Towards distributed geolocation for large scaledisaster management

Florent Coriat∗†, Luciana Arantes∗†‡, Olivier Marin∗†‡, Anne Fladenmuller∗†, Nicolas Hidalgo§ and Erika Rosas§∗ Sorbonne Universites, UPMC Univ Paris 06, UMR 7606, LIP6, F-75005, Paris, France

E-mail: [florent.coriat, anne.fladenmuller, luciana.arantes, olivier.marin]@lip6.fr† CNRS, UMR 7606, LIP6, F-75005, Paris, France

‡ Inria, REGAL project-team, Paris-Rocquencourt, France§ DIINF, University of Santiago, Chile

E-mail: [nicolas.hidalgo, erika.rosas]@usach.cl

Abstract—In the aftermath of major disasters such as earth-quakes, locating individuals is crucial for passing on vitalinformation, for example warnings and safety announcements.However, large scale disasters cause extensive damage to thenetwork infrastructures and a generalized loss of communicationsin the chaos that ensues. This position paper presents a prelim-inary study for a geolocation service that relies on inter-deviceconnections: mobile devices exchange positions of previouslyencountered devices when they come into contact. Every devicethus builds a partial map of device locations and can use it toenforce geographic routing protocols that are resilient to largescale disasters.

Index Terms—disaster management, geolocation, mobile de-vices

I. INTRODUCTION

When dealing with large scale disasters, real-time geoloca-tion to identify danger zones and to track/contact survivorsand rescuers constitutes a major challenge. Indeed, disasteroccurrences generally lead to severe communication failuresdue both to physical damage incurred by the network infras-tructures and to overloads caused by increased traffic: officialsources striving to propagate emergency announcements whileindividuals try to communicate with their relatives to exchangenews about their safety and well-being.

Continuous geolocation in this context offers many benefits.Persons who are unaware of the danger can be notifiedand oriented safely towards rescue centers. Competent andable-bodied rescuers can be teamed up and directed towardshelpless victims. Other survivors that made it to safety can betracked to simplify relief efforts. Obviously, all of the aboverequire sustained network routing.

We envision a best effort approach to allow continuedrouting of data: it builds upon weakly connected networkson top of user devices to share information about devicelocations. User mobility thus contributes to the generationand the propagation of information. Upon every temporaryconnection, a mobile device sends its current location alongwith a list of the locations received from other devices duringprevious encounters. In this manner, every mobile devicebuilds a partial map of the disaster area. Our approach assumesthe existence of rescue centers where servers can aggregate

partial maps and correlate them to build a global map, whichin turn facilitates routing.

The present paper constitutes a preliminary study towardssuch an approach. Section II describes the model we designedto simulate the movement of mobile devices in the midst oflarge scale disasters. Section III presents the evaluation metricsand a simplistic geolocation protocol we intend to use in orderto assess our own geolocation scheme. Section IV comparesour work with other related research. Section V concludes anddetails the remaining steps towards the completion of the workproposed in this position paper.

II. MODELLING A LARGE SCALE DISASTER

In the context of large scale disasters, our ultimate goalis to build a solution that facilitates the navigation of deviceusers towards their destination and that enables the routingof messages from one point to another in the absence oflegacy network infrastructures. To this end, we intend toconceive communication protocols among mobile devices:exchanges of positioning data lead to the generation of aglobal view of device locations, which in turn provides a basisfor geographic routing. Our first step consists in designinga model that allows assessments and comparisons betweensuch communication protocols. The second step, presented insection III, is to conduct these assessments and comparisonsby running simulations on top of this model.

A. Model assumptions

We consider mobile device users that can travel freely alongthe paths of a map towards their destination. We restrict ourstudy to pedestrians: device users move at different speedswithin the range between standing still and running. Users canbecome incapacitated and will remain motionless. Accidentsthat block pathways on the map may occur at any moment.

The behavior of device users corresponds to their level ofawareness with respect to the situation. We identified threesuch levels:

1) Oblivious. Oblivious users either have no awareness ofthe danger they are in or are clueless regarding thebehaviour they should adopt towards ensuring their own

safety. As an example, a global power outage combinedwith the leakage of a toxic gas can lead to the formercase: many individuals will remain oblivious until theyare warned of the crisis. The latter case could correspondto users who have heard of the leakage on the radio buthave not yet acquired information concerning the rescuecenters where they will be safe.

2) Alarmed. Users become alarmed when they receivewarning of the disaster and acquire at least one targetdestination to reach for safety. We consider two meansof notification: an alarmed user will notify an obliviousupon encounter; otherwise an oblivious user may catchnews of the danger and acquire safety information athome as the authorities start delivering such informationover the radio.

3) Safe. Users consider themselves safe once they havereached their designated target destination.

Mobile devices interact via Bluetooth links, with a range of10 meters. We chose Bluetooth over WiFi as it is far moreenergy-efficient, and we believe this is an essential criteriaduring disasters. Every device is associated with a uniqueidentifier, and incorporates both a physical clock and somepositioning feature such as GPS.

Geolocation servers, called ”hotspots”, are accessible atthe appointed rescue zones on the map. These locationsare fixed and are considered common knowledge: mobiledevices store these locations in advance, and orient the deviceowner towards a selected hotspot upon receiving a notificationof a disaster occurence. Every hotspot dedicates significantCPU and storage resources to the geolocation service. It cancommunicate with mobile devices via Bluetooth but cannotexchange data with other hotspots since we assume that legacynetwork infrastructures are down.

B. Emergency Movement Model

Given the above assumptions, we designed a movementmodel to represent the mobility of device users in the midstof a large scale crisis.

Our Emergency Movement model reflects the different be-haviours of a device user during a crisis. Figure 1 gives thestate diagram that regulates these behaviors. Users either startas oblivious or alarmed. Oblivious users can start inside oroutside their homes. Those outside walk randomly around themap for some time, and then head home. When inside theirhomes, oblivious users remain there until they receive warningof the danger. Warnings come from two possible sources:either an alarmed user passes by and notifies the oblivious userwhen they connect, or the oblivious user becomes self-warned(with a statistical probability that can be set). Alarmed usersknow of at least one safe target destination and try to reach it.Our model allows accidents that block pathways dynamically,so an alarmed user may become stranded if there are nomore accessible paths left towards its target. To ensure thatsimulations based on our model do end, we do not requireall users to reach the safe state. Once a set percentage ofusers are safe, all remaining users become stranded. Finally,

start

At homeWalk to

evacuation center

Safe at evacuation

center

walkProb

walkTime

receive a danger message

get toan evacuation

center

Walk randomly

selfwarnedProb

prewarnedProb

receive a danger message

Stranded(SOS)

no path to the evacuation

center

the percentageof nodes to save

(percentageToSave)has been reached

Fig. 1. Emergency movement model - state diagram.

upon reaching its target safety destination a safe user remainsstationary somewhere within the bounds of the safety area.

DangerMovement

SwitchableMovement

HomeMovement

ShortestPathMapBasedPoiMovement

EvacuationCenterMovement

RandomPathMapBasedMovement

MapBasedMovement

SosMovement

Figure 1 – Diagramme de classes des 5 sous modèles de DangerMovement

2.1.1 Sous-modèles de mouvements

Nous décrivons dans ce chapitre les di�érents sous-modèles de mouvements,ainsi que les changements de sous-modèles opérés par le modèle de mouvement dedanger. La figure 3 montre les interactions entre ces sous-modèles de mouvement.

Marche aléatoire

Le sous-modèle de mouvement de marche aléatoire est activé au début de lasimulation selon la probabilité walkProb. Une destination est tirée aléatoirementet le nœud se dirige vers celui-ci. Lorsqu’il atteint cette destination, il en choisitune autre aléatoirement et se rend vers celle-ci, jusqu’à ce que le temps de marchealéatoire walkTime soit écoulé. À la fin de sa marche aléatoire, le nœud élit domicileà la position qu’il atteint et change son modèle de mobilité vers celui représentantle mouvement au domicile. Si le nœud est alerté du danger par un autre nœudpendant sa marche aléatoire, alors il rejoint un centre d’évacuation et change sonmodèle de mobilité vers celui représentant la marche vers un centre d’évacuation.

Le sous-modèle de marche aléatoire est implémenté par la classe RandomPath-MapBasedPoiMovement.

Mouvement au domicile

Le sous-modèle de mouvement au domicile est décrit par une position fixedu nœud à son domicile. Il est activé au début de la simulation pour tous les

3

Emergency

At home

Walk to evac.center

Walk randomly

Stranded (SOS)

Safe at evac.center

Fig. 2. Emergency movement model - class diagram.

The Emergency Movement model combines a couple ofclassic movement models with a few of our own. Each ofthese models corresponds to a user state. Figure 2 representsthe class diagram associated to the state diagram of fig-ure 1. Oblivious users regain their homes with a random pathmap-based movement (RandomPathMapBased). At home,they levy-walk randomly within a very limited area (Home).Alarmed users try to reach their designated evacuation centerwith a variation of the shortest path map-based POI movement(ShortestPathMapBasedPoi): since the shortest pathmay be cut off at any moment, it is recomputed every timethe user stumbles upon a dead-end. Once they reach theirdesignated safety area, safe users pick a location at randomwithin the area, walk straight to it and remain stationaryfrom then on (EvacuationCenter). Finally, alarmed userswho are stranded for lack of any path towards their safetydestination stop moving altogether (SOS).

Figure 3 shows an example of our model applied to asimulation. This portion of a larger map contains only one

Fig. 3. Emergency movement model - Simulation example.

safety area; several pathways are inaccessible due to accidents.In this snapshot of the simulation, two device users (p100 andp43 – not very readable because they’re standing very closeto each other) have actually reached safety. p22 and p42 arestill unaware of the disaster; p22 awaits notification at homewhile p42 is walking around randomly. p3, p4, p61, and p99are alarmed and trying to reach the safety area. p30 and p84are stranded between two accidents that prevent any possibleescape.

III. ASSESSING GEOLOCATION SCHEMES

Along with our Emergency Movement model, we need a toolto implement and evaluate distributed protocols to geolocatedevices during large scale disasters. The ONE simulator [1]fits our requirements: it is specifically designed for evaluat-ing Delay-Tolerant Network (DTN) routing and applicationprotocols, it provides implementations for a wide range ofclassic movement models, and it can import real-world mapsas well as mobility data from real-world traces to producemore realistic simulations.

The question remains about how to determine the efficiencyand appropriateness of a geolocation scheme. The presentsection discusses metrics for this purpose, and assesses a naıveprotocol that we deem worthy of comparison.

A. Evaluation metrics

In our opinion, a geolocation scheme must aim towardspiecing together information that is both accurate and up todate. Another goal is to make this information as availableas possible to all users, whilst minimizing the global cost ofcommunications. With this in mind, we propose the followingmetrics to measure the performance of any given geolocationprotocol.• Exploration speed. This first metric computes the pro-

portion of devices that have exchanged locations with atleast one other device as time passes. It reflects efficiency

when locating individual devices. Ideally, a geolocationscheme ought to find as many devices as physicallypossible over a given timespan.

• Graph coverage. Our second metric calculates the pro-portion of links of acquaintance created since the startof the experiment. A new link of acquaintance ΛA→B

is spawned every time a device A acquires a locationfor a previously unknown device B. This metric verifieshow much knowledge of the global map is actuallydisseminated among the devices. Given a scenario withN devices, the total number of possible links of acquain-tance is N2 (by assumption, every device is constantlyaware of its own location). A good geolocation schemeshould tend towards a large graph coverage as fast aspossible.

• Positioning accuracy. The error associated with the linkof acquaintance from a device B to a device A equals theeuclidian distance between the stored location on B andthe real-world location of A. When A sends its location toB and then both devices move on in separate directions,the error associated with ΛB→A increases as A and Bproceed farther away from each other. Our third metriccomputes the average error for all links of acquaintance.Thus it represents how close the average partial view ofa device is to the real situation on the map: the smallerthe accuracy value, the better.

• Network load. Our final metric calculates the totalnumber of messages, as well as the total quantity ofdata exchanged amongst mobile devices in order to carryout the geolocation protocol. It measures the cost of theprotocol on top of the DTN: a protocol that generatestoo much overhead has a negative impact on energyconsumption and hinders the availability of the networkfor other applications.

B. Comparison with a naıve geolocation scheme

We have defined a first, simplistic geolocation protocolbased on the epidemic propagation of the estimated locationsamong mobile devices. Every device maintains a counter that itincrements upon every connection, and a table that associatesa counter value and a location with every link of acquaintance.When two devices connect, they exchange their locations, theircounter values and their tables. This allows every device tobring its table up to date: it inserts new links of acquaintanceacquired during the encounter into its table, and it updateslocations associated with a known device but with a freshercounter value.

Additionally, devices maintain a list of locations of un-expected roadblocks: every time a device comes across aroadblock it adds it to it list and will propagate the informationupon every ulterior connection. This reduces the time it takesfor device users to reach safety zones.

We implemented this protocol for simulation purposes onthe ONE, and ran experiments detailed hereafter. Our ex-perimental setup assumes that there are no bounds on thestorage capacity of mobile devices and on the amount of

information that can be exchanged between devices during asingle connection. We are aware that these assumptions are notrealistic, however they provide an experimental upper boundon the exploration speed and on the graph coverage.

Fig. 4. Experimental setup - Map of Santiago de Chile.

We used a map of a portion of Santiago de Chile (figure 4)for our experiments, set the real addresses of medical facilitiesas safety zones and ran simulations with 500, 1000, and 2000nodes initially distributed at random throughout the map. Weproduced our results by running each simulation 20 times andby computing the average of every output.

0

10 20

30 40

50

60 70

80 90

100

0 1000 2000 3000 4000 5000 6000 7000 8000

Disc

over

ed d

evic

es (%

)

Time (s)

Proportion of discovered devices (SantiagoCenter)

500 nodes1000 nodes2000 nodes

Fig. 5. Epidemic propagation - Exploration speed.

Figure 5 shows the exploration speeds obtained with ourepidemic propagation protocol for different densities of users.As expected, the larger the number of devices on the map, thefaster they explore the map and encounter isolated devices.An interesting result is that all three curves converge (> 90%)after an hour, which does not seem that long given thesignificant surface of the map (29km2). The remaining 10%of devices either take much longer to come across a firstdevice or fail to encounter any device at all, producing atail effect visible on all three curves. This small fraction of

devices prevents the map exploration from reaching 100%. Toensure that simulation runs will actually end, our experimentalsettings include a threshold for the number of devices thatmake it to safety. Above this threshold, all remaining devicesare set to stranded and the simulation stops.

0

5

10

15

20

25

30

35

40

0 2000 4000 6000 8000 10000

Link

s of

acq

uain

tanc

e (%

)

Time (s)

Graph coverage (SantiagoCenter)

500 nodes1000 nodes2000 nodes

Fig. 6. Epidemic propagation - Graph coverage.

Figure 6 illustrates the progression of the graph coverageobtained with our epidemic propagation protocol for differentdensities of users. Here also, the results confirm intuitiveexpectations: the higher the density of devices, the fasterthey propagate knowledge among themselves. A result thatis less trivial is that a higher density of devices guaranteesa higher upper bound for the coverage, regardless of thetime that passes. We imagine that there is a critical mass ofdevices beyond which this upper bound cannot increase, butconfirming this intuition will require further experimentations.

0

100

200

300

400

500

600

700

800

900

0 2000 4000 6000 8000 10000 12000 14000

Dist

ance

(m)

Time (s)

Positioning accuracy (SantiagoCenter)

500 nodes1000 nodes2000 nodes

Fig. 7. Epidemic propagation - Positioning accuracy.

Figure 7 provides the progression of the positioning accu-racies obtained with our epidemic propagation protocol fordifferent densities of users. Obviously, even though the resultsare better for higher densities, they are overall very bad.Devices do not discard links of acquaintance that they acquiredin the distant past. In a more realistic setup, our epidemicprotocol could prioritize the locations to exchange reverselyto their order of encounter. Indeed, if device A meets B andthen C, the location for C is more likely up to date than thelocation for B. Coupled to a simple LRU cache policy, webelieve it could improve these results and plan on integratingthis solution in our next implementation.

IV. RELATED WORK

Geographic routing protocols prevent the maintenance ofrouting tables on mobile nodes and thus appear as a goodrouting alternative for DTNs. These protocols rely on anaccurate location service to retrieve the exact position of thedestination. Quantifying the accuracy of the position estimatethus constitutes a first step to evaluate the performances ofrouting protocols in a crisis scenario.

Several geographic routing protocols have been proposedin the litterature. In these kinds of systems, location serversuphold the mapping between nodes and location: they updatethe map periodically and handle lookups from other nodes.Each node usually determines its own geographic locationwith some global positioning system like GPS. As we aimto tackle reliability and scalability issues, we are particularlyinterested in location services which are decentralized. Byexploiting node locations, effective and scalable global ge-ographic routing protocols can be implemented and presentbetter performance than other types of solutions: mobile adhoc networks (MANETs) topology-based protocols such asDSR, DSDV, AODV, intermittent connectivity network (DTN)protocols that often use store-carry-forward or contact-basedsolutions or vehicular ad-hoc network (VANET) protocols.

Distributed location service protocols for MANETs cansteer towards different approaches: a quorum approach ([2]and [3]), a greedy approach such as GPSR [4] or (GPCR) [5],a hierarchical-based location service such as GLS [6], HLS[7] or a hash-based approach like SLURP [8] or GHLS [9].

In a quorum-based location service [2] [3], each nodeupdates its position by sending the information to a subset(update quorum) of nodes. When a node wants to know thelocation of a target node, it requests its location from a subset(query quorum) of nodes. The two subsets must be designedso as to have a non empty intersection. Node location is, inaverage, disseminated to O(

√N) other nodes.

Greedy Perimeter Stateless Routing (GPSR) [4] is a rout-ing protocol for wireless networks that uses router positionsand message destination to make decisions about messageforwarding. GPSR makes greedy forward choices based onlyon information about the locations of nodes that are one hopaway from the routers in the network topology. Thus, in caseof topology changes due to node mobility, GPSR can use localtopology information to compute new routes quickly. Anotherexample of routing protocol that uses a greedy algorithmto forward messages is the Greedy Perimeter CoordinatorRouting (GPCR) [5]. However, both protocols can forwardmessages efficiently provided that the underlying network isfully connected.

In a hierarchical-based location service like GLS [6], thearea is recursively divided into a hierarchy of grids. For eachnode, one or more nodes in each grid are chosen as its locationserver. Therefore, the density of location servers for a node ishigh in areas close to the node and decreases exponentiallyas the distance to the node increases. Each mobile nodeperiodically updates its current location on a set of location

servers which are determined according to the geographic gridand the node identifier. GLS disseminates each node locationto O(log N) other nodes.

In a hashing-based approach (SLURP [8], GHLS [9]), aknown function maps each node identifier to a fixed homeregion composed of one or more nodes. The nodes of theregion store the location information about the node and,by hashing the node identifier, other nodes can obtain itshome region and therefore its location. Each node locationis disseminated to O(1) location servers.

All the above location service protocols are decentralized.However they target MANETs, so their assumptions are incon-sistent with large scale disasters: relatively small communica-tion latency, no network partitioning, and a mobility modelwhich does not characterize crisis/disaster scenarios.

The DTN-based Price [10] protocol proposes a hybridstrategy that firstly exploits a geographic approach wheremessages are propagated towards nodes along the directionof the destination. It switches to a contact-based approachwhen the message is close to the destination. Price predictsfuture contacts based on history of past contacts and assumesrepetitive human behaviors. In [11], the authors also argue thatthe short-term future locations of vehicles can be predictedand develop a geographic routing protocol based on thisassumption. GeOpps [12] is also a delay tolerant networkrouting algorithm that uses GPS information from vehicles.It exploits routes suggested by the GPS of every vehicle topiggyback message routing information on the vehicles. OtherVANET examples that use node positions are GPSR-DTN[13], which adds a store-carry-forward mechanism to GPRS,and GeoDTN+Nav [14], an extension of VANET Cross LinkCorrected Routing (VCLCR) [15] which improves VCLCRby exploiting the vehicular mobility and on-board vehicularnavigation systems. Notice that all these systems are eitherfor VANETs or do not consider any mobility model in acrisis/disaster scenario.

V. CONCLUSION

This position paper presents a preliminary study for adistributed geolocation service that is resilient to large scaledisasters. Mobile devices maintain local lists of the positionswhere they encountered roadblocks and other devices; uponevery connection, devices share their respective lists in orderto improve their partial views of the global map. As partof this undergoing work, we designed a mobility model thatis specific to large scale disasters, we formulated metrics toassess the performance of a distributed geolocation service,and we devised a very simplistic algorithm based on epidemicpropagation to test our model and to calibrate our futureexperimentations.

In the context of this project, our next step is to conceivea full-fledged protocol that: proves resilient, explores themap and discovers devices efficiently, and positions devicesaccurately without inducing excessive overhead. Our ideas forsuch a protocol include using the hotspots located in safe areasto centralize information and build a more accurate version of

the global map that can then be used for optimized geographicrouting. We shall evaluate these properties by implementingand running the protocol on our experimental testbed. Concur-rently, we intend to implement a revised version of GLS [6]for comparison with our own solution.

ACKNOWLEDGEMENT

This work is supported by the STIC-AmSud project RE-SPOND, Reputation and Energy-Aware Search for SupportingNatural Disasters.

REFERENCES

[1] A. Keranen, J. Ott, and T. Karkkainen, “The ONE Simulator for DTNProtocol Evaluation,” in Proc. of ICST SIMUTools, 2009.

[2] I. Stojmenovic and B. Vukojevic, “A routing strategy and quorum basedlocation update scheme for ad hoc wireless networks,” University ofOtawa, Tech. Rep. TR-99-10, 1999.

[3] Z. J. Haas and B. Liang, “Ad hoc mobility management with uniformquorum systems,” IEEE/ACM Transactions on Networking, vol. 7, no. 2,pp. 228–240, 1999.

[4] B. Karp and H. T. Kung, “Gpsr: Greedy perimeter stateless routingfor wireless networks,” in Proceedings of the 6th Annual InternationalConference on Mobile Computing and Networking, ser. MobiCom ’00,2000, pp. 243–254.

[5] C. Lochert, M. Mauve, H. Fusler, and H. Hartenstein, “Geographicrouting in city scenarios,” SIGMOBILE Mob. Comput. Commun. Rev.,vol. 9, no. 1, pp. 69–72, Jan. 2005.

[6] J. Li, J. Jannotti, D. S. J. De Couto, D. R. Karger, and R. Morris, “Ascalable location service for geographic ad hoc routing,” in Proceedingsof the 6th Annual International Conference on Mobile Computing andNetworking, ser. MobiCom ’00, 2000, pp. 120–130.

[7] W. Kies, H. Fusler, J. Widmer, and M. Mauve, “Hierarchical locationservice for mobile ad-hoc networks,” SIGMOBILE Mob. Comput. Com-mun. Rev., vol. 8, no. 4, pp. 47–58, Oct. 2004.

[8] S.-C. M. Woo and S. Singh, “Scalable routing protocol for ad hocnetworks,” Wirel. Netw., vol. 7, no. 5, pp. 513–529, Sep. 2001.

[9] S. M. Das, H. Pucha, and Y. C. Hu, “Performance comparison of scalablelocation services for geographic ad hoc routing,” in INFOCOM 2005.24th Annual Joint Conference of the IEEE Computer and Communica-tions Societies, 13-17 March 2005, Miami, FL, USA, 2005, pp. 1228–1239.

[10] M. Boc, A. Fladenmuller, M. D. de Amorim, L. Galluccio, andS. Palazzo, “Price: Hybrid geographic and co-based forwarding in delay-tolerant networks,” Comput. Netw., vol. 55, no. 9, pp. 2352–2360, Jun.2011.

[11] Y. Zhu, R. Jian, J. Yu, and M. Li, “Geographic routing based onpredictive locations in vehicular ad hoc networks,” EURASIP Journalon Wireless Communications and Networking, vol. 1, p. 134, 2014.

[12] I. Leontiadis and C. Mascolo, “Geopps: Geographical opportunisticrouting for vehicular networks,” 2013 IEEE 14th International Sym-posium on ”A World of Wireless, Mobile and Multimedia Networks”(WoWMoM), vol. 0, pp. 1–6, 2007.

[13] L. Zhao, F. Li, and Y. W. 0003, “Hybrid position-based and dtnforwarding in vehicular ad hoc networks,” in Proceedings of the 76thIEEE Vehicular Technology Conference, VTC Fall 2012, Quebec City,QC, Canada, September 3-6, 2012, 2012, pp. 1–5.

[14] P.-C. Cheng, K. C. Lee, M. Gerla, and J. Harri, “Geodtn+nav: Ge-ographic dtn routing with navigator prediction for urban vehicularenvironments,” Mobile Networks and Applications, vol. 15, no. 1, pp.61–82, 2010.

[15] K. Lee, L. Chen P. Weng, J. Tung, and M. Gerla, “Vclcr: a practicalgeographic routing protocol in urban scenarios.” UCLA, Tech. Rep. TR-080009, 2008.