How$AI$is$helping
Smart$Cities
SK#Reddy
Chief#Product#Officer#AI
skreddy99
skreddy99
Confidential2
11"variables,"including"
whether"the"restaurant"had"
previous"violations,"how"long"
it"has"been"in"business"(the"
longer,"the"better),"the"
weather"(violations"are"more"
likely"when"it’s"hot),"even"
stats"about"nearby"burglaries"
(which"tells"something"about"
the"neighborhood,"though"
analysts"aren’t"sure"what)
How$to$prioritize$restaurant$inspections
https://www.wsj.com/articles/the7rise7of7the7smart7city71492395120
Confidential3
https://blogs.worldbank.org/opendata/artificial8intelligence8smart8cities8insights8ho8chi8minh8city8s8spatial8development
Land%Use%Classification%from%Satellites%in%Ho%Chi%Minh%City
Confidential4
Los$Angeles’s$Clean$Streets$program
https://www.wsj.com/articles/the7rise7of7the7smart7city71492395120
Confidential5
Listens'to'gun'shot'sounds…..
Confidential6
New$Orleans$mapped$the$combined$risk$of$missing$smoke$alarms$and$fire$deaths
https://www.wsj.com/articles/the7rise7of7the7smart7city71492395120
Confidential7
AR#experience
Confidential8
Smart&City&Use&Cases
• Parking1availability1
• How1long1a1bus1will1take?
• What1is1the right1length1for1a1stop1light at1every1given1minute
• Real&time&data&about&traffic&and&accidents
• Check1if1vehicles1parked1there1have1the1right1permits
• License1plate1recognition
• Detecting&human&action&using&videos
• Preventive1maintenance1of1the1road1infrastructure
• Buses1and1trains1communicate1with1each1other1and1the1general1public
• Face&recognition
• Find1stolen1cars
• Locate1gunfire1based1on1a1sensor1network
• Localized1warnings1about1possible1natural1disaster
• Count&vehicles and1pedestrians1
• Recognize&faces
• Track1the1speed1and1movements1of1vehicles1
• Count&cars&in&a&parking&lot&or1track1road1use
• Emergency1dispatch1
• Ambulance1call1
https://www.techemergence.com/smartIcityIartificialIintelligenceIapplicationsItrends/
Data
• Volume:1terabytes1(TB),1petabytes1
(PB),1zettabytes1(ZB)
• Variety:1types1of1data1(sensors,1
devices,1social1networks,1the1web,1
mobile1phones,1etc.)
• Velocity:1how1frequently1the1data1is1
generated1(every1millisecond,1
second,1minute,1hour,1day,1week,1
month,1or1year
Confidential9
Online&data*processing&architecture
https://arxiv.org/pdf/1709.01363.pdf
Main<components<of<a<Storm<cluster
Apache<Flink’s execution<model
Confidential10
Text%processing%Framework
https://arxiv.org/pdf/1709.03406.pdf
• Lowercasing
• Lemmatization
• Tokenization
• TransformingDrepeatedDcharacters:D(e.g.D"loooool"DtoD
"loool”)
• PunctuationDremoval
• CleaningDEntitiesDandDNumericalDSymbols
• RemovingDURLs,DuserDmentions,DhashtagsDandDdigitsDfromD
theDtextDmessages
• StopDandDshortDwordsDremoval
• GeographicDDistributionsD
• TemporalDFrequencies
• volumeDofDtweetsDpostedDperDhour,DperDday,DasDwellD
asDtheDactivityDbyDdayWofWtheWweekDorDhourWofWtheWdayD
• ContentDComposition
• metadataD(hashtags,DuserDmentions,DURLsDandD
mediaDattached)DexplorationDforDaddedWvalueD
information
Confidential11
Data-streaming-
Infrastructure+architecture
https://arxiv.org/pdf/1703.02755.pdf
Confidential12https://arxiv.org/pdf/1712.01432.pdf
Large&scale*video*management
Confidential13https://arxiv.org/pdf/1506.02640v5.pdf
Count&cars&in&a&parking&lot&
https://medium.com/the?downlinq/car?localization?and?counting?with?overhead?imagery?an?interactive?exploration?
YOLO2
Confidential14
Traffic'flow'detection
https://arxiv.org/pdf/1801.03317.pdf
Example=ray=tracing=scenario=illustrating=the=shadowing=
effects=caused=by=a=SUV=passing=the=proposed=
classification=system
Confidential15
(a).0The0spatial0component0uses0a0
local0CNN0to0capture0spatial0
dependency0among0nearby0
regions.0The0local0CNN0includes0
several0convolutional0layers.0A0fully0
connected0layer0is0used0at0the0end0
to0get0a0low0dimensional0
representation.0
(b).0The0temporal0view0employs0a0
LSTM0model,0which0takes0the0
representations0from0the0spatial0
view0and0concatenates0them0with0
context0features0at0corresponding0
times.0
(c).0The0semantic0view0first0
constructs0a0weighted0graph0of0
regions0(with0weights0representing0
functional0similarity)
Deep$Multi*View$Spatial*Temporal$Network$for$
Taxi$Demand$Prediction
https://arxiv.org/pdf/1802.08714.pdf
Confidential16
http://openaccess.thecvf.com/content_cvpr_2017/papers/Wang_Learning_to_Detect_CVPR_2017_paper.pdf
Learning(to(Detect(Salient(Objects(
with(Image7level(Supervision
Confidential17
http://openaccess.thecvf.com/content_cvpr_2017/papers/Chen_Beyond_Triplet_Loss_CVPR_2017_paper.pdf
Deep$quadruplet$network$for$
Person$re3identification
Confidential18
https://arxiv.org/pdf/1704.08063.pdf
Comparison*of*open-set*and*closed-set*face*recognition
SphereFace:<
Deep<Hypersphere<Embedding<for<Face<Recognition
Confidential19
https://arxiv.org/pdf/1704.00389.pdf
Analyzing)videos)of)
human)actions
TOP:?MOTIONNET.?BOTTOM:?TRADITIONAL?TEMPORAL?STREAM.?M?IS?THE?
NUMBER?OF?ACTION?CATEGORIES.?STR:?STRIDE.?CH?I/O:?NUMBER?OF?
CHANNELS?OF?INPUT/OUTPUT?FEATURES.?IN/OUT?RES:?INPUT/OUTPUT?
RESOLUTION
Confidential20https://arxiv.org/pdf/1706.07911.pdf
Large&Scale*Mapping*using*Geo&Tagged*Videos
Training<MotionNet to<learn<optimal<optical<flow<involves<minimizing<the<following<three<objective<
functions:
1. A<standard<pixelGwise<reconstruction<error<
function
2. A<smoothness<loss<to<address<the<ambiguity<of<
estimating<motion<in<nonGtextured<regions<(the<
aperture<problem)
3. A<structural<similarity<(SSIM)<loss<that<helps<
MotionNet learn<the<structures<of<frames.<It<is<
calculated<as
The<overall<loss<is<a<weighted<sum<of<the<pixelGwise<
reconstruction<loss,<the<pixelGwise<smoothness<loss<and<
the<regionGbased<SSIM<loss
Spatial*analysis*of*the*46th*San*Francisco*Pride*parade*in*2016.*
Left:*official*parade*route.*Right:*m ap*of*classified*videos.*Note*the*
correlation
Confidential21
Thank&you
SK&Reddy
Chief&Product&Officer&AI
skreddy99
skreddy99

Practical implementation of AI solutions for Smart Cities