Spatio-temporal multidimensional collective data analysis for providing comfortable living anytime

17 Pages·2017·3.24 MB·English

SIP(2018),vol.7,e4,page1of17©TheAuthors,2018.
ThisisanOpenAccessarticle,distributedunderthetermsoftheCreativeCommonsAttributionlicence(http://creativecommons.org/licenses/by/4.0/),whichpermits
unrestrictedre-use,distribution,andreproductioninanymedium,providedtheoriginalworkisproperlycited.
doi:10.1017/ATSIP.2018.4
industrial technology advances
Spatio-temporal multidimensional collective
data analysis for providing comfortable living
anytime and anywhere
naonoriuedaandfutoshinaya
Machinelearningisapromisingtechnologyforanalyzingdiversetypesofbigdata.TheInternetofThingserawillfeaturethe
collectionofreal-worldinformationlinkedtotimeandspace(location)fromallsortsofsensors.Inthispaper,wediscussspatio-
temporalmultidimensionalcollectivedataanalysistocreateinnovativeservicesfromsuchspatio-temporaldataanddescribethe
coretechnologiesfortheanalysis.Wedescribecoretechnologiesaboutsmartdatacollectionandspatio-temporaldataanalysis
andpredictionaswellasanovelapproachforreal-time,proactivenavigationincrowdedenvironmentssuchaseventspacesand
urbanareas.Ourchallengeistodevelopareal-timenavigationsystemthatenablesmovementsofentiregroupstobeefficiently
guidedwithoutcausingcongestionbymakingnear-futurepredictionsofpeopleflow.Weshowtheeffectivenessofournavigation
approachbycomputersimulationusingartificialpeople-flowdata.
Keywords: Spatio-temporaldataanalysis,IoT,Smartcities,Proactivenavigation,Machinelearning
Received8August2017;Revised19January2018;Accepted19January2018
I. INTRODUCTION As envisioned by MLC, the research and development
visionsforbigdataanalysistechnologiesintheIoTeraare
Machinelearningisapromisingtechnologyforanalyzing shown in Fig. 1. From this perspective, regression analy-
diversetypesofbigdata.However,sinceasinglecoretech- sis, which is representative of conventional data analysis,
nologyisinsufficientforachievinginnovativeservicesthat aimstodescribeobjectivevariablesusingmultipleexplana-
exploitbigdata,acompositetechnologythatcombineskey tory variables. That is, it has an analysis technology for
technologiesisessential.Furthermore,toconducttechnol- determining whether an objective variable (e.g. sales) can
ogy trials in the field and uncover business needs, close be expressed as a function of explanatory variables (fac-
discussions are needed with specialists in applied fields. tors).IntheIoTera,however,thereisaneedfortechnology
Against this background, the Machine Learning and Data thatcanextractlatentinformationspanningheterogeneous
ScienceCenter(MLC)wasestablishedinApril2013atNTT multidimensional datasets that cannot be discovered by
Laboratoriesasaninter-laboratorycollaborativeorganiza- individually analyzing different sets of data. To this end,
tionfocusedonbigdataanalysis[1]. we developed a technique called “multidimensional data
It has been more than a decade since the concept of analysis” that enables multiple heterogeneous sets of data
“bigdata”wasfirstproposed.Unfortunately,definitionsof to be simultaneously analyzed whose usefulness has been
big data and its analysis were unclear from the start. In demonstratedthroughactualfieldtrials[2–4].
time, though, advances in sensor technologies will enable With the improvement of sensing technology and the
the use of sensors in all sorts of fields including social rapid spread of smartphone applications and IoT devices,
infrastructure, medicine and healthcare, transportation, various big data on car and object movements, human
and agriculture. An environment is currently evolving in behavior,andenvironmentalchangescanbemeasuredand
which massive amounts of data can be collected and ana- collectedeverywhereinrealtime.Bymakingmovingenti-
lyzedinrealtime:thebirthofaconceptcalledtheInternet ties such as people, cars, and things function as sensors,
ofThings(IoT).ThroughIoTthetruenatureofbigdatais wecanefficientlycollectavarietyoffine-graineddatacol-
finallybeingrevealed. lectionofreal-worldinformationlinkedtotimeandspace
(location)frommanykindsofsensors.Thistypeofinfor-
NTT Communication Science Laboratories, NTT Corporation, 2-4 Hikaridai, mation is called “spatio-temporal data”. In this paper, we
Seika-cho,Soraku-gun,Kyoto,Japan.Phone:[phone]
alsodescribeourtrialsthatcollectedanddetectedcity-wide
Correspondingauthor: eventsbygovernmentvehiclesinFujisawacity,Kanagawa,
N.Ueda
Japanasmobileenvironmentalsensors[5–8].
Email:[email protected]
1
https://doi.org/10.1017/ATSIP.2018.4 Published online by Cambridge University Press
2 naonori ueda and futoshi naya
(cid:129) Navigationshouldcopewiththeavoidanceofcongestion,
while also takinginto account the intentionsof individ-
uals as much as possible. That is, it must enable many
peopletoreachtheirdestinationsasquicklyaspossible.
Our challenge is to develop a real-time navigation system
that satisfies the above ideal navigation characteristics. In
thispaper,wedescribeournavigationsystemandshowits
effectivenessbycomputersimulationusingartificialpeople-
flowdata.
Therestofthepaperisorganizedasfollows.InSection
II, we describe our smart data collection system to detect
Fig.1. CoretechnologiesforeraofbigdataandInternetofThings. city-wide events. In Section III, we briefly explain our
machinelearningtechnologiesforspatio-temporalmultidi-
mensionalcollectivedataanalysis.InSectionIV,weexplain
ReferringtoFig.1,time-seriesanalysismodelsthetem- real-time,proactivenavigation,whichisacoretechnology
poral interaction or cause-and-effect relationship among forprovidingcomfortablelivinganytimeandanywhere.
data, but spatio-temporal analysis constructs models that
alsoconsiderthespatialdynamicsamongdata.Toanalyze
II. SMART CITY-WIDE DATA
thespatio-temporalbehaviorofpeopleandthingsandpre-
COLLECTION USING
dict when, where, and what in real time, we established a
GOVERNMENT VEHICLES
research theme called “spatio-temporal multidimensional
collective data analysis”. This form of analysis addresses
In the following, we describe our trials that collected and
timeandspacealongmultidimensionalaxesandusespast
detected city-wide events using government vehicles in
datafromacertainperiodoftimetolearnaboutthemutual
Fujisawacityasmobileenvironmentalsensors.
relationships between time and space with respect to the
“flow”ofpeople,things,information,etc.
Theglobaltrendofurbanizationhasincreasedthecon- A) City-widedatacollectionbysensorized
centrationofpeopleinmetropolitanregions[9].Thistrend governmentvehicles
willintensifycongestionatmasseventsaswell.Pedestrians
One of the most important concerns of citizens in urban
oftengetlostinunfamiliarurbanenvironmentsorinlarge,
areasisthehealthimpactofairpollution.TheWorldHealth
indoor spaces. Even if people are familiar with their sur-
Organizationestimatesthat4.3milliondeathsoccurannu-
roundings,theywilllikelyhavedifficultyevacuatingduring
allyfromexposuretoindoorairpollutionand3.7million
emergencies. During Japan’s most recent natural disaster,
to outdoor pollution [12]. Although cities in developing
2011’sGreatEastJapanEarthquake,manypeoplecouldnot
countriesaregenerallymorepollutedthanfirst-worldcities,
return to their homes due to heavy congestion caused by
millionssufferworldwidefromallergiestocedarpollenand
damage to the infrastructure of highways and train lines.
PM concentration.Accumulatingfine-grainedairquality
Insuchsituations,eitheroutdoorsorindoors,guidanceon 2.5
informationanddeliveringittothepublicareimportantto
demand must be provided that is individually tailored to
assessurbanairconditionstoprotectthehealthofcitizens.
humanneeds.
However,acquiringdetailedairqualityinformationisusu-
As one critical application of spatio-temporal collec-
allydifficultbothspatiallyandtemporallybecauseexpen-
tive data analysis, we describe our novel approach for
sive atmospheric observation stations are limited even in
real-time and proactive navigation in such crowded envi-
cityareasandthesamplingratefrequencyisaslowasjust
ronments as event spaces and urban areas where many
onceashour.
peoplearesimultaneouslymovingtowardtheirdestinations
Tocopewiththeaboveissue,wehavebeeninvestigating
[10,11].Theidealnavigationsystemrequiresthefollowing
city-wideeventdetectiontechnologyusingenvironmental
characteristics:
data collected by car-mounted sensors. We have installed
dozens of sensors on garbage trucks that operate daily all
(cid:129) Navigation should be optimized not for individuals, but overFujisawacity[5,6].Figure2showsasensorizedgarbage
foroverallhumannavigationsothatnewcongestionwill truckandablockdiagram.Thetruckisequippedwithvar-
notbecausedbyindependentindividualnavigations. ious environmental sensors, such as NO , CO, O , PM ,
2 3 2.5
(cid:129) Navigation should be performed in real time because pollen,temperature,humidity,luminance,UV,twotypesof
countermeasuresmustbefacilitatedbasedonthepeople- air contaminants, and a microphone for measuring ambi-
flowchangesfrommomenttomoment,dependingonthe entnoise.Thesesensordataaresampledeveryfewseconds
people-flowcondition. bythesensortype,andmanysensorswithglobalposition-
(cid:129) Navigation should be proactive; it should be achieved ingsystem(GPS)geolocationinformationaretransmitted
before congestion occurs by predicting near-future toaremoteserverbyamobileInternetconnectionservice
people-flowconditions. every30s.Suchfine-grainedenvironmentaldatahelpdetect
https://doi.org/10.1017/ATSIP.2018.4 Published online by Cambridge University Press
spatio-temporal multidimensional collective data analysis 3
Fig.2. Sensorizedgarbagetruckandblockdiagram.
Fig.3. HeatmapsofNO2andambientnoiselevelsofFujisawacity.
spatio-temporaleventsinmoredepth,e.g.theemergenceof services in the future. Even though developing countries
airpollutionhotspotsandthegenerationofambientnoise. are transitioning toward better waste management, they
Figure3showstheheatmapsofNO andtheambientnoise have insufficient collection and inefficient waste disposal.
2
levelsofFujisawacitybydatacollectedfrom1month.Some Sincerapidpopulationgrowththroughurbanizationcreates
partsofthemainroadsarenoisierthanotherroadsaswell a huge amount of solid waste, appropriate waste manage-
assomeparticularhotspots. ment is required based on future population projections
Normally,fordesigningsuchasensingsystem,thedata for each region. To this end, we have been conducting
sampling intervals and communication frequencies are experimentstoevaluateregionalsolid-wasteproductionby
fixed.However,theseparametersforcapturingchangesin motionsensorsmountedongarbagetrucks[8].
sensordatawithfinergranularityareoftenunknownuntil Inconventionalwastemanagement,theamountofsolid
thedataarecollectedandanalyzedonce.Theyalsodepend wasteforagivenareaissummarizedbyweighingthetotal
ontheservicerequirementsfromcitizensandlocalmunic- waste delivered to incineration plants by garbage trucks
ipalities. To address this, we developed a remote program assigned to the region. In fact, one garbage truck collects
relocation function that allows us to remotely change the solidwasteinmultipleseparatedareastobalanceworkloads
program on the sensor nodes on demand. Detailed infor- amongmultipletrucks.Therefore,estimatingtheweightof
mationisavailable[6,7]. regionalwasteisdifficultbasedonthesummarizedweight;
it is roughly estimated based on human intuition, which
includes sizable error. From interviews with municipality
B) Smartwastemanagement
officials, we concluded that a method that accurately esti-
Another important issue in cities across the world is matestheamountofwasteforeachareaisrequiredforplan-
waste management. Efficient waste management is essen- ninggarbagecollectionschemesbasedonregionalseasonal
tialforlocalmunicipalitiestosustainstablewastecollection variations in the amount of solid waste. This information
https://doi.org/10.1017/ATSIP.2018.4 Published online by Cambridge University Press
4 naonori ueda and futoshi naya
Fig.4. Accelerationmeasurementpatternsandtheirspectrogramswhilecollectinggarbage,driving,andidling.
willenablemunicipalitiestominutelyestimatefuturesolid- blue, green) correspond to estimated garbage collection
wasteamounts,thusincreasingtheefficiencyofwasteman- spots by different operator companies and/or municipal-
agement.Theregionalgarbageamountinformationisalso ity sections. Currently, 60 of 100 garbage trucks are sen-
important to promote citizen awareness and reduce the sorized. The bottom-left of Fig. 5 shows choropleth maps
amountofwaste. of the total amount of solid waste summarized weekly by
Tomeettheaboverequirements,weproposedamethod town region. By accumulating long-term spatio-temporal
forestimatingtheregionalamountsofsolidwastebasedon garbage amount data, we can analyze spatial and seasonal
thegarbagecollectiondurationineachareaandthewaste tendenciesalongwithcensusstatisticsofthecity.Further-
weightmeasuredatincinerationplants.Thegarbagecollec- more,bycreatingasystemtoprovidefeedbackoftheabove
tion duration at garbage collection place in each house or information, we can expect to quantitatively evaluate the
apartment building is estimated from the garbage truck’s effectivenessofourapproachforwastereduction.
vibration that is caused when it scoops up garbage with
a rotating plate. We mounted motion sensors on garbage III. SPATIO-TEMPORAL
trucks to measure and detect specific vibration patterns MULTIDIMENSIONAL DATA
whencollectinggarbage,driving,idling,andsoon.Figure4 ANALYSIS
showsexamplesoftheaccelerationsensordataofthehor-
izontal axis parallel to the truck’s traveling direction and In this section, we describe our two representative results
their spectrograms. Peak acceleration frequencies differ ofspatio-temporalmultidimensionaldataanalysis.Thefirst
dependingonthesituationofthegarbagetrucks. isamethodforpatterndiscoveryandmissingvaluecom-
We annotated the sensor data with labels based on pletionfromspatio-temporalcountingdata.Thesecondis
their situations by listening to the sounds recorded by amethodforestimatingpeopleflowfromspatio-temporal
microphones attached to the garbage trucks. By applying populationdata.
astandardsupervisedlearningmethod,suchasasupport
vectormachine,totheextractedspectrogramfeaturesofthe
A) Spatio-temporalpatterndiscoveryand
accelerometerdata,wecanclassifyandspottemporaldura-
missingvaluecompletion
tionsthatcorrespondtogarbagecollection.Theestimated
temporal durations of the garbage collection are recorded Variouskindsofspatio-temporaldata,suchastrafficflow,
with the truck’s locations from a GPS receiver. When the purchasingrecords,andweatherdata,areobservedbysen-
garbage truck arrives at the incineration plant, the total sormonitoringsystemsinsmartcities.Thesedataareuti-
weight of its garbage is measured. We then calculate the lizedtounderstandthedynamicsofacity’shumanactivities
amountofsolidwastebyregionbydistributingtheweight anditsenvironmentalconditions[13].However,asthenum-
fromtheoperatingdurationoftherotatingplate.Herewe berofsensornodesincreases,itbecomeshumanlyimpos-
assumethattheoperatingdurationoftherotatingplateis sible to check every dimension of these spatio-temporal
proportionaltotheamountofsolidwaste. data.Currentmonitoringsystemsarenotalwaysstableand
Figure 5 shows a screenshot of our regional garbage often fail to observe data, because of problems with sen-
amount visualization system. Its right side shows a map sor nodes and data transmission errors. Thus, in practice
that compares the estimated regional solid-waste collect- we need to deal with missing values that are included in
ing spots for two different weeks. The color dots (orange, spatio-temporaldata.
https://doi.org/10.1017/ATSIP.2018.4 Published online by Cambridge University Press
spatio-temporal multidimensional collective data analysis 5
Fig.5. Screenshotofregionalgarbageamountvisualizationsystem.
Non-negativetensorcompletion(NTC)[14],whichwas trafficdataprovidedbyCityPulsecollectedinAarhus,Den-
proposed for missing value completion, is an extension mark [16]. The horizontal and vertical axes in Fig. 6(a)
of non-negative tensor factorization (NTF) that simulta- correspondtothetimesandthevehiclecounts.Figure6(a)
neously extracts latent factors from observed values and shows the daily periodicity of the vehicle counts, and we
infers the missing values. Tensor is a generalization of a identifiedatemporalcontinuityforthembetweentwoadja-
matrix that can represent a high-dimensional data array centtimes.ThearrowsinFig.6(b)correspondtothestart
without any loss of data attributes. NTF decomposes a and the end points of each monitoring site. The red and
non-negativevaluetensorintosparseandreasonablyinter- bluearrows,respectively,correspondtothelargeandsmall
pretablelatentfactorsandiswidelyusedformanyapplica- vehiclecountsatthesites.FromFig.7(b),weconfirmedthe
tions [15]. However, the conventional NTC does not fully spatialcontinuityofthevehiclecountsamongneighboring
utilize such spatio-temporal data properties as the orders siteswithsimilardirections.
oftimepointsandthetopologyofsensors,althoughthese Examples of extracted spatial and temporal factor seg-
structuresshouldbeutilizedtoregularizethelatentfactors. ments from CityPulse of our proposed method and NTC
Toaddressthisproblem,weproposedanewNTCmethod withtheEucliddistanceareshowninFig.7.Byutilizingour
[4]. graph-based regularizer, our proposed method extracted
Asadiscrepancymetricandspatio-temporalstructures, interpretablefactors.Withdayandtimemodefactors,dif-
the generalized Kullback–Leibler (gKL) divergence was ferent temporal dynamics were extracted in Figs 7(a) and
employed as regularizers. We used a gKL that fits partic- 7(b). With the site mode factors, different spatial flows of
ularly well if the data are non-negative integer values. To vehicles were extracted in Figs 7(c), 7(d), and 7(e), show-
incorporate spatio-temporal structures, we introduced a ing that factor 2 revealed a traffic pattern that occurred
graphLaplacian-basedregularizeranduseditforinducing onweekdaymorningswherevehicleswereheadingtoward
latentfactorstorepresentauxiliarystructures.Thismodifi- downtownAarhus.
cationenabledustoextractmoremeaningfullatentfactors The above result only reflects the count data, but the
thantheconventionalmethodswithoutconsideringspatio- method is extendable to multidimensional data. Figure 8
temporalstructures.Theformalizationdetailsaregivenin illustrates the conceptual idea for multidimensionalcases.
Appendix.Moreover,wedevelopedacomputationallyeffi- Variouskindsofinputdataaresimultaneouslyanalyzedto
cient learning algorithm that only needs to solve linear find the latent structure over different kinds of observed
equations[4]. data.Suchanalysisisimportantinthesensethatitcandis-
Figure 6(a) shows the total vehicle counts for 2 weeks, coverlatentfactorsthatcouldnothavebeenobtainedfrom
and Fig. 6(b) shows them at each site of the automobile individualdata.
https://doi.org/10.1017/ATSIP.2018.4 Published online by Cambridge University Press
6 naonori ueda and futoshi naya
Fig.6. CityPulsedataset.
Fig.7. ExtractedlatentfactorsforCityPulsedataset.
B) People-flowestimationfrom
population figures in 500m grid squares in Japan’s urban
spatio-temporalpopulationdata
areas, calculated from mobile network operational data.
DuetotheprevalenceofmobilephonesandGPSdevices, Toeffectivelyutilizepopulationdata,weestimatethepeo-
spatio-temporalpopulationdatacannowbeobtainedeas- ple flow from the data. Of course, if we could obtain the
ily.Forexample,mobilespatialstatistics[17]containhourly tracking data of individuals, then we would not need to
https://doi.org/10.1017/ATSIP.2018.4 Published online by Cambridge University Press
spatio-temporal multidimensional collective data analysis 7
Fig.8. Latentstructureextractionbymultidimensionalcomplexdataanalysis.
Fig.9. Spatio-temporalpopulationdata:(a)TokyoonJuly1,2013and(b)OsakaonAugust8,2013.Darkercolorsrepresenthigherpopulationdensitiesineachgrid
cell.
estimate the people flow. However, such privacy data are of aggregated information about many individuals, are
often hard to obtain. Therefore, estimating people flow aggregatedtopreserveprivacyorbecauseofthedifficultyof
onlyfrompopulationdataischallenging.Estimatedpeople trackingindividualsovertime.Forinstance,mobilespatial
flowscanbeusedforawidevarietyofapplications,which statistics [17] are preprocessed for privacy protection, and
include simulating people movements during a disaster, no one can follow a particular user, which allows mobile
detectinganomalouspeoplemovements,predictingfuture phoneoperatingcompaniestopublishpopulationdatathat
spatialpopulationsgiventhecurrentspatialpopulation,and are calculated based on information of about 60 million
designingtransportationsystems. mobile phone users. If the trajectories for each individual
Thespatio-temporalpopulationdataweuseasinputare are given, people flow can be estimated straightforwardly
the populations in each grid cell over the time shown in by counting the number of people who move between
Fig.9.Thespatio-temporalpopulationdata,whichconsist gridcells,thatis,thetransitionpopulation.However,with
https://doi.org/10.1017/ATSIP.2018.4 Published online by Cambridge University Press
8 naonori ueda and futoshi naya
Fig.10. Estimatedpeopleflowsforeachtime-of-dayclusterforTokyoandOsakadata.Arrowsdenoteadirection.
aggregated data, we can directly determine the size of the cellitself.Figure10showspeopleflowsestimatedbythepro-
transition population. This is why we call this “collective posedmethod.WiththeTokyodata,frommidnightto6:00,
dataanalysis”. theflowsaresparse,whichisreasonablesincemostpeople
To overcome this problem of modeling individual aresleeping.From8:30to11:00,peoplemovetothecenterof
behavior given aggregated data, we proposed a mixture Tokyofromthesuburbsforwork.Flowscontinuetothecity
of collective graphical models [18]. Our proposed model centerfrom14:30to15:00.From18:30to23:00,flowsfrom
assumes that individuals move by transition probabilities thecitycentertothesuburbswereestimatedsinceatthat
that depend on their locations and time points. Since the timepeoplearereturninghomefromtheiroffices.Withthe
transitionpopulationsarenotgiven,wetreatthemashid- Osakadata,similartemporalflowpatternswereextracted.
denvariables.Thehiddentransitionpopulationsarerelated Wealsoevaluatedthepredictionoferrorsforpeopleflows
totheobservedpopulationsineachcell;thepopulationina thatwasaveragedoverallthetimepointsandfoundthatour
cellequalsthesumoftransitionpopulationsfromit,andthe proposed method outperformed the conventional related
populationinacellinthenexttimepointequalsthesumof methods[18].
thetransitionpopulationstoit.Byusingtheserelationsthat Although our results are encouraging, our framework
representflowconservationasconstraints,thehiddentran- can be improved in a number of ways. For example, we
sitionpopulationsandtransitionprobabilitiesareinferred should incorporate the spatial correlation of flows using
simultaneously. location-dependent mixture models. Moreover, we want
The proposed model can handle people-flow changes to extend the proposed model to include information
over time by segmenting time-of-day points into multiple about the day of the week and study its effectiveness
clusters, where different clusters have different transition withdifferentgridsizes,timeintervals,andneighborhood
probabilities. For our task that models people flow, incor- settings.
porating time information is crucial. For example, people
move from the suburbs to the city center for work in the
morning,returntothesuburbsintheeveningafterwork,
IV. REAL-TIME AND PROACTIVE
andrarelytravelinthemiddleofthenight.Theproposed
NAVIGATION BASED ON SPATIO-
modelisanextensionofcollectivegraphicalmodels[19,20]
TEMPORAL PREDICTIONS
for handling changes over time. Since it does not use tra-
jectoryinformation,whenwehaveenoughpopulationdata
A) Spatio-temporalpredictions
thatapproximatetheactualpopulationdistribution,wecan
estimatethepeopleflow. To achieve the proactive navigation mentioned in Section
We evaluated the proposed method using real-world I, our approach detects future congestion with a spatio-
spatio-temporalpopulationdatasetsobtainedinTokyoand temporal statistical method that predicts people flow. The
Osaka(Fig.9).Inallofthedatasets,thegridcellsaresquare, peopleflow(orvelocity)canbeusedtomeasurecongestion
and the neighbors are the surrounding eight cells and the by extrapolating velocities to determine where congestion
https://doi.org/10.1017/ATSIP.2018.4 Published online by Cambridge University Press
spatio-temporal multidimensional collective data analysis 9
Fig.11. Processingflowforfutureprediction.
willoccurinthenearfuture.Weassumethatasetoftrajec- above spatio-temporal kriging approach to obtain dense
toriesisobservedinthecurrenttimeslotsbyapositioning statistics (Fig. 11(b)). Then we estimated a kernel regres-
sensorsuchasGPS.Thepositioningtrajectorydataaresim- sionfunctionoverthenear-pasttimedomainbyfittingaset
pletimeseriesofcoordinatevalueswithwhichweconstruct ofkernelfunctionstoeachofthenear-pastdensestatistics
spatio-temporal data. The observed spatio-temporal data (Fig.11(c)).Sinceweusedaradialbasisfunctionasakernel
areusuallyverysparse,andtheratioofthenumberofobser- function, the obtained regression function is a mixture of
vations to the time stamps over a certain time interval is radialbasisfunctions.
minute for each region. Therefore, we need to interpolate Oncewehavethetime-consecutivekernelfunctions,we
velocitiesthatarenotobservedovertime. estimatethenear-futurekernelparametersfromthenear-
More specifically, for each region we interpolate and pastkernelparametersbylinearregression.Wefoundthat
extrapolateeight-directionalvelocitiesduringacertaintime kernel parameter prediction is more robust and accurate
interval.Interpolationmeansanestimationoftheunknown than the direct prediction of statistics because the kernel
pastvelocities,whileextrapolationmeansthepredictionof function is smoother than the statistics themselves. Since
near-futurevelocities.Forthispurpose,wefirst employed thekernelfunctioniscontinuous,wecanalsoestimatethe
aspatio-temporalstatisticalapproachbasedonthekriging statisticsforanyplaces.
approach,whichiscommonlyutilizedinfieldslikegeology Since this approach does not assume a second-order
and climatology [21] and is equivalent to Gaussian pro- stationarycondition,itcanprovidemoreaccurateextrapo-
cess prediction in machine learning. As for extrapolation, lationresults.Weappliedourmethodtothepopulationdata
wealsocomparethekrigingapproachwiththerepresenta- collected in an actual event venue (about 110m×360m),
tivetime-seriesanalysis,whichisthevectorauto-regression whichmorethan20000peoplevisitperday.Thepopula-
(VAR)model[10].Inourexperiment,thekrigingapproach tiondatawerecollectedonceaminuteby24Wi-Fi1 access
slightlyoutperformedVAR[22]. pointsinstalledinthevenue.Thepredictionperformance
However, we found that these methods are inadequate comparison between an existing method (Gaussian pro-
for cases where people flows rapidly change because the cess)andtheproposedmethodisshowninFig.12.Thepro-
observed values are assumed to be second-order station- posedmethodobtainedbetterpredictionresults,especially
aryrandomvariablesintheconventionalmethods.Thatis, forthe20and30minpredictions.
the mean and the variance of the observed values should
be constant, which is unreasonable for actual people-flow B) Learningmulti-agentsimulation
data. To solve this problem, we developed a new method
basedonamachinelearningapproach.Asetofkernelfunc- Weexaminedtheeffectivenessofourspatio-temporalpre-
tions is trained using observed people-flow data for each dictionmethodusingactualpeople-flowdata.Byexploiting
timestepforthespatialinterpolation,andthenthetrained ourmethod,wecanpredictthenear-futurecongestionrisk
time-serieskernelparametersareusedforextrapolation. justfromnear-past,people-flowobservations.Thenextstep
Figure 11 shows the processing flow for our spatio- isdesigninganoptimalnavigationplanforagroupofpeo-
temporal predictions. Figure 11(a) shows an example of pletoavoidsuchcongestionrisksinrealtime.Totacklethis
observedstatisticssuchaspeoplecounts.Inpractice,since problem, we have been developing a learning multi-agent
observationlocationsarelimitedandoftensparseinspatio-
temporal domains, we performed interpolation using the 1Wi-FiisaregisteredtrademarkofWi-FiAlliance.
https://doi.org/10.1017/ATSIP.2018.4 Published online by Cambridge University Press
10 naonori ueda and futoshi naya
Fig.12. Predictionperformancecomparisonandheatmapofpeople-flowprediction.
simulation(MAS).Startingwithreal-spaceinformationon commonlyusedmodelconsiderstheaveragewalkingspeed
people or vehicles collected from real-time observations, (e.g.4km/h)relatedtosuchattributesofindividualspecta-
we input this information online into a simulation envi- torsasageandgenderaswellastheattenuationofthatwalk-
ronmentincyberspace,modelitsspatio-temporalfeatures, ingspeedinproportiontothecongestionconditions(crowd
instantaneouslypredicttheimmediatecongestionrisk,and density).Crowdcontrolmanuals[23]statethatovertaking
pre-emptively and optimally navigate the crowd to avoid becomesdifficultandwalkingspeedbeginstodecreaseata
thatrisk. crowddensityof1.2persons/m2andthatmovementgrinds
MASisawidelyusedtechniqueformodelingtheindi- to a halt at 4persons/m2. Crowd density is also called the
vidual behavior of such autonomous entities (=agents) as “servicelevel”[24],andtheroadwidth,space,andflowrate
people,cars,andanimalsbymodelingtheirmicrointerac- canbesecuredinpedestrianlanesthatmaintainacertain
tionswiththesurroundingenvironmentandanalyzingand servicelevelforsatisfyingsafetystandards.
predictingmacrophenomenathatdevelopfromtheinter- InconventionalanalysismethodsusingMAS,itiscom-
actionsamongmultipleagentsandtheirinteractionswith montomanuallydesignsuchparametersaswalkingspeed
theirenvironment.AnalyticaltechniquesusingMAShave andnavigationplans(includingpedestrianroutesandflow
been widely investigated in fields called complex systems, rate)byconductingsimulationstoevaluatetheireffectsin
where such overall macro behaviors as social activities, advance. Such methods have been applied when actually
brainactivities,andweatherphenomenacannotbebroken implementing crowd control (Fig. 13(a)). However, these
downintothebehavioroftheindividualelementsofpeople parameters and navigation plans are limited to a few pre-
withwhichsocietyiscomprised,neuronsthatconstitutethe determinedcombinationsanddonotnecessarilymatchthe
neuralnetworksinthebrain,andmoleculesandatomsin observation results of actual person movement or naviga-
the atmosphere. These techniques are recently being used tionoperations.
in sensor networks, smart grids, and intelligent transport Today,however,whereIoTandsensingtechnologiesare
systems as well as in evacuation guidance simulations for rapidlyadvancing,localpeople-flowandcongestioncondi-
disastercountermeasures. tionsintherealworldcanbemeasuredinrealtimeusing
As an example, consider a navigation scene of tens of a variety of positioning means, such as surveillance cam-
thousandsofspectatorsexitingastadium.Inthiscase,the eras, GPSs, Wi-Fi, and beacons. With this in mind, NTT
individual spectators are agents who are leaving the sta- laboratories continue to develop a learning-MAS system
diumfromtheircurrentlocationsandmovingtowardtheir thattransferseventsintherealworld,suchaspeopleflow
trainstationdestinations.Inthemovementofspectators,a observedinrealtimetoasimulationenvironmentincyber
https://doi.org/10.1017/ATSIP.2018.4 Published online by Cambridge University Press