Predicting Fine-Grained Traffic Conditions via Spatio...

Research ArticlePredicting Fine-Grained Traffic Conditions viaSpatio-Temporal LSTM

Xiaojuan Wei Jinglin Li Quan Yuan Kaihui Chen Ao Zhou and Fangchun Yang

State Key Laboratory of Networking and Switching Technology Beijing University of Posts and Telecommunications Beijing China

Correspondence should be addressed to Jinglin Li jllibupteducn

Received 27 September 2018 Revised 6 December 2018 Accepted 24 December 2018 Published 14 January 2019

Guest Editor Qingchen Zhang

Copyright copy 2019 Xiaojuan Wei et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Predicting traffic conditions for road segments is the prelude ofworking on intelligent transportationMany existingmethods can beused for short-term or long-term traffic prediction but they focusmore on regions than on road segmentsThe lack of fine-grainedtraffic predicting approach hinders the development of ITS Therefore MapLSTM a spatio-temporal long short-term memorynetwork preluded bymap-matching is proposed in this paper to predict fine-grained traffic conditions MapLSTM first obtains thehistorical and real-time traffic conditions of road segments via map-matchingThen LSTM is used to predict the conditions of thecorresponding road segments in the future Breaking the single-index forecasting MapLSTM can predict the vehicle speed trafficvolume and the travel time in different directions of road segments simultaneously Experiments confirmed MapLSTM can notonly achieve prediction for road segments based a large scale of GPS trajectories effectively but also have higher predicting accuracythan GPR and ConvLSTM Moreover we demonstrate that MapLSTM can serve various applications in a lightweight way such ascognizing driving preferences learning navigation and inferring traffic emissions

1 Introduction

Traffic prediction of road segments is a fundamental issuein the Intelligent Transportation Systems (ITS) which canbe hopefully used for planning optimal driving routes [1]urban computing [2] balancing traffic control [3 4] andenhancing driving comfort [5] It is necessary to explorethe traffic dynamics and analyze the evolution pattern oftraffic flow Due to the generation of industrial IoT bigdata network infrastructures and computational modelshave been equipped and applied [6ndash8] If the global trafficinformation is not recognized accurately and timely ITS willbe not successfully deployed or the deployed system will beparalyzed sooner or later

In general the power of effectively predicting the futuretraffic conditions for road segments comes from the historicaland real-time traffic information According to the durationfor the future 3-10 days 1-3 days within 1 day and no morethan 15 minutes traffic flow forecast usually is included long-term recent-term short-term and short-time [9] Most ofthe existing methods present prediction trend either by usingprobability and statistics of the time-dependent evolution

of current road or only using the pure spatial relationshipsamong various road segments Although available spatiotem-poral information is combined to model the traffic networkpattern the information does not play out its full potential

Traffic network possesses complicated spatio-temporalrelationship The prediction methods should have accu-racy robustness adaptability and portability as the trafficflow is a high-dynamic high-dimensional non-linear andnon-stationary random process Traffic conditions of roadsegments are influenced inevitably by the spatio-temporalinformation in the traffic network Deep learning can be usedtomodel high-level abstractions by using multiple non-lineartransformations while the learning network has rarely takenthe overall spatio-temporal dynamic pattern into accountIt is not convincing to achieve accurate traffic predictionmerely by spatial relations between regions or road segmentsHence the prediction results perform not well at certaintimes which occur especially when there are insufficientGPS trajectories through road segments Based on this it isproper to consider more supplementary aspects such as map-matching technology used to recognize traffic conditions forroad segments accurately and finely

HindawiWireless Communications and Mobile ComputingVolume 2019 Article ID 9242598 12 pageshttpsdoiorg10115520199242598

2 Wireless Communications and Mobile Computing

In this paper we propose a fine-grained and lightweightapproach for traffic predicting of road segments namedMapLSTM a spatio-temporal long short-term memorynetwork (LSTM [10]) preluded by map-matching [11]MapLSTMonly requires vehiclesGPS and not need to deployspecialized traffic sensors in urban and not use the unob-tainable data from ground loop MapLSTM first obtains thehistorical and real-time traffic conditions of road segmentsviamap-matchingThenLSTM is utilized to predict the trafficconditions of the corresponding road segments in the futureBreaking the single-index forecasting MapLSTM can predictmultiple traffic conditions for road segments simultaneouslyTo summarize the major contributions of this paper consistof the following aspects

(1) Breaking through the difficulty of obtaining segment-based traffic data we perform the cognizing of road-grainedtraffic conditions via map-matching technology

(2) Based on a large scale of taxi GPS trajectorieswe propose MapLSTM to extract features from the high-dynamic high-dimensional non-linear and non-stationarytraffic flow And we confirm that MapLSTM have a higherpredicting accuracy than GPR [1] and ConvLSTM [12]

(3)We demonstrate MapLSTM can serve to various prag-matic applications cognizing driving preferences learningnavigation and inferring traffic emissions

The remainder of this paper is organized as followsSection 2 reviews the literature on traffic prediction Section 3describes the materials and gives details of our mechanismMapLSTM The next Section 4 demonstrates the effective-ness and applications This paper ends in Section 5 withconclusion on our work

2 Literature Review

Traffic condition prediction cannot only be used as the designbasis of signal control of ITS but also provide decision supportfor dynamic route guidance Whereas there are still somebottlenecks in short or long term traffic prediction througha lot of real spatio-temporal data

Spatio-temporal semi-supervised learning model pro-posed in [13] can infer the volume of each road with real-world data collected from 155 loop detectors and 6918 taxisover 17 days There are totally 19165 road segments in theurban area but only 155 road segments are equipped withloop detectors which results there in an inherent deviationin acquiring the city-wide traffic volumes Although theconstructed affinity graph can characterize the similaritiesamong roads based similar speed patterns the factors influ-encing traffic flow are not only the speed of vehicles but alsothe road topology road structure the regional characteristicsand so on Traffic conditions of each road segment cannotbe predicted accurately only by the spatial relations on themacro

A vehicle speed is influenced by many factors the vehicletype the traffic conditions and the driverrsquos behaviour A datadriven model is proposed in [14] for vehicle speed predictionwhere the average traffic speed is estimated based on histor-ical traffic data at first and then the statistical relationshipwith individual vehicle speed is presented by hidden markov

models Finally the individual vehicle speed is predicted byforward-backward algorithm Another mechanism proposedin [15] is a cooperative method which combines with fuzzymarkov model and auto-regressive model These machinelearning approaches for vehicle speed prediction do not careabout the basic data sources and focus more on the accuracyof prediction algorithms rather than the accuracy of segment-based traffic prediction

DeepSense [16] is a typical deep learning approach fortraffic prediction with Taxi GPS traces DeepSense gainsthe prediction results based on sufficient dataset by usingRestricted Boltzmann Machine Due to the night data istoo sparse so DeepSense made a prediction based fillingaccording to the data of the same time in history But thisprediction-based prediction approach may lose credibility indeep learning In addition DeepSense extract and classify thespeed only on 0 sim 60119896119898ℎ to reflect the traffic congestion orsmooth which lacks universality in some other regions

Understanding traffic density from large-scale images isanother way to recognize the traffic status Reference [17]as a related work selects a region of interest in a videostream at first then counts the number of vehicles in theregion for each frame so the density is calculated by dividingthat number by the region length Reference [18] is anotherimage-based learning to measure traffic density using a deepconvolutional neural network These vision-based cognitivemethods mainly play a role in local regions which candedicate to the operational control but cannot make anefficient decision in tactical planning with in the long run

In addition the predicted object is univocal in the existingmethods more is traffic volume or speed which can merelyinfer the traffic state of the road segment is congestion slownormal moderate and unimpeded It is necessary to explorefine-grained and accurate perception in a simple way

3 Materials and MapLSTM

In this section we first provide materials on GPS trajec-tory map-matching and LSTM Then we depict MapLSTMdesigned for traffic prediction

31 Materials

311 GPS Trajectory Taxis can be considered as ubiquitousmobile sensors constantly probing a cityrsquos rhythm and pulseBeing inherent characteristic GPS-based taxies have provento be an extremely useful data source for uncovering theunderlying traffic behaviour So far the taxi GPS data havebeen used for urban computing detecting hot spots mapreconstruction finding routes and so on [2 19]

The GPS records of a large number of taxis in a cityare routinely saved to a log file 119871 resulting in a very largedata set 119871 = 11990111 11990112 1199011119899 1199011198941 1199011198942 119901119894119899 Fieldsfor each GPS record generally contains TaxiID Location(Longitude Latitude) Speed Event GPS state Bearing and6 commonly used timing units YYYY-MM-DDHHMMSSFigure 1 shows an example of GPS log and trajectories Atrajectory 119879 is a time series of GPS points with the timeinterval between any consecutive GPS points not exceeding

Wireless Communications and Mobile Computing 3

middotmiddotmiddotmiddotmiddotmiddot

middotmiddotmiddot

P21

P22

P23

P24

P25 P26

P12

P11

P13

P14

P15

P17

Pi1

Pi2

Pi3

Pi4 Pi5

P16

T1

T2

Ti

Figure 1 An example of GPS log and GPS trajectories

Map-Matching

timetjtit1 t2t0

titjtit1 t2tt0

before after

Figure 2 Map-matching when before and after

a certain threshold 119905 (usually 119905 ge 1 min) ie 119879119894 1199011198941 997888rarr1199011198942 997888rarr 1199011198943 997888rarr sdot sdot sdot 997888rarr 119901119894119899

312 Map-Matching Map-matching is the process of align-ing a sequence of observed GPS positions with the roadnetwork on a digital map [20] As a preprocessing stepof MapLSTM map-matching can effectively improve theexisting huge amount of low-sampling-rate GPS trajectoriesin data set

As shown in Figure 2 map-matching can be performedwith the same or different time interval as theGPS pointsTheGPS points without map-matching can only be mapped tothe road network Not all GPS points can be mapped to theircorresponding segments due to the GPS positioning errorBut after map-matching all GPS points can be corrected tothe corresponding road segments

Input The networkrsquos input in current time 119909119905the initial weight matrix119882and bias units 119887 about gates 119868 119874 119865

Output The forget gate input gate cell state in differenttime output gate and the cell output

(1) 119891119905 = 120590(119882119891 sdot [ℎ119905minus1 119909119905] + 119887119891)(2) 120590 represent Sigmoid function(3) 119894119905 = 120590(119882119894 sdot [ℎ119905minus1 119909119905] + 119887119894)(4) 119862119905 = tanh(119882119862 sdot [ℎ119905minus1 119909119905] + 119887119862)(5) 119862119905 = 119891119905 lowast 119862119905minus1 + 119894119905 lowast 119862119905(6) 119900119905 = 120590(119882119900 sdot [ℎ119905minus1 119909119905] + 119887119900)(7) ℎ119905 = 119900119905 lowast tanh(119862119905)(8) return 119891119905 119894119905 119862119905 119862119905 119900119905 ℎ119905

Algorithm 1 For calculating each element of LSTM

313 LSTM LSTM [10] is a time recurrent neural networkwhich is the most widely used method to process and predictevents with relatively long intervals in time series LSTM canlearn about long-term reliant information by input gate 119868output gate 119874 and forget gate 119865 where 119868 determines howmuch of the network input at the current time 119909119905 is saved tothe cell state 119888119905 119874 determines how much of the control unitstate 119888119905 is output to the current output value ℎ119905 of LSTM 119865determines howmuch of the cell state from the previous time119888119905minus1 remains to the current time 119888119905 In short the input 119883 atdifferent time determines the cell state119862 at the correspondingtime and the current cell state 119888119905 will be affected by theprevious cell 119888119905minus1

The calculation of each element of LSTM is shown inAlgorithm 1 At the current time 119905 119891119905 denotes forget gate 119894119905represents input gate obtained by the previous output ℎ119905minus1and the current input 119909119905 119862119905 denotes the cell state and 119862119905denotes the cell state at the previous time 119900119905 denotes theoutput gate and ℎ119905 denotes the cell output LSTM can notonly save information long ago under the control of 119865 butalso avoid the current irrelevant content into memory basedthe gate 119868

32 MapLSTM MapLSTM is fine-grained and lightweightway It only requires sampled GPS points of vehicles and notneed to deploy expensive traffic sensors in urban and not usethe unobtainable data from ground loop In this section wedescribe MapLSTM in detail


FC

Y

LSTM LSTM LSTM

(a) Map-Matching (b) Data processing

(c) LSTM predictingTime

Inputs

HiddenLayer

HiddenLayer

Outputs

1 2 3 4 5

-

-

- - -

Vehicle speed

Traverse time

Traffic volume

hellip

hellip

hellip hellip

1=Speed 2=Time 3=Volume



MapLSTM works better with

environment of collaboration

computing

1

2 3

4

56

7

8

9

10

11

12

13 1

4

Figure 3 MapLSTM framework for traffic prediction It consists of three processes Map-matching data processing and LSTM predicting

321 Framework Figure 3 shows the framework ofMapLSTMwhich consists of three processes map-matchingdata processing and LSTM predicting

(a) Map-Matching A large number of sampled GPS pointsstored in GPS log need to be matched to road segments Inorder to facilitate the operation it is necessary to manuallyredivide road segments based on the road network beforematching Generally the division is based on the intersec-tion or no redivision just based on the inherent segmentsstructure in road network if the calculation resources androad segments information are sufficient and detailed Wedo our best to maintain the original topography relationshipbetween the divided road segments After map-matchingall GPS points can be shifted to the corresponding roadsegments

(b) Data Processing The road segments experienced map-matching also mean the information has been extendedwhere the road segments and vehicle information are pairedoff according to their ID and location Therefore we can haveinformation statistics including vehicle speed traverse timeand traffic volume taking one road segment as a unit (iesegment-based)The traverse time can be counted in differentdirections fromwest to east from east to west from north tosouth and from south to north The processed data are sentto prediction model LSTM as the training set and testing set

(c) LSTM Predicting The traffic data of vehicle speed the tra-verse time and traffic volume based road segments are inputto LSTM concurrently for predicting task The hidden layersof LSTM can control the long-term or short-term impact onthe current state After output layer of LSTM it goes througha full connected network with three layers in which thepurpose is to better explore the implied relationships betweenstates

MapLSTM enables cognition of road segment-basedtraffic conditions in a lightweight way For the real-timecognition of global situations MapLSTM is still valid bycollaboration computing where a groups of cells worktogether to accomplish a relatively large task Edge computingafter cloud computing is a typical collaborative computingenvironment and has been widely used [21 22]

322 Map-Matching Algorithm Before map-matching it isnecessary to have a information understanding about roadsand vehicles Table 1 describes an example with a sampleof the information All the information about roads andvehicles can be correlated based the auxiliary information(ID longitude and latitude)

ST-Matching [20] is a pathway with candidate computa-tion and spatio-temporal analysis for low-sampling-rate GPStrajectories We follow ST-Matching analysis architectureand make a map-matching work on a real digital mapin Beijing As described in Algorithm 2 for the available


Table 1 An example with a sample of the main information about road and vehicle

Name eMain Fields

Road ID MapID PathName Pathclass Oneway Width Length Direction Meters59565200918 595652 Xing Fu Xi Jie 4 F 30 0284 2 368

Vehicle ID Bearing Speed State Longitude Latitude Event Time Positioning6409 84 46 1 3973633 11633100 1 20160916182046 GPSBeiDouMix

Input Beijing Road network 119877 Coordinate axis 119860Trajectories 119879 where 119879 = 1199051 1199052 1199053 119905119899119905119894 = 1199011198941 997888rarr 1199011198942 997888rarr 1199011198943 997888rarr sdot sdot sdot 997888rarr 119901119894119899119901119894119895 is a GPS sampling point 119894 119895 isin [1 119899]

Output The one-to-one results of road segments andvehicles information1198721198791015840

(1) Initialize 119862119875119904119890119905 = 0(2) Repeat 119894 = 1 2 3 119899(3) For 119895 = 1 to 119899(4) 119878119894119895 = 119866119890119905119862119886119899119889119894119889119886119905119890119875119900119894119899119905119904(119901119895 119877)(5) 119862119875119904119890119905119886119889119889(119878119894119895)(6) End for(7) 119896 = 1(8) While 119896 le 119862119875119904119890119905 do(9) 119881119904 = 119866119890119905119878119901119886119881119886119897(119877 119860 119862119875119904119890119905(119896) 119889119894119904119905(119901 119862119875119904119890119905119901))(10) 119881119905 = 119866119890119905119879119890119898119881119886119897(119877119860 119878119901119890119890119889 119905119894119898119890(119901119894119895 119901119894119895+1))(11) 119872119879 = 119872119886119905119888ℎ119878119890119902(119881119904 119881119905)(12) 119896+ = 1(13) End while(14) Visualized119872119879[1198721198791015840(15) return 1198721198791015840

Algorithm 2 Map-matching algorithm

historical trajectories GPS sampling points in the trajectoriesare traversed to get the candidate point set which waitingto be corrected For all candidate points the spatial valuecan be reached by combining with the information of roadnetwork longitude latitude and distance and the temporalvalue can be reached by adding the time information Afterspatial analysis and temporal analysis matching results canbe accomplished

After map-matching roads information where the vehi-cles are located can be easily obtained and the traffic dataabout the roads can also be clearly gained after statistics inturn

323 Training Data Generating The raw trajectory datacannot be used directly for our predicting task It is necessaryto match and statistics at first If we want to get the trafficstatus prediction of road segments we need to make asegment-based statistics about the traverse time in differentdirections the vehicle speed and the traffic volume

The data of traverse time in different directions theaverage vehicle speed and traffic volume of road segmentscan be generated by Algorithm 3 When map-matching isdone more fine-grained data can also be obtained such as theaverage speed and traffic volume under different directionsof road segments The data after map-matching and statistics

Input Road Log 119877119897 GPS Log 119881119897OutputThe traffic data about the road segments(1) Initialize 119879119879 119908119890 119879119879 119890119908 119879119879 119904119899 119879119879 119899119904 119878119901119890119890119889

119862119900119906119899119905 119908119890 119862119900119906119899119905 119890119908 119862119900119906119899119905 119899119904119862119900119906119899119905 119904119899(2) GetVehicleColumns(ID Time V Speed V lon V lat)(3) GetRoadColumns(RoadID R lon R lat)(4) Repeat 119877119900119886119889119868119863(5) 119865119897119886119892 = 0(6) For 119894 = 119878119905119886119903119905119879119894119898119890 to 119864119899119889119879119894119898119890(7) if(8) 119905119894 119881 119897119900119899 lt 119877119908 119897119900119899 119881 119897119886119905 lt 119877119899 119897119886119905(9) 119905119894+1 119881 119897119900119899 ge 119877119908 119897119900119899 119881 119897119886119905 ge 119877119899 119897119886119905(10) 119905119899+119894 119881 119897119900119899 lt 119877119890 119897119900119899 119881 119897119886119905 lt 119877119904 119897119886119905(11) 119905119899+119894+1 119881 119897119900119899 ge 119877119890 119897119900119899 119881 119897119886119905 ge 119877119904 119897119886119905(12) then(13) 119865119897119886119892 = 1(14) 119862119900119906119899119905 119908119890+ = 1(15) 119862119900119906119899119905 119899119904+ = 1(16) 119879119879 119908119890+ = 119905119899+119894+1 minus 119905119894+1(17) 119879119879 119899119904+ = 119905119899+119894+1 minus 119905119894+1(18) Imitate GetCountData(119862119900119906119899119905 119890119908 119862119900119906119899119905 119904119899)(19) Imitate GetTTData(119879119879 119890119908 119879119879 119904119899)(20) if 119865119897119886119892 = 1 then 119878119901119890119890119889+ = 119881 119878119901119890119890119889(21) End for(22) 119860V119892 119879119879 119908119890 = 119879119879 119908119890119862119900119906119899119905 119908119890(23) Imitate GetData(119860V119892 119879119879 119890119908 119860V119892 119879119879 119904119899

119860V119892 119879119879 119899119904)(24) 119881119900119897119906119898119890 = sum119908119890119899119904119901119902 119862119900119906119899119905 119901119902(25) 119860V119892119878119901119890119890119889 = 119878119901119890119890119889119881119900119897119906119898119890(26) return 119860V119892 119879119879 119908119890 119860V119892 119879119879 119890119908 119860V119892 119879119879 119904119899

119860V119892 119879119879 119899119904 119860V119892119878119901119890119890119889 119881119900119897119906119898119890

Algorithm 3 To generated the segment-based traffic data

can be used which also mean that the training data and thetesting data of prediction network are generated

4 Experiment

We compare the following experiments to verify the perfor-mance of MapLSTM

(1) Gaussian Process Regression (GPR) [1] It is one of themost popular used prediction algorithms and often used tocompare performance as a baseline

(2) ConvLSTM [12] It extends LSTM to have convolutionalstructures in both the input-to-state and state-to-state transi-tions and captures spatiotemporal correlations better

(3) ConvLSTM+ It is ConvLSTM increased epoch numbers


TT_ns(s)TT_sn(s)TT_we(s)TT_ew(s)Speed(kmh)Count(EA)

40

30

20

10

0

0

100

200

300

400

500

20 30 40 50 60 70 80 90 100 110 120 130 14010

X = road segment

30 40 50 60 70 80 90 100 110 120 130 140

100

80

60

40

20

0

0

100

200

300

400

500

3875

489

118

300

465 4935

10 20

Figure 4 Traffic data about road segments at 800 on November 1 2012

41 Datasets A large scale of real taxi trajectory data are usedin our predicting task The data package of GPS log includesover 400000 taxicabsrsquo trajectories inNovember 2012 BeijingAnd full-scale entries are contained during 24 hours foreach day We use data between 800 sim 2000 in weekdaysas the traffic pattern can be learned better in the daytimeWe can get dataset 22 times 13 times 30 when the time interval is2 minutes

There are too many segments in road network 119877 so wemanually redivide the road segments based 119877 to verify thefeasibility of MapLSTM The road segments after redivisionis stored to set 119877119904119904 Figure 4 depicts the traffic data of 119877119904119904on November 1 2012 at 8 orsquoclock including the traverse timein different directions 119879119879119882larrrarr119864119879119879119873larrrarr119878 the average vehiclespeed and the traffic volume

42 Training In MapLSTM the obtained dataset is dividedinto training set and test set in an 8 2 ratio The predictionmodel has 119887119886119905119888ℎ 119904119894119911119890 = 20 119897119903 119889119890119888119886119910 = 093 ℎ119894119889119889119890119899 119904119894119911119890 =250 and 119899119906119898 119904119905119890119901119904 = 6 (the size of window that means usingdata from the previous 6 time units to predict the next one)The sizes of the three full connection layers are 180 times 150250times180 and 250times250 which is related to the total numberof road segments

ConvLSTM has the same dataset as MapLSTM and themodel has 119894119899119901119906119905 = 21 lowast 21 119887119886119905119888ℎ 119904119894119911119890 = 8 119899119906119898 119904119905119890119901119904 =6 119896119890119903119899119890119897 = 5 lowast 5 119891119894119897119905119890119903119904 = 10 and 119898119886119909 119890119901119900119888ℎ = 70ConvLSTM+ is iterated 20 times more than ConvLSTM

43 Performance Evaluation Mean absolute error (MAE) isthe most commonly used criteria in predictive algorithmsand is employed to evaluate the proposed MapLSTM

119872119860119864 = 1119873

119873

sum119894=1

1003816100381610038161003816(119891119894 minus 119910119894)1003816100381610038161003816 (1)

where 119891119894 is the predicted value and 119910119894 is the observed valueThe smaller the MAE the stronger the predictable ability ofalgorithms

As shown in Table 2 whether it is the MAE of vehiclespeed traffic count or travel time in different direction119879119879 119882119864 (from west to east) 119879119879 119864119882 (from east to west)119879119879 119878119873 (from south to north) and 119879119879 119873119878 (from north tosouth) MapLSTM is smaller than GPR and ConvLSTMFor a certain algorithm the closer the value of ldquoTrainrdquo andldquoTestrdquo of each parameter is the more robust it is The resultsof ConvLSTM are similar to MapLSTM but do not exceedMapLSTM That is because ConvLSTM with the ability tocapture spatiotemporal correlations is good at predictingrelatively single spatial pattern but the spatial patterns of roadtraffic are complex In the future we will focus on complexspatial correlations in traffic environment Compared toConvLSTM some parameters of ConvLSTM+ are slightlybetter because ConvLSTM+ increased the number of epoch

It is important to note that MAE is affected by theaccuracy of the raw data and it will decline if the dataset islarge enough

44 Applications

441 Cognizing Driving Preference Different drivers havedifferent preferences about different types of roads and theyalso have different impulse to reroute roads due to theirdifferent tolerance about the cost expectations of currentcongestion For example the drivers with low tolerance maychoose a highway bypass which have a lower congestioncost expectations but have more traffic lights Tolerance ofdrivers changes dynamically with various spatial-temporalconditions such as travel distance congestion time andarrival time Therefore a large deviation between the trafficoptimization results and the actual expectation of driverswill lead to failure of traffic scheduling Quite a few drivers


Table 2 MAEs comparison of GPR ConvLSTM and MapLSTM

Algorithms Speed Count TT WE TT EW TT SN TT NSGPR 7079 4643 6681 707 632 6608

ConvLSTM Train 1878 718 1875 1827 1777 1901Test 1927 732 1871 1811 1814 1929

ConvLSTM+ Train 1944 713 1875 1859 1782 1946Test 1891 689 1871 1796 1785 1893

MapLSTM Train 1833 559 1642 165 1694 1862Test 1853 705 1691 1721 17 1857

Time (s)Count (EA)Speed (kmh)

Distance (m)

20 40 60 80 100 120 1400ID

050

100150200250300350400

2500

2000

1500

1000

500

Figure 5 Segment-based traffic information at a certain time

choose a looked like shortest road only to find the route iscongested by many vehicles whose drivers make a similardecision

The traditional route planningmethods aremore inclinedto train driversrsquo basic selection tendency and do not havepersonalized features The participants in these methods areconsidered the rational contenders perfectly The plannedresult is the purely rational optimal solution and does notexpress the noncomplete rational decision-making prefer-ence for drivers in the actual routing decisions Although thequestionnaire may be a handy pathway for cognizing drivingpreferences it lacks efficiency and comprehensiveness

The premise of learning driving preferences is to obtainan understanding about the roads conditions The more weaware of road properties the more satisfied we cognise thepersonalized preferences MapLSTM can have a fine-grainedcognition of road traffic conditions so we can learn thedriving preferences easily For drivers of vehicles there aretwo preferences getting themost attention time and distance

Figure 5 shows the traffic information about the travel timedistance vehicle density and speed of each road segments in119877119904119904 where the vehicle is driving from place A to place B Inorder to compare the preference in driving the full drivingroutes based different driving preferences including averagespeed vehicle count distance and travel time are shown inFigure 6

442 Learning Navigation Navigating vehicles to their des-tination is an important service for ITS In addition tousing historical and real-time traffic conditions the state-of-the-art systems take into account the impact on the futuretraffic conditions which can be obtained by predicting Forexample the method in [23] has the ability of learningexperience-based autonomous navigation based the globaltraffic dynamic and the method in [1] is another dynamicplanning scheme based on situation awareness where the citysensors are deployed to maintain an up-to-date view of thecityrsquos current traffic state


149

2

1

34

56

7

89

11

1012

1314

1516

1718

19

2021

2223

24

25

26

27

28 29 30 31 32 33 34 35

36 3738 39 40 41 42 43 44

4546 47 48 49 50 51 52 53

54

55 56 57 58 59 60 61

62

63 64 65 66 67 68 69 70

71

72

73

74

75

76 77 7879

80 81 82

83 8485

8687 88

89 9091

92 93

9495

96 97

98

99 100101 102 103 104 105 106 107

108 109 110 111 112 113 114 115

116117 118 119 120 121 122 123 124

125126

127128

129130

131132

133134

135136 137 138

139140

141142

143

144

145

146147 148

A

B

Distance

Time

Speed

Count

104 8192 91 74 4866 57 47 46

104 101103 102 89 7480 73 56 47 46

104 8192 91 73 4672 5565

46104 8192 91 74 56 47

Figure 6 Routes with different driving preferences from A to B

Asmentioned above the existing methods are still labori-ous for lightweight fine-grained and accurate prediction SoweproposeMapLSTM topredict traffic conditions effectivelyWe analyze and compare the use about the predicted trafficconditions in navigation planning as in Table 3 the lower thecomputing complexity the lighter the planning algorithmthe higher the navigation accuracy the better the navigationperformance perdurability represents the sustainability of atransportation system the higher the perdurability the moresustainable the transportation system

443 Inferring Traffic Emissions In theCOPERTmodel [28]hot emissions are one of the key essentials about trafficemissions Hot emissions occur when the engine of vehicleis at its normal mode Hot emission factor 119864119865 the amountof pollutant a single vehicle emits per kilometer (gkm) iscalculated as a function of travel speed V(119896119898ℎ) [29]

119864119865 =(119886 + 119888V + 119890V2)(1 + 119887V + 119889V2)

(2)

where 119886 119887 119888 119889 119890 are the pollution emission parameters ofCOPERTmodel these values are given in [29] to calumniatedifferent kinds of emissions and gas consumption COHydrocarbon Nox Fuel Consumption (FC)

As in Figure 7 we infer different kinds of traffic emissionsand gas consumption of 126th road segment at 1000 in thenext five days the average of CO is about 05 Hydrocarbon is004 Nox is 009 FC is 424 CO2 is 429 and PM25 is 0007

As for other pollutants like CO2 and PM25 their emissionfactors are proportional to FC

1198641198651198881199002 = 318 lowast 119864119865119865119862

11986411986511987511987225= 3 lowast 10minus5 lowast 119864119865119865119862

(3)

444 Other Applications Table 4 compares the applicationsabout traffic prediction in recent two years It can be seenfrom Table 4 that the traffic prediction methods is moreinclined to usemachine learning and deep learning algorithmto achieve more accurate and larger regional predictionthe advance cannot be separated from the rapid devel-opment of machine learning and deep learning in recentyears

5 Conclusions

Urban road traffic system is the lifeblood of a city whichensures its operation Predicting traffic conditions for roadsegments is the prelude of working on intelligent trans-portation In this paper we proposed MapLSTM a trafficpredicting mechanism for road segments to promote thedevelopment of ITS MapLSTM can accelerate the landingof many applications in a lightweight and fine-grained wayIn the future autonomous humanlike driving based on roadtopography is worth concern and we will focus on complexspatial correlations in traffic environment


Table3Ap

plications

andcomparis

onsa

bout

thep

redicted

traffi

ccon

ditio

nsin

navigatio

nplanning

Literature

Rawda

tasource

Object-ba

sed

Pathway

Corea

lgorith

mCom

plexity

Accuracy

Save

time

Perdurability

[1]

smartsensors

city

self-aw

are

GaussianProcessR

egression

middle

middle

middle

low

[23]

GPS

points

region

autono

mou

sVa

lueIteratio

nNetwo

rkmiddle

middle

high

middle

[24]

street-v

iewim

ages

intersectio

nautono

mou

sCN

N+R

L+A⋆

high

middle

high

low

[25]

GPS

points

region

agents

Ant

Colon

y+RL

middle

middle

middle

middle

[26]

vehicles

sharing

city

RIS

statistic

slow

low

middle

low


Table4Ap

plications

andcomparis

onsa

bout

thetrafficp

rediction

Year

Literature

Basicd

atasource

Target

Term

Corea

lgorith

mCom

plexity

Granu

larity

Object-ba

sed

2018

[18]

webcamera

traffi

cdensity

short

Con

volutio

naln

euraln

etwo

rkhigh

fine-grained

intersectio

n[27]

anop

endataset

traffi

cflow

long

sho

rtGenerativea

dversaria

lnetwo

rkhigh

coarse-grained

freew

ay

2017

[17]

webcamera

traffi

cdensity

short

Fully

convolutionaln

etwo

rks

high

fine-grained

restr

ictedarea

[15]

anexperim

entalcar

vehiclespeed

short

Auto-regressivem

odel

middle

fine-grained

road

segm

ent

[14]

floatingcar

vehiclespeed

short

HMMs+SU

MO

middle

coarse-grained

motorway

[13]

Loop

Detector

traffi

cvolum

eshort

STsemi-sup

ervisedlearning

low

fine-grained

road

segm

ent

[1]

traffi

cloo

pstraffi

cflow

long

Gaussianprocessregression

low

coarse-grained

region


126 CO Nox Hydrocarbon Fuel Consumption

+1 +3 +4 +5+2day

0

1

2

3

4

5

6

7

00

01

02

03

04

05

Figure 7 Interring traffic emissions of 126th road segment

Data Availability

Weused the source code of ConvLSTM in our paper theURLis ldquohttpsgithubcomcarlthometensorflow-convlstm-cellrdquoMoreover we used the dataset ldquoT-Drive Taxi Trajectoriesrdquoreleased by MSRA the URL is ldquohttpswwwmicrosoftcomen-usresearchprojecturban-computingrdquo There is just oneweek of data in released dataset Although one week of datacan also conduct secondary analyses we used one month ofdata of ldquoT-Drive Taxi Trajectoriesrdquo in our experiments forbetter performance in which data was from the previouscooperation project

Conflicts of Interest

The authors declare that they have no conflicts of interest

Acknowledgments

This work is supported by the Natural Science Foundationof Beijing under Grant no 4181002 and the Natural ScienceFoundation of China under Grant no 61876023

References

[1] T Liebig N Piatkowski C Bockermann and K MorikldquoDynamic route planning with real-time traffic predictionsrdquoInformation Systems vol 64 pp 258ndash265 2017

[2] Y Zheng Y Liu J Yuan and X Xie ldquoUrban computing withtaxicabsrdquo in Proceedings of the 13th International Conferenceon Ubiquitous Computing (UbiComp rsquo11) pp 89ndash98 ToulouseFrance September 2011

[3] G-R Iordanidou I Papamichail C Roncoli and M Papa-georgiou ldquoFeedback-based integrated motorway traffic flowcontrol with delay balancingrdquo IEEE Transactions on IntelligentTransportation Systems vol 18 no 9 pp 2319ndash2329 2017

[4] L Li K Ota and M Dong ldquoHumanlike driving empiricaldecision-making system for autonomous vehiclesrdquo IEEE Trans-actions on Vehicular Technology vol 67 no 8 pp 6814ndash68232018

[5] J Li G LuoN Cheng et al ldquoAn end-to-end load balancer basedon deep learning for vehicular network traffic controlrdquo IEEEInternet of Things Journal 2018

[6] S Choudhury ldquoCellular automata and wireless sensor net-worksrdquo in Emergent Computation pp 321ndash335 Springer 2017

[7] Q Zhang L T Yang Z Chen P Li and F Bu ldquoAn adaptivedroupout deep computation model for industrial iot big datalearning with crowdsourcing to cloud computingrdquo IEEE Trans-actions on Industrial Informatics 2018

[8] Q Zhang L T Yang A Castiglione Z Chen and P LildquoSecure weighted possibilistic c-means algorithm on cloud forclustering big datardquo Information Sciences 2018

[9] I Lana J Del Ser M Velez and E I Vlahogianni ldquoRoadtraffic forecasting recent advances and new challengesrdquo IEEEIntelligent Transportation Systems Magazine vol 10 no 2 pp93ndash109 2018

[10] S Hochreiter and J Schmidhuber ldquoLong short-termmemoryrdquoNeural Computation vol 9 no 8 pp 1735ndash1780 1997

[11] S S Chawathe ldquoSegment-based map matchingrdquo in Proceedingsof the IEEE Intelligent Vehicles Symposium pp 1190ndash1197 Istan-bul Turkey June 2007

[12] X Shi Z Chen H Wang D-Y Yeung W-K Wong and W-C Woo ldquoConvolutional LSTM network A machine learningapproach for precipitation nowcastingrdquo in Proceedings of the29th Annual Conference on Neural Information Processing Sys-tems (NIPS rsquo15) pp 802ndash810 December 2015

[13] C Meng X Yi L Su J Gao and Y Zheng ldquoCity-wide trafficvolume inference with loop detector data and taxi trajectoriesrdquoin Proceedings of the 25th ACM SIGSPATIAL InternationalConference pp 1ndash10 Redondo Beach Calif USA November2017

[14] B Jiang and Y Fei ldquoVehicle speed prediction by two-level datadriven models in vehicular networksrdquo IEEE Transactions onIntelligent Transportation Systems vol 18 no 7 pp 1793ndash18012017

[15] J Jing D Filev A Kurt E Ozatay JMichelini andU OzgunerldquoVehicle speed prediction using a cooperative method of fuzzyMarkovmodel and auto-regressivemodelrdquo in Proceedings of theIEEE Intelligent Vehicles Symposium (IV rsquo17) pp 881ndash886 LosAngeles Calif USA June 2017

[16] X Niu Y Zhu and X Zhang ldquoDeepSense A novel learningmechanism for traffic prediction with taxi GPS tracesrdquo inProceedings of the IEEE Global Communications Conference(GLOBECOM rsquo14) pp 2745ndash2750 Austin TX USA December2014

[17] S ZhangGWu J PCosteira and JMMoura ldquoUnderstandingtraffic density from large-scale web camera datardquo in Proceedingsof the IEEE Conference on Computer Vision and Pattern Recog-nition (CVPR rsquo17) pp 4264ndash4273 Honolulu HI USA July 2017

[18] J Chung and K Sohn ldquoImage-based learning to measuretraffic density using a deep convolutional neural networkrdquo IEEETransactions on Intelligent Transportation Systems vol 19 no 5pp 1670ndash1675 2018

[19] P S Castro D Zhang C Chen S Li and G Pan ldquoFromtaxi GPS traces to social and community dynamicsrdquo ACMComputing Surveys vol 46 no 2 pp 1ndash34 2013


[20] Y Lou C Zhang Y Zheng X Xie W Wang and Y HuangldquoMap-matching for low-sampling-rateGPS trajectoriesrdquo inPro-ceedings of the 17th ACM SIGSPATIAL International Conferenceon Advances in Geographic Information Systems pp 352ndash361ACM Seattle WA USA November 2009

[21] Y C Hu M Patel D Sabella N Sprecher and V YoungldquoMobile edge computinga a key technology towards 5grdquo ETSIWhite Paper vol 11 no 11 pp 1ndash16 2015

[22] Q Zhang M Lin L T Yang Z Chen S U Khan and PLi ldquoA double deep q-learning model for energy-efficient edgeschedulingrdquo IEEE Transactions on Services Computing 2018

[23] S Yang J Li J Wang Z Liu and F Yang ldquoLearning UrbanNavigation via Value Iteration Networkrdquo in Proceedings ofthe IEEE Intelligent Vehicles Symposium (IV rsquo18) pp 800ndash805Changshu Suzhou China June 2018

[24] S Brahmbhatt and J Hays ldquoDeepNav Learning to NavigateLargeCitiesrdquo inProceedings of the IEEEConference onComputerVision and Pattern Recognition (CVPR rsquo17) pp 3087ndash3096Honolulu HI USA July 2017

[25] A Eydi S Panahi and I iNakhai Kamalabadi ldquoUser-basedvehicle route guidance in urban networks based on intelligentmulti agents systems and the ant-q algorithmrdquo InternationalJournal of Transportation Engineering vol 4 no 3 pp 147ndash1612017

[26] T Yamashita K Izumi and K Kurumatani ldquoCar navigationwith route information sharing for improvement of traffic effi-ciencyrdquo in Proceedings of the 7th International IEEE Conferenceon Intelligent Transportation Systems (ITSC rsquo04) pp 465ndash470Yokohama Japan October 2004

[27] A Koesdwiady and F Karray ldquoNew results on multi-step trafficflow predictionrdquo Artificial Intelligence 2018 httpsarxivorgabs180301365

[28] L Ntziachristos Z Samaras S Eggleston et al ldquoCopert iiicomputer programme to calculate emissions from road trans-portmethodology and emission factors (version 21)rdquoEuropeanEnergy Agency 2000

[29] J Shang Y Zheng W Tong E Chang and Y Yu ldquoInferringgas consumption and pollution emission of vehicles throughouta cityrdquo in Proceedings of the 20th ACM SIGKDD InternationalConference pp 1027ndash1036 New York NY USA August 2014

International Journal of

AerospaceEngineeringHindawiwwwhindawicom Volume 2018

RoboticsJournal of

Hindawiwwwhindawicom Volume 2018


Active and Passive Electronic Components

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawiwwwhindawicom

Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Control Scienceand Engineering

Journal of



Journal ofEngineeringVolume 2018

SensorsJournal of



RotatingMachinery


Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation


Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia

Submit your manuscripts atwwwhindawicom


In this paper we propose a fine-grained and lightweightapproach for traffic predicting of road segments namedMapLSTM a spatio-temporal long short-term memorynetwork (LSTM [10]) preluded by map-matching [11]MapLSTMonly requires vehiclesGPS and not need to deployspecialized traffic sensors in urban and not use the unob-tainable data from ground loop MapLSTM first obtains thehistorical and real-time traffic conditions of road segmentsviamap-matchingThenLSTM is utilized to predict the trafficconditions of the corresponding road segments in the futureBreaking the single-index forecasting MapLSTM can predictmultiple traffic conditions for road segments simultaneouslyTo summarize the major contributions of this paper consistof the following aspects

(1) Breaking through the difficulty of obtaining segment-based traffic data we perform the cognizing of road-grainedtraffic conditions via map-matching technology

(2) Based on a large scale of taxi GPS trajectorieswe propose MapLSTM to extract features from the high-dynamic high-dimensional non-linear and non-stationarytraffic flow And we confirm that MapLSTM have a higherpredicting accuracy than GPR [1] and ConvLSTM [12]

(3)We demonstrate MapLSTM can serve to various prag-matic applications cognizing driving preferences learningnavigation and inferring traffic emissions

The remainder of this paper is organized as followsSection 2 reviews the literature on traffic prediction Section 3describes the materials and gives details of our mechanismMapLSTM The next Section 4 demonstrates the effective-ness and applications This paper ends in Section 5 withconclusion on our work

2 Literature Review

Traffic condition prediction cannot only be used as the designbasis of signal control of ITS but also provide decision supportfor dynamic route guidance Whereas there are still somebottlenecks in short or long term traffic prediction througha lot of real spatio-temporal data

Spatio-temporal semi-supervised learning model pro-posed in [13] can infer the volume of each road with real-world data collected from 155 loop detectors and 6918 taxisover 17 days There are totally 19165 road segments in theurban area but only 155 road segments are equipped withloop detectors which results there in an inherent deviationin acquiring the city-wide traffic volumes Although theconstructed affinity graph can characterize the similaritiesamong roads based similar speed patterns the factors influ-encing traffic flow are not only the speed of vehicles but alsothe road topology road structure the regional characteristicsand so on Traffic conditions of each road segment cannotbe predicted accurately only by the spatial relations on themacro

A vehicle speed is influenced by many factors the vehicletype the traffic conditions and the driverrsquos behaviour A datadriven model is proposed in [14] for vehicle speed predictionwhere the average traffic speed is estimated based on histor-ical traffic data at first and then the statistical relationshipwith individual vehicle speed is presented by hidden markov

models Finally the individual vehicle speed is predicted byforward-backward algorithm Another mechanism proposedin [15] is a cooperative method which combines with fuzzymarkov model and auto-regressive model These machinelearning approaches for vehicle speed prediction do not careabout the basic data sources and focus more on the accuracyof prediction algorithms rather than the accuracy of segment-based traffic prediction

DeepSense [16] is a typical deep learning approach fortraffic prediction with Taxi GPS traces DeepSense gainsthe prediction results based on sufficient dataset by usingRestricted Boltzmann Machine Due to the night data istoo sparse so DeepSense made a prediction based fillingaccording to the data of the same time in history But thisprediction-based prediction approach may lose credibility indeep learning In addition DeepSense extract and classify thespeed only on 0 sim 60119896119898ℎ to reflect the traffic congestion orsmooth which lacks universality in some other regions

Understanding traffic density from large-scale images isanother way to recognize the traffic status Reference [17]as a related work selects a region of interest in a videostream at first then counts the number of vehicles in theregion for each frame so the density is calculated by dividingthat number by the region length Reference [18] is anotherimage-based learning to measure traffic density using a deepconvolutional neural network These vision-based cognitivemethods mainly play a role in local regions which candedicate to the operational control but cannot make anefficient decision in tactical planning with in the long run

In addition the predicted object is univocal in the existingmethods more is traffic volume or speed which can merelyinfer the traffic state of the road segment is congestion slownormal moderate and unimpeded It is necessary to explorefine-grained and accurate perception in a simple way

3 Materials and MapLSTM

In this section we first provide materials on GPS trajec-tory map-matching and LSTM Then we depict MapLSTMdesigned for traffic prediction

31 Materials

311 GPS Trajectory Taxis can be considered as ubiquitousmobile sensors constantly probing a cityrsquos rhythm and pulseBeing inherent characteristic GPS-based taxies have provento be an extremely useful data source for uncovering theunderlying traffic behaviour So far the taxi GPS data havebeen used for urban computing detecting hot spots mapreconstruction finding routes and so on [2 19]

The GPS records of a large number of taxis in a cityare routinely saved to a log file 119871 resulting in a very largedata set 119871 = 11990111 11990112 1199011119899 1199011198941 1199011198942 119901119894119899 Fieldsfor each GPS record generally contains TaxiID Location(Longitude Latitude) Speed Event GPS state Bearing and6 commonly used timing units YYYY-MM-DDHHMMSSFigure 1 shows an example of GPS log and trajectories Atrajectory 119879 is a time series of GPS points with the timeinterval between any consecutive GPS points not exceeding



middotmiddotmiddot

P21

P22

P23

P24

P25 P26

P12

P11

P13

P14

P15

P17

Pi1

Pi2

Pi3

Pi4 Pi5

P16

T1

T2

Ti


Map-Matching

timetjtit1 t2t0

titjtit1 t2tt0

before after













FC

Y

LSTM LSTM LSTM



Inputs

HiddenLayer

HiddenLayer

Outputs

1 2 3 4 5

-

-

- - -

Vehicle speed

Traverse time

Traffic volume

hellip

hellip

hellip hellip






computing

1

2 3

4

56

7

8

9

10

11

12

13 1

4











Name eMain Fields













119860V119892 119879119879 119899119904)(24) 119881119900119897119906119898119890 = sum119908119890119899119904119901119902 119862119900119906119899119905 119901119902(25) 119860V119892119878119901119890119890119889 = 119878119901119890119890119889119881119900119897119906119898119890(26) return 119860V119892 119879119879 119908119890 119860V119892 119879119879 119890119908 119860V119892 119879119879 119904119899

119860V119892 119879119879 119899119904 119860V119892119878119901119890119890119889 119881119900119897119906119898119890



4 Experiment







40

30

20

10

0

0

100

200

300

400

500

20 30 40 50 60 70 80 90 100 110 120 130 14010

X = road segment

30 40 50 60 70 80 90 100 110 120 130 140

100

80

60

40

20

0

0

100

200

300

400

500

3875

489

118

300

465 4935

10 20







119872119860119864 = 1119873

119873

sum119894=1

1003816100381610038161003816(119891119894 minus 119910119894)1003816100381610038161003816 (1)




44 Applications









Distance (m)

20 40 60 80 100 120 1400ID

050

100150200250300350400

2500

2000

1500

1000

500








149

2

1

34

56

7

89

11

1012

1314

1516

1718

19

2021

2223

24

25

26

27

28 29 30 31 32 33 34 35

36 3738 39 40 41 42 43 44

4546 47 48 49 50 51 52 53

54

55 56 57 58 59 60 61

62

63 64 65 66 67 68 69 70

71

72

73

74

75

76 77 7879

80 81 82

83 8485

8687 88

89 9091

92 93

9495

96 97

98

99 100101 102 103 104 105 106 107

108 109 110 111 112 113 114 115

116117 118 119 120 121 122 123 124

125126

127128

129130

131132

133134

135136 137 138

139140

141142

143

144

145

146147 148

A

B

Distance

Time

Speed

Count

104 8192 91 74 4866 57 47 46

104 101103 102 89 7480 73 56 47 46

104 8192 91 73 4672 5565

46104 8192 91 74 56 47




119864119865 =(119886 + 119888V + 119890V2)(1 + 119887V + 119889V2)

(2)




1198641198651198881199002 = 318 lowast 119864119865119865119862


(3)


5 Conclusions



Table3Ap

plications

andcomparis

onsa

bout

thep

redicted

traffi

ccon

ditio

nsin

navigatio

nplanning

Literature

Rawda

tasource

Object-ba

sed

Pathway

Corea

lgorith

mCom

plexity

Accuracy

Save

time

Perdurability

[1]

smartsensors

city

self-aw

are

GaussianProcessR

egression

middle

middle

middle

low

[23]

GPS

points

region

autono

mou

sVa

lueIteratio

nNetwo

rkmiddle

middle

high

middle

[24]

street-v

iewim

ages

intersectio

nautono

mou

sCN

N+R

L+A⋆

high

middle

high

low

[25]

GPS

points

region

agents

Ant

Colon

y+RL

middle

middle

middle

middle

[26]

vehicles

sharing

city

RIS

statistic

slow

low

middle

low


Table4Ap

plications

andcomparis

onsa

bout

thetrafficp

rediction

Year

Literature

Basicd

atasource

Target

Term

Corea

lgorith

mCom

plexity

Granu

larity

Object-ba

sed

2018

[18]

webcamera

traffi

cdensity

short

Con

volutio

naln

euraln

etwo

rkhigh

fine-grained

intersectio

n[27]

anop

endataset

traffi

cflow

long

sho

rtGenerativea

dversaria

lnetwo

rkhigh

coarse-grained

freew

ay

2017

[17]

webcamera

traffi

cdensity

short

Fully

convolutionaln

etwo

rks

high

fine-grained

restr

ictedarea

[15]

anexperim

entalcar

vehiclespeed

short

Auto-regressivem

odel

middle

fine-grained

road

segm

ent

[14]

floatingcar

vehiclespeed

short

HMMs+SU

MO

middle

coarse-grained

motorway

[13]

Loop

Detector

traffi

cvolum

eshort

STsemi-sup

ervisedlearning

low

fine-grained

road

segm

ent

[1]

traffi

cloo

pstraffi

cflow

long


low

coarse-grained

region



+1 +3 +4 +5+2day

0

1

2

3

4

5

6

7

00

01

02

03

04

05


Data Availability




Acknowledgments


References

































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




middotmiddotmiddot

P21

P22

P23

P24

P25 P26

P12

P11

P13

P14

P15

P17

Pi1

Pi2

Pi3

Pi4 Pi5

P16

T1

T2

Ti


Map-Matching

timetjtit1 t2t0

titjtit1 t2tt0

before after













FC

Y

LSTM LSTM LSTM



Inputs

HiddenLayer

HiddenLayer

Outputs

1 2 3 4 5

-

-

- - -

Vehicle speed

Traverse time

Traffic volume

hellip

hellip

hellip hellip






computing

1

2 3

4

56

7

8

9

10

11

12

13 1

4











Name eMain Fields













119860V119892 119879119879 119899119904)(24) 119881119900119897119906119898119890 = sum119908119890119899119904119901119902 119862119900119906119899119905 119901119902(25) 119860V119892119878119901119890119890119889 = 119878119901119890119890119889119881119900119897119906119898119890(26) return 119860V119892 119879119879 119908119890 119860V119892 119879119879 119890119908 119860V119892 119879119879 119904119899

119860V119892 119879119879 119899119904 119860V119892119878119901119890119890119889 119881119900119897119906119898119890



4 Experiment







40

30

20

10

0

0

100

200

300

400

500

20 30 40 50 60 70 80 90 100 110 120 130 14010

X = road segment

30 40 50 60 70 80 90 100 110 120 130 140

100

80

60

40

20

0

0

100

200

300

400

500

3875

489

118

300

465 4935

10 20







119872119860119864 = 1119873

119873

sum119894=1

1003816100381610038161003816(119891119894 minus 119910119894)1003816100381610038161003816 (1)




44 Applications









Distance (m)

20 40 60 80 100 120 1400ID

050

100150200250300350400

2500

2000

1500

1000

500








149

2

1

34

56

7

89

11

1012

1314

1516

1718

19

2021

2223

24

25

26

27

28 29 30 31 32 33 34 35

36 3738 39 40 41 42 43 44

4546 47 48 49 50 51 52 53

54

55 56 57 58 59 60 61

62

63 64 65 66 67 68 69 70

71

72

73

74

75

76 77 7879

80 81 82

83 8485

8687 88

89 9091

92 93

9495

96 97

98

99 100101 102 103 104 105 106 107

108 109 110 111 112 113 114 115

116117 118 119 120 121 122 123 124

125126

127128

129130

131132

133134

135136 137 138

139140

141142

143

144

145

146147 148

A

B

Distance

Time

Speed

Count

104 8192 91 74 4866 57 47 46

104 101103 102 89 7480 73 56 47 46

104 8192 91 73 4672 5565

46104 8192 91 74 56 47




119864119865 =(119886 + 119888V + 119890V2)(1 + 119887V + 119889V2)

(2)




1198641198651198881199002 = 318 lowast 119864119865119865119862


(3)


5 Conclusions



Table3Ap

plications

andcomparis

onsa

bout

thep

redicted

traffi

ccon

ditio

nsin

navigatio

nplanning

Literature

Rawda

tasource

Object-ba

sed

Pathway

Corea

lgorith

mCom

plexity

Accuracy

Save

time

Perdurability

[1]

smartsensors

city

self-aw

are

GaussianProcessR

egression

middle

middle

middle

low

[23]

GPS

points

region

autono

mou

sVa

lueIteratio

nNetwo

rkmiddle

middle

high

middle

[24]

street-v

iewim

ages

intersectio

nautono

mou

sCN

N+R

L+A⋆

high

middle

high

low

[25]

GPS

points

region

agents

Ant

Colon

y+RL

middle

middle

middle

middle

[26]

vehicles

sharing

city

RIS

statistic

slow

low

middle

low


Table4Ap

plications

andcomparis

onsa

bout

thetrafficp

rediction

Year

Literature

Basicd

atasource

Target

Term

Corea

lgorith

mCom

plexity

Granu

larity

Object-ba

sed

2018

[18]

webcamera

traffi

cdensity

short

Con

volutio

naln

euraln

etwo

rkhigh

fine-grained

intersectio

n[27]

anop

endataset

traffi

cflow

long

sho

rtGenerativea

dversaria

lnetwo

rkhigh

coarse-grained

freew

ay

2017

[17]

webcamera

traffi

cdensity

short

Fully

convolutionaln

etwo

rks

high

fine-grained

restr

ictedarea

[15]

anexperim

entalcar

vehiclespeed

short

Auto-regressivem

odel

middle

fine-grained

road

segm

ent

[14]

floatingcar

vehiclespeed

short

HMMs+SU

MO

middle

coarse-grained

motorway

[13]

Loop

Detector

traffi

cvolum

eshort

STsemi-sup

ervisedlearning

low

fine-grained

road

segm

ent

[1]

traffi

cloo

pstraffi

cflow

long


low

coarse-grained

region



+1 +3 +4 +5+2day

0

1

2

3

4

5

6

7

00

01

02

03

04

05


Data Availability




Acknowledgments


References

































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



FC

Y

LSTM LSTM LSTM



Inputs

HiddenLayer

HiddenLayer

Outputs

1 2 3 4 5

-

-

- - -

Vehicle speed

Traverse time

Traffic volume

hellip

hellip

hellip hellip






computing

1

2 3

4

56

7

8

9

10

11

12

13 1

4











Name eMain Fields













119860V119892 119879119879 119899119904)(24) 119881119900119897119906119898119890 = sum119908119890119899119904119901119902 119862119900119906119899119905 119901119902(25) 119860V119892119878119901119890119890119889 = 119878119901119890119890119889119881119900119897119906119898119890(26) return 119860V119892 119879119879 119908119890 119860V119892 119879119879 119890119908 119860V119892 119879119879 119904119899

119860V119892 119879119879 119899119904 119860V119892119878119901119890119890119889 119881119900119897119906119898119890



4 Experiment







40

30

20

10

0

0

100

200

300

400

500

20 30 40 50 60 70 80 90 100 110 120 130 14010

X = road segment

30 40 50 60 70 80 90 100 110 120 130 140

100

80

60

40

20

0

0

100

200

300

400

500

3875

489

118

300

465 4935

10 20







119872119860119864 = 1119873

119873

sum119894=1

1003816100381610038161003816(119891119894 minus 119910119894)1003816100381610038161003816 (1)




44 Applications









Distance (m)

20 40 60 80 100 120 1400ID

050

100150200250300350400

2500

2000

1500

1000

500








149

2

1

34

56

7

89

11

1012

1314

1516

1718

19

2021

2223

24

25

26

27

28 29 30 31 32 33 34 35

36 3738 39 40 41 42 43 44

4546 47 48 49 50 51 52 53

54

55 56 57 58 59 60 61

62

63 64 65 66 67 68 69 70

71

72

73

74

75

76 77 7879

80 81 82

83 8485

8687 88

89 9091

92 93

9495

96 97

98

99 100101 102 103 104 105 106 107

108 109 110 111 112 113 114 115

116117 118 119 120 121 122 123 124

125126

127128

129130

131132

133134

135136 137 138

139140

141142

143

144

145

146147 148

A

B

Distance

Time

Speed

Count

104 8192 91 74 4866 57 47 46

104 101103 102 89 7480 73 56 47 46

104 8192 91 73 4672 5565

46104 8192 91 74 56 47




119864119865 =(119886 + 119888V + 119890V2)(1 + 119887V + 119889V2)

(2)




1198641198651198881199002 = 318 lowast 119864119865119865119862


(3)


5 Conclusions



Table3Ap

plications

andcomparis

onsa

bout

thep

redicted

traffi

ccon

ditio

nsin

navigatio

nplanning

Literature

Rawda

tasource

Object-ba

sed

Pathway

Corea

lgorith

mCom

plexity

Accuracy

Save

time

Perdurability

[1]

smartsensors

city

self-aw

are

GaussianProcessR

egression

middle

middle

middle

low

[23]

GPS

points

region

autono

mou

sVa

lueIteratio

nNetwo

rkmiddle

middle

high

middle

[24]

street-v

iewim

ages

intersectio

nautono

mou

sCN

N+R

L+A⋆

high

middle

high

low

[25]

GPS

points

region

agents

Ant

Colon

y+RL

middle

middle

middle

middle

[26]

vehicles

sharing

city

RIS

statistic

slow

low

middle

low


Table4Ap

plications

andcomparis

onsa

bout

thetrafficp

rediction

Year

Literature

Basicd

atasource

Target

Term

Corea

lgorith

mCom

plexity

Granu

larity

Object-ba

sed

2018

[18]

webcamera

traffi

cdensity

short

Con

volutio

naln

euraln

etwo

rkhigh

fine-grained

intersectio

n[27]

anop

endataset

traffi

cflow

long

sho

rtGenerativea

dversaria

lnetwo

rkhigh

coarse-grained

freew

ay

2017

[17]

webcamera

traffi

cdensity

short

Fully

convolutionaln

etwo

rks

high

fine-grained

restr

ictedarea

[15]

anexperim

entalcar

vehiclespeed

short

Auto-regressivem

odel

middle

fine-grained

road

segm

ent

[14]

floatingcar

vehiclespeed

short

HMMs+SU

MO

middle

coarse-grained

motorway

[13]

Loop

Detector

traffi

cvolum

eshort

STsemi-sup

ervisedlearning

low

fine-grained

road

segm

ent

[1]

traffi

cloo

pstraffi

cflow

long


low

coarse-grained

region



+1 +3 +4 +5+2day

0

1

2

3

4

5

6

7

00

01

02

03

04

05


Data Availability




Acknowledgments


References

































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




Name eMain Fields













119860V119892 119879119879 119899119904)(24) 119881119900119897119906119898119890 = sum119908119890119899119904119901119902 119862119900119906119899119905 119901119902(25) 119860V119892119878119901119890119890119889 = 119878119901119890119890119889119881119900119897119906119898119890(26) return 119860V119892 119879119879 119908119890 119860V119892 119879119879 119890119908 119860V119892 119879119879 119904119899

119860V119892 119879119879 119899119904 119860V119892119878119901119890119890119889 119881119900119897119906119898119890



4 Experiment







40

30

20

10

0

0

100

200

300

400

500

20 30 40 50 60 70 80 90 100 110 120 130 14010

X = road segment

30 40 50 60 70 80 90 100 110 120 130 140

100

80

60

40

20

0

0

100

200

300

400

500

3875

489

118

300

465 4935

10 20







119872119860119864 = 1119873

119873

sum119894=1

1003816100381610038161003816(119891119894 minus 119910119894)1003816100381610038161003816 (1)




44 Applications









Distance (m)

20 40 60 80 100 120 1400ID

050

100150200250300350400

2500

2000

1500

1000

500








149

2

1

34

56

7

89

11

1012

1314

1516

1718

19

2021

2223

24

25

26

27

28 29 30 31 32 33 34 35

36 3738 39 40 41 42 43 44

4546 47 48 49 50 51 52 53

54

55 56 57 58 59 60 61

62

63 64 65 66 67 68 69 70

71

72

73

74

75

76 77 7879

80 81 82

83 8485

8687 88

89 9091

92 93

9495

96 97

98

99 100101 102 103 104 105 106 107

108 109 110 111 112 113 114 115

116117 118 119 120 121 122 123 124

125126

127128

129130

131132

133134

135136 137 138

139140

141142

143

144

145

146147 148

A

B

Distance

Time

Speed

Count

104 8192 91 74 4866 57 47 46

104 101103 102 89 7480 73 56 47 46

104 8192 91 73 4672 5565

46104 8192 91 74 56 47




119864119865 =(119886 + 119888V + 119890V2)(1 + 119887V + 119889V2)

(2)




1198641198651198881199002 = 318 lowast 119864119865119865119862


(3)


5 Conclusions



Table3Ap

plications

andcomparis

onsa

bout

thep

redicted

traffi

ccon

ditio

nsin

navigatio

nplanning

Literature

Rawda

tasource

Object-ba

sed

Pathway

Corea

lgorith

mCom

plexity

Accuracy

Save

time

Perdurability

[1]

smartsensors

city

self-aw

are

GaussianProcessR

egression

middle

middle

middle

low

[23]

GPS

points

region

autono

mou

sVa

lueIteratio

nNetwo

rkmiddle

middle

high

middle

[24]

street-v

iewim

ages

intersectio

nautono

mou

sCN

N+R

L+A⋆

high

middle

high

low

[25]

GPS

points

region

agents

Ant

Colon

y+RL

middle

middle

middle

middle

[26]

vehicles

sharing

city

RIS

statistic

slow

low

middle

low


Table4Ap

plications

andcomparis

onsa

bout

thetrafficp

rediction

Year

Literature

Basicd

atasource

Target

Term

Corea

lgorith

mCom

plexity

Granu

larity

Object-ba

sed

2018

[18]

webcamera

traffi

cdensity

short

Con

volutio

naln

euraln

etwo

rkhigh

fine-grained

intersectio

n[27]

anop

endataset

traffi

cflow

long

sho

rtGenerativea

dversaria

lnetwo

rkhigh

coarse-grained

freew

ay

2017

[17]

webcamera

traffi

cdensity

short

Fully

convolutionaln

etwo

rks

high

fine-grained

restr

ictedarea

[15]

anexperim

entalcar

vehiclespeed

short

Auto-regressivem

odel

middle

fine-grained

road

segm

ent

[14]

floatingcar

vehiclespeed

short

HMMs+SU

MO

middle

coarse-grained

motorway

[13]

Loop

Detector

traffi

cvolum

eshort

STsemi-sup

ervisedlearning

low

fine-grained

road

segm

ent

[1]

traffi

cloo

pstraffi

cflow

long


low

coarse-grained

region



+1 +3 +4 +5+2day

0

1

2

3

4

5

6

7

00

01

02

03

04

05


Data Availability




Acknowledgments


References

































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




40

30

20

10

0

0

100

200

300

400

500

20 30 40 50 60 70 80 90 100 110 120 130 14010

X = road segment

30 40 50 60 70 80 90 100 110 120 130 140

100

80

60

40

20

0

0

100

200

300

400

500

3875

489

118

300

465 4935

10 20







119872119860119864 = 1119873

119873

sum119894=1

1003816100381610038161003816(119891119894 minus 119910119894)1003816100381610038161003816 (1)




44 Applications









Distance (m)

20 40 60 80 100 120 1400ID

050

100150200250300350400

2500

2000

1500

1000

500








149

2

1

34

56

7

89

11

1012

1314

1516

1718

19

2021

2223

24

25

26

27

28 29 30 31 32 33 34 35

36 3738 39 40 41 42 43 44

4546 47 48 49 50 51 52 53

54

55 56 57 58 59 60 61

62

63 64 65 66 67 68 69 70

71

72

73

74

75

76 77 7879

80 81 82

83 8485

8687 88

89 9091

92 93

9495

96 97

98

99 100101 102 103 104 105 106 107

108 109 110 111 112 113 114 115

116117 118 119 120 121 122 123 124

125126

127128

129130

131132

133134

135136 137 138

139140

141142

143

144

145

146147 148

A

B

Distance

Time

Speed

Count

104 8192 91 74 4866 57 47 46

104 101103 102 89 7480 73 56 47 46

104 8192 91 73 4672 5565

46104 8192 91 74 56 47




119864119865 =(119886 + 119888V + 119890V2)(1 + 119887V + 119889V2)

(2)




1198641198651198881199002 = 318 lowast 119864119865119865119862


(3)


5 Conclusions



Table3Ap

plications

andcomparis

onsa

bout

thep

redicted

traffi

ccon

ditio

nsin

navigatio

nplanning

Literature

Rawda

tasource

Object-ba

sed

Pathway

Corea

lgorith

mCom

plexity

Accuracy

Save

time

Perdurability

[1]

smartsensors

city

self-aw

are

GaussianProcessR

egression

middle

middle

middle

low

[23]

GPS

points

region

autono

mou

sVa

lueIteratio

nNetwo

rkmiddle

middle

high

middle

[24]

street-v

iewim

ages

intersectio

nautono

mou

sCN

N+R

L+A⋆

high

middle

high

low

[25]

GPS

points

region

agents

Ant

Colon

y+RL

middle

middle

middle

middle

[26]

vehicles

sharing

city

RIS

statistic

slow

low

middle

low


Table4Ap

plications

andcomparis

onsa

bout

thetrafficp

rediction

Year

Literature

Basicd

atasource

Target

Term

Corea

lgorith

mCom

plexity

Granu

larity

Object-ba

sed

2018

[18]

webcamera

traffi

cdensity

short

Con

volutio

naln

euraln

etwo

rkhigh

fine-grained

intersectio

n[27]

anop

endataset

traffi

cflow

long

sho

rtGenerativea

dversaria

lnetwo

rkhigh

coarse-grained

freew

ay

2017

[17]

webcamera

traffi

cdensity

short

Fully

convolutionaln

etwo

rks

high

fine-grained

restr

ictedarea

[15]

anexperim

entalcar

vehiclespeed

short

Auto-regressivem

odel

middle

fine-grained

road

segm

ent

[14]

floatingcar

vehiclespeed

short

HMMs+SU

MO

middle

coarse-grained

motorway

[13]

Loop

Detector

traffi

cvolum

eshort

STsemi-sup

ervisedlearning

low

fine-grained

road

segm

ent

[1]

traffi

cloo

pstraffi

cflow

long


low

coarse-grained

region



+1 +3 +4 +5+2day

0

1

2

3

4

5

6

7

00

01

02

03

04

05


Data Availability




Acknowledgments


References

































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia









Distance (m)

20 40 60 80 100 120 1400ID

050

100150200250300350400

2500

2000

1500

1000

500








149

2

1

34

56

7

89

11

1012

1314

1516

1718

19

2021

2223

24

25

26

27

28 29 30 31 32 33 34 35

36 3738 39 40 41 42 43 44

4546 47 48 49 50 51 52 53

54

55 56 57 58 59 60 61

62

63 64 65 66 67 68 69 70

71

72

73

74

75

76 77 7879

80 81 82

83 8485

8687 88

89 9091

92 93

9495

96 97

98

99 100101 102 103 104 105 106 107

108 109 110 111 112 113 114 115

116117 118 119 120 121 122 123 124

125126

127128

129130

131132

133134

135136 137 138

139140

141142

143

144

145

146147 148

A

B

Distance

Time

Speed

Count

104 8192 91 74 4866 57 47 46

104 101103 102 89 7480 73 56 47 46

104 8192 91 73 4672 5565

46104 8192 91 74 56 47




119864119865 =(119886 + 119888V + 119890V2)(1 + 119887V + 119889V2)

(2)




1198641198651198881199002 = 318 lowast 119864119865119865119862


(3)


5 Conclusions



Table3Ap

plications

andcomparis

onsa

bout

thep

redicted

traffi

ccon

ditio

nsin

navigatio

nplanning

Literature

Rawda

tasource

Object-ba

sed

Pathway

Corea

lgorith

mCom

plexity

Accuracy

Save

time

Perdurability

[1]

smartsensors

city

self-aw

are

GaussianProcessR

egression

middle

middle

middle

low

[23]

GPS

points

region

autono

mou

sVa

lueIteratio

nNetwo

rkmiddle

middle

high

middle

[24]

street-v

iewim

ages

intersectio

nautono

mou

sCN

N+R

L+A⋆

high

middle

high

low

[25]

GPS

points

region

agents

Ant

Colon

y+RL

middle

middle

middle

middle

[26]

vehicles

sharing

city

RIS

statistic

slow

low

middle

low


Table4Ap

plications

andcomparis

onsa

bout

thetrafficp

rediction

Year

Literature

Basicd

atasource

Target

Term

Corea

lgorith

mCom

plexity

Granu

larity

Object-ba

sed

2018

[18]

webcamera

traffi

cdensity

short

Con

volutio

naln

euraln

etwo

rkhigh

fine-grained

intersectio

n[27]

anop

endataset

traffi

cflow

long

sho

rtGenerativea

dversaria

lnetwo

rkhigh

coarse-grained

freew

ay

2017

[17]

webcamera

traffi

cdensity

short

Fully

convolutionaln

etwo

rks

high

fine-grained

restr

ictedarea

[15]

anexperim

entalcar

vehiclespeed

short

Auto-regressivem

odel

middle

fine-grained

road

segm

ent

[14]

floatingcar

vehiclespeed

short

HMMs+SU

MO

middle

coarse-grained

motorway

[13]

Loop

Detector

traffi

cvolum

eshort

STsemi-sup

ervisedlearning

low

fine-grained

road

segm

ent

[1]

traffi

cloo

pstraffi

cflow

long


low

coarse-grained

region



+1 +3 +4 +5+2day

0

1

2

3

4

5

6

7

00

01

02

03

04

05


Data Availability




Acknowledgments


References

































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



149

2

1

34

56

7

89

11

1012

1314

1516

1718

19

2021

2223

24

25

26

27

28 29 30 31 32 33 34 35

36 3738 39 40 41 42 43 44

4546 47 48 49 50 51 52 53

54

55 56 57 58 59 60 61

62

63 64 65 66 67 68 69 70

71

72

73

74

75

76 77 7879

80 81 82

83 8485

8687 88

89 9091

92 93

9495

96 97

98

99 100101 102 103 104 105 106 107

108 109 110 111 112 113 114 115

116117 118 119 120 121 122 123 124

125126

127128

129130

131132

133134

135136 137 138

139140

141142

143

144

145

146147 148

A

B

Distance

Time

Speed

Count

104 8192 91 74 4866 57 47 46

104 101103 102 89 7480 73 56 47 46

104 8192 91 73 4672 5565

46104 8192 91 74 56 47




119864119865 =(119886 + 119888V + 119890V2)(1 + 119887V + 119889V2)

(2)




1198641198651198881199002 = 318 lowast 119864119865119865119862


(3)


5 Conclusions



Table3Ap

plications

andcomparis

onsa

bout

thep

redicted

traffi

ccon

ditio

nsin

navigatio

nplanning

Literature

Rawda

tasource

Object-ba

sed

Pathway

Corea

lgorith

mCom

plexity

Accuracy

Save

time

Perdurability

[1]

smartsensors

city

self-aw

are

GaussianProcessR

egression

middle

middle

middle

low

[23]

GPS

points

region

autono

mou

sVa

lueIteratio

nNetwo

rkmiddle

middle

high

middle

[24]

street-v

iewim

ages

intersectio

nautono

mou

sCN

N+R

L+A⋆

high

middle

high

low

[25]

GPS

points

region

agents

Ant

Colon

y+RL

middle

middle

middle

middle

[26]

vehicles

sharing

city

RIS

statistic

slow

low

middle

low


Table4Ap

plications

andcomparis

onsa

bout

thetrafficp

rediction

Year

Literature

Basicd

atasource

Target

Term

Corea

lgorith

mCom

plexity

Granu

larity

Object-ba

sed

2018

[18]

webcamera

traffi

cdensity

short

Con

volutio

naln

euraln

etwo

rkhigh

fine-grained

intersectio

n[27]

anop

endataset

traffi

cflow

long

sho

rtGenerativea

dversaria

lnetwo

rkhigh

coarse-grained

freew

ay

2017

[17]

webcamera

traffi

cdensity

short

Fully

convolutionaln

etwo

rks

high

fine-grained

restr

ictedarea

[15]

anexperim

entalcar

vehiclespeed

short

Auto-regressivem

odel

middle

fine-grained

road

segm

ent

[14]

floatingcar

vehiclespeed

short

HMMs+SU

MO

middle

coarse-grained

motorway

[13]

Loop

Detector

traffi

cvolum

eshort

STsemi-sup

ervisedlearning

low

fine-grained

road

segm

ent

[1]

traffi

cloo

pstraffi

cflow

long


low

coarse-grained

region



+1 +3 +4 +5+2day

0

1

2

3

4

5

6

7

00

01

02

03

04

05


Data Availability




Acknowledgments


References

































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



Table3Ap

plications

andcomparis

onsa

bout

thep

redicted

traffi

ccon

ditio

nsin

navigatio

nplanning

Literature

Rawda

tasource

Object-ba

sed

Pathway

Corea

lgorith

mCom

plexity

Accuracy

Save

time

Perdurability

[1]

smartsensors

city

self-aw

are

GaussianProcessR

egression

middle

middle

middle

low

[23]

GPS

points

region

autono

mou

sVa

lueIteratio

nNetwo

rkmiddle

middle

high

middle

[24]

street-v

iewim

ages

intersectio

nautono

mou

sCN

N+R

L+A⋆

high

middle

high

low

[25]

GPS

points

region

agents

Ant

Colon

y+RL

middle

middle

middle

middle

[26]

vehicles

sharing

city

RIS

statistic

slow

low

middle

low


Table4Ap

plications

andcomparis

onsa

bout

thetrafficp

rediction

Year

Literature

Basicd

atasource

Target

Term

Corea

lgorith

mCom

plexity

Granu

larity

Object-ba

sed

2018

[18]

webcamera

traffi

cdensity

short

Con

volutio

naln

euraln

etwo

rkhigh

fine-grained

intersectio

n[27]

anop

endataset

traffi

cflow

long

sho

rtGenerativea

dversaria

lnetwo

rkhigh

coarse-grained

freew

ay

2017

[17]

webcamera

traffi

cdensity

short

Fully

convolutionaln

etwo

rks

high

fine-grained

restr

ictedarea

[15]

anexperim

entalcar

vehiclespeed

short

Auto-regressivem

odel

middle

fine-grained

road

segm

ent

[14]

floatingcar

vehiclespeed

short

HMMs+SU

MO

middle

coarse-grained

motorway

[13]

Loop

Detector

traffi

cvolum

eshort

STsemi-sup

ervisedlearning

low

fine-grained

road

segm

ent

[1]

traffi

cloo

pstraffi

cflow

long


low

coarse-grained

region



+1 +3 +4 +5+2day

0

1

2

3

4

5

6

7

00

01

02

03

04

05


Data Availability




Acknowledgments


References

































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



Table4Ap

plications

andcomparis

onsa

bout

thetrafficp

rediction

Year

Literature

Basicd

atasource

Target

Term

Corea

lgorith

mCom

plexity

Granu

larity

Object-ba

sed

2018

[18]

webcamera

traffi

cdensity

short

Con

volutio

naln

euraln

etwo

rkhigh

fine-grained

intersectio

n[27]

anop

endataset

traffi

cflow

long

sho

rtGenerativea

dversaria

lnetwo

rkhigh

coarse-grained

freew

ay

2017

[17]

webcamera

traffi

cdensity

short

Fully

convolutionaln

etwo

rks

high

fine-grained

restr

ictedarea

[15]

anexperim

entalcar

vehiclespeed

short

Auto-regressivem

odel

middle

fine-grained

road

segm

ent

[14]

floatingcar

vehiclespeed

short

HMMs+SU

MO

middle

coarse-grained

motorway

[13]

Loop

Detector

traffi

cvolum

eshort

STsemi-sup

ervisedlearning

low

fine-grained

road

segm

ent

[1]

traffi

cloo

pstraffi

cflow

long


low

coarse-grained

region



+1 +3 +4 +5+2day

0

1

2

3

4

5

6

7

00

01

02

03

04

05


Data Availability




Acknowledgments


References

































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




+1 +3 +4 +5+2day

0

1

2

3

4

5

6

7

00

01

02

03

04

05


Data Availability




Acknowledgments


References

































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia















RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia


Date post:	23-Jan-2019
Category:	Documents
Upload:	trinhdan
View:	214 times
Download:	0 times

Predicting Fine-Grained Traffic Conditions via Spatio...

Documents