms (vetea = ms (onizontal 1.8; SEF 2.38 1.68 x ems horzortal) Gipeny= 17 (iter SEP ratios derives rom te aoromen- toned ratios. (@ 8 channel paralle1 receiver @ Low power consumption, 0.5 W at 5 VDC @ Compact size (1.4"x23" x 0.5°/96x 58x 12mm) @ Resists the urban canyon syndrome @ Standard NMEA-0183 output data @ Differential GPS ready in SC104 format © Supplied in a shield case for better protection against interference @ 1 PPS (pulse per second) output available FOR FURTHER INFORMATION PLEASE CONTACT —— FURUNO ELECTRIC CO.,LTD. 9.52 Aahhorscho, Nhinomya 062, Japan Tologhone #3 79863-0182 Tolls: +61-798-66:0281 {evn set 0° lgpsshoplgpsanop. ha (oma: apes fans oni Circle 39 nny 8 PE WORD as

You might also like