Institut EurecomDepartment of Mobile Communications

2229, route des CretesB.P. 193

06904 Sophia-AntipolisFRANCE

Research Report RR-05-137

A Framework for Mobility Models Generation and itsApplication to Inter-Vehicular Networks

April 5th, 2005

Jerome Harri, Fethi Filali and Christian Bonnet

Tel : (+33) 4 93 00 26 26Fax : (+33) 4 93 00 26 27

Email : {Jerome.Haerri,Fethi.Filali,Christian.Bonnet}@eurecom.fr

1Institut Eurecom’s research is partially supported by its industrial members: Bouygues Telecom,Fondation d’entreprise Groupe Cegetel, Fondation Hasler, France Telecom, Hitachi, Sharp, ST Mi-croelectronics, Swisscom, Texas Instruments, Thales.

A Framework for Mobility Models Generation and itsApplication to Inter-Vehicular Networks

Jerome Harri, Fethi Filali and Christian Bonnet

Abstract

Without using a realistic mobility model, performance results obtainedfrom simulations of mobile ad hoc networks may not correlate well with per-formance in a real deployment. A realistic mobility model must be based ontopological maps, must include a traffic generation model, and must takeinto account preferential movements or destinations. In this paper, we pro-pose a novel concept map for mobility models and use it in a short survey ofexisting proposals. We then review necessary requirements, and define keycomponents for the generation of mobility models adapted to vehicular adhoc networks (VANETs). Based on this, we first adapt our concept map tovehicular motion, then present a framework for the generation of vehicularmobility models that include all parameters vehicles experience while mov-ing, and finally propose two derived mobility models at the stage of research.

Index Terms

Mobility modeling, framework, concept map, traffic generator, motioncontraints, vehicular ad hoc networks, VANET.

Contents

1 Introduction 1

2 Improved Concept Map and Short Survey of Existing Mobility Models 22.1 Bi-polar Concept Map for Mobility Models . . . . . . . . . . . . 22.2 Survey of Existing Mobility Models with Regards to the Proposed

Concept Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

3 Identification of Vehicular Mobility Models Components 6

4 Framework for Vehicular Mobility Model Generation 94.1 Vehicles-Adapted Concept Map . . . . . . . . . . . . . . . . . . 94.2 Framework General Description . . . . . . . . . . . . . . . . . . 94.3 Two Derived Models from the General Framework . . . . . . . . 12

4.3.1 A Simplified Architecture . . . . . . . . . . . . . . . . . 124.3.2 A more detailed concept using maps and attraction points . 14

5 Conclusion 14

v

List of Figures

1 Concept map of actual mobility models . . . . . . . . . . . . . . 32 Proposed concept map of mobility model generation for inter-vehicle

communications . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 Graphical representation of the switches between the Random Walk

and the Random Waypoint according to pp and pt . . . . . . . . . 13

vi

1 Introduction

In the past few years, we have seen the emergence of technologies providingnetwork connectivity to mobile users. These technologies are based on a back-bone of access points, which mobile devices can connect to. Examples of suchsystems are the cellular network or WiFi networks. Yet, a growing demand on in-creased bandwidth and improved communication quality made engineers chooseto decrease the transmission range of mobile remote devices. Consequently, thebackbone had to be re-designed with an increased number of access-points. Forexample, in order to obtain the advertised throughput out of a WiFi network, anaccess point basically needs to be set every 100m, and the third generation of cel-lular network needs as much as twice the amount of access points (base stations) asto the second generation of cellular networks. Therefore, the infrastructure-basedapproach is not always most effective and is naturally supplemented by direct com-munication between terminals, also called ad-hoc communication.

One emerging new type of ad hoc networks is vehicular ad hoc networks(VANETs), in which vehicles constitute the mobile nodes of the network. Enhance-ments in transportation technologies have to consider, besides traditional aspectssuch as security and driving conditions, the ability of vehicles to communicate.It also covers the inter-networking of vehicles to the Internet. Connecting vehi-cles to the Internet provides users with the possibility to have an access to webservices. However, offering this capability in an efficient way requires resolvingseveral technical challenges going from gateways optimal placement on roadsidesto the handover management between gateways. On the other hand, several appli-cations may be provided for inter-vehicle communications. Indeed, vehicles mayexchange multimedia information about traffic, people inside vehicle are able toplay distributed games, etc. In order to evaluate the performance of communica-tions protocols designed for inter-vehicle communications, one has to use simula-tion tools like ns-2 since it is very hard to set up a real testbed of vehicles especiallywhen one of the performance parameters to be taken into account is the scalability.

Besides, an increased motivation for the development of vehicular ad hoc rout-ing protocols comes from particular vehicular capabilities such as the availabilityof GPS/GALILEO positioning and motion speed. These features actually occultcertain traditional concerns with mobile nodes, like power efficiency. Vehicularmotion patterns are also different from simply mobile motion pattern. Vehicles areusually constraints to follow predefined routes, directions and velocities.

More generally, in simulating mobile systems, it is important to use mobilitymodels that reflect as close as possible the reality of their behavior in real-world.Indeed, the mobility model is one important part when trying to evaluate the per-formance of communications protocols for mobile ad hoc networks. It describeshow nodes move during the simulation by specifying at each time the velocity andthe trajectory of each node in the network. Mobility pattern affect the performanceresults of routing, transport and MAC protocols when using network simulators.

It is quite obvious that each mobile environment could have a different pattern

1

depending on the mobile objects (nodes) and their freedom to move. Hence, auto-mobiles and pedestrian have different mobility patterns. Moreover, the ideal wouldbe to have a real mobility model obtained after the analysis of a large measure-ment campaign. However, this kind of model does not exist in the open literature.Therefore, researchers often use random mobility models, and have to adapt themto specific environments such as vehicular ad hoc networks.

In this paper, we propose a novel concept map for mobility models and use itin a survey of existing models that are or could be adapted to vehicular motion.We then identify basic properties that can be found in mobility models, and listmissing features that should be considered for vehicular mobility models. Basedon this, we adapt our concept map to vehicular motions and define a frameworkthat vehicular mobility models should follow in order to correctly simulate reallife vehicular motion. Finally, we propose two enhanced mobility models that arecompliant with our framework.

The rest of this paper is organized as follows. In Section 2, we present a newconcept map and survey existing mobility models, while in Section 3, we iden-tify necessary components for a vehicular mobility model. Section 4 describes avehicular-adapted concept map, then proposes a framework for a proper vehicularmobility model, and finally exposes two possible compliant mobility models. Last,in Section 5, we draw some concluding remarks on key issues on vehicular motion,we propose an acceptable solution at this state of progress, and we give suggestionsfor the implementation of realistic mobility models for inter-vehicular networks1.

2 Improved Concept Map and Short Survey of ExistingMobility Models

In this section, we first introduce an original concept map of mobility models,that besides being simple and easy to understand, is also able to categorize mostrecent mobility models. It should therefore be the basis for future developmentof more realistic mobility models. Then, using the criteria of our concept map,we briefly survey some of the common mobility models that can be found in theliterature.

2.1 Bi-polar Concept Map for Mobility Models

Actual mobility models usually fall in some of those categories.

• bounded vs. non-bounded domain

• constrained vs. random motion liberty

• convex vs. non convex domain1The terms inter-vehicular networks and vehicular ad hoc networks will be used interchangeably

throughout the rest of the paper.

2

• social vs. selfish behavior

• random vs. pattern vs. trace-based motion behavior

Yet, two major denominations can be straightforwardly observed in all recentmodels. The categories composing mobility models are distributed between thosetwo components as depicted in Figure 1. Although being a simplified model com-pared to the model proposed in [4], our proposed concept map finds its originalityin the bi-polarity between the Domain model and the Node model.

Mobility Model Domain Model

Bounded Non

bounded

Constrained Free Space

Social

Selfish

Node Model

Pattern Motion

Erratic Motion

Wrapping Reflection

Non convex

Convex

Trace Motion

I n d

i v i d

u a

l b e

h a

v i o

r

Border behavior

1 1 1 n

Motion freedom

Motion randomness

G e

o m

e t r

i c a

l p

r o p

e r t

y

Figure 1: Concept map of actual mobility models

Bettstetter described in [4] a concept map where the two major components arethe Randomness Degree and the Level of Detail. At the time his paper was written,a strong concern was set on macroscopic randomness. Nodes were randomly react-ing to their environments, in other words nodes were enduring their environmentand had no proactive actions.

In real life, this is not generally the case, and in recent years, an increasingnumber of research papers have tried to model nodes mutual dependencies and non-random behaviors. A clearer picture can be drawn showing that mobility modelsare actually composed of two different components: a Domain model and a Nodemodel. The Domain model describes the simulation domain along with its motionconstraints, while the Node model depicts nodes motion patterns. Our originalapproach is quite logical in a sense that when developing a mobility model, one firstmodels the domain, then nodes evolvement within that domain. This is particularlytrue when we want to model different classes of nodes.

While most of the contributions on mobility models have been performed onthe Domain model, the Node one seems to have drawn an increasing attentiononly in recent years. Indeed, a growing interest is carried out on Node motionmodeling, both in macroscopic and microscopic point of views. In the former

3

case, we can observe real life pattern motions, such as preferred paths or socialmotion. In the latter case, a simple analysis of vehicular motions allows us to seethat nodes individual behavior is correlated not only to their simulation domain,but also to other neighboring nodes. The approach taken by Bettestetter in [4] isto consider neighboring nodes as being part of the simulation domain. We thinkthat this should not be the case and motion correlation due to other nodes shouldbe described separately. This approach is therefore able to clarify mobility modelsin order to better shape them to real behavior.

2.2 Survey of Existing Mobility Models with Regards to the ProposedConcept Map

Several surveys of models for mobility of nodes in a network has been pre-sented in the past, including those from Camp et al. [7] and Bettstetter [3], Helmyet al. [6], and Le Boudec et al. [8]. We briefly summarize below some of the com-mon mobility models.

The Random Walk with Reflection mobility model is a paradigm, where thedomain model is bounded and non-constrained, while the node model is selfishand erratic. This mobility model was developed to mimic irregular movement innature. A mobile node (MN) moves for a specified time from its current location toa new location by randomly choosing a direction and speed from particular speedand direction distributions. When this time ends, the whole process is repeated allover again. If it reaches a simulation boundary, it bounces off the simulation borderwith an angle determined by the incoming direction. The MN then continues alongthis new path.

The Random Walk with Wrapping mobility model is similar to the standardRandom Walk mobility model, with the difference that it is not bounded. Thiscan be seen as a Random Walk on a torus simulation area. When mobile nodeshit the boundaries they are wrapped to the other side of the simulation area fromwhere they continue their trip. A slightly different mobility model is the RandomDistance mobility model, where nodes move until they reach a randomly chosendistance from the simulation boundary. This mobility model is therefore boundedby nature.

The most commonly used mobility model in the mobile ad hoc wireless re-search community is the Random Waypoint model [9]. Its domain model is usuallyconvex and bounded, while its node model is erratic, although several declinationshave appeared in recent years [8]. In this model, each node individually choosesa random destination within the simulated network boundary, and also determinesa motion speed randomly chosen between a minimum and maximum limit. Basedon this, it moves toward its destination at its determined velocity. Once the destina-tion has been reached, each node stops for a randomly chosen time interval. Afterthat pause time, it then repeats the process by choosing another random destina-tion and a random speed. The characteristics and properties of this mobility modelhave recently been studied in detail in [1, 2]. Bettstetter [3] modified this model so

4

that instead of suddenly choosing a new random speed when a node has reached adestination, the node accelerates to reach a targeted speed and, before reaching thedestination, it decelerates to come to a halt. Many declinations of this model maybe found and are described in [8]. We find the Random Waypoint in non-convexsimulation areas. In this mobility model, the straight-line trips encounter obstaclesthat trigger detours. Hence, nodes have to find alternate paths yet with the objectiveto reach the initial destination. Note that constrained mobility models (like the citysection) are special cases of such models. Another particular case of such model isthe Random Waypoint on a sphere. This mobility model may be used to representmobility of airplanes or satellites around the globe.

Of all models described so far neither represent real motions. Realistic motionis a framework containing the Social motion subset from the Node model andthe Constrained Motion subset from the Domain model. A model is consideredrealistic if it is compliant with at least one set contained in this framework. To thebest of our knowledge, only one model, the Weighted Waypoint mobility model [16]includes both sets.

An example of Social motions is the Reference Group mobility model [10].Nodes are gathered in groups in which mobile motions are not independent butgoverned by the motion of a reference point for the group. This reference pointcould be a leader, or simply a guideline. Nodes within those groups experiencesome degree of liberty but have to follow the group. This is very similar to ashepherd or a sheepdog guiding a herd of sheep. Each sheep may have its degree ofliberty but is constrained to move with the herd. Another social mobility model isthe Social Networks Based mobility model [12]. The authors noted that movementsare strongly affected by the needs of humans to socialize in one form or another. Inparticular, this model allows collections of hosts to be grouped together in a waythat is based on social relationships among the individuals.

Constrained motions use a different approach. Mobile nodes have limitedchoices for their destination, or for the path to their destination. For example,Jardosh et al. [11] proposed a Space Graph mobility model where the simulationarea is composed of graph vertices, and where nodes are constrained to followedges connecting these vertices. Consequently, a mobile node starts by randomlychoosing an initial vertex and moves along the shortest path to another randomlychosen vertex. The authors also introduced obstacles into the scenario such thatboth nodes mobility and wireless transmissions are restricted.

The last kind of Constrained mobility models we list here are the Manhattanand the City Section [13] mobility models. While the former models a down-town urban area with horizontal and vertical streets, the City Section relaxes thishorizontal and vertical shape to model all kind of roads and streets. Saha andJohnson [14] even proposed to use real map extracted from the US Census Bu-reau which records detailed street maps from the entire United States, based onthe Bureau’s TIGER (Topological Integrated Geographic Encoding and Referenc-ing) database [15]. Actually, to the best of our knowledge, this is the first freelyavailable work analyzing the characteristics of a realistic street mobility model.

5

Yet, mobility models that restrict nodes to move on a bounded area, or do nottake precautions for the initial distribution of nodes positions, have the character-istic that, over time, nodes tend to congregate toward the center of the simulationarea. Nodes also can encounter subtle problems such as decay of average speed asthe simulation progresses, or a difference between the long term distribution andthe initial one, or even model instability. Leboudec et al. [8] proposed a method toobtain time stationarity in the distribution of nodes locations and speeds for a largerange of mobility models. They adapted their technique, called Perfect Simulation,to almost all previously described mobility models in order to avoid the describedside effects that usually provide unsound basis for simulation studies.

We also have to mention that Lu et al. [5] proposed to create hybrid mobilitymodels by mixing the Random Waypoint and the Manhattan model, for instance.The authors also defined contractions and expansions, which are particular pointsthat attract or repulse mobile nodes. These proposed models cover scenarios inwhich nodes merge, scatter, or switch to different movement patterns over time.The interesting point here is that, although the domain model may be constrainedor not (by using Manhattan or the Random Waypoint), the node model is alsodefined by motion patterns.

Finally, although all these proposals have tried to create realistic mobility mod-els, they are all based on random mobility. Recently, some teams became interestedin non-random patterns that can be experienced in real life. Among them, Hsu etal [16] proposed a Weighted Waypoint mobility model which captures preferencesin choices of destinations of pedestrian mobility patterns in a campus environment.The authors estimated the parameters of this model using mobility survey data fromthe campus of the University of Southern California. They also proposed in [18] amethod to gather real traces of movements of people in this campus. Their methodcan help capture group statistics in more details, or extract actual users’ directionpatterns over time and place. Yet, this method is difficult to scale and a largeamount of data is necessary to obtain satisfactory results. Nevertheless, their ap-proach is promising since their Weighted Waypoint’s domain model is constrainedby the campus structure, while the node model is based on real motion traces, andindividual behaviors are described by social motions. This method is therefore thefirst mobility model that fits the closest to real pedestrian mobility.

3 Identification of Vehicular Mobility Models Components

Now that we have listed most known and used mobility models and shown aneasier concept map for mobility models, let us discuss here particular requirementsmobility models need to manage in order to accurately describe vehicular motions.We shall also see in the following section that, by its generality, the concept mapwe proposed in Section 2 readily adapts itself to vehicular mobility models.

Vehicular mobility models have to be considered in both microscopic andmacroscopic approaches [17]. When focusing on a macroscopic point of view,

6

we consider motion constraints such as roads, streets, crossroads, and traffic lights.We also assess traffic generation such as traffic density, traffic flows, and initialvehicles distributions.

In contrast, in the microscopic approach, the movement of each individual ve-hicle and its behavior with respect to other vehicles is determined. For example, theinter-distance between vehicles, the driver’s potential aggressiveness toward othercars, and its ability to follow speed limits or speed decays on a change of direction.Yet this micro-macro approach has to be generalized. Indeed, this particularityalso exists for obstacle. For instance, macro-obstacles are buildings and non-LOSwireless communications, while micro-obstacle are car shadowing, traffic-jam, andhidden nodes for wireless transmission.

However, this micro-macro approach is more a way to analyze a mobilitymodel. It is not what composes a mobility model. As a matter of fact, a vehic-ular mobility model is constituted of two blocks: Motion Constraints and TrafficGenerator. Besides, both blocks may be analyzed macroscopically and micro-scopically, as we will see later. The Motion Constraints part describes how eachvehicle moves (its respective degree of freedom) and is usually obtained from atopological map. The more precise it is, the more accurate and realistic vehicularmotions are modeled. Macroscopically, motion constraints are streets or building,but microscopically, constraints are modeled by neighboring cars, pedestrians, orby limited roads’ diversities either due to cars’ type or drivers’ habits. The Traf-fic Generator, on the other hand, generates different kind of cars (SUVs, Com-pact Cars), and deals with interactions between cars, traffic regulation, and trafficsign considerations. Again, macroscopically, it models traffic densities or trafficflows, while microscopically, the traffic generator deals with properties like inter-distances between cars.

All actual freely available mobility models only deal with fully random micro-mobility, which we find unrealistic. For example, recent mobility models simulat-ing vehicular motions (like the City Section) use maps in order to remove macro-scopic randomness, yet they completely ignore the micro-mobility. To our opinion,this is a particular field we need to focus on. In fact, solutions for microscopic ve-hicular motion exist but are usually for commercial use only, especially vehiclemanufacturers that make use of it to determine the lifetime of certain part of acar. One well known and well validated is the Daimler Chrysler internal driverbehavior simulator tool called FARSI. This simulator models realistic microscopicassets such as speeds or distances, and macroscopic properties like traffic flow andlane usage. It is regularly employed to generate traffic simulations and productdevelopment. Unfortunately, this system is not available for public use at this time.However, some other models have been developed and simulated by Helbing andTreiber ([23, 25, 24]), which are freely available and could be used in order toimprove this micro-mobility.

All recent contributions to vehicular mobility models proposed to constrainvehicles mobility. Bettstetter [3] proposed to smoothly accelerate and deceleratebetween waypoints, Helmy [5] developed attraction, repulsion, or empirical mod-

7

eling [18], Musolesi [12] founded a mobility model on social network theory, andSaha [14] suggested to use real topological maps. In fact, a mobility model shouldinclude all these solutions. A realistic mobility model should therefore include

• True and accurate topological maps: They should include different cate-gories of streets as well as different velocities given the kind and configura-tion of those streets.

• Smooth deceleration and acceleration: Each change of direction or veloc-ity should be smooth.

• Obstacles: They should be included both for mobility constraints and forwireless communication (non line of sight, shadowing)

• Attraction points: Vehicles do not randomly use roads. Some are preferred,depending on the driver’s habit. Cars usually take a beltway instead of goingthrough a city’s downtown.

• Simulation time: It should not be freely adjustable. Simulations should beperformed on particular driving patterns

– Morning and evening rush hours;

– Lunch Break: cars leaving the office for a restaurant and vice and versa;

– Night life;

All these different patterns do not take place at the same time. Thus it wouldbe wrong to include them all in a single simulation. Instead, we either sim-ulate each one (or a part of it) or simulate a typical day by simulating eachone, one after the other.

• Non-random distribution of vehicles: Cars do not appear by magic on astreet. They usually stay at home and the owners use them to go to the office.Accordingly, the initial distribution of cars should be done between homes,offices, or shopping malls; in other words: center of interests.

• If a topological map deals with macro-mobility, we definitely need to have atraffic generator that control vehicles mutual interactions such as overtak-ing, traffic jam, preferred path.

Moreover, it might sound strange to be interested in mobility predictions for mo-bility models. Usually, mobility predictions are extracted from mobility modelsin order to obtain non-random motion patterns that would improve routing strate-gies. Indeed, mobility prediction does not seem of a particular use for macroscopicmobility. Yet, when we look deeper in the microscopic case, we see that mobilityprediction is indeed of a particular interest, and might even be a key asset to createclever driving behaviors. Positions, velocity information, as well as predictions

8

can be used for example to regulate the inter-distance between vehicles. Consider-ing real drivers, we see that we do mobility predictions (or estimations since we donot have precise information of other vehicles) every time we drive. For example,we estimate the relative distance between our car, the next car and the car com-ing on the opposite lane in order to decide upon an overtaking attempt. Similarly,we make predictions when engaging on a highway, or how soon we need to slowdown in order not to bump the car in front of us. All this has to do with estimations,which can be extended to predictions when cars communicate with each other’s.This is what makes people considering themselves as clever drivers, and this isparticularly what differentiates experienced drivers from rookies.

4 Framework for Vehicular Mobility Model Generation

In this section, we first propose a concept map for mobility models adaptedto vehicular motion. Then, we describe a framework for the generation of realis-tic vehicular mobility models. Finally, we present two derived models from ourframework.

4.1 Vehicles-Adapted Concept Map

Figure 2 illustrates the main needed components for mobility model genera-tion: Motion Constraints, Time Patterns, and Traffic Generator. The MotionConstraints set includes all components needed to describe the simulation domainin which vehicles are moving. This set may be compared to the Domain modelin Figure 1, but adapted to mobility constraints vehicles may experience while inmotion. Usually, it is composed of a precise domain map extracted from a topo-logical map and enhanced by obstacles, or attraction points, to name only a few.Similarly, the Traffic Generator set is a more specific case of the Node Motionmodel in Figure 1 also adapted to deal with erratic cars and drivers’ particularities.Finally, as the Mobility Constraints set describes determinism in space domain ofa vehicular mobility model, the Time Patterns set represents determinism in timedomain. For example, an initial point will be considered as an attraction point ora repulsion point depending on the time pattern. Time patterns describe differentconfigurations of a day or a week, when a particular motion pattern may be ob-served. The Time Patterns component also exists in the concept map of regularmobility models. However, for the sake of clarity, we did not include it on Figure 1since to best of our knowledge, no actual mobility model uses it. Therefore, we cansee that our proposed concept map for vehicular mobility model is a specializationof the concept map proposed in Section 2.

4.2 Framework General Description

As we mentioned in Section 3 and also could see in Figure 2, the two mostimportant aspects of a mobility model are a Topological Map and a Traffic Gen-

9

1

n n

Mobility Model

Motion Constraints

Time Patterns

Traffic Generator

Topological Maps Obstacles

Attraction/ Repulsion

Points

Speed Constraints

Car Generation Engine

Driver Behavior Engine

Car’s type and particularities

Mobility Predictions

Social Habits

Driver Danger Assessments

Alter

1

1

1

Describe

Improve

Compose

Centers of Interest

Determine

initial location

Describe cars

capabilities

Determine

preferred motion

Describe mutual behaviors

Compose

Figure 2: Proposed concept map of mobility model generation for inter-vehiclecommunications

erator. Let us first consider how to obtain accurate Topological Map. The mainissue here is to be able to digitalize true maps in order to obtain an input for con-strained traffic. To our knowledge, the only freely available solution at this time isthe TIGER database from the US Census Bureau that unfortunately only containsmaps of US cities. Yet, it is possible to use this as a starting point and work on newsolutions for worldwide cities. Such maps should take into consideration:

• Street heterogeneity: Streets, roads, path, and alleys are several categoriesof streets that have to be modeled by these maps. Moreover, several lanesshould be included, as well as bi-directional and one-way roads.

• Street capacity heterogeneity: Each kind of street or road should be ableto accept only particular classes of vehicles. For instance, an alley cannot bea preferred path for a truck, and mopeds are allowed to enter highways.

• Speed heterogeneity: Each road should also have a particular speed limi-tation, due to its configuration and its class, or due to the possible drivingpitfall composing it. For example, one cannot drive at 100km/h in an alley.

• Radio obstacle: A mobility model should include the blocking of signaltransmissions by objects such as high-rise buildings in the city. In otherwords, for a pair of vehicles to communicate directly, they must have a lineof sight to each other, in addition to being in range of one another.

At this time, to the best of our knowledge, no actual mobility model includes allthese features in its configuration. The M-Grid [19, 20] mobility model includessome of them, but lacks by its rigorous squared modeling of streets, and true topo-logical maps. The City Section [14], on another hand, solves this problem but does

10

not consider the rest of these requirements. Therefore, by grouping the main fea-tures of both approaches, we could obtain a good starting point for developing arealistic vehicular mobility model.

Traffic Generator may be an easier task to perform. Indeed, several modelshas been developed for microscopic traffic simulation such as the Driver BehaviorModel [21], the Optimal-Velocity Model [22], the Intelligent Driver Model [23],the Intelligent Driver Model with Memory [25], or the Human Driver Model [24]to name only a few. the Driver Behavior Model does not only take the charac-teristics of cars into account but it also includes a model of the driver’s behavior,for instance, lane changing and passing decisions, traffic regulation and traffic signconsiderations, or decreasing speed in curves, to name only a few. A traffic gener-ator should take into account

• Cars characteristics heterogeneity.

• Lane changing and passing decisions including inaccuracies and anticipa-tions using mobility predictions.

• Finite reaction time including memory and frustration effects due to con-gested traffic.

• Enroute diversion behavior including familiarity with potential alternateroutes, social habits, and drivers individual danger assessments.

Then, the FARSI simulator tool for macroscopic modeling might be freelyavailable in the future, and combining both approaches would be a key asset for re-alistic simulators. Meanwhile, we could have a look at the FASTCARS project [26].The authors discussed the development and implementation of FASTCARS (Free-way and Arterial Street Traffic Conflict Arousal and Resolution Simulator), aninteractive microcomputer-based animated simulator designed for in-laboratoryexperimentation and data collection to assist in the estimation and calibration ofpredictive models of driver behavior under the influence of real-time information.Although being an old study, the authors concluded that several factors impact en-route diversion behavior including:

• familiarity with potential alternate routes;

• visual information gathered from variable message signs;

• changes in travel speed;

• drivers individual risk preferences;

We can see that their conclusions are still very contemporary.Finally, an interesting study [27] has been performed in the framework of the

FleetNet [28] project that focused on the impact of vehicular dynamics on proto-cols for ad-hoc networks. The authors used traffic theory in order to model average

11

vehicle distances, average velocities and other traffic parameters. Among theircontributions, they investigated possible communication durations between vehi-cles. This is particularly relevant to a mobility model if we want to allow vehiclesto share information to improve microscopic traffic modeling. They also analyzedthe dynamics of topology changes that are caused by vehicles joining or leavingcommunication. Their conclusions should therefore be considered when develop-ing a microscopic mobility model.

4.3 Two Derived Models from the General Framework

4.3.1 A Simplified Architecture

If we do not have access to traffic generators or topological maps, we proposehere a simplified architecture derived from a basic Stationary Random WaypointModel proposed by Le Boudec et al. [8] that includes perfect sampling, i.e. theinitial mobility state can be set in steady state at the beginning of the simulation,so that annoying transient artifacts are avoided. Yet, we enhance it to better fit tovehicular traffic.

We initiate the model with time stationary distributions of locations within thesimulation domain as proposed in Section V I.B of [8]. We also want to includesmooth transitions between speed changes. Therefore, similarly to [3], we de-fine a node’s targeted speed as V target

node. A targeted speed is uniformly chosen in

[Vmin, Vmax]. Then, the node samples an acceleration from an uniform distribu-tion between [0, αmax]. Then, each period of time ∆t, the node’s velocity increasesaccording to

v(t) = v(t) + α(t) · ∆t

until the node reaches V targnode

. Then, the acceleration is set to 0 and the node moveswith constant speed until the next speed change. A speed changes occur when ap-proaching to a waypoint2 . Consequently, it follows the same procedure to smoothlydecelerate before reaching the waypoint and pausing, yet using an acceleration uni-formly chosen in [αmin, 0].

Our contribution here is that we also assume that at each waypoint, a node re-mains in the same trajectory with a probability 1 − pt (and changes its trajectorywith a probability pt), and stop with probability pp. Each waypoint may thereforebe seen as a particular point, before which a node decelerates and changes its tar-geted speed. It can stop with a probability pp, and change its direction due to avirtual crossroad with another probability pt. This simulates vehicles’ behaviorswhen confronted to traffic signs and crossroads.

A graphical representation of the proposed mobility model is depicted in Fig-ure 3. From a microscopic point of view, when approaching a waypoint, a nodeonly changes its targeted speed with probability 1 − pp, or decelerates and stopswith a probability pp. In that case, it samples a pause time from a uniform den-sity fpause, on the expiration of which it accelerates again to reach a new target

2A waypoint is considered here as any point where either speed or direction may be altered

12

speed. This represents vehicles behavior when reaching a crossroad, a traffic light,or simply entering a street section where speed limitations change.

Yet, as it can be seen in the same Figure, from a macroscopic point of view,when a node keeps its trajectory (with a probability 1−pt), the model switches to aRandom Walk with Wrapping mobility model. Accordingly, rather than samplinga new destination, it keeps the same direction and samples a new targeted speedand a residual trip duration from an exponential distribution of parameter λ. Whenthis time expires, we fall back again to the case where, first it chooses to pause withprobability pp, then with a probability 1 − pt it samples a new residual time withthe same direction, or change its trajectory with probability pt. If so, the modelswitches back to the Random Waypoint mobility model, the node samples a newdestination and starts heading to it. If a trajectory makes a node reaches the domainboundary, it is wrapped to the other side of the domain, where it continues its tripwith the same trajectory. The domain used by this hybrid model can be seen as aTorus. By doing so, we try to simulate infinite domains, which is typically whatcan be experienced by vehicles in real motions.

Stop Acceleration Deceleration

Random Waypoint

Motion

Random Walk Motion

On expiration

P t

1-p t

Approaching a waypoint

Approaching

a waypoint

p p

1-p p

Figure 3: Graphical representation of the switches between the Random Walk andthe Random Waypoint according to pp and pt

Therefore, the two macroscopic configuration probabilities pt and pp simulatevirtual crossroads, the microscopic exponential distribution parameter λ describesroad lengths, and the pause time density function fpause represents traffic relatedmobility disturbance such as traffic lights, traffic jams. Thanks to this, the modelneither relies on any topological maps nor on traffic generators, yet it is able tokeep a quite generic aspect. Furthermore, the moving area, which may have anyshape, also can have one, two or three dimensions. Our model might consequentlyalso be applied for aircrafts simulations or for simulating hybrid environments.

13

4.3.2 A more detailed concept using maps and attraction points

The first consideration we will make here is that we assume that a node remainsfollowing a constraints path from one repulsion point to an attraction point. At eachintersection, it decelerates, then accelerates afterwards similarly to what we wrotein Section 4.3.2. Yet, V targ

node is now uniformly chosen in [V streetmin , V street

max ], whereV street

max is the maximum speed allowed on a particular street section.Along the path, it follows the streets and traffic regulations allowed by the map.

However, drivers’ danger assessment determines how they react when confrontedto traffic regulation signs. For example, how far from the traffic light a driver startsdecelerating, or how close to the speed limitations it drives.

After having reached an attraction point, a driver then chooses a new destina-tion from a set of all attraction points and makes a pause time with a probabilitypp.

As mentioned before, our proposed model includes topological maps obtainedby the TIGER database from the US Census Bureau. Yet, we add all proposed fea-tures from the M-Grid model. Thanks to the TIGER database, our topological mapsstraightforward include Street Heterogeneity and Speed Heterogeneity. We thenadd attraction and repulsion points. For example, Residential areas are consideredas repulsion points in the morning but attraction points in the evening. Similarly,Shopping Malls, Work places, Restaurants are also considered as attraction pointsor repulsion points depending on simulation time patterns.

Therefore, contrarily to the City Section, destinations are not randomly chosen,and drivers do not specifically choose the shortest path. Driver’s habits are takeninto accounts. Similarly, different from the M-Grid model, paths are not restrictedto horizontal or vertical streets. As M-grid, we also include radio obstacles fornon-LOS communications.

Thanks to the key assets taken from the City Section, the M-grid model, as wellas some new concepts of our own, our model is compliant with our framework andfits with realistic vehicular motions.

5 Conclusion

In this paper we proposed a framework for a realistic mobility model for Inter-Vehicular Networks. We reviewed actual mobility models, proposed an originalconcept map, and identified key features that should be included in a vehicular mo-bility model in order to obtain realistic motions. Such model should be self-drivenfrom the moment we set the proper parameters. Randomness should be limitedto jitters, traffic regulation liberty, or proactive routing. But vehicles distribution,paths, and destinations should have nothing to do with randomness.

We also described a general random limited mobility model that is fully com-pliant with our framework. Nevertheless, some parts of it are none-trivial tasks andeither are not freely available or simply not feasible at this time. Consequently, welikewise presented a simplified fully random mobility model that is compliant with

14

parts of our framework but which does not implement all features we described inthis paper.

We finally proposed to use mobility predictions in order to obtain Clever Driv-ing features, precise Topological Maps for accurate motion constraints, and Pointsof Interests as a way to better fit with social motions that can be experienced inmetropolitan areas. Besides, such approaches were proposed for pedestrian mobil-ity. Consequently, why couldn’t that be the case also for the vehicular movement?

To conclude, in the case we have access to topological maps, a good startingpoint to describe basic inter-vehicular motions is to merge characteristics of theM-Grid and the City Section mobility models to include Topological patterns. Inthe negative case, we propose to use an Hybrid mobility model as described in Sec-tion 4.3.1. As mentioned in [14], the Random Waypoint mobility model exhibitssimilar characteristics as the City Section for initial results. And our contribution,in addition to that of [3], further improves it with the objective to better describevehicular motions.

At this point of research, we are currently implementing the mobility modelproposed in Section 4.3.1 and are considering doing so with the more detailedone from Section 4.3.2. Yet, we suggest to investigate other solutions proposed inthis paper in order to improve the realism of vehicles particular motion patterns,among which are Mobility Prediction, Point of Interests, Map Digitalization andNon-Random Traffic Distribution.

Acknowledgment

This research is part of the NewCom Project A, JPA3 assignment.

15

References

[1] Christian Bettstetter, Hannes Hartenstein, and Xavier Perez-Costa. ”Stochas-tic properties of the random waypoint mobility model: Epoch length, direc-tion distribution, and cell change rate”. In Proc. of the Fifth ACM Interna-tional Workshop on Modeling Analysis and Simulation of Wireless and Mo-bile Systems, pages 7-14, 2002.

[2] Christian Bettstetter and Christian Wagner. ”The spatial node distribution ofthe random waypoint mobility model”. In Proc. of the First German Work-shop on Mobile Ad Hoc Networks (WMAN), pages 41-58, March 2002.

[3] Christian Bettstetter. ”Smooth is better than sharp: A random mobility modelfor simulation of wireless networks”, In Proc. of the Fourth ACM Interna-tional Workshop on Modeling, Analysis and Simulation of Wireless and Mo-bile Systems, pages 19-27, 2001.

[4] Christian Bettstetter. ”Mobility Modeling in Wireless Networks: Categoriza-tion, Smooth Movement, and Border Effects”, In ACM Mobile Computingand Communications Review, vol. 5, no. 3, pp. 55-67, July 2001.

[5] Y. Lu, H. Lin, Y. Gu, A. Helmy, ”Toward Mobility-Rich Performance Analy-sis of Routing Protocols in Ad Hoc Networks: Using Contraction, Expansionand Hybrid Models”, in IEEE International Conference on Communications(ICC), June 2004.

[6] F. Bai, A. Helmy, ”A Survey of Mobility Modeling and Analysis in Wire-less Ad hoc Networks”, Book Chapter in the book on ”Wireless Ad Hoc andSensor Networks”, Kluwer Academic Publishers. June 2004.

[7] T. Camp, J. Boleng, and V. Davies. ”A survey of mobility models for adhoc network research”, In Wireless Comm. and Mobile Computing (WCMC),2(5):483–502, 2002.

[8] Jean-Yves Le Boudec and Milan Vojnovic, ”Perfect Simulation and Station-arity of a Class of Mobility Models”, to appear in Proc. of the Infocom’05,USA, 2005.

[9] Josh Broch, David A. Maltz, David B. Johnson, Yih-Chun Hu, and JorjetaG.Jetcheva. ” A Performance Comparison of Multi-Hop Wireless Ad HocNetwork Routing Protocols”, In Proc. of the Fourth Annual ACM/IEEE Inter-national Conference on Mobile Computing and Networking (MobiCom’98),pages 85-97. ACM/IEEE, October 1998.

[10] Xiaoyan Hong, Mario Gerla, Guangyu Pei, and Ching-Chuan Chiang, ”Agroup mobility model for ad hoc wireless networks”, In Proc. of the SecondACM International Workshop on Modeling, Analysis and Simulation of Wire-less and Mobile Systems, pages 53-60, 1999.

17

[11] Amit Jardosh, Elizabeth M. Belding-Royer, Kevin C. Almeroth, and Sub-hash Suri. ”Toward realistic mobility models for mobile ad hoc networks”,In Proc. of the Ninth Annual International Conference on Mobile Computingand Networking, pages 217-229, 2003.

[12] Mirco Musolesi, Stephen Hailes and Cecilia Mascolo, ”An Ad Hoc MobilityModel Founded on Social Network Theory”, In Proc. of the 7th ACM/IEEEInternational Symposium on Modeling, Analysis and Simulation of Wirelessand Mobile Systems (MSWiM 2004), Venezia, Italy. October 2004.

[13] V. Davies, ”Evaluating mobility models within an ad hoc network”. in Mas-ter’s thesis, Colorado School of Mines, 2000.

[14] Amit Kumar Saha and David B. Johnson, ”Modeling Mobility for VehicularAd Hoc Networks”. Appeared as a poster in the First ACM Workshop onVehicular Ad Hoc Networks (VANET 2004), Philadelphia, Pennsylvania, Oct2004.

[15] United States Census Bureau, ”Topologically Integrated Ge-ographic Encoding and Referencing (TIGER) Database”,http://www.census.gov/geo/www/tiger/tiger2003/tgr2003.html

[16] W. Hsu, K. Merchant, H. Shu, C. Hsu, A. Helmy, ”Weighted Waypoint Mo-bility Model and its Impact on Ad Hoc Networks”, Appeared as a poster toMobiCom 2004,Philadelphia, September 2004.

[17] D. Helbing, ”Traffic and related self-driven many particle systems”, In Rev.Modern Physics, Vol. 73, pp. 1067-1141, 2001.

[18] D. Batacharjee, A. Rao, C. Shah, M. Shah, ”Empirical Modeling of Campus-wide Pedestrian Mobility: Observation on the USC Campus”, in Proc. IEEEVehicular Technology Conference (VTC), September 2004.

[19] K. J. Wong, B. S. Lee, B. C. Seet, G. Liu, and L. Zhu, ”BUSNet: Modeland Usage of Regular Traffic Patterns in Mobile Ad Hoc Networks for Inter-Vehicular Communications”, in Proc. ICT 2003, Thailand, April, 2003.

[20] B. S. Lee, K. J. Wong, B. C. Seet, L. Zhu, and G. Liu, ”Performance of MobileAd Hoc Network in Constrained Mobility Pattern”, in Proc. InternationalConference on Wireless Networks (ICWN’03), Las Vegas, USA, Jun. 2003.

[21] T. Benz, L. Schafers, C. Stiller, and D. Vollmer, ”Feasibility study ontruck planning on european motorways”, Deliverable D08.1 of ITS projectPROMOTE-CHAUFFEUR, 1999.

[22] M. Bando, K. Hasebe, K. Nakayama, A. Shibato, and Y. Sugiyama, ”Dy-namical model for traffic congestion and numerical simulation”, In PhysicalReview, E 51, pp. 1035-1042, 1995.

18

[23] M. Treiber, A. Hennecke, and D. Helbing, ”Congested traffic states in empir-ical observations and microscopic simulations” In Physical Review, E 62, pp.1805-1824, 2000.

[24] M. Treiber, A. Kesting, and D. Helbing, ”Delays, inaccuracies, and anticipa-tion in microscopic traffic models”, submitted to Physical Review E, 2005.

[25] M. Treiber and D. Helbing, ”Memory effects in microscopic traffic modelsand wide scattering in flow-density data”, In Physical Review E 68, 2003.

[26] J.L. Adler, W.W. Recker, and M.G. McNally, ”Using Interactive Simulation toModel driver Behavior Under ATIS”, Working Paper University of California,Berkley, 1992.

[27] M. Rudack, M. Meincke, and M. Lott, ”On the Dynamics of Ad Hoc Net-works for Inter Vehicle Communications (IVC)”, in the 2002 InternationalConference on Wireless Networks ICWN 2002, Las Vegas, USA, June 2002.

[28] FleetNet Project, http://www.et2.tu-harburg.de/fleetnet/index.html.

19