Kyun queue: a sensor network system to monitor road traffic queues

14 Pages·2011·3.64 MB·English

Kyun Queue: A Sensor Network System To Monitor Road Traffic
Queues
RijurekhaSen,AbhinavMaurya,BhaskaranRaman,RupeshMehta,
RamakrishnanKalyanaraman,NagamanojVankadhara,SwaroopRoy, PrashimaSharma
{riju, ahmaurya,br,rupeshm,krishnan,nagamanojv}@cse.iitb.ac.in,
{bswarooproy,prashimasharma}@gmail.com
IndianInstituteof Technology,Bombay
Abstract 1 Introduction
Unprecedentedrate of growth in the numberof vehicles The problem of road congestion is currently plaguing
has resulted in acute road congestion problems worldwide. most cities in the world. Long traffic queues at signalized
Better traffic flow management, based on enhanced traffic intersectionscause unpredictabletraveltimesandfuelinef-
monitoring, is being tried by city authorities. In many de- ficiency [1, 2]. The problem is felt more acutely in grow-
velopingcountries,thesituationisworsebecauseofgreater ing economies like India, primarily because infrastructure
skewingrowthoftrafficvstheroadinfrastructure. Further, growth is slow compared to the growth in vehicles due to
theexistingtrafficmonitoringtechniquesperformpoorlyin spaceandcostconstraints.
thechaoticnon-lanebasedtraffichere. Secondly, automated traffic management solutions, de-
Inthispaper,wepresentKyun1 Queue,asensornetwork veloped for western traffic, are at best an accidental fit for
systemforrealtimetrafficqueuemonitoring. Comparedto roads in developing countries. Traffic in many developing
existing systems, it has several advantages: it (a) works in regionsisdistinctlydifferentfromwesternroadsintwoways
chaotic traffic, (b) does not interrupt traffic flow during its (see Fig. 1): it is (1) non-lane based and (2) highly het-
installation and maintenance and (c) incurs low cost. Our erogeneous. Four wheeler heavy vehicles like buses and
contributions in this paper are four-fold. (1) We propose trucks, four wheeler light vehicles like cars, three wheeler
a new mechanism to sense road occupancy based on vari- auto-rickshaws and two-wheeler motorcycles ply the same
ation in RF link characteristics, whenline of sightbetween road, intermingled with each other, without any lane disci-
atransmitter-receiverpairisobstructed.(2)Wedesignalgo- pline[3].2
rithmstoclassifytrafficstatesintocongestedorfree-flowing Anidealtrafficsensingsystemfordevelopingregionshas
at time scales of 20 secondswith above90% accuracy. (3) thefollowingstringentrequirements:(1)itshouldsenseroad
We design and implement the embedded platforms needed occupancyevenif traffic is chaotic and non-lanebased, (2)
to dothe sensing, computationand communicationto form should be deployable without interrupting traffic flow, (3)
a networkofsensors. Thisnetworkcancorrelatethetraffic becapableofrealtimesensingandclassificationtosupport
state classificationdecisionsofindividualsensors, to detect applications like traffic signal control and (4) should have
multiple levels of traffic congestion or traffic queue length lowinstallationandmaintenancecosts. Section2discusses
on a givenstretch of road, in real time. (4) Deploymentof the shortcomingsof existingsensing systems like magnetic
oursystemonaMumbairoad,aftercarefulconsiderationof loops, camera, infraredsensors, acoustic sensors and probe
issues like localization and interference, gives correct esti- sensors,vis-a-vistheserequirements.
mates of traffic queue lengths, validated against 9 hours of In this paper, we present Kyun Queue, a sensor network
image-basedgroundtruth. Oursystemcanprovideinputto systemforrealtimetrafficqueuemonitoring.Webuildanew
severaltrafficmanagementapplicationsliketrafficlightcon- mechanismtosenseroadoccupancy,basedonhowRFlink
trol,incidentdetection,andcongestionmonitoring. qualitysuffersinabsenceoflineofsight. Oursystemcom-
prises of an IEEE 802.15.4 transmitter-receiver pair across
1kyun,pronouncedas’few’startingwith’k’insteadof’f’,means’why’
the road, where the transmitter continuously sends packets
inHindi
and the receiver measures metrics like signal strength and
packet reception ratio. We show that these metrics show a
strong correlationwith the occupancylevelon the road be-
tweenthem. Weinvestigatewhichwirelesslinkcharacteris-
tics, whattime windowsand whatalgorithmsgivea highly
Permissiontomakedigitalorhardcopiesofallorpartofthisworkforpersonalor accuraterealtimeclassificationoftrafficstates. Oursystem
classroomuseisgrantedwithoutfeeprovidedthatcopiesarenotmadeordistributed givesabove90%binaryclassificationaccuracyon16hours
forprofitorcommercialadvantageandthatcopiesbearthisnoticeandthefullcitation
onthefirstpage.Tocopyotherwise,torepublish,topostonserversortoredistribute ofdatafromtworoadsinMumbai.
tolists,requirespriorspecificpermissionand/orafee.
SenSys’12,November6–9,2012,Toronto,ON,Canada. 2 [3] has several representative videos of chaotic traffic at
Copyright(cid:13)c 2012ACM978-1-4503-1169-4...$10.00 http://www.cse.iitb.ac.in/˜riju/rss-videos/
127
Figure1:Chaotictraffic(source[4]) Figure2:Lanebasedtraffic Figure3:Issueswithinfrared
Building on this binary classification of traffic states us- withoutaffectingtheflowoftraffic.Wecansenseroadoccu-
ing one pair of sensors, we constructa linear array of such pancyeveniftrafficischaoticandnon-lanebased.Ourtech-
sensors. This can detect the length of a queue or the de- nique is cheap, with each sensor pair costing about $200.
gree of congestion on a given road stretch. We design and This is the cost of prototype using off-the-shelf modules,
buildappropriateembeddedplatformstodo(1)sensingand whichcanfurtherreducewithcustomizeddesign. Eachpair
computationforbinaryclassification,(2)communicationto gives a binary classification of traffic state as free-flow or
combinethedecisionsofmultiplesensorstoknowthelength congested. To know length of queue, we need an array of
oftraffic queueand(3)communicationofthe queuelength sensorpairs.Wecancoverabout200metersofroadatabout
valuestoaremoteserver. $1200.
Oursystem,deployedonaMumbairoadfor9hours,pe- Image sensors - Video surveillance based traffic moni-
riodically (every 30 secs in our deployment) reports traffic toringisfairlycommon[8].[17]givesacomprehensivesur-
queuelengthvaluestoaremoteserver. Manualexamination veyof the major computervision techniquesused in traffic
of the corresponding9 hours of image-based ground truth, applications. Butthetraditionalsettingforwhichvisional-
givesmaximumaccuracyofourqueueestimatestobe96% gorithmsexistcanbeseeninFig.2(b). Forusabilityinde-
and minimum accuracy to be 74%. These queue estimates velopingcountries,algorithmsareneededforscenarioslike
can be potentially used in a range of applications such as: Fig.1.[18]isapreliminaryworkonimageprocessingalgo-
automatingtrafficlightcontrol,detectingbottlenecksorcre- rithmsforchaotictrafficsensing.Thealgorithmsareoffline,
atingcongestionmapsofacity. sothetrade-offbetweencomputationandcommunicationis
The rest of the paper is organizedas follows. Section 2 notyetunderstood.Alsothesensingaccuracyitselfhasbeen
discussestheexistingtrafficsensingsolutions.Ourproposed tested on only 2 minutes of video clip. [19] is another re-
sensing technique and the real time traffic state classifica- cent work to use low quality images from CCTV for traf-
tionalgorithmsbasedonit,arediscussedinSection3. The ficsensing.Butcomputationaloverhead,real-timelinessand
design of our sensor network is detailed in Section 4 and accuracyofthedesignedalgorithmsareyettobeevaluated.
hardwareprototypesaredescribedinSection5. Wepresent Thus thoughimage-basedsensing shows promise, with ad-
thedeploymentbasedevaluationofoursysteminSection6. vantagessuchaseaseofdeploymentandpotentiallowcost,
Finallywediscusstheissuesofsensorlocalization,interfer- thereareseveralaspectsthatneedcarefulevaluationandval-
enceandpowerconsumption,beforeconcludingthepaper. idation,especiallyinchaotictrafficconditions.
In comparison, we present an alternative traffic sensing
2 Related work
systemthatworksinchaoticnon-lanebasedtraffic.Wehave
Automated traffic management solutions have been de- thoroughly evaluated the sensing accuracy on 16 hours of
ployedin manydevelopedcountriessince severalyears. In traffic in Mumbai. The algorithms are very low overhead
thissection,wediscussthetrafficsensingmethodsprimarily andwehaveimplementedthemonlowendembeddedsensor
usedtoprovideinputtothetrafficmanagementsystemsand platforms. Sensing and computation are done on the road.
theirusabilityinchaotictraffic. Only the traffic queue length values are communicated to
Loopdetectors-Therearealotofresearchprojects[5,7] theremoteserver,avoidingthecommunicationoverheadof
andseveraldeployedsystems[6]thatdovehiclecountingus- videotransfer.
ingmagneticloopsunderthe road. Avehicleloopdetector Infraredsensors-Activeinfraredsensorsplacedacross
costs $700 for a loop, $2500 for a controller, $5000 for a the road suffer beam cuts by vehicular movement in be-
controller cabinet and 10% of the original installation cost tween. [10] is a commercial product that measures speed,
forannualmaintenance[1]. Loopsneedtobeplacedunder length and lanes of vehicles from beam cuts. However, in-
the road. Digging up roads to install and maintain the in- fraredpropagatesasaray,andhenceisstrictlylineofsight
frastructureintrudesintotrafficflow.Alsore-layingtheroad (LOS) based. This makes it overly sensitive to even small
surface needs re-laying the sensors. Moreover, loops have obstacles. WhenwetriedusingitonMumbairoads,pedes-
beentraditionallyplacedundereachlaneasseeninFig.2(a). trianswhofrequentlyusetheroads(Fig.3(a))orirregulari-
Howshouldtheplacementbeinabsenceoflanesisyettobe tiesoftheroadsurface(Fig.3(b))wouldnotallowthetx-rx
explored. pairtocommunicate.Also[10]hasbeentestedonlyinlane-
Incomparison,oursensorsaretobeplacedonroadside, basedtraffic. Beamcutsinhighdensity,non-lanebasedand
128
Technique Drawbacks
loopdetectors[5,6], highcost,under-the-roadinstallation,
magneticsensors[7] donotcurrentlysupportchaotictraffic
videoandimages[8,9] attractiveoptionforexploration,
butnoprovensolutionsforchaotictrafficthusfar
potentiallyhighcomputationand/or
communicationoverheadforchaotictraffic
IRsensors[10] pedestrians,roadsurfaceirregularitiesandnon-lane
basedvehicularmovementcausespuriousbeamcuts
acousticsensors[11,12,3] hightrainingoverhead,slowresponse
probesensors[1,13,14,15,16] usefulforcomplementaryapplications,
buttoosparseandnoisytoallowmicromanagementofroadnetworks
Table1: Drawbacksofexistingtechniquesforchaotictrafficmanagement
heterogeneoustraffic as seen in Fig. 1, would need careful ableusingprobedata.
studytobeusedforchaotictrafficmeasurement. Our work is aimed at providing strictly periodic, accu-
The setup of sensor pair across road in our system uses rateinputtotrafficmanagementsystems. Oursystem,based
802.15.4radioswhichhaveaspreadpropagationmodel,in- on static sensors, would potentially be used at some key
steadofraypropagationmodelofinfrared. Thismakesour intersections of a city, where intelligent and adaptive con-
technique robust to noise and thus suitable for disorderly trol would positively affect traffic flow. This is orthogonal
roadconditionssuchasinFig.1. Thisrobustnessisevident and complementary to the commuter applications handled
from the rigorousempirical evaluation of our binary traffic byprobesensing.
stateclassificationover16hoursofchaotictrafficdata. We Thedrawbacksofexistingtechniquesaresummarizedin
alsoenhancethepairwisesensingtoanarrayofsensorsand Table 1. Some techniques like loops and acoustic sensing
detect length of queues, an aspect not examined in the IR- have clear disadvantages. Some like probe sensors are ill-
basedsystems. suited for the particular applications that we are targeting
Acousticsensors-Somerecentresearchisbeingdoneto in this paper. Some others like image and infrared sensing
useacousticsensorsfortrafficstateestimation,especiallyin might work after adaptation for chaotic traffic, but there is
developingregions,wheretrafficbeingchaoticisnoisy[12]. noworkingadaptationtilldate.
But standing traffic does not have any uniform sound sig-
3 Traffic sensing withwireless across road
nature and dependslargely on the driver behaviorand type
of vehicles. Forexample, if vehiclesare standingin a long Inthispaper,wewishtodesignasystemfortrafficqueue
queueawaitingagreensignal,manydriversmightjustshut lengthmeasurementinchaoticroadconditions.Thishassev-
downthe enginesandwaitsilentlyormightblowhonkim- eral potential applicationslike traffic light control, incident
patiently,producingtwoverydifferentsoundsignaturesfor detection, and congestion monitoring. An important point
same traffic state. This reduces the sensing accuracy [3]. tonotehereisthat,atalmostallintersectionsindeveloping
Also,toattainavalidsignaturecorrelatingtoatrafficstate, countries, a variety of vehicles stand as a coagulated mass
sensing time window of at least a minute or so is needed, waiting for green signal(see Fig. 1). On getting greensig-
makingthesesystemsslowinresponse. nal, they all move forward almost bumper to bumper, with
In comparison, our technique works in the order of 10- little variability in speeds. So queue length is proportional
20seconds. Theclassificationaccuracyisalsomuchhigher to traffic volume, the road width being the proportionality
thanacoustic,asvehiclesbeingheavymetallicobstaclesal- constant. We thus assume, as is intuitive, that finer granu-
waysaffectRFsignalquality.Thishighaccuracyalsomakes larityinformationlikeexactvehiclecountsareunnecessary.
it possible, to take the binary classification to a multilevel Totunetrafficlightsproportionaltowaitingtrafficvolumes,
classificationusinganarrayofRFsensorstoknowlengthof queue lengths are sufficient. Thus we aim to detect road
queueonaroad. Achievingthiswithlowaccuracyacoustic occupancy in chaotic traffic, and build upon that to detect
sensorswouldbetricky. lengthofqueues.
Probe sensors - Significant research is being done to Itiswellknowninthewirelessnetworkingdomain,that
leverage cell phones and participatory sensing to generate characteristicsofwirelesslinkslike802.11and802.15.4are
trafficrelatedinformation. Focushasbeengiventoresearch affectedintheabsenceofLOS[25]. Obstaclescausereflec-
bothinnetworkingissues[20,21,22]andmachinelearning tion, absorption and scattering of the RF signal, degrading
issues [23, 1, 15]. But participatory sensing data is inher- thelinkquality.Researchershaveexploitedthesevagariesof
entlynoisy[24].Alsoprobevehiclesmightnotbepresentat RFlinksindifferentkindsofapplicationscenarios,themost
agivenintersectionatalltimes.Traveltimeestimates,transit commonbeingthatforindoorlocalization[26,27,28,29].
vehicleinformationandcongestionmapstobedisseminated The effects of vehicular traffic on wireless links were
to commuters for route selection, can tolerate aperiodicity brieflystudiedin[30],wheretheauthorsseekedtoquantify
andnoise. Suchcommuterapplicationsarethushighlysuit- the link quality in harsh deploymentscenarios, like tunnels
129
1 1
0.9 5:37-5:42pm 0.9 5:37-5:42pm
5:43-5:48pm 5:43-5:48pm
bility 00..78 55::4595--56::5040ppmm bility 00..78 55::4595--56::5040ppmm
ative proba 000...456 66666:::::0112340620-----66666:::::0122395175pppppmmmmm ative proba 000...456 66666:::::0112340620-----66666:::::0122395175pppppmmmmm
Cumul 00..23 666:::344628---666:::445173pppmmm Cumul 00..23 666:::344628---666:::445173pppmmm
0.1 6:54-7:00pm 0.1 6:54-7:00pm
7:00-7:05pm 7:00-7:05pm
0 0
-95 -90 -85 -80 -75 -70 -65 -60 -55 0 5 10 15 20 25 30 35 40 45
RSSI (dBm) No. of packets received in 1 sec out of 25 packets sent
Figure4: Wirelesscommunication
acrossroad Figure5: CDFofRSSI(dBm) Figure6:CDFofPRR
withtrafficmovingthroughthem. Inthispaper,wewantto of RSSI are around -93dBm in congestion and -78dBm in
applythiseffectoftrafficonwirelesslinksforacompletely free-flow. (b)the60th and80th percentilesofreceptionrate
novelapplication,tomeasureroadtrafficdensity. Thebasic arearound0packets/secincongestionand24packets/secin
intuitionisasfollows. Ifawirelesstransmitter-receiverpair free-flow.Weseesimilartrendsin16hoursofdatacollected
iskeptacrossaroadandmadetocommunicate,thelinkchar- overthree weeks, from two Mumbairoads. Sample videos
acteristics,observedatthereceiver,shouldbeaffectedbythe offree-flowingandcongestedtrafficareavailableat[31].
vehicles on the road. But unlike IR, these wireless signals
follow the spread model of propagation, and hence should
notbeoverlysensitiveto smallobstacles. Can trafficstates
bequantitativelyinferredfromRFlinkcharacteristics? Can
classificationsbedoneinorderoftensofsecondstosupport
applicationsliketrafficlightcontrol?Whatwirelessfeatures
andalgorithmscanbeusedforhighclassificationaccuracy?
Inthissection,weseektoanswerthesequestions.
3.1 Does roadtraffic affect wirelesslinks?
Figure10:Mismatchofsensingandtrafficstate
To answer this, we start with some proof-of-conceptex-
periments. We create a setup shown in Fig. 4 on Adi
ShankaracharyaMarg, a road in Mumbai, about 25m wide 3.2 Real-timeclassificationoftraffic states
in each direction. We keep two 802.15.4compliant Telosb Asseenabove,wirelesslogsof5minutesdurationshow
motes acrossthe road, one as transmitter (tx) and the other good visual differencein the distributionsof wireless char-
asreceiver(rx), ona line perpendicularto the lengthof the acteristics, betweenfree-flowingandcongestedtraffic. Ap-
road.Thetxsends25packetspersecond,eachhavingapay- plicationslikecongestionmapsandbottleneckdetectioncan
loadof 100bytes, at −25dBmtransmitpower. Therx logs be handled at a time scale as large as 5 minutes, but adap-
thenumberofpacketsreceivedandReceivedSignalStrength tivetrafficlightcontrolwouldintuitivelyneedfasterinputs.
Indicator(RSSI) andLinkQuality Indicator(LQI)foreach From our observations of Mumbai and Bengaluru traffic
receivedpacket. Onepersonstandsontheroadsidefootpath lights,trafficsignalcyclestypicallylastforabouta minute.
holdingtherxandanotherstandsacrosstheroad,ontheroad This one minute cycle time is divided into slots, in which
divider,withthetx,bothrxandtxbeingataheightofabout different contending flows get their respective green times.
0.5mfromtheground. Thesetwopersonsalsoobservethe Greentimeforanyflowlastsforabout10-30secs,thoughit
roadtonotethegroundtruthofthetrafficsituation. Wecol- cangooveraminuteforcriticalflows.
lect14logsof5minuteseachfromabout5:30pmto7pm. Anysystemlikeours,aimingtoprovidetrafficstateinfor-
Figures 5 and 6 show the CDF of RSSI and packet re- mationtotrafficlights,wouldneedaninputparameterabout
ception rate respectively. Each graph shows 14 plots: each the frequency at which traffic queue estimates are needed.
aCDFcalculatedover5minutes. Asseenfromthefigures, Wewishtodeterminethelowestclassificationtimewindow
thecurvesineachgraphcanbeclassifiedintothreedistinct at which a sufficiently high accuracy(say 90%) can be ob-
groups – Group1 between 5:37-6:21pm, Group2 between tained. Verylowtimewindowsgivenoisypredictions. This
6:22-6:27pmandGroup3between6:30-7:05pm.Theground noise comes from two sources - (a) the inherent stochastic
truthoftrafficstatenotedisfree-flowingtill6:20pm,slowfor nature of wireless links which causes link quality to be in-
about5minutesandthenheavilycongestedtilltheendofthe termittentlybadthoughthetx-rxarein perfectline ofsight
experiment. ThusGroup1correspondsto free-flowingtraf- (b)theinstantaneoustrafficconditionbetweenthetx-rxare
fic, Group2 to slowly moving traffic, intermediate between contrary to the actual traffic state. For e.g. tx-rx may be
free-flowingandcongestedandGroup3toheavilycongested inlineofsightbetweenseveralstandingvehiclesinconges-
traffic. ThehighcorrelationoftheCDFswithtrafficstateis tion (Fig. 10(a))or severalheavy vehiclesmay pass the tx-
apparentvisually. Fore.g.,(a)the50th and70th percentiles rx pairsimultaneouslyin free-flowobstructingline of sight
130
20 10000 20 10000 10000
18 NKum-MbeSeaVr noMsf DTTrraaatiiannp EEoirrnrrootsrr 18 NKum-MbeSeaVr noMsf DTTrraaatiiannp EEoirrnrrootsrr 20 RRSSSSI,I R,P SLRQSRII
8000 8000 18 RSSI, LQI, PRR 8000
Error (%) 11110246 46000000 mber of Datapoints Error (%) 11110246 46000000 mber of Datapoints Error (%) 111246 Number of Datapoints 46000000 mber of Datapoints
Nu Nu Nu
8 2000 8 2000 10 2000
6 6 8
0 0 0
10 20 30 40 50 60 10 20 30 40 50 60 10 20 30 40 50 60
Prediction Time Window (seconds) Prediction Time Window (seconds) Prediction Time Window (seconds)
Figure7: FCalgorithmtrainingerror Figure8:SCalgorithmtrainingerror Figure9:Choiceoffeatures
(Fig.10(b)). Thusweneedtochoosetheclassificationtime accuracyandlabelingoverhead.
window,henceforthreferredtoast,carefully. FeatureClassifier(FC)algorithm-Inthisalgorithm,the
3.3 Labeled data-set forevaluation wirelessdataforacertaintistransformedintoafeaturevec-
tor comprising 9 features: from the set of RSSI values of
To evaluate our choices of t, features and algorithms
packets received in a time window, the nine percentile val-
for real time traffic state classification, we need a data-set
uescorrespondingto10th,20th,...,90thpercentilearedrawn
with labeled ground truth. For this purpose, we use the
asfeatures.Thegroundtruthforthecorrespondingtimewin-
16 hours data-set, collected using the setup in Fig. 4. The
dowisappendedtothegeneratedfeaturevector.Ifnopacket
dataiscollectedfromtwoMumbairoads,a25mwideAdi
isreceivedinatimewindow,adummypackethavingRSSI
Shankaracharya Marg, henceforth referred to as wide-road
of-95dBm,closetotheradiosensitivitylevel,isconsidered
andanotherroad,8minwidth,henceforthreferredtoasnar-
tohavebeenreceived. Thecollectionofalldatapointsthus
rowroad.Specificallywehave13676secsofwide-roaddata
obtainedcomprisesthetrainingdataset. Thisisusedtotrain
labeled as free-flow and 14992 secs labeled as congested.
classificationmodelseitherusingSVMorK-Means. Inthe
Similarly,wehave13486secsofnarrow-roaddatalabeledas
testingphase,similarfeaturevectorsarecreatedfromwire-
free-flowand16678secslabeledascongested.Thelabeling
lessdataoverthesametimewindowt.Theneachtislabeled
wasdoneatalargertime-scaleof5minutestoreduceman-
asfree-floworcongestedbasedonthetrainingmodel.
ualoverheadandalltheonesecondlongwindows,belonging
SignalClassifier (SC) algorithm - This meta-classifier
to the same 5 minutewindowwere uniformlylabeled. The
alternativetoFeatureClassifier,usesmajorityvotingonper-
roads and the times of day of data collection were chosen
packetcongestionpredictions. Inthetrainingphase,weob-
in a way that traffic states did not toggle within 5 minutes.
tainaper-packetclassifier usingK-meansorSVM,bycon-
Thus the error in groundtruth observation, even if present,
sidering only the packet RSSI as a feature. In the testing
isverysmall. Representativevideos,showingfree-flowand
phase, for each t, we employ the per-packet classifier ob-
congestedtrafficonthewideroad,canbefoundat[31].
tained in the training phase on each received packet, ob-
3.4 Classificationalgorithms
(t)
tainingalabelforeachpacket. Considercount and
We use machine learning algorithms to do traffic state congestion
(t)
classification. In this paper, we concentrate on binary traf- count to be the number of packet-level congestion
freeflow
ficstateclassification: congestedtraffic,whenvehicleshave and free-flow predictions in a time slot t. We predict con-
tobrakeandstopvsfree-flowingtraffic,whenvehiclesmove (t)
gestion in the time slot if count is greater than or
accordingtothedriver’sintendedspeed,boundedbytheroad congestion
(t)
speed limit. We show that this binary classification is suf- equal to count and free-flow otherwise. We evalu-
freeflow
ficient for queue length estimation, using binary decisions ate the four algorithms:1) FeatureClassifier using SVM, 2)
fromalineararrayofmultiplesensorpairs. FeatureClassifier using K-Means, 3) SignalClassifier using
Todecidewhatbinaryclassifiertouse,thetrade-offisbe- SVM,and4)SignalClassifierusingK-Means,onourlabeled
tween(1)accuracyofclassification,(2)implementabilityon dataset. Fig. 7 and Fig. 8 show the training errors of the
alowendembeddedplatform,(3)complexityoftheclassi- FeatureClassifierandtheSignalClassifieralgorithmsrespec-
fier modelsand (4) overheadof modeltraining. Linear hy- tively. Aswecansee,FeatureClassifierusingK-Meansout-
perplaneclassifiers are simple enoughto implementon our performs the other algorithms in terms of accuracy. Also,
platform(TI’sC5505)andtotrainandtestinnearrealtime. the classification errorof SignalClassifier is notaswell be-
SVMandK-Means-basedclassifiersbelongtothiscategory havedasFeatureClassifier,andissurprisinglyhigherforthe
and are state-of-the-art supervized and unsupervized learn- supervizedSVMalgorithmthantheunsupervizedK-Means
ing algorithmsrespectively. K-Means, being unsupervized, algorithm. Thisis duetothe increasedlabelnoisein trans-
hastheadditionaladvantageofminimalmanuallabelingof lating the ground truth of a 5 minute time window to each
training data. We build four possible algorithms based on packetreceivedinthatwindow.
these two classifiers andsubsequentlychooseone basedon The accuracy for FeatureClassifier using K-Means is
131
1 1
0.9 0.9
bility 00..78 bility 00..78
a a
ob 0.6 ob 0.6
ative pr 00..45 112550mmm N LLLOOOSSS ative pr 00..45 112550mmm N LLLOOOSSS
Cumul 00..23 222055mmm NN LLLOOOSSS Cumul 00..23 222055mmm NN LLLOOOSSS
0.1 30m LOS 0.1 30m LOS
30m NLOS 30m NLOS
0 0
0 5 10 15 20 25 30 35 40 -95 -90 -85 -80 -75 -70 -65 -60 -55 -50 -45
No. of packets received in 1 sec out of 20 packets RSSI (dBm)
Figure11:Sensingonnarrow
road Figure12:CDFofPRRwithd”variation Figure13:CDFofRSSIwithd”variation
above90%, when the classification time windowis at least road,thetx-rxpairhavetobeplaceddiagonallyacrossthat
20 seconds, shown by a vertical line in Fig. 7. Hence, we road,insteadofbeingperpendicular,toincreasetheeffective
chooseFeatureClassifierusingK-Meansandt as20seconds distancebetweenthem. Butsuchroadspecificsensortopol-
in our Kyun Queue system. The graphs presented here are ogyadjustmentscan be avoided, if RSSI alone is used as a
forthewide-roaddataset,butweobservedsimilarresultsfor feature.
thenarrow-roaddatasetaswell.
3.6 Classificationmodels: summary
3.5 Choiceoffeatures
Thus we use FeatureClassifier algorithm with K-Means
RSSI,LQIandPacketReceptionRate(PRR),allshowed
model built over 20 secs of RSSI percentiles in our Kyun
goodvisualdifferenceindistributionsbetweenfree-flowing
Queuesystem. ThedeployedKyunQueuesystem,described
and congested road traffic. But there are several IEEE
802.15.4compliantradioslikeXBEE3,whichdonotreport in Section 6, uses the training model built from the 8 hour
datasetcollectedfromthe sameroad,with approximately4
LQIvalues.Thusweseektounderstandtheeffectofnotus-
hours of free-flowing and 4 hours of congested data. Effi-
ingLQIfortraffic classification. Using thewide-roaddata,
cient training model building for varieties of roads and to
weplotthetrainingerrorofanSVMclassifierinFig.9. In-
detectdriftsinmodelwithtimeandenvironmentalchanges
terestingly, RSSI percentile-based features yield better ac-
areinterestingaspectsoffuturework.
curacies than using additional percentile features based on
LQIandPRR. Thereasonforthisnon-intuitiveobservation
4 Designofthe Kyun Queue system
is that RSSI is muchmore stronglycorrelatedwith line-of-
Asseenintheprevioussections,onepairoftx-rxacross
sightthanLQIorPRR.
roadcan inferthe roadoccupancylevelbetweenthemwith
ThisstrongcorrelationofRSSI withline-of-sightisalso
significant accuracy. Next we seek to extend this pairwise
evidentfromanexperimentthatweperformonthenarrow-
sensingtobuildanarrayofsensors, whichcanperformco-
road. The goal of the experiment is to observe the effect
ordinated sensing and detect length of traffic queue on a
of distance between transmitter and receiver, on the mea-
givenstretchofroad.
sured wireless link characteristics. The experimentalsetup
is shown in Fig. 11. Directly measuringd′ is difficulton a (1) Three pairs of transmitter (Ti) - receiver (Ri) are shown
busyroadwithvehiclespassingby.Sowemeasured′′. as example. Each pair performs sensing and computation
toknowthetrafficconditionbetweenthem. Thenumberof
Keepingthetransmitterfixed,wemovethereceiverfrom
pairstobeplacedonagivenroadstretch,woulddependon
d′′=5mtod′′=40minstepsof5mandletthereceiverlogfor
theworstcaselengthoftrafficqueuesonthatroad.
5 minutesateachposition. We dothisbothin free-flowing
(2)TheindividualobservationofeachT-R pair,canbecom-
andcongestedconditions. TheCDFofreceptionandRSSI, i i
municatedtoacentralcontrollerunit(C),thatmayresideon
for both traffic states, are shown in Fig. 12 and 13 respec-
thetrafficlight.C,uponreceivingtheroadoccupancyobser-
tively.LOSindicatesline-of-sightconditioninfree-flowand
vationvaluesfromsensorsoneachincominglane,cancom-
NLOS,nonline-of-sightconditionincongestion. We show
pute the optimal green light distribution. This can handle
the plots only for 15m, 20m, 25m and 30m to prevent the
applicationslocaltoaparticularintersection. e.g. minimiz-
figuresfromgettingcluttered.
ingworstcasewaitingdelays.
As we can see, both free-flow (LOS) and congested
(NLOS) traffic show almost identical PRR for d′′ < 25 m. (3)Ccancommunicatewithaserver(S),thatmayresideina
remotetrafficcontroloffice.Theserver,uponreceivingroad
This is because, signal quality is quite good at small dis-
occupancy information from several controller units of the
tances such that NLOS does not have much effect on the
city-wide road network, can implement other applications
reception of packets. RSSI, which directly measures sig-
likeco-ordinatedsignalcontrol,bottleneckidentificationand
nalquality,however,showsgooddifferencebetweenthetwo
congestionmapping.
traffic states even for small d′′ (Fig. 13) . Thusif PRR has
to be useful as a feature to detect traffic state on a narrow 4.1 Architecture
3WeuseXBEEradiosinourhardwareprototype. TheproposedsystemarchitectureisshowninFig.14.
132
We keep a Telosb mote stationary which transmits 25
packets per second. Another mote is kept at distance x m
from the transmitter; we vary x=30 m to x = 100 m, in
steps of 10 m, logging number of packets received for 5
minutes at each distance. Both motes are 0.5 m above the
ground. Theyareplacedalongtheroad,onthesidewalkof
AdiShankaracharyaMarg,insteadofbeingplacedacrossthe
roadlikeinpreviousexperiments,sincewewanttoevaluate
thecommunicationandnotthesensinglinks. Wedotheex-
Figure14:SystemArchitecture perimentonceat6aminthemorning,whenthesidewalkis
emptyandagainat8pmintheevening,whenthesidewalk
iscrowdedwithpedestrians(seeFig.15)andrepeatbothex-
4.2 WhichRF technology to use forsensing perimentsoverfourdays.
Thoughwestartwith802.15.4tosensetrafficaswehave
easyaccesstoTelosbmotes,wealsoexperimentwith802.11
a, b and g radiosand Bluetooth. The experimentalsetup is
similarinallcases.Wekeepatx-rxpairacrossroad,atabout
0.5mfromthegroundandmeasureRSSIandpacketrecep-
tionrateattherx. Powerrequirementsfor802.11a/b/gare
higherthan802.15.4[32]. Bluetoothshowsanissueofpoor
rangeandlinksarehardtoestablishforwideroadsof25m.
Secondly,beingaconnectionorientedprotocol,ithasissues
oflongdelaysinre-establishingconnectionfollowingpacket Figure16:Sensorpairsetup
losses. Thus,thoughallthetechnologiesareaffectedbycon-
gestion to some extent and can perhaps be used for traffic The median packet reception ratio (PRR) is 60% at 60
sensing–basedonaccuracy,power,range,form-factoretc., m at 6 am. The link quality degradesat 8 pm whenpedes-
802.15.4seemstobeabetterchoicethantheothers. trians on the sidewalk block the line of sight between the
motes.MedianPRRisatmost40%atanydistanceabove30
4.3 Sensing andcommunication conflicts
m. SuchlowPRRwouldmakeourcommunicationlinksun-
Inoursystem, we havetwo typesofwireless links– (1)
reliable, andhenceunsuitableforapplicationslike adaptive
sensinglinks, acrossthe road, fromT to Rand(2) commu-
trafficsignalcontrol,wheretimelinessiscritical.
nicationlinks,alongtheroad,fromoneRtoanotherR.The
To resolvethis issue, we choose to use two 802.15.4ra-
sensing links should be affected by the traffic flow on the
diosinourreceiver(R)units- oneXBEEradioforsensing
road,sothatwirelesslinkcharacteristicsmeasuredonthem
andaCC2520radioforcommunication. Thesetupforatx-
reflect the traffic volume. This necessitates the radioson T
rx pair is shownin Fig. 16. T is the tx unitacrossthe road
andRtobeatalowheightofabout0.5mfromtheground
containingamicrocontrollerandanXBEEradiototransmit
level,suchthatthepacketsareblockedbythebodyoftheve-
sensing packets. R is the rx unitwith a microcontroller,an
hicles. Ontheotherhand,inatypicaldeploymentscenario,
XBEEradiotoreceivesensingpacketsandaCC2520radio
the R units will be mountedon road-sidelamp-posts. Inter
tocommunicatetootherR’s. TheCC2520antennaisplaced
lamp-postdistanceisintheorderof30m.Evenifweputour
higherusing RF cable, atabout2 m fromthe ground,clear
unitsoneachlamp-post,anynetworkfaultmightneedunits
ofpedestriansonthefootpath.Otherpossibledesignchoices
on alternate lamp-posts to communicate. Thus, we should
tohandlethesensing-communicationconflict,arediscussed
keep provisionfor at least 60 m long communicationlinks.
inSection8.
These links have to be reliable as well. The question thus
arises: can a single radio handle both sensing and commu- 4.4 Software protocol
nicationlinkswithconflictingrequirements?Thisisverified TocorrelatethetrafficstatedecisionsofdifferentRunits
next,throughasetofexperiments. and calculate the queue length, we have to ensure that
all R’s perform sensing and computation simultaneously,
so that their individual measurements are co-ordinated in
time. Since our architecture has a C unit, we choose to
use centrally controlled measurement cycles, triggered by
C, to achieve this. Thus we do not need any explicit time-
synchronizationmechanismamongtheunits.
Onepossiblewaytodesignthesoftwareflowofournet-
work is outlinedin Fig. 20. The R unitsremain in receiver
readymodeofCC2520radio(C-RDY)onpowerup,waiting
fora controlmessage. C,onpowerup, asksforthecurrent
Figure15:Obstaclesonsidewalk:pedestriansand
time from a server over GPRS and initializes its real time
impatientmotorcycliststoo!
clock. C then sends a control message which each R re-
133
Figure17:Transmitter(T) Figure18:Receiver(R) Figure19:Controller(C)
work. TDMA needs strict time-synchronization among
units.Ontheotherhand,bydesign,bothourcontrolanddata
messagesaretransmittedsequentiallybyoneRfollowedby
thenextR.ThussimpleCSMA-CAcanhandleourMACis-
sues, and this is what we use in our network. To increase
reliability, all messages are transmitted four times. If there
isstillamessageloss,ourdesignhandlesitbyusingatime-
out in D-RDY state. Upon timeout, the unit goes back to
C-RDYstateandparticipatesinthenextmeasurementcycle
whenthenextcontrolmessagecomes. Thusthefixednum-
ber of retransmissions allows us to achieve a good balance
betweenresiliencetostraywirelesslossesandimplementa-
tioncomplexity.
C keeps track of the currentmeasurementcycle number
Figure20:Softwareflow bygeneratingandinsertingasequencenumberinthecontrol
message. IfCreboots,itlooksupthelastsequencenumber
fromtheMicro-SDcardandgeneratesthenextone. IfanR
ceivesandtransmitstothenextR.Aftertransmission,each reboots, it simply waits in C-RDY and copiesthe sequence
RstartssensingtheincomingpacketsfromTontheXBEE numberofthefirstcontrolmessageitgets,asthecurrentse-
radioandcomputesitsbinarydecisionoftrafficstatebased quence number. None of the R’s can generate a sequence
onthealgorithmdescribedinSection3. Oncethisdecision number. This ensures that though the sequence numbers
isready,eachRentersreceiverreadymodeofCC2520radio wraparoundafter0-255,thereisnostale sequencenumber
(D-RDY),waitingforthedatamessagefromnextR.Thelast in the network. Thus units can confidently rejectmessages
Rcreatesadatamessagetosenditsdecisiontotheprevious containingsequencenumbersalreadyseenorwhichareout
R.EachR,uponreceivingthedatamessagefromnextR,ap- oforder,asretransmittedmessages.
pendsitsowndecisiontothemessageandtransmitsittothe IftheCandRnodesarearrangedalongaroad,asshown
previousR. in Fig. 14, this software protocoluses links between C and
WhenthedatamessagereachesC,itwritesthedatames- R1,thefirstRialongtheroad,andthenbetweeneachRiand
sage, along with the time of its reception, in the Micro-SD Ri+1. Oursystem,usingdualradio,hasprovisionforlonger
card. This ensures retention of data and control state in- communication links between Ri and Ri+2, which the cur-
formation,in case C rebootsorserver communicationgoes rent software does not utilize. All the messages are routed
downtemporarily.Ccancomputethetrafficsignalschedule alongthehardwiredpathofconsecutive{Ri,Ri+1}pairs.The
fromthequeuelengthinformationinthedatamessage.4 All longer links may be used for RSSI based self-localization,
R unitsgo back to C-RDY after transmittingdata message. whichwediscussinSection7. Alsoinfuture,faultswherea
Cappropriatelysendsthenextcontrolmessage,whenitin- particularRifailsorlinkbetweenanyconsecutive{Ri,Ri+1}
tendstostartthenextmeasurementcycleandthisgoesonin fails,canbedetectedorcorrectedusingthelongerlinks.
a loop. While theR’s dosensingandcomputation,Creads 5 Hardwareprototypes
theMicro-SDcardandsendsthedatamessageoftheprevi-
Basedonthedesignchoicesoutlinedabove,wehaveim-
ous cycle to the server overGPRS. The server can use this
plementedthreehardwareprototypesforT,RandC.TheT
informationfromdifferentCunitsinthecityforco-ordinated
units have (1) an Atmel AT89C51RD2microcontrollerand
signalcontrolorforvisualizationofcongestionmaps.
(2)anXBEEradioonUARTtotransmitsensingpackets. In
Next we consider the MAC protocol to use in our net-
theRunits,wehaveused(1)aTIC5505ezdspstickforthe
computation, (2) a TI CC2520 radio connected on SPI for
4Theitalicizedpartsofthesoftwareflowliketrafficlightschedulecom-
communicationwithotherRunitsandtheCunitand(3)an
putationandco-ordinatedtrafficsignalcontrolhavenotbeenimplemented.
Negotiationsforthesearecurrentlyunderprogresswith[9]. XBEE radioon UART for sensing. Four 1.5Vbatteriesare
134
connectedinseriestoprovide6Vpowertoeachcircuit. The fline fromthe8 hours(approximately4 hourseach offree-
Cunithas(1)aTIC5515ezdspstickforprocessingwith(2) flowingandcongestedtraffic)dataset,collectedonthesame
an integrated Micro-SD card for data storage. There is (3) roadaboutoneyearback. Usingthestoredmodel,wethen
aTICC2520radioonSPItocommunicatewiththeRunits run the online FeatureClassifier algorithm on R, to classify
and (4) a SIMCOM SIM300C GPRS modem to communi- chunksof20secondsofRSSIpercentilesofpacketscoming
cate to a server. An 11.1V, 1 Amp peak current battery is from T. We experimentfor 50 minutesin each of free-flow
usedforpoweringupthiscircuit. and congested traffic. 149 out of 150 data points are cor-
ThechoiceoftheTIezdspsticksarefortheirsmallform rectly classified as free-flow and all of 150 data points are
factors, low cost, low power requirements and convenient correctly classified as congested. Thus overall accuracy of
pinoutsofseveralinterfaceslikeUARTandSPI.Thesemake onlineclassificationonthisparticularsetofunseentestdata
hardwareintegrationeasy. ThevastresourceofTIchipsup- pointsis99.67%.
portlibraryfunctionsandup to 100MHz clockrates make
programming on these platforms convenient as well. The
twoIEEE802.15.4compliantradios,neededintheRunits,
are chosen to be CC2520 and XBEE. This is because they
use two differentinterfaces of SPI and UART respectively,
whichsimplifiesprototypedesignandimplementation.
AC5505ezdspstickcosts$50,aC5515ezdspstickcosts
$80,aGPRSmodemcosts$70,anXBEEradiocosts$18,a
CC2520 radio costs $50, an AT89C51RD2 microcontroller
costs $4. With interfacingPCB’s, connectorsand batteries,
Figure22:ViewfromAndroidPhone
areceiver(R)-transmitter(T)paircostsabout$200andthe
controller(C)costsabout$250.
Nextwe seek to evaluate the accuracyand timeliness of
6 Deployment based evaluation
queueestimatesgivenbyournetworkofsensors.Wedeploy
What is the accuracyof online traffic state classification oneCunitona lamp-postnearto thetraffic signalandfive
byasinglesensorpair?Whatistheaccuracyofqueuelength T-Rpairsonthenextfiveconsecutivelamp-posts. Allunits
estimatesgivenbyanetworkofsuchpairwisesensors? Are are packaged in ABS plastic boxes with holes to bring out
thequeueestimatesavailableinrealtimeattheserver? Can the XBEE and CC2520 radio antennas. The C and the R
weseeanypatterninthequeuelengthsoveradayorbetween units are clamped to lamp-posts on the sidewalk, at about
days? Weseektoevaluatethesequestionsthroughadeploy- 0.6mabovetheground. TheirCC2520antennas,connected
mentofoursensornetwork.Weperformallourexperiments totheradioexternalportswithRFcables,aretiedvertically
on a stretch of Adi-ShankaracharyaMarg road in Mumbai. tothelamp-postsat2mabovetheground.TheCunithasan
This road is about25m wide and has fair amount of traffic additionalGPRSantenna. TheTunitsareclampedtolamp-
throughoutthedayasitconnectstwoexpresshighways. postsonthedivider,perpendicularlyoppositetotheRunits,
Toevaluatetheaccuracyofonlineclassification,weplace such that a T-R pair face each other across the road. The
a T unit on a divider lamp-post and an R, perpendicularly deployedunitsandthedeploymentsiteareshowninFig.21.
across the road on the sidewalk lamp-post. R has a stored
K-Means clustering model using nine RSSI percentile val-
uescorrespondingto10th,20th,...,90thpercentile,measured
7
over t =20 secs, as features. This model is computed of- ber) 6 bouancdtueadl sseennsseedd ddaattaa
num 5 ground truth
e
od 4
n
gth ( 3
n
e le 2
eu 1
u
Q
0
18 18.25 18.5 18.75 19 19.25 19.5 19.75 20 20.25 20.5 20.75 21
Time of day in hours
Figure23:6-9pm,Nov17,2011,Saturday
We implement our designed software for this deployed
network, such that C initiates a new measurement cycle
throughacontrolmessageevery30secs. TheRunitssense
for20secondseachandcomputeindividualbinarydecisions
aboutthetrafficstate. Thesedecisionsarecommunicatedto
Cinadatamessage. CstoresthemessageinMicro-SDcard
Figure21:DeployedUnits
andalsosendsitoverGPRStoaremoteserver.Themessage
135
Date #ofdetections %Exactmatches %Errorof1unit %Errorof2units %Errorof3units
(about30m) (about60m) (about90m)
Nov17 359 74.09 20.05 4.45 1.39
(fp=14.76,fn=5.29) (fp=2.23,fn=2.23) (fp=1.39,fn=0)
Nov18 359 96.37 1.67 1.39 0.56
(fp=1.67,fn=0) (fp=1.39,fn=0) (fp=0.56,fn=0)
Nov19 353 90.93 7.64 1.97 0
(fp=3.39,fn=4.24) (fp=0.28,fn=1.13) (fp=0,fn=0)
Overall 1071 87.11 9.8 2.4 0.65
(fp=6.62,fn=3.17) (fp=1.3,fn=1.12) (fp=0.65,fn=0)
Table2:Accuracyanderrorbreakupsofdeploymentresults
isintheformofanarrayof5binaryvalues,eachsignifying spectively.Toaidvisualcomparisonwithgroundtruthwhich
decisionbyanR,inincreasingorderofthelamp-postsfrom coversuptounit3,weplotboththeactualqueuevaluesre-
the C unit. The server logs these updates coming every 30 portedbyoursystem,termedasactualsenseddatainthefig-
secondsand computesthe queue length as the unitnumber ures, and the min(3,actualsensed data), termedas bounded
of the last R reportingcongested state. Thus our measured senseddatainthefigures.
queuelengthscantake6discretevalues:0,whereallR’sre-
portfree-flow,1whenonlyR1reportscongestionwhilethe
7
othersreportfree-flow,2whenR1andR2reportcongestion ber) 6 bouancdtueadl sseennsseedd ddaattaa
whileothersreportfree-flow,3,4andfinally5,whenallR’s num 5 ground truth
reportcongestion.WerunthisdeployednetworkonNov17, ode 4
Thohuurrssdeavye,ryNdoavy,1b8e,tFwreideany6a-9ndpmNo.v19,Saturday,2011,for3 ngth (n 3
To know the accuracy of our measurements, we use an eue le 12
u
image-based manual verification scheme. We run an An- Q
0
droidapplicationonaSamsungGoogleNexusphonetocap- 18 18.25 18.5 18.75 19 19.25 19.5 19.75 20 20.25 20.5 20.75 21
Time of day in hours
ture an image every 30 second and store it. The phone is
placed on the roof of a four storeyed building by the road- Figure24:6-9pm,Nov18,2011,Saturday
side. The phone can cover T-R pairs 1, 2 and 3. We tried
different apartment buildings by the roadside and different
orientationsandzoom-levelsonthephone,butthiswas the AswecanseefromTable2,theaccuracyisupto96%on
maximumnumberofsensorsthatwecouldcover. Theview Nov18andNov19.Ofthethreedays,Nov17hasminimum
fromthephoneisshowninFig.22.Inthefigure,Cisfurther accuracy: 74%. But as seen from Fig. 23, queue buildup
totheleftofsensorpair1andT-Rpairs4,5arefurthertothe andclearingwasveryrapidonthatday,increasingthechal-
rightofsensorpair3. Theimagesforthethreedayscanbe lenge of deciding queue length by manual observation. So
viewedat[31]. Onepersonobservestheimagesofflineand our low accuracy is likely a combined effect of our errors
estimatesthequeuelengthsmanually. Incase thisobserver anderrorsofmanualgroundtruthestimation. Forexample,
findsitdifficulttoestimatethelengthfromtheimage,asec- theinstanttheimageistaken,thequeuemighthavecleared
ondobserverisconsultedandthequeuelengthisascertained but it might have been present for most of the 20 seconds
bytheirmutualconsent. ofsensing,leadingtoafalsepositive. Acaseoffalsenega-
The accuracy and break up of errors of our deployment tivewouldoccurifaqueuebuildsuptheinstanttheimageis
resultsaresummarizedinTable.2. Errorof1unitindicates taken,whilemostofthesensingtime,trafficisfree-flowing.
thatourqueueestimateandthegroundtruthdifferby1.This A better way of ground truth estimation would be to take
in turn can be a case of false positive (fp) or false negative continuousvideo,insteadofaninstantimage,andweaccept
(fn). Thefalse positivesand false negativesaredetermined thistobea limitationofthiswork. Falsenegativesarerare
in the following way. The viewer of the image determines fortheerrorvaluesof2and3,asinstantgrowthandreduc-
thecurrentqueuelengthbyseeingtheimage.Thisisconsid- tion of long queues is non intuitive. But even on Nov 17,
eredasthegroundtruth. Ifthequeuelengthreportedbyour almostalltheerrorsareonlyofoneunitandhighererrorsof
sensor system is more than the ground truth queue length, 2-3units,giveninthelasttwocolumnsofTable2,arevery
thatisconsideredasfalsepositive,asweareoverestimating low. The above results indicate that, the Kyun Queue sys-
thequeue.Similarly,ifthereportedqueuelengthislessthan temisaccurateinestimatingtrafficqueuelengths,andshows
thegroundtruthqueuelength,thatisconsideredasfalseneg- goodpromiseforuseinapplicationssuchasautomatedtraf-
ative,asweareunderestimatingthequeue. Errorofupto3 ficsignalcontrol.
unitscanoccur,asimagescoverunits0-3. An interesting point to note from the figures is: queue
Fig.23,Fig.24andFig.25showthelengthofthequeue lengthcanbefairlyvariable(1)overthreehoursonasingle
asmeasuredbyoursystemonthethreeconsecutivedaysre- day, as seen in Fig. 23, (2) between two week days at the
136