Luận án tiến sĩ: Mobile Phone Based Context-Aware Traffic State Estimation

This dissertation proposes a novel mobile phone based context-aware traffic state estimation MC-TES framework which consists ofnotable solutions for difficulties mentioned about.Among ac

Context-Aware Traffic State

Estimation

Tran Minh QuangGraduate School of Engineering and Science

Shibaura Institute of Technology

A thesis submitted for the degree of

Doctor of Engineering (Dr.Eng.)

2012 September

I feel tremendously lucky to have the opportunity to work with Prof.Eiji Kamioka on the ideas in this dissertation Prof Kamioka instilledin me a love for researching on mobile multimedia communications andubiquitous computing, agreed to take me on as a graduate student,and encouraged me to immerse myself in something I had a passionfor He is always beside me, understands my difficulties and worksside by side with me in finding the right direction He incubated mein research method, scientific writing skill, etc He even took me inhis car to collect real-time traffic data for evaluations I have nevermet a professor more generous with his time and experience.

I am grateful to Prof Shuji Kubota at Shibaura Institute of nology (SIT), Prof Shigeki Yamada at National Institute of Infor-matics (NII), Prof Hiroaki Morino and Prof Takumi Miyoshi atSIT for serving on my oral committee Prof Kubota also serves asa co-supervisor who has inspired me to new ideas proposed in thisdissertation Prof Yamada has advised me precious experiences onresearch approaches and gave several useful comments on the effec-tiveness of the proposed approaches when I visited and presented atNII Prof Morino has provided me both ideas and fruitful commentson my path of doing this research His vigorous discussions alwaysinspire me to something new

Tech-I would like to thank Prof Takashi Watanabe at STech-IT, Mr HiroyukiIshizaki at International Center for Social Entrepreneurship (ICSE),Mr Takashi Yoshida at PLANS, Ltd., who gave me constructivecomments related to the feasibility of the mobile phone based trafficstate model proposed in this thesis

when I study in Japan I still remember Ms Reiko Kageyama onthe first day when she picked me up at Narita Airport Mrs YukikoSonoi has spent a lot of time to help me find suitable apartment Mrs.Midori Yabe has supported me on any daily difficulty, from teachingme Japanese, interpreting documents issued by the city hall, to ex-plaining the worries coming from the Great Tohoku Earthquake tome I am grateful to Ms Yukari Moriuchi, Ms Ayumi Maemoto, ,at the Graduation School Section who always kindly support and ad-vise the procedures related to my research duty at SIT I would liketo thank Prof Ayao Tsuge, Prof Masato Murakami, Prof ChiakiNakayama for their great Human Resource Fostering Program whichpositively influences the way I do this research I also wish to thankSIT’s Hybrid Twining Program (HBT), Japanese Government (Mon-bukagakusho: MEXT) scholarship, and Dept of Computer Scienceand Engineering (HCM Uni of Tech Vietnam) for their supportswhen I study in Japan for my Ph.D degree.

I would like to acknowledge the fine work of the other individualswho have contributed to this research Mr Pawel Gora (from Univ.of Warsaw, Poland) provided traffic simulation software, namely thetraffic simulation framework (TSF), to generate traffic synthetic data.He also generously donated his time and expertise in customizing theTSF for special requirements on data generation in this research Mr.Issei Yajima at SIT donated his time taking me on his motorbike tocollect surrounding noise data on streets which is used for evaluatingthe proposed vehicle classification model

I would like to thank all my wonderful friends and colleagues at MobileMultimedia Communications Laboratory Mr Hiroki Murata, Mr.Keizaburo Nishina, Mr Kaoru Koshimizu, Mr Shigeki Nakamura,Mr Tomoki Takemura, Mr Yuji Yoshimura have spent a lot oftime instructing me on the campus life We have also had enjoyable

Ariff Buharudin and Mrs Effiyana Binti Ghazali Nurzal.Finally, I would like to thank my family, my parents, my wife, and mylittle daughter for their love and support My parents give me the lifeand have provided me with steady guidance, knowledge, and encour-agement My wife, Nguyen Thi Hang, has made countless sacrifices tobe always with me giving me her great support and encouragement.My little daughter, Tran Thi Minh Tam (Chip-chan) always gives mehappy laughs which ultimately helps me to forget the work wheneverit was at a standstill My parents, my wife and my little daughter arelevers for me to accomplish the thesis This dissertation is dedicatedto them.

Traffic congestion causes several social issues such as economic loss, airand noise pollution, quality of life degradation, and so forth Trafficstate estimation (TES) is one of the most important fields in intelli-gent transportation systems (ITS) aiming at alleviating traffic conges-tion Conventional TES systems which utilize road-side fixed sensorssuch as loop detectors, RFID readers, video cameras, etc., confrontessential issues of coverage limitation, real time effect and invest-ment/maintenance cost, since it is impractical to install a huge num-ber of sensors at every street In the recent years, with the advancesof mobile phone technologies, mobile devices such as mobile phones,PDA, etc., are utilized as traffic probes to overcome the aforemen-tioned road-side fixed sensor systems’ disadvantages However, mobilephone based traffic state estimation (M-TES) approach introducesemerging challenges which relate to (1) Processing lack informativedata reported by “on-the-shelf” mobile phones; (2) Managing limitedresources such as low computational capacity and narrow bandwidth;and (3) Solving difficulties rooted from low and uncertain penetrationrate This dissertation proposes a novel mobile phone based context-aware traffic state estimation (MC-TES) framework which consists ofnotable solutions for difficulties mentioned about.

Among achieved results, most important ones include: (1) Inventing anotable traffic state quantification (TSQ) model by which traffic stateis granularly quantified; (2) Proposing a novel data collection policybased on the “3R” philosophy (the Right data is collected at the Righttime by the Right devices) to control data transmission load and min-imize data redundancy; (3) Designing a unique vehicle classification

model is applied in developing countries where several vehicle typesparticipate in road networks; (4) Investigating thoroughly the effectof low penetration rate on traffic state estimation to figure out theso-called “acceptable” penetration rate for a predefined “expected”accuracy; (5) Inventing two novel velocity-density inference circuits(VDICs), namely the adaptive and the adaptive feedback VIDCs, toimprove the traffic state estimation effectiveness even if the pene-tration rate is low; (6) Proposing a notable genetic algorithm (GA)based velocity-density estimation (GA-VDEM) model to optimize thepreviously proposed VDICs; (7) Introducing a practical artificial neu-ral network (ANN) based prediction model to assure the traffic stateestimation accuracy even when the penetration rate becomes unac-ceptably low, namely just several percent or even zero; (8) Discussingand proposing suitable context-aware approaches to the difficulties of“less informative” data remaining in each issue mentioned above.Each of the proposed solutions mentioned above was evaluated thor-oughly using both the real field experimental data and numerous sim-ulated data Evaluation results reveal the effectiveness, the robustnessas well as the scalability of the proposed solutions For example, theproposed TSQ accurately provides quantitative traffic state of evenirregular traffic flows where velocity and density do not relate to eachother in a common way The GA-VDEM provides traffic state esti-mation as high accuracy as around 90% when the penetration rateis low but still relevant, namely larger than 18% The ANN basedprediction approach ensures the accuracy as around 73% even if thepenetration rate is unacceptably low, namely just several percent oreven zero Consequently, this thesis opens a new research directionon the mobile probe based traffic state estimation, accelerating theM-TES, or more general the mobile phone based ITS (M-ITS), intorealization.

Contents vi

1.1 Motivation 1

1.2 Existing Technologies and Challenges 2

1.3 Mobile Phone Based Context-Aware Traffic State Estimation TES) 10

(MC-1.4 Contributions and Thesis Organization 14

2 Mobile Phone Based Traffic State Quantification 192.1 Road Segmentation 20

2.2 Traffic State Indicators 21

2.3 Traffic State Quantification 25

2.4 Summary 31

3 Controlling Communication Cost and Ensuring Estimation curacy 343.1 Introduction 34

Ac-3.2 The “3R” Data Collection Policy 36

3.3 Vehicle Classification 39

3.4 Traffic State Quantification with Different Participating VehicleModes 48

3.5 Evaluation 51

3.5.1 Effecticveness of the “3R” Data Collection Policy 51

3.5.2 Effectiveness of the Vehicle Classification Method 55

4.3.1 Conventional Velocity-Density Inference Models 67

4.3.2 The Adaptive Velocity-Density Inference Circuit 69

4.3.3 The Adaptive Feedback Velocity-Density Inference Circuit 754.4 Evaluation 77

4.4.1 Experimental Environment 77

4.4.2 Effect of Penetration Rate on the Estimation Accuracy 78

4.4.3 Effectiveness of the Velocity-Density Inference Circuits 83

4.5 Summary 85

5 Synergistic Approaches to Optimize the Velocity-Density ence Model 875.1 Introduction 87

Infer-5.2 The Modified VDEM 89

5.3 Intelligent Context-Ware Approach to Velocity-Density Estimation 905.3.1 Overview of the Genetic Algorithm 91

5.3.2 GA-based Velocity-Density Inference Mechanism 92

5.3.3 GA-Based Velocity-Density Estimation Model (GA-VDEM) 955.4 Traffic State Prediction Under Unacceptably Low Penetration Rate 985.4.1 The Proposed ANN-based Prediction Model 99

5.4.2 Narrowing the Computational Space 100

5.5 Evaluation 104

5.5.1 Effectiveness of the GA-VDEM 104

5.5.1.1 The Detailed Optimization Capacity 104

5.5.1.2 The Overall Effectiveness of the Proposed VDEM 105

GA-5.5.2 Effectiveness of the ANN-based Prediction Model 107

5.5.3 Effect of Related Road Segments on Prediction Accuracy 1095.6 Summary 110

6 Discussion 1136.1 Traffic State Quantification Model 114

6.2 Solutions for Low and Uncertain Penetration Rate Issues 116

6.3 Implementation Difficulties in MC-TES 117

6.4 System Extension 119

6.5 Potential Business Model 120

7 Conclusion and Future Work 1237.1 Conclusion 123

7.2 Future Work 125

1.1 The architecture of a context-aware system 11

1.2 The overall architecture of the MC-TES 12

1.3 Organization of the thesis 16

2.1 Road segmentation: Principle and practice 20

2.2 Capacity, Ck,i= Qi0+ Qk,i, of a road segment i with 2 lanes 24

2.3 The traffic-state quadrant space 26

2.4 Better qualifying traffic states based on the Goodness metric 272.5 Quantifying traffic states based on the Goodness metric 28

2.6 A visual view of identifying traffic state level based on the ness metric 28

Good-3.1 A vehicle reports its information based on its velocity change rate 373.2 The boundary of a walker mobile phone at point P0 38

3.3 Data uploading decision in a walker mobile phone 39

3.4 Vehicle classification model 41

3.5 Vehicle classification at the server side based on GPS data 42

3.6 Vehicle classification based on acceleration patterns 42

3.7 An ANN model for vehicle classification using relative accelerationand surrounding noise patterns 43

3.8 Two connected biological neurons (Source: “Neural Network sign”, T.H Martin, at el [90]) 45

De-3.9 Modeling an artificial neuron 46

3.10 Activate functions 47

3.11 A Feed-forward, multilayer neural network 48

3.12 The velocity change calculating and data upload timing 53

3.13 AAThe error rate of pedestrian estimation 55

3.14 Data transmission load under the “3R” data collection policy 56

3.15 Vehicle classification result using an ANN with the combination ofacceleration and surrounding noise patterns as contextual input data 584.1 Vehicles that report data are denoted as the car-shape ones Here,p = 8, q = 15 and ρ=8/15 64

4.2 Effect of penetration rate on velocity and density estimations 67

4.3 The adaptive Velocity-Density inference circuit 70

4.4 Relation between the meanspeed capacity and the velocity tion error (both are estimated from sensed data) 74

estima-4.5 The adaptive Feedback Velocity-Density inference circuit 76

4.6 The road segmentation in TSF 78

4.7 The relation between the penetration rate and the everage velocityand density estimation errors 80

4.8 Comparing the effect of penetration rate on velocity estimationerror with the Herring work 81

4.9 Effectiveness of the velocity-density inference circuits in estimatingthe average velocity 84

4.10 Effectiveness of the velocity-density inference circuits in estimatingthe density 84

5.1 The flowchart of a GA’s process 91

5.2 The pseudo code of the Genetic Algorithm 92

5.3 A single cross-point crossover 92

5.4 The GA-Based Velocity-Density Estimation Model (GA-VDEM) 965.5 The ANN-based prediction model dealing with unacceptably lowpenetration rate 100

5.6 Related road segments of a considered road segment i 101

5.7 Effectiveness of the GA-VDEM - The case of average velocity timation 106

es-5.8 Effectiveness of the GA-VDEM - The case of density estimation 106

5.9 Effectiveness of the ANN-based prediction model compared to thatof the GA-VDEM 108

5.10 The effect of different order related road segments on predictionaccuracy 110

6.1 Business model of the T24 system 121

Transportation [1] and road traffic [2] are important parts of any economy allover the world Therefore, research in Intelligent Transportation Systems (ITS)[3], [4] has attracted a numerous number of researchers in various fields includ-ing transportations, civil engineering, statistical study, computational science,communication engineering, and so forth [5], [6], [7], [8], [9] However, besidesadvanced achievements in ITS traffic congestion still remains as a serious issuein almost every big city across the world Traffic jam is not only the cause ofeconomic loss but also the source of pollution (air, noise pollution, etc.), violenceand other social issues [10], [11] The Urban Mobility Report [12] reported in2007 that traffic congestion causes 4.2 billion hours of extra travel time requiring2.9 billion extra gallons of fuel, which cost the United State tax-payers additional$78 billion [13] The Ministry of Land Infrastructure and Transport of Japanreported in 2006 that the economic loss caused by traffic jam is around $100billion annually [14] In addition, the situations where ambulances are hopelesslystuck on the way to hospitals; shops along the road sides have to be closed; stu-dents, teachers, workers, officers cannot go to school/work in time because oftraffic jams, are not unusual in the modern cities Such uncomfortable and evendangerous traffic environments are declining the citizen’s quality of life (QoL)

This situation becomes more and more serious in developing countries such

as China, Thailand, Vietnam, Brazil, etc., where the transportation ture does not catch up the demands of economic growth [15] These countries’governments are investing a large amount of money for building transportationpolicy, developing road infrastructures as well as launching computerized projectsto control this issue However, traffic jam was not decreased but it seems to bemore and more serious year by year For example, Vietnamese government hasstarted to solve the traffic jam issue from the last decade, but the situation is notbetter if not to say worse One of the primary reasons was that relevant comput-erized traffic state estimation systems were not deployed to provide commutersupdated traffic information Consequently, commuters cannot avoid entering theheavy traffic areas revealing worse traffic conditions.

infrastruc-This thesis aims at proposing a novel traffic state estimation model using bile phones as traffic probes The proposed model must be ubiquitous (anywhere),large coverage, low investment and maintenance cost, and be viable in any trafficinfrastructure, namely even in developed or developing countries where the trafficinfrastructures are modern or not The existing technology accompanying withchallenges, the overview of the proposed approach as well as the contributionsand the organization of the dissertation will be presented in the remainder of thischapter

As mentioned above, traffic state estimation is one of the most important fields inITS researches [16], [17] aiming at alleviating traffic congestion Obviously, accu-rately updated traffic information helps commuters to avoid unconsciously enter-ing the heavy traffic areas so that traffic congestions can be alleviated Existingtraffic state estimation systems such as VICS (Vehicle Information and Commu-nication System) [18], NAVITIME [19] in Japan, the ITS project at Kansas, USA[20] majorly rely on road-side fixed sensors for traffic data collection The trafficdata such as time, vehicle image, etc., are collected anytime when a transponder(embedded on a vehicle) passes over a road-side fixed sensor such as a loop de-tector [21], [22], a RFID reader [23], [24], [25], [26], a video camera [27], [28], [29],and so forth

Loop detectors [30] can be considered as a matured technology which hasbeen widely used for providing basic traffic data such as traffic volume, presence,occupancy, and headway The velocity of vehicles can be inferred using two-loop speed trap or using single loop with a sophisticated velocity estimationalgorithm whose inputs are loop length, average vehicle length, time, and thenumber of vehicles However, vehicle types, which are useful for traffic stateestimation especially in estimating density of a traffic flow, cannot be extractedfrom common loop detectors [31] Disadvantages of loop detectors also consistof high cost for installation and repair, their high failure ratio, and traffic stateconversion complexity [32] These drawbacks can be overcome by applying on-roadway sensor systems Video image processor, microwave radar, active andpassive infrared, FRIDs, ultrasonic, and passive acoustic array are representativetechnologies These systems do not require the installation of sensors directly onthe road surface, but the sensors are mounted on the road side.

Video image processor (VIP) systems analyze images extracted from sive frames to identify basic traffic data such as the presence of vehicles, speed,density, etc [33], [34] The image processing algorithms have to deal with theissues of image filtering to remove gray level variations caused by weather con-dition, shadow, and nighttime artifacts retaining only the objects that identifyvehicles Obviously, images or video clips convey many useful data for trafficstate estimation such as traffic volume (density), speed, vehicle type, etc How-ever, there still remain several challenges which need to be thoroughly investigatedin the VIP technology More concretely, this technology is vulnerable to viewingobstructions, inclement weather (e.g rain, strong wind), shadows, vehicle projec-tion into adjacent lanes, vehicle/road contrast, vehicle distinction, day-to-nighttransition, and so forth

succes-Microwave radar [35], RFID [25], active and passive infrared [36] technologiescan complement the VIP technology for some of aforementioned disadvantages.The most important advantages of these technologies are their insensitivity toinclement weather, direct measurement of speed, multi-lane gathering of trafficdata, and cost effectiveness These road-side fixed sensor techniques, however,still disclose their essential weakness in terms of coverage limitation as well asinvestment cost since it is impractical to install a huge number of sensors on every

street.In order to overcome the overage issues, vehicle-probe based approach [37], [38]has been proposed The GPS receivers equipped in the existing car navigationsystems can provide the location based data including time, position, speed,heading of vehicles This data is useful for traffic state estimation However,car navigation systems target on showing the way from the starting point to thedestination, thus the data they need is the digital map and the current positionof a particular vehicle In contrast, traffic state estimation systems using vehicleprobes require data reported from every vehicle travelling on the road network.Meanwhile, to the best of our knowledge there is no collaboration between vehiclemakers and between navigation system providers by which the GPS data fromdifferent car navigation systems can be integrated In addition, not every carequips a car navigation system leading to poorer GPS data collection Generallyspeaking, the existing car navigation systems could not be used for traffic stateestimation unless there is a robust traffic state inference model which can provideaccurate traffic state information even if a small portion of data is collected.Standing on an opposite point, the traffic state estimation system is useful forcar navigations systems since traffic information can be used for finding optimalroutes (in terms of travel time).

Ad-hoc network technology [14], [39], [40] theoretically helps to improve thecoverage.Nevertheless, all of the difficulties discussed in apply car navigation sys-tems as traffic probes mentioned above can be easily seen in this approach More-over, this technology is not matured enough to be applied in real-world applica-tions

In the recent years, with the advances of mobile phone technologies, mobiledevices have been utilized as traffic probes for collecting traffic data [41], [42], [43],[13], [44] Since mobile phones are available everywhere and the mobile phonenetwork has already been deployed [45], [46], [47], the aforementioned issues oncoverage limitation, real-time effect, investment and maintenance cost can beovercome [48] Consequently, the mobile phone based ITS (M-ITS) research isentering a new stage of realizing a SAFE and GREEN (no traffic jam) trafficenvironment to improve the citizen’s QoL

However, paralleling with the advantages of the mobile phone technologies

such as ubiquity, large coverage, real-time effect, etc., are the emerging issueswhich require to be investigated more The difficulties come from three primarylimitations, namely 1) the limitations of the sensors equipped on the commercial“on-the-shelf” mobile phones [49], [50], [51]; 2) the limitations in the mobilephone technology resources such as the computational capacity of mobile phonesand the bandwidth of the mobile phone network [52], [53], [54], [55]; and 3)the difficulties rooted from low and uncertain penetration rate (the portion ofthe number of vehicles that report data to the estimation server out of the totalnumber of vehicles traveling in the considered road segment) [56], [42], [57] Thesedifficulties and the corresponding solutions are briefly summarized as follows:

Firstly, the GPS (Global Positioning System) [58], [59] receiver is only thecommon sensor equipped on the commercial “on-the-shelf” mobile phones whichcan be used for traffic data collection [60], [61], [62], [63] However, GPS data suchas position, direction, velocity, etc., of individual vehicles are less informative, interms of traffic state estimation [64], [65], [66] For example, traffic state canbe estimated based on the average velocity and density of the traffic flow on theconsidered road segment Here, the density, for instance, is affected by not onlythe number of traveling vehicles but also by the types of vehicles such as bus,truck, car, motorbike, etc However, classifying vehicle type based on the GPSdata reported by mobile phones is definitely a challenging issue

Secondly, the limitations in mobile phone computational capacity and in thebandwidth of the mobile phone networks [52], [53] require a sound policy to con-trol the data transmission load and the data redundancy Obviously, the higherthe data uploading (from the mobile devices to the estimation server) rate is,the more data is collected, and thus the higher accuracy the estimation modelcan reach However, a higher data uploading rate requires a higher communica-tion resource and it may lead the data redundancy Identifying the appropriatedata uploading timing which can trade of the communicational efficiency and theestimation effectiveness is also a challenging difficulty

Thirdly, in the mobile phone based traffic state estimation traffic state isestimated based on the data reported by mobile phones The more data arecollected, the higher accuracy the estimation is, and vice versa [56], [42] How-ever, there is no way to compel every mobile phone user to report the data to

the estimation server That is to say, the penetration rate, namely the portionof the number of vehicles that report data to the estimation server out of thetotal number of vehicles traveling in the considered road segment, is commonlylow, especially when the system has just been launched, which affects the trafficstate estimation accuracy significantly In addition, the “actual” penetration rateis uncertain at the estimation time implying difficulties in employing optimiza-

tion models for traffic state estimation To the best of our knowledge, the lowand uncertain penetration rate issues are the most difficult issues impeding the

realization of mobile phone based traffic state estimation systems [67].The Mobile Millennium Project (MMP) [41] is closely related to this researchwhich employed GPS-enabled mobile phones as traffic probes This project, how-ever, did not discuss the issue of classifying vehicle such as bus, truck, car, mo-torbike, etc As a result, the traffic state level might not be estimated accurately.Furthermore, the MMP did not mention the difficulty of detecting and process-ing data reported concurrently by different mobile phones on the same vehicle.It is clear that 50 data points reported from a bus (i.e 49 passengers and thedriver) are different from those reported from 50 buses (i.e from 50 bus drivers).This thesis proposes a novel mobile phone based traffic state estimation (M-TES)model by which all the issues mentioned above are thoroughly resolved Moreconcretely, whether the data is reported by mobile phones on the same vehicle aswell as the type of the carrying object, namely “walker” or “vehicle” (if vehicle,what type is it, namely bus, truck, car or motorbike) the reporting mobile phonesbelong to, are clarified These clarifications contribute significantly on improvingthe effectiveness of the M-TES model compared to the existing ones

Another fundamental requirement for a traffic state estimation system is toaccurately estimate the traffic state level The MMP employed a dynamicaltheory for analyzing vehicle flows on the road network to estimate traffic statelevels [68], [69] However, the density inference based on the estimated velocity ofthe vehicle flow on a long road does not assure an accurate estimation J Yoon,et al [45], proposed a method of dividing a road network into separate roadsegments and quantifying traffic states based on complete traces in road segments.This approach, however, could not be applied in the cases of serious congestionswhere no complete traffic trace can be obtained Our previous work [42] where

traffic state could be granularly quantified using the data reported from mobilephones is considered as one of the beginning bricks constituting the backgroundfor this dissertation In this thesis, a notable traffic state quantification (TSQ)

model is proposed in detail thereby both mean speed capacity and density of a

traffic flow are estimated independently and directly from the real-time trafficdata before being appropriately integrated in the estimation In the proposed

TSQ model, not only the qualitative traffic information such as “where the traffic

congestion occurs” but also the quantitative traffic state such as “how serious thetraffic congestion is” is provided.

Neither the MMP [41] nor the J Yoon [45] projects mentioned above discussthe issues of data transmission load and data redundancy Obviously, to controlthe data transmission load while keeping the completeness of data collection(no missing of useful data), the data collection and uploading timing shouldbe investigated thoroughly The work in [43] proposed to preset Virtual TripLines (VTLs) on the road network by which traffic data is collected and reportedonly when a vehicle passes a VTL Intuitively, this approach reduces the numberof data collections, and thus alleviates the data transmission load However,the useful data may be missed since there is no relationship between VTLs andthe places where traffic congestion actually occurs Moreover, the VTL settingcriteria such as where on the road a VTL should be set, how far should twoconsecutive VTLs be, how to set the VTLs in different road types such as arterialroad, highway, urban street, and so forth, are matters of argument Our previouswork in [48] introduced the so-called “3R” approach by which only the Rightdata is collected at the Right time by the Right mobile devices For example,the mobile phones carried by walker will not report the data since walkers do not

affect the traffic state This approach reduces both the data transmission loadand the data redundancy significantly However, the effect of low and uncertain

penetration rate on the estimation effectiveness was not discussed in that work.Instead, it still assumed that the penetration rate is always relevant

To the best of our knowledge, there is rare of existing researches on mobilephone based traffic state estimation that investigate the issues of low penetrationrate thoroughly The work in [13] focused on ensuring the traffic estimation ef-fectiveness even if the penetration rate becomes unacceptably low The authors

proposed to apply a statistical learning model to estimate traffic state in terms of

travel time and congestion state Historical data was utilized to train the

statisti-cal learning model so that it can estimate/forecast the travel time and congestionstate of the considered road segment when the current observed data is applied.The work claimed that the logistic regression model works effectively even if thepenetration rate is quite low For example, even if the penetration rate is as lowas 5% the estimation error is lower than 30% However, there are several issuesremaining that need to be thoroughly discussed and clarified First, the workutilized the VTL concept proposed in [41] while, as mentioned above, the effec-tiveness of the VTL itself is still a matter of argument Second, the work couldnot estimate the density of a traffic flow, thus this important factor could notbe taken into account when estimating traffic state Third, congestion state was

defined as a “binary” indicator which accepts only two states, namely “congested”and “not-congested” Obviously, this setting biases the estimation accuracy since

any “blind” guessing approach can also has the opportunity to reach the racy of 50% Fourth, the work employed the Paramics simulator [70] to generatesynthetic data, which gave information about every vehicle, for evaluations Toimitate a low penetration rate dataset, namely 5% for example, a large portionof data (95%) was removed In fact, this process could not generate the appro-priate low penetration rate dataset as it was defined in the work Therefore, theestimation error versus the penetration rate obtained in that research should beclarified

accu-Our previous works in [71], [67] developed a background and new entrance on

solving the issues of low and uncertain penetration rate These works investigate

the effect of low penetration rate on estimation effectiveness, and then proposedappropriate solutions Consequently, two major solutions, namely, the “velocity-density inference circuits” (VDICs) and the ANN-based prediction model wereproposed The VDIC models aim at improving the effectiveness of the M-TESwhen the penetration rate is low using only the real-time traffic data reportedby mobile phones However, as any inference model which relies on the real-time data, VDICs cannot properly work when the quality of the real-time datadeclines, that is to say, when the penetration rate becomes unacceptably low TheANN-based prediction model complements the weakness of the VDIC approaches

coping with the issue of unacceptably low penetration rate.As mentioned in the beginning of this section, besides the advantages of themobile phone technologies such as ubiquity, large coverage, real-time effect, etc.,several issues are waiting to be solved in order to make the proposed M-TESeffective, robust and realistic The challenges here can be summarized as follows:The M-TES is expected to provide accurate and detailed information about traf-fic state under the condition of less informative data, in terms of traffic stateestimation Meanwhile, the obtained data is considered to be low qualified since

it is affected by the issues of low and uncertain penetration rate To solve this

dilemma, the proposed M-TES model is built on the background of context-awaresolutions [72] More concretely, the M-TES model is aware the contexts, which

are extracted from the data sensed by mobile devices, surround the considered

road segment to improve the efficiency and the effectiveness of the estimation.For example, in order to improve the accuracy of the density estimation, vehicletype must be recognized effectively However, data reported from mobile phonesincludes neither the image nor the size of the vehicle, thus the vehicle classifi-cation becomes more difficult To resolve this difficulty, the M-TES utilize thecontextual data, namely the acceleration pattern [42], [73] extracted from theGPS data and the surrounding noise data [48] sensed by the mobile phone’s mi-cro phone, for instance, to be aware about the type of vehicle which carries themobile phone More concretely, the heavy vehicle (e.g bus, truck) may takelonger time to accelerate its velocity compared to that of the light vehicle (car,motorbike) and the surrounding noise around a motorbike is commonly largerthan that around an automobile

The proposed model is named “Mobile Phone Based Context-Aware TrafficState Estimation” (MC-TES) The next section will briefly describe the overviewof the proposed MC-TES

1.3Mobile Phone Based Context-Aware Traffic

State Estimation (MC-TES)

The notion of context has been observed in numerous areas including linguistics,

knowledge discovery and presentation [72], artificial intelligent [74], informationretrieval [75], [76], reasoning, robotics, theory of communication [77], [78], and so

forth As a high level of abstraction, context is defined as “that which surrounds,

and gives meaning to, something else.” [72], [79] In this definition “something”

can be an artifact, a building, a person, a computer system or even an assertionin logic, etc Dey [80] gives a more detailed and practical definition of context

as “Context is any information that can be used to characterize the situation of

an entity An entity is a person, place, or object that is considered relevant tothe interaction between a user and an application, including the user and appli-cations themselves.” The “situation” or the “status” of an entity may be clearly

represented by the entity itself The entity status may also a variable which mustbe “inferred” The inference based on information surrounding the entity at theestimation time can be interpreted as a “context-aware inference” model

As mentioned, entity and considered status are application specific An entitymay have several status based on the point of view For example, in the MC-TES, the entity is the considered road segment and the status is its traffic state.Entities may also occur in subsystems For example, to improve the effectivenessof the MC-TES, vehicle mode should be well classified The entity here is thevehicle which carries the mobile phone and the considered status is its mode

Seng Loke [72] describes a context-aware system as having three basic

func-tionalities, namely sensing, thinking, and acting as illustrated in Fig. 1.1 The

sensing component retrieves the data around the considered entities and

pro-vides such data to the computer system for processing Sensors provide a meansto acquire data about the physical world, thus they can be viewed as a bridgebetween the physical world and the virtual world There is a large variety of sen-sors including motion sensors, smoke sensors, temperature sensors, touch sensors,FRID tags or smart labels, and so forth [81] Moreover, many other devices canalso be viewed as sensors such as clock in the computer, microphones in mobile

phones, etc The thinking component processes data sensed by sensors to

PresentingDisplaying

Sensors

Acting

Thinking

Sensing

Figure 1.1: The architecture of a context-aware system

vide the acting component accurate status/information of/about the consideredentity via reasoning More concretely, raw information/data is perceived throughthe senses, and then reasoning is employed to infer more knowledge before act-ing The reasoning can also be cascaded which means the output of a reasoningmodel can be served as the sensed data for the higher layers of reasoning for

further knowledge The acting component provides actions to users based on therecognized situations provided by the thinking component An action may be an

automatically completed action such as turning on the fire alarm when a relevantvolume of smoke is detected, warning drivers if two cars are too close with eachother on a highway, and so forth It may also be as simple as displaying the en-tity’s status while the final action deferred to users Actions are also performingfurther sensing

Depends on the application, a context-aware system may consist of complexsensors but performs simple data processing before acting In contrast, othersystems may utilize little sensing but perform much more sophisticated reasoning.In the context-aware system design, designer has to analyze suitable contexts

which are useful for an optimal inference/reasoning model Too few contextsmay result in lacking of relevant data, while too many contexts, in contrast, maybias the inference effectiveness since some contexts may be un-useful and interferecontributions of other contexts In addition, the operational feasibility in sensingcontexts under the condition of sensor availability in the proposed system shouldbe investigated thoroughly The proposed MC-TES aims at providing accuratetraffic state information based on the data reported by mobile phones in whichnot many types of sensors are available Therefore reasoning in the MC-TESmight be sophisticated The overall architecture of the MC-TES is illustrated inFig 1.2.

Mobile phones, PDA or Navigation devices

Estimating System

DatabaseCollecting real-time

traffic data, notifying current traffic state

informationUpdating the data

Preprocessing: noise filtering, refreshing data, Estimating traffic

situationsStore the real-time traffic data and digital Map for road systems

Tier 1

Tier 2

Tier 3

Figure 1.2: The overall architecture of the MC-TES

As shown in Fig 1.2, the architecture of the MC-TES consists of three tiers.Mobile devices such as mobile phones or PDAs at the 1st tier play at the sensingto collect real-time traffic data The mobile devices also decide whether to report

data to the server based on their roles (i.e “walker” mobile phones or “driver”mobile phones), decide when data should be reported (appropriate data upload-ing timing) to ensure only useful data is collected based on the knowledge theyinferred from appropriate contexts Mobile devices also serve as terminals for re-ceiving and displaying traffic state information The estimation server at the 2nd

tier processes the data, employs reasoning to effectively estimate and disseminatetraffic state information to mobile devices or to the Internet The server alsoupdates new traffic states to the database server at the 3rd tier for the furtherusage The database server stores real-time traffic data and a digital map whichis necessary for the traffic estimation and dissemination processes

The primary aim of this model is to provide accurate traffic state information,not only qualitatively but also quantitatively, using the data reported by mobilephones However, as mentioned before, there is a limitation on sensing providedby the commercially “on-the-shelf” mobile devices Therefore, suitable context-aware approaches are applied For example, to control the data transmission load,the data timing should be sound managed In the MC-TES, the data uploadingis taken place only when the velocity of the vehicle changes (increases/decreases)significantly since the change of vehicles’ velocities reflects the change of trafficstate In turn, the change of velocity can be detected by analyzing the GPSdata reported by mobile phones Another example of applying the context-awareapproach was the one given at the end of the previous section where vehicle typecan be classified based on the velocity change rate (or the acceleration pattern)and the surrounding noise The last but not least example is to solve the issue of

uncertain penetration rate In practice, the “actual” penetration rate cannot be

known at the estimation time so that optimal inference models cannot be applied.However, there are some ways to select the appropriate optimization model basedon related contexts such as the calculated average velocity and density (using thereal-time data), time of the day, road segment physical contexts such as thelength, number of lanes, maximum capacity, limited velocity, etc., and so forth.These issues will be thoroughly discussed and resolved in detail in chapter tochapter The remaining of this chapter is to summarize the contributions andthe organization of this dissertation

1.4Contributions and Thesis Organization

The primary contribution of this thesis is proposing a notable traffic state tion model using mobile phones as traffic probes Here, not only the advantagescome from mobile phone technologies such as ubiquity, large coverage, real-timeeffect, etc., are utilized but also the emerging issues in mobile phone based trafficstate estimation such as the lack of necessary information for traffic state estima-

estima-tion, the uncertain about the data reported (i.e low and uncertain penetration

rate) are appropriately resolved In order to resolve all the aforementioned ficulties effectively, appropriate context-aware approaches should be employed.As a result, a novel mobile phone based context-aware traffic state estimationmodel (MC-TES) was proposed The main contributions of this dissertation canbe summarized as follows:

dif-1 A notable traffic state quantification model was proposed under the dition of less informative traffic data sensed by mobile phones More concretely,contributions in this subject are as follows:

con A notable traffic state quantification model was proposed thereby not onlythe qualitative information of the positions where congestions occur but also thequantitative information of the congestion levels will be provided This modelalso improves the effectiveness of the traffic state estimation model since both the

velocity and the density of a traffic flow are estimated independently and directly

from the real-time traffic data before being integrated in the estimation.- A novel “3R” philosophy for real-time traffic data collection was introducedto control the data transmission load (saving the limited communication resource)and reduce the data redundancy The “3R” philosophy means only the “Right”data are collected by the “Right” players/devices at the “Right” time

- A unique vehicle classification method was proposed to classify ing vehicles such as bus, truck, car, motorbike, etc., using the data automaticallycollected by mobile phones This method helps to improve the traffic state esti-mation accuracy in cases there are several vehicle types participate in the roadnetwork, especially in developing countries

participat-2 Low and uncertain penetration rate issues were thoroughly resolved.

- The effect of low penetration rate on the traffic state estimation

effective-ness was thoroughly investigated by which suitable solutions can be figured out.In addition, based on this analysis, the so-called “acceptable” penetration ratefor a predefined ”expected” accuracy can be recognized This recognition helpsto set a pre-condition in order to ensure an “expected” accuracy of the trafficstate estimation For example, if the expected accuracy is 70% the acceptablepenetration rate must be at least 40% The “acceptable” penetration rate alsoserves as a trigger to notify the commuters on the confidence of the estimate incases of low penetration rate.

- Proposed two novel velocity-density inference circuits (VDICs) based onthe “adaptive” and the adaptive “feedback” approaches to improve the trafficstate estimation’s effectiveness These models can also minimize the “acceptable”penetration rate for a given expected estimation accuracy This contribution ispractically meaningful since it helps to assure the effectiveness of the estimationmodel even with a small portion of mobile phones participate in the system

- A notable genetic algorithm (GA) based optimization mechanism was posed to optimize the effectiveness of the two VDICs mentioned above

pro Introduced a reasonable artificial neural network (ANN) based predictionmodel to assure the MC-TES’s accuracy even when the penetration rate becomesunacceptably low, namely just several percent or even zero This contributionis important since it helps to assure a level of estimation accuracy in the MC-TES, even no penetration rate is detected For example, with the ANN-basedprediction model, the MC-TES can assure that the estimation accuracy is around70% regardless of penetration rate

The organization of this thesis is illustrated in Fig 1.3 and described asfollows:

Chapter 1: Introduction The motivation and background of this research

are described in this chapter Moreover, literature review on existing technologieswas presented by which the necessary of the proposal in this dissertation (i.e.to propose a mobile phone based context-aware traffic state estimation model -MC-TES) was figured out The primary contributions of this research were alsoconcretely summarized in this chapter

The two main topics, namely “traffic state quantification model” and“solving the issues of low and uncertain penetration rate”, contributed

Chapter 2:Mobile Phone Based Traffic State

QuantificationTraffic State Quantification Model Solving the issues of low and

uncertain penetration rate

Chapter 3:Controlling Communication Cost and

Ensuring Estimation Accuracy

Chapter 4:Adaptive Approaches to Low Penetration

Rate Issues

Chapter 5:Synergistic Approaches to Optimize the

Velocity-Density Inference Model

Chapter 7:Conclusion and Future Work

Chapter 8:Summary Chapter 1:Introduction

Chapter 6:Discussion

Figure 1.3: Organization of the thesis

by this thesis are thoroughly resolved in detail from chapter 2 to chapter 5.Chapter 6 discusses, chapter 7 concretely concludes the work and figures outfuture work directions Chapter 8 summarizes this dissertation.

Chapter 2: Mobile Phone Based Traffic State Quantification This

chapter proposes a notable traffic state quantification model thereby traffic state

can be quantitatively estimated In this model, a so-called traffic-state quadrant

was proposed which can help to integrate both average velocity and density ofthe traffic flow in an appropriate way for traffic state quantification

Chapter 3: Controlling Communication Cost and Ensuring

Esti-mation Accuracy This chapter is dedicated to solve the issues of

communica-tion limitacommunica-tion and other related issues to improve the effectiveness of a mobilephone based traffic state model A novel data collection policy named the “3Rdata collection policy” was proposed by which only the Right data is collectedat the Right time by the Right players This data collection policy reduces datatransmission load and data redundancy significantly without compromising thecompleteness of useful data In addition, a unique vehicle classification modelusing only the data reported from mobile phones was proposed This model canidentify the type such as bus, truck, car, motorbike, etc., of vehicles travelingin the considered road segment Therefore, it helps to improve the effectivenessof the density estimation and hence the effectiveness of the overall traffic stateestimation

Chapter 4: Adaptive Approach to Low Penetration Rate Issues In

this chapter, the effect of low penetration rate on the traffic state estimation tiveness was thoroughly investigated leading to suitable solutions Consequently,this chapter proposes two novel velocity-density inference circuits (VDICs) basedon the “adaptive” and the adaptive “feedback” approaches to improve the trafficstate estimation’s effectiveness even in the cases of low penetration rate

effec-Chapter 5: Synergistic Approaches to Optimize the Velocity-Density

Inference Model This chapter proposes a notable genetic algorithm (GA)

based velocity-density estimation model (GA-VDEM) to optimize the ness of the two VDICs proposed in chapter 4 The GA-VDEM improved theeffectiveness of the traffic state estimation model significantly when the penetra-tion rate becomes low However, if the penetration rate becomes unacceptably

effective-low, the GA-VDEM cannot work properly This difficulty is solved by a sonable artificial neural network (ANN) based prediction model proposed in thischapter The ANN-based prediction model assures the MC-TES’s accuracy toaround 73% regardless of penetration rate.

rea-Chapter 6: Discussion This chapter discusses the work investigated and

solutions proposed in this dissertation by which advantages as well as the ing issues will be summarized In addition, system implementation difficulties,system extension, and the potential business model for the proposed system arediscussed

remain-Chapter 7: Conclusion and Future Work This chapter concludes the

dissertation by which advantages as well as remained difficulties were discussed.Finally, research directions of great interest for the future work were drawn out

Chapter 8: Summary This chapter concretely summarizes the

disserta-tion

Mobile Phone Based Traffic StateQuantification

One of the essential requirements for a reliable traffic state estimation systemis that it not only qualitatively recognizes where traffic congestions occur but itmust have the ability of providing more detailed traffic state information, namelyin traffic state level This chapter introduces a notable traffic state quantificationmodel, which makes this work different from existing researches, based on datareported by mobile phones

Traffic congestion is a condition on road networks that occurs as use increasescharacterizing by slower speeds, longer trip times, and increased vehicular queu-ing [82] Generally speaking, traffic congestion occurs when demand approachesthe capacity of a road segment Therefore, traffic state estimation should beperformed on each road segment by which the so-called “capacity” can be wellmeasured based on the road segment’s physical characteristics, and hence traffic

state can be more accurately estimated In addition, velocity and density of a

traffic flow directly reflect traffic condition of the considered road segment This

chapter investigates traffic state related indicators such as velocity and density,

their estimation in the environment of the mobile phone based traffic state mation (M-TES), as well as an appropriate integration these factors in granularlyquantifying traffic state As a result, a notable traffic state quantification (TSQ)model based on data reported by mobile phones is proposed

esti-2.1Road Segmentation

Traffic characteristics commonly vary from road segments to road segments, thustraffic state should be estimated based on a road segment basis In this work, theroad segment is defined in a suitable way that motivates to improve the mobilephone based traffic estimation effectiveness Obviously, road-land marks such asintersection, crosswalk, curved place, place where the width or the number of laneschange, etc., would be the starting points of a road segment For the highwayswhere characteristics mentioned above do not change frequently, road segmentsare divided in each kilometer Figure2.1shows the road segmentation where Fig

2.1a depicts the principle and Fig 2.1b shows the practical application in theTSF (Traffic Simulation Framework) [83] It should be noted that road segmentsare divided based on the road direction instead of the number of lanes Hence, a

stretch of 2-way road consists of two different segments regardless of the number

of lanes in each direction

a road segmentation principle

b road segmentation in TSF

Figure 2.1: Road segmentation: Principle and practice

2.2Traffic State Indicators

This section proposes a novel traffic state quantification model by which trafficstate is not only qualitatively qualified but also granularly quantified so thattraffic information becomes more useful for the user Considering a road network

with a total of N road segments, the set of all road segments is denoted as

V={i|i = 1 N} For any road segment i ∈ V, traffic data is available at any

time t However, the obtained data (GPS data) is the event-based data which

cannot be directly transformed into traffic state Therefore, traffic state should be

aggregated in predefined time intervals, namely in t-second windows Concretely,traffic state can be estimated at times k = 0, t, 2t, , where t is the aggregation

time mentioned above The task here is how to effectively estimate traffic state

of the considered road segment i at time interval k based on data reported by

mobile devices

Obviously, velocity and density of a traffic flow directly reflect traffic conditionof a road segment Therefore, both velocity and density of a traffic flow should be

considered in estimating traffic state level Existing researches commonly tried

to estimated velocity using the sensed data (the real data reported by sensor

systems or by mobile phones in this research) and then applied some inferencemodels such as the Greenshields’s velocity-density inference model [84] to infer

the density using the estimated velocity However, this inference approach mayreveal some systematic errors if the estimated velocity contains any error In

addition, there are several types of traffic state in which the relationship between

velocity and the density could neither be known in advance nor follow common

rules For example, commonly the increasing of density causes the slowing downof velocity In practice, there are cases, however, where few vehicles travel in a

low speed because of the bad physical condition on road networks In this case,

although the velocity is low, the density is still low, thus the traffic state is not in

congestion On the other hand, there are several cases where many vehicles travel

with good speeds In these cases, the traffic flows are in high velocity with high

density and they are also not in congestion As a result, if only the velocity or only

the density is considered, there may be some biases in the estimation To solve

these issues, a notable traffic state quantification model should be proposed by

which velocity and density of a traffic flow should be independently and directly

estimated using real-time data before being integrated in an appropriate way toidentify traffic state in a granular level In order to reach this target, traffic staterelated indicators and their estimations should be clearly defined

Definition 2.1: The average velocity of a traffic flow in the road segment iduring time k, denoted as Vk,iAvg, is the average velocity of all vehicles traveling inthe considered road segment

The average velocity can be formally expressed in equation (2.1) Here, Vk,itm,jis the velocity of any individual vehicle j (j = 1 q) detected at time tm(m = 1,2, , r) within the time interval k ([k-1]t ≤ tm≺ kt), and qr is the total number

of vehicles q multiplied by the total number of detection times r during timeinterval k.

VAvgk,i =

Pq

j=1Vtm,jk,iqr, (k − 1)t ≤ tm≺ kt (2.1)

Intuitively, the average velocity defined above can describes how good or howbad the traffic flow is in terms of travel time In practice, however, the limit speedvaries from road segments to road segments due to the physical feature of eachroad segment, the average velocity defined above may not correctly represent thetraffic condition in terms of travel time As a part of this work, a new term,

namely the mean speed capacity [42], [48], to better present the travel time of a

traffic flow at a specific road segment was proposed as follows:

Definition 2.2: The mean speed capacity of the road segment i during time

k, denoted as Mk,iV , is defined as the average velocity divided by the limit speedof the considered road segment

The mean speed capacity can be mathematically expressed in equation (2.2),

where Vk,iAvgis the average velocity defined in definition 2.1, and Vi

M ax is the

limited speed of the road segment i.

MVk,i = V

k,iAvg

ViM ax

(2.2)

Observation 2.1: The mean speed capacity, Mk,iV , represents traffic state inthe manner of travel time It is clear that the higher the mean speed capacity is,

the better the traffic state is, and vice versa In this research, the threshold of

MVis set to 0.6 for a good traffic condition in terms of travel times.In addition to travel time represented by mean speed capacity, the density of

the traffic flow also reflects traffic condition of the considered road segment Thedensity of a traffic flow is defined as follow:

Definition 2.3: The density of a traffic flow in the road segment i duringtime k, denoted as Dk,i, is the fraction of the number of the vehicles traveling

through the considered road segment during time k out of the capacity of the

considered road segment.Equation (2.3) describes the density calculation, where qk,iis the total number

of vehicles traveling through the road segment i during time k, and Ck,i is the

capacity of the road segment i Here, qk,i is estimated based on the sensed data

(the real data reported by mobile phones), while Ck,i is defined in definition 2.4

Dk,i = qk,i

Definition 2.4: The flow capacity of the road segment i during time k,denoted as Ck,i, is the maximum number of vehicles passing through the road

segment i during time k under the best traffic condition (i.e vehicles can reach

the limited speed of the considered road segment) The flow capacity can becalculated in equation (2.4)

where, Qi

0 is the maximum number of vehicles that can be arranged (without

moving) in the road segment i and Qk,i is the volume of the traffic flow which

can pass the down-stream boundary of the road segment i during time k in thebest traffic condition The static volume, Qi

0, can be calculated in equation (2.5)

while the dynamic volume, Qk,i, requires a further investigation

Qi0 = ml

In this equation m and l are the number of the lanes and the length of theroad segment i, respectively, and lc is the average length of a car which is set to

5m in this work [48], [85] The value of 1.5 is the coefficient describing the space

which must be yielded between two cars in the worst congested area (i.e it must

be 0.5 the average length of a car) Figure 2.2 illustrates these parameters

0.5lcarl

Down-stream boundaryQ0

Qk,im=2

The dynamic volume, Qk,i, of a traffic flow is defined as the number of vehicles

passing the down-stream boundary of road segment i during time k The dynamic

volume can be estimated in equation (2.6) Here, ¯t is the average elapse timebetween two consecutive vehicles, namely vehicles ith and (i+1)th, passing the

down-stream boundary; Vi

M ax, and m are limit speed and the number of lanes ofroad segment i, respectively; and lc is the average length of a car as mentionedbefore

Qk,i= mk

¯t= mk

ViM ax

1.5lc

(2.6)

For convenience in estimating traffic state in the manner of density, the term

called FREE space ratio is defined as follow:Definition 2.5: The FREE space ratio of the road segment i during time k,denoted as σk,i

s , is calculated in equation (2.7), where Dk,iand Ck,i were definedin definitions 2.3 and 2.4, respectively

Observation 2.2: The FREE space ratio, σk,i

s , represents traffic state in the

manner of density It is clear that the higher the FREE space ratio is, the betterthe traffic state is, and vice versa In this work, the threshold of σsis set to 0.4

for a good traffic condition

After analyzing the effect of mean speed capacity and FREE space ratio to

traffic state separately, these two factors should be integrated in an appropriateway in order to obtain a better estimation This integration in order to buildup an effective traffic state quantification model will be thoroughly discussed indetail in the next section

2.3Traffic State Quantification

As discussed in the previous section, each factor, namely mean speed capacity

(MVk,i) and FREE space ratio (σk,i

s), represent traffic state in different manners,

namely in travel time and density (i.e the influence of a traffic flow), respectively.

However, how to integrate them to draw out a clear figure about traffic state bywhich traffic state can be granularly quantified is still a challenge In this work,

the so-called “traffic-state” quadrant space is proposed in order to effectively

es-timate and comprehensively present traffic state Figure 2.3 shows the quadrant

space which is built up by two axes representing MVk,iand σk,i

s , respectively ues in each axis range from 0 to 1 representing from the worst to the best trafficconditions in corresponding manners, namely in travel time and in density, re-

Val-spectively The “traffic-state” quadrant space is created by two new axes that are

orthogonal to the original axes at the point representing threshold values, namely

0.6 and 0.4 for MVk,iand σk,i

s, respectively This “traffic-state” quadrant space can

be used to qualitatively identify the traffic state of the considered road segment

The 1st quadrant (the rightmost one) represents good traffic states where both

MVk,iand σk,i

s are greater than their thresholds for a good traffic condition, namely0.6 and 0.4, respectively as mentioned in the previous section In contrast, the

3rd quadrant represents the bad states since both the two aforementioned factors

are lower than their thresholds Meanwhile, the 2ndand 4th quadrants represent

the threat traffic conditions The 2nd quadrant expresses situations where thereare few vehicles travelling slowly These situations may occur due to some specialphysical conditions such as road construction, bad weather, etc., by which vehi-cles cannot move faster They are considered as threat situations (i.e not good

but not bad) in terms of traffic state The 4th quadrant represents the oppositesituations (i.e good speed but in high density) In this case even the mean speed

capacity is good, if something happens by which the drivers in the front ate then the traffic condition becomes worse quickly The merit of the quadrantspace is that it is viable to estimate traffic situations that the conventional meth-ods could not estimate More concretely, the Greenshields-like models [84], [86],[87] cannot estimate, even in qualitative manner, the traffic states that fall in

deceler-quadrants 2 and 4.

1

0.60.4

1

10

2

A road segment is mapped to a node on the quadrant space

Figure 2.3: The traffic-state quadrant space

The next question is how to quantify traffic states since the two “bad” (andcorrespondingly, two “good”) traffic states (e.g at two different road segments)do not represent the same real traffic condition Intuitively, a point that is farapart from the new origin (the origin forms the quadrant space - Fig 2.3) upwardand right-warded represents a better state, and vice versa Here, the term called

the Goodness value is defined to quantify the traffic state as follow:Definition 2.6: The Goodness value of the road segment i during the esti-mation time k, denoted as Gk,i(MVk,i, σk,i

s ), is calculated in equation (2.8), where

MV 0and σs0are the thresholds of mean speed capacity and FREE space ratio,

respectively

Gk,i(MVk,i, σsk,i) = (MVk,i− MV 0) + (σk,is− σs0) (2.8)

Observation 3: The Goodness value is a continuous value, ranging from -1to 1, representing from the worst to the best traffic states, respectively.

For example, the Goodness values of the worst traffic state (presented by theoriginal origin, namely at the point (0,0)) and of the best traffic state (presented

by the rightmost upward point on the quadrant space) are calculated in equations(2.9) and (2.10), respectively

Gworst= G(0, 0) = (0 − 0.6) + (0 − 0.4) = −1 (2.9)

Gbest= G(1, 1) = (1 − 0.6) + (1 − 0.4) = 1 (2.10)

MV





Figure 2.4: Better qualifying traffic states based on the Goodness metric

Using the Goodness value, traffic state can be granularly quantified as shown

in Fig 2.4 and Fig 2.5 Figure 2.4 shows that the Goodness values, namely

Gk,i(MV, σs), of all the points on the cross line are identical and equal to 0 Thepoints upper the cross line represent good (1st quadrant) or fairly good (a

portion of the 2ndand 4th quadrants) traffic states while the points under thisline represent the bad and the threat traffic conditions More concretely, Fig

2.5shows the Goodness values of some selected points in the quadrant space Thevertical axis represents the Goodness values of traffic states which are mapped

into the quadrant space on the ground of the chart Figure 2.6 shows a visual

view of identifying traffic state level based on the Goodness value For instance,the Goodness value of the traffic state which is mapped to the rightmost upwardpoint on the quadrant space is 1, hence this traffic state is considered as the best

one It should be recalled that the Goodness value is continuous so that it is