ROAD safety is always the highest concern for people.

Transcription

1 1038 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 19, NO. 4, APRIL 2018 SAINT+: Self-Adaptve Interactve Navgaton Tool+ for Emergency Servce Delvery Optmzaton Ywen Shen, Student Member, IEEE, Jnho Lee, Student Member, IEEE, Hohyeon Jeong, Student Member, IEEE, Jaehoon Jeong, Member, IEEE, Eunseok Lee, Member, IEEE, and Davd H. C. Du, Fellow, IEEE Abstract Ths paper proposes an evolved Self-Adaptve Interactve Navgaton Tool (SAINT+) to reduce the delvery tme of emergency servces and to mprove navgaton effcency for the vehcles nfluenced by accdents. To the best of our knowledge, SAINT+ s the frst attempt to optmze the delvery of emergency servces as well as the navgaton routes of vehcles around accdent areas. Based on the congeston contrbuton model of SAINT and aggregated nformaton from vehcles n the vehcular cloud, we propose a vrtual path reservaton strategy for emergency vehcles to guarantee a fast emergency servce delvery. We also develop an accdent area protecton scheme based on an adjusted congeston contrbuton matrx and protecton zones to evacuate vehcles n the accdent area. To further reduce travel delay of neghbor vehcles n the accdent area, we also present a dynamc traffc flow control model. Through extensve smulatons wth a real-world map, SAINT+ outperforms other state-of-the-art schemes for the travel delay of emergency vehcles. In scenaros wth a hgh vehcle densty, SAINT+ reduces the travel delay of emergency vehcles by 42.2%. Index Terms Navgaton, path plannng, road emergency servce, road accdent, vehcular networks, self-adaptve, nteractve. I. INTRODUCTION ROAD safety s always the hghest concern for people. Accordng to the Traffc Safety Fact 2014 [1], publshed by the U.S. Natonal Hghway Traffc Safety Admnstraton (NHTSA) n 2016, more than 32 thousand people were klled, and more than 3.2 mllon people were njured n the estmated 6 mllon vehcle traffc crashes n the U.S.. A smlar stuaton happened n European Unon countres n 2013, where road crashes took about 26 thousand lves [2]. Road Manuscrpt receved July 13, 2016; revsed January 18, 2017 and Aprl 4, 2017; accepted May 28, Date of publcaton June 19, 2017; date of current verson March 28, Ths work was supported n part by the Natonal Research Foundaton of Korea through the Mnstry of Scence, ICT, and Future Plannng, Basc Scence Research Program under Grant and n part by the Next Generaton Informaton Computng Development Program under Grant The Assocate Edtor for ths paper was Z. Dng. (Correspondng author: Jaehoon Jeong.) Y. Shen, J. Lee, H. Jeong, and E. Lee are wth the Department of Computer Scence and Engneerng, Sungkyunkwan Unversty, Seoul 16419, South Korea (e-mal: chrsshen@skku.edu; jnholee@skku.edu; jeonghh89@skku.edu; leees@skku.edu). J. Jeong s wth the Department of Interacton Scence, Sungkyunkwan Unversty, Seoul 16419, South Korea (e-mal: pauljeong@skku.edu). D. H. C. Du s wth the Department of Computer Scence and Engneerng, Unversty of Mnnesota, Mnneapols, MN USA (e-mal: du@cs.umn.edu). Color versons of one or more of the fgures n ths paper are avalable onlne at Dgtal Object Identfer /TITS emergency servces (e.g., Emergency Medcal Servce (EMS) and frefghtng squad) provde lfe-and-death rescue servces to people who are n road crashes. The delvery tme of emergency servces s one of the most mportant aspects to reduce fataltes n road crashes. The delvery tme s nfluenced by many factors (e.g., response tme, tranng level, road traffc and stuaton on ste), and some of the factors can be mproved through ntensve exercses. However, a real-world and dynamc road traffc mxed wth randomness challenges the effcent delvery of emergency servces. In addton, snce the vcnty of the accdent s affected by the road accdent, the overall traffc qualty tends to be downgraded. Wth the advancement of Vehcular Ad Hoc Networks (VANET), vehcles equpped wth an On-Board Unt (OBU) can share drvng nformaton wth each other va Vehcle-to-Vehcle (V2V) communcaton, and report road condtons to a vehcular traffc center va a Road Sde Unt (RSU) [3] or a base staton of cellular networks (e.g., Evolved Node B (enodeb) n 4G-LTE [4], [5]) by a Vehcle-to-Infrastructure (V2I) data forwardng scheme. A Traffc Control Center (TCC) [6], whch s a vehcular traffc center, can dssemnate drvng nstructons to vehcles by an Infrastructure-to-Vehcle (I2V) data forwardng scheme. Eventually vehcles and TCC can share nformaton n realtme, whch enables vehcles to have a global sense of the traffc stuaton. Lots of V2I [7], [8] and I2V data forwardng schemes [9] [11] for VANET have been proposed to mprove the effcency of nformaton sharng among vehcles and RSUs. The merts to nstall OBUs n vehcles and RSUs n transportaton nfrastructures, respectvely, can justfy the nstallaton cost. Meanwhle, publc servce departments n varous natons have been consderng to recommend or mandate the nstallaton of OBUs and RSUs. For example, the U.S. Department of Transportaton (USDOT) has been evaluatng the polcy of mandatory nstallaton of OBUs and RSUs, and conductng a seres of feld tests snce It s expected that a new mandatory regulaton for nstallng V2V and V2X devces n new vehcles can be approved n the U.S. n near future [12]. On the other hand, the technology of cloud computng and bg data processng has provded a powerful tool for dealng wth huge volumes of data n a cloud server. A cloud server dedcated to traffc montorng wth a bg data processng engne can process real-tme traffc nformaton IEEE. Personal use s permtted, but republcaton/redstrbuton requres IEEE permsson. See for more nformaton.

2 SHEN et al.: SAINT+ FOR EMERGENCY SERVICE DELIVERY OPTIMIZATION 1039 collected from road networks. Also, the wdely equpped GPS module enables vehcles to obtan moblty nformaton such as geographc locaton, speed and drecton. A vehcle can report the nformaton to other vehcles and a TCC. Furthermore, the GPS navgator, ether a dedcated one (e.g., Gamn [13] and TomTom [14]) or a smartphone App (e.g., Waze [15] and Navfree [16]), can mprove the drvng effcency by suggestng drvng routes to a drver. However, the current GPS road navgators manly suggest a tme-wse shortest path usng the Djkstra algorthm [17] along wth a real-tme road traffc. Bascally, they provde vehcles wth a non-nteractve navgaton servce. When the road traffc abruptly shfts, a vehcle may follow an neffcent navgaton n a road network. Our prevous work, SAINT [18], proposed a new model, called congeston contrbuton, whch estmates the congeston level at each road segment. Through the nteracton between vehcles and the traffc cloud, the runnng vehcles can have a global optmzed navgaton route. Nevertheless, SAINT lacks a dedcated mechansm to handle the delvery of emergency servces, where emergency vehcles may take a bg detour on a heavy traffc condton that contradcts the purpose of emergency servces. And, the deterorated traffc near an accdent ste s not consdered by SAINT ether. Ths paper proposes an evolved Self-Adaptve Interactve Navgaton Tool (SAINT+) for the delvery of emergency servces and accdent area protecton. The goal of ths work s to reduce the delvery tme of emergency servces and the traffc congeston resultng from a road accdent. An accdent road segment wll be solated from the road network by ntentonally ncreasng the congeston level correspondng to the road segment. As a result, other vehcles that wll use the accdent road segment n the near future wll be requred to reroute so that the nfluence of the accdent can be mnmzed. Also, SAINT+ protects the route of an emergency vehcle by restrctng other vehcles to use the route, whch can maxmally guarantee the effcency of the emergency servce delvery. Furthermore, because the accdent road segment usually causes congeston n the vcnty, SAINT+ creates protecton zones for the accdent road segment and controls the traffc flow of the zones. The man dfference between SAINT+ and SAINT s that SAINT+ dentfes and mproves an effcency ssue on emergency servces delvery n modern ntellgent transportaton systems havng wreless communcatons between vehcles and nfrastructures. To the best of our knowledge, SAINT+ s the frst attempt to optmze emergency servce delvery and to effcently reroute the traffc around an accdent area n a road network through an nteracton between vehcles and traffc cloud. The man contrbutons of ths paper are as follows: An archtecture for emergency servce delvery and road accdent area protecton: We propose an archtecture that provdes an optmzed delvery of emergency servces and road accdent protecton n an nteractve manner. An optmzed road emergency servce delvery scheme: We propose an optmzed scheme that can mnmze the travel tme of emergency vehcles va vrtual road reservaton. The scheme also provdes an effcent reroutng for vehcles around an accdent area. Ths strategy dynamcally adjusts vrtual congeston on the path of an emergency vehcle, and can reduce the traffc nfluence from the emergency vehcle. A novel dynamc traffc flow control model for accdent protecton area: We derve a traffc flow control model based on the congeston contrbuton model, whch can adjust the traffc n the vcnty of an accdent area. The dynamc traffc flow control strategy consders nflow and outflow rates to form a traffc ndcator. Ths traffc ndcator along wth the average congeston contrbuton of road segments outsde the protecton zones decdes the congeston level nsde the protecton zones. The rest of ths paper s organzed as follows. Secton II summarzes and analyzes related works. Secton III explans the archtecture of SAINT+, assumptons, goals, and challenges. In Secton IV, we descrbe the travel delay model and SAINT navgaton system. Secton V elaborates the desgn of emergency servce delvery and accdent area protecton n SAINT+. Secton VI descrbes the workng procedure of SAINT+. Secton VII demonstrates the effectve performance of SAINT+ by comparng t wth legacy navgaton schemes. Fnally, n Secton VIII, we conclude ths paper along wth future work. II. RELATED WORK A lot of commercal vehcle dedcated GPS navgators (e.g., Garmn [13], TomTom [14], and NAVI [19]) or smartphone-app navgators (e.g., Waze [15], Navfree [16], Skobbler [20], and Tmap [21]) use tme-wse shortest path algorthms, whch utlze traffc statstcs or real-tme traffc nformaton. These legacy navgators can to some extent provde vehcles wth good navgaton servces when the traffc s not heavy. However, durng rush hours n commute areas or an accdent area, many vehcles use the same routes at the same tme, the tme-wse shortest navgaton path at a certan moment can not provde those vehcles wth a effcent navgaton. Ths s because the legacy navgators do not consder an nstant congeston caused by an accdent. Many research results [22] [25] related to mtgatng road traffc congeston have been reported. Wang et al. [22] proposed a method called Next Road Reroutng (NRR) to mtgate unexpected urban traffc congeston. NRR reroutes vehcles based on a routng cost functon combnng occupancy, travel tme, dstance to destnaton, and geographc closeness to the congeston. Pan et al. [23] proposed a dstrbuted vehcular traffc reroutng system (DIVERT). DIVERT enables vehcles to make reroutng decsons collaboratvely, whch moves the bulk of reroutng computaton from servers to vehcles. Km et al. [24] explored the Markov decson process model to solve a dynamc vehcle routng problem. They utlzed a neuro-dynamc programmng algorthm to avod the hgh dmensonalty. Chen and Chang [25] proposed a new traffc control framework. Ths framework ams at optmzng the throughput of road networks and the travel tme of vehcles by consderng the jont passng rate and traffc lghts at adjacent ntersectons.

3 1040 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 19, NO. 4, APRIL 2018 Pan et al. [26] proposed fve strateges to proactvely reroute vehcles for lower travel tme (called PVTR). The fve strateges are the dynamc shortest path (DSP), A* shortest path wth repulson (AR*), random k shortest path (RkSP), entropybalanced ksp (EBkSP), and flow-balanced ksp (FBkSP). The fve strateges can be categorzed nto two groups, sngle shortest path strateges and multple shortest path strateges. DSP and AR* belong to the frst group, and RkSP, EBkSP, and FBkSP belong to the second. DSP s a classcal shortest path method that reschedules the routes of vehcles accordng to the current road traffc. AR* s a modfcaton of the A* search algorthm [27], whch lets the heurstc functon of A* nclude a repulsve force R(x) that consders the nfluence of other vehcles on the optmal selecton of alternatve paths. RkSP, EBkSP, and FBkSP are all based on the k shortest path algorthm [28]. The dfference s that RkSP randomly selects alternatve paths, whereas EBkSP explots an entropy model and FBkSP consders traffc flow, respectvely, n order to select alternatve paths. Ther proposed system compares the above fve strateges and selects the best paths for the users. SAINT+ s also based on the Djkstra algorthm and k shortest path algorthm. However, wth a new congeston contrbuton model, t provdes not only a short travel tme for general vehcles as PVTR dd [26], but also the assurance of accdent area protecton and fast emergency servce delvery. Wang et al. [29] presentd a real-tme path (RTP) plannng for a transportaton system based on a hybrd network combnng the cellular network and VANET. A vehcle-traffc server feeds real-tme path plannng wth traffc nformaton collected from RSUs or cellular networks. RTP utlzes the measurement of traffc nflow and outflow metrcs, the road capacty, and the vrtual queues that mantan the count of the buffered vehcles stayng at each ntersecton, whch are classfed by dfferent destnatons. Especally, t consders recurrent congeston and non-recurrent congeston [29]. When a traffc congeston occurs at a road segment, the vehcles around the congested area wll obtan suggestons for ther new paths from the vehcle-traffc server through an RSU or a base staton. The new suggested paths depend on the selecton of the maxmum weght derved from the average turnng cost for each related canddate ntersecton. Dfferent from the local metrcs of RTP s vrtual queues at each ntersecton along wth dstance-based turnng cost, SAINT+ employs a congeston contrbuton matrx wth a traffc predcton to globally plan the routes of vehcles that are requested to detour due to an accdent or congeston. Moreover, SAINT+ s dedcated to the protecton of an accdent area and emergency servce delvery. For an accdent area, t constructs protecton zones based on the realstc ratonale that traffc congeston at an accdent area can affect the traffc of the vcnty. Chen et al. [30] proposed a novel emergency vehcle dspatchng system that ncludes automatc emergency resource assgnment and a path plannng scheme [30]. The proposed system learns traffc condton from traffc hstory and suggests a lane reservaton scheme for emergency vehcles. It utlzes RSU as a dssemnaton mean to tell other vehcles to adjust ther velocty or swtch to another lane when an emergency vehcle moves along each road segment. In the path plannng scheme of ths system, the Djkstra algorthm s used to schedule the tme-wse shortest navgaton path of an emergency vehcle. Ths scheme may fal to perceve future congeston, whch s possble that a route based on current traffc condtons or hstorc traffc statstcs may be congested n the near future. Thus, ths approach cannot guarantee the effcent delvery of emergency vehcles. SAINT+ plans the routes for emergency vehcles based on a predcton model, and schedules globally optmal routes for other vehcles by usng ths predcton model. Therefore, SAINT+ can guarantee not only the effcent delvery of emergency vehcles, but also mnmze the mpact of the road accdent on the traffc flow n neghborng areas. Tan et al. [31] explored an algorthm that combnes the dfferental evoluton (DE) and the partcle swarm optmzaton (PSO) to schedule rescue vehcles for extngushng forest fres. The algorthm mnmzes both the extngushng tme of fres and the number of dspatched vehcles by the DE and the PSO. They partcularly consdered the nfluence of the fre spread speed. Tan et al. [32] also proposed multobjectve optmzaton models for placng vehcle nspecton statons. The optmzaton models mnmze the total transportaton cost and tme of customers subject to the constrants that the customers shall arrve at the nspecton locatons wthn a specfc tme and wth a certan cost. The optmzaton models employ a teachng-learnng-based optmzaton algorthm, and combne t wth a mult-objectve optmzaton method. Instead of optmzng the on-ste tme and the number of dspatched vehcles, and the locaton of vehcle nspecton statons, SAINT+ focuses on optmzng the travel tme of emergency vehcles n an urban scenaro wth background traffc. SAINT+ utlzes the congeston contrbuton model of SAINT to dynamcally control traffc flow of the emergency event area. Malvya et al. [33] proposed two new classes of approxmate technques for dynamc route plannng n a contnuous query system. The two technques are K-paths and proxmty measures, whch can speed up the processng of desgnated routes n a contnuous route query envronment. Xu et al. [34] proposed two effcent route search strateges, called ncremental route search (IRS) and herarchcal route search (HRS), n herarchcal dynamc road networks. IRS strategy computes a partal path towards some ntermedate destnatons, and HRS strategy computes the fastest path on a small graph based on a generated herarchcal road network. Unlke the technques from Malvya et al. and Xu et al. usng real-tme traffc update, SAINT+ reserves paths for emergency vehcles based on the traffc congeston predcton n order to detour vehcles that wll use the paths of the emergency vehcles n a short-term tme duraton. Also, SAINT+ constructs protecton zones for a vehcle accdent and suggests a dynamc flow control strategy for the accdent area. III. PROBLEM FORMULATION Ths secton descrbes the archtecture, assumptons, and desgn challenge for road emergency servce delvery. Based on our prevous work [18], we try to mprove the effcency

4 SHEN et al.: SAINT+ FOR EMERGENCY SERVICE DELIVERY OPTIMIZATION 1041 s under an enodeb s coverage. enodeb s a supplementary component for vehcles to communcate wth TCC when they cannot connect wth RSUs. Vehcle: A vehcle s equpped wth a GPS navgator, a DSRC communcaton devce [35], [36] and a 4G-LTE communcaton module [4]. It can communcate wth an RSU or an enodeb to report current moblty nformaton (e.g., speed, locaton, and acceleraton) and the planned navgaton route. In ths paper, we categorze vehcles nto two types: emergency vehcles (EVs) and general vehcles (GVs). Fg. 1. Archtecture and Scenaro of Emergency Servces Delvery. of emergency rescue servces for a road traffc accdent, and mantan the global optmzed routes for other vehcles,.e., guarantee a fast delvery of emergency servces and reduce the nfluence of the traffc accdent on the travel tme of other vehcles. A. Road Emergency Servce Delvery Archtecture Ths secton descrbes the archtecture of road emergency rescue and the component nodes of the vehcular cloud. Fg. 1 shows the archtecture for road emergency servce delvery. Ths archtecture ncludes the followng components: Traffc Control Center (TCC): A TCC s an urban traffc dedcated management server complex that can nclude hgh performance computng clusters, wred and wreless networks, data storage, etc. TCC can collect real-tme road traffc nformaton such as average speed, vehcle densty, and traffc flow. It has a relable connecton wth an emergency center to whch a vehcle accdent event can be reported. TCC s also n charge of the mantenance of the congeston contrbuton matrx and calculaton of the optmal navgaton route. Emergency Center (EC): An EC s an urban emergency event response hub that s responsble for the emergency event processng and the dspatch of emergency vehcles. Usually, t s located nsde a hosptal that can provde emergency medcal treatment. EC can receve notfcaton of a road emergency from TCC. Road-Sde Unt (RSU): An RSU s a wreless access pont located at each ntersecton. RSU communcates wth vehcles n vehcluar ad hoc networks va DSRC and s connected to the Internet. In a target urban road network, RSUs are connected wth each other va ether wred or wreless networks. Evolved Node B (enodeb): enodeb s a base staton n 4G-LTE cellular networks. A subscrbed moble devce can use voce, text message and data access servces f t B. Assumptons Ths secton lsts assumptons for SAINT+ as follows: A vehcle can communcate wth the TCC by ether DSRC [35], [36] or 4G-LTE [4]. The default communcaton mode s va DSRC. If a vehcle cannot communcate wth an RSU by DSRC, 4G-LTE communcaton module can be used to exchange nformaton wth the TCC. The communcaton delay between vehcles and the TCC can be gnored because the order of communcaton delay va DSRC [37] or 4G-LTE [38] s much smaller than the tme that a vehcle takes to pass through a road segment. Thus, a vehcle and a TCC can exchange nformaton n tme. Loop detectors are connected wth RSUs at road segments. Loop detectors can count the number of vehcles enterng and leavng the road segment, and RSUs can know the road traffc accurately. The traffc nformaton s reported to the TCC for the navgaton servce of SAINT+. Vehcles agree to report ther postons and destnatons to TCC whenever necessary for SAINT+. The nformaton transmsson between vehcles and TCC shall be encrypted, and TCC shall not dsclose the drvers moblty and dentty nformaton to any 3rd party for ther prvacy protecton. C. Scenaro and Challenges We consder a scenaro that a car crash happens n an urban area, as shown n Fg. 1. The crashed cars report the accdent to the TCC va an RSU or an enodeb, and TCC nforms the nearest EC of the accdent. EC dspatches an EV toward the accdent ste. The blue dash lne s a selected path wth the mnmum travel tme. Although ths path s the shortest one, along ths path there are many other vehcles currently drvng (e.g., V 1, V 2, V 3 ). When the EV s passng, other vehcles may stop or pull up on the road sde n order to leave space for the EV. Meanwhle, vehcles on neghborng roads (e.g., V 4 and V 5 ) may also need to use the path of the EV. The vcnty of the car crash s facng a possble traffc or low-speed drvng caused by the accdent. As a result, the travel tme for vehcles around the accdent wll be ncreased. To mprove the delvery effcency of emergency servces and also protect an accdent ste, we have followng challenges: Even though the path of an EV s tme-wse shortest for the current traffc condton, vehcles on the path of the EV may hnder the penetraton of the EV.

5 1042 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 19, NO. 4, APRIL 2018 Vehcles that wll use the path of the EV may also affect the effcency of the emergency servce delvery. The slow traffc flow around the accdent ste may cause extra traffc to other vehcles. In the next secton, we wll explan our soluton for the above challenges, whch s based on the SAINT navgaton scheme. IV. TRAVEL DELAY PREDICTION AND SAINT NAVIGATION Ths secton explans the modelng of travel delay on both road segments and End-to-End (E2E) travel paths, based on our early work [9], [39], as well as the SAINT navgaton scheme [18]. Frst, a defnton s gven as follows. Defnton 1 (Road Network Graph): A Road Network Graph s a drected graph G = (V, E) for a road map such that V s the set of vertces (.e., ntersectons), denoted as V (G), and E s the set of drected edges e j (.e., road segments) for, j V, denoted as E(G). A. Travel Delay on Road Segments and End-to-End Delay The prevous research [9], [39] showed that n a scenaro of a lght road traffc, the travel delay of a vehcle of a certan length follows a Gamma dstrbuton [40]. The travel delay on a road segment s defned as d, whch follows a Gamma dstrbuton, d Ɣ (κ,θ ),whereκ and θ are the shape parameter and scale parameter, respectvely [41]. κ and θ can be derved from the mean travel delay μ and the varance of travel delay σ 2 on road segment, whereμ and σ 2 can be collected from ether a real-tme drvng report of vehcles or loop detectors on each road segment [42], [43]. Assume that the travel delays of road segments n a path are ndependent from each other and follow the same Gamma dstrbuton. That s, the collecton of all random varables d n the travel delay D of ths path s ndependent and dentcally dstrbuted (..d). Thus, the mean and varance of the travel delay of ths path can be estmated as the cumulatve mean and varance of the road segments along the path [9], [39]. Hence, supposng that, from the source poston to the destnaton poston, there are n ntersectons and n 1roadsegments, the mean E[D] and varance Var[D] of travel delay D of ths path can be expressed as: n 1 n 1 E[D] = E[d ]= μ, (1) =1 n 1 =1 n 1 Var[D] = Var[d ]= =1 =1 σ 2. (2) Therefore, the E2E travel delay D can be modeled as a Gamma dstrbuton, D Ɣ (κ D,θ D ),whereκ D and θ D are formulated from E[D] and Var[D] [9]. In practce, varous methods can be used to measure the delay of a sngle road segment or an E2E path, e.g., a camera or loop detectors. As mentoned n Secton III-B, each road segment can be nstalled wth loop detectors at ts entrance and ext to measure travel delay accurately n real-tme. Fg. 2 Fg. 2. Installaton of Loop Detector. shows the nstallaton of loop detectors. Assume that for a sample perod T, the count of vehcles travelng on a road segment (v,v j ) s n, where(v,v j ) represents an edge wth two vertces v and v j, the tme for each vehcle V to travel on ths road segment s t V, and the mean travel delay d (v,v j ) of road segment (v,v j ) can be calculated as follows: n t V =1 d (v,v j ) = n > 0, n (3) l (v,v j ) + d TL n = 0, v L where l (v,v j ) s the length of the road segment (v,v j ), v L s the speed lmt of ths road segment, and d TL s the mean watng tme of traffc lght. In (3), two cases are consdered. The frst case s when there are vehcles drvng on a road segment. In ths case, d (v,v j ) s the mean travel tme for all these vehcles. The second case s no vehcle moved on a road segment durng the samplng tme. In ths case, d (v,v j ) s the travel tme for the road segment by the speed lmt plus the mean traffc lght watng tme. Note that the tme t V ncludes the watng tme for traffc lghts at an ntersecton before the next road segment. For a vehcle V wth a route R V contanng a sequence of ntersectons: R V = v 1,v 2,...,v n, vk V (G), thee2e delay can be formulated as: D V n n 1 = k=1 d (vk,v k+1 ), (4) where R V s an ntersecton set ncludng all vertces n the route of V. In the next secton, we ntroduce our early work, called the SAINT navgaton scheme [18]. B. SAINT Navgaton Scheme Ths secton ntroduces the SAINT [18] navgaton scheme from two aspects: () a Congeston Contrbuton model and () a Congeston Contrbuton model-based shortest path algorthm. 1) Congeston Contrbuton Model: Basedontheroad segment delay and E2E delay defned n the prevous secton, each vehcle has a travel delay on each road segment (called lnk delay) along ts route, as well as E2E delay for the whole route. For a partcular vehcle V j, the Congeston Contrbuton (CC)

6 SHEN et al.: SAINT+ FOR EMERGENCY SERVICE DELIVERY OPTIMIZATION 1043 Fg. 3. Example of a Step Functon. Fg. 4. Example of a Congeston Contrbuton Matrx. c V j [18] s modeled as: c V j = 1 DV j D V, (5) j n where D V j n s the E2E delay of a vehcle for a travel path wth n vertces,.e., the travel delay from the source ntersecton 1 to the destnaton ntersecton n,andd V j s the sub-route delay from the source ntersecton 1 to an ntermedate ntersecton : 1 D V j d (vk,v = k+1 ) for 2, k=1 (6) 0 for = 1, where d (vk,v k+1 ) s the travel delay for a road segment (v k,v k+1 ) n the route. Note that D V j 1 s defned as 0 snce t s travel delay at the begnnng of the route, and the correspondng CC c V j 1 s 1. The c V j on each road segment of a route s mantaned as nvarant, so we defne the Congeston Contrbuton Step Functon (CCSF) C V j (x) for the sub-route delay x from a vehcle s start poston to an ntermedate poston on ts trajectory: C V j (x) = c V j u(x D V j ), (7) where u(x D V j ) s a shfted unt step functon defned as: u(x D V j ) { 1 x D V j 0 x < D V j for (1, n). (8) For example, Fg. 3 shows the step functons of two vehcles mapped to ther routes. Red vehcle V 1 and blue vehcle V 2 have dfferent scales for the CC value on the y-axs. Both the vehcles have a CC value 1 (ndcatng the njecton of one vehcle nto a road segment) at the entrance of ther 1st road segment n ther route, and then the CC value lnearly decreases, whle on each road segment t s mantaned as constant. For a road network graph G wth n vertces (ntersectons), a Congeston Contrbuton Matrx (CCM) s defned as: 0 m 1, m 1,n m 2,1 0 m, j. M =....., (9).. m n 1,n m n, m n,n 1 0 where, j V (G) and m, j s the cumulatve lnk congeston contrbuton of road segment e, j,.e., the sum of congeston contrbuton from all vehcles that are passng and wll pass through edge e, j. After a vehcle that contrbutes to the congeston passes through the edge, the correspondng congeston contrbuton of ths vehcle wll be subtracted from the correspondng entry n the matrx CCM. Note that the CCM s mantaned by TCC n the vehcular cloud and dynamcally updated va V2I and I2V communcatons between vehcles and RSUs. Wth ths CCM, we know that the upper bound of any element n the CCM s the number of vehcles (denoted by N) runnng n the road network f all N vehcles pass through the same road segment e uv. As shown n Fg. 4, two vehcles V 1 and V 2 are movng n a target road network and they have dfferent routes leadng to dfferent destnatons, but there s a common road segment e j that they wll pass through. Assume that at the current tme, there are no other vehcles drvng n ths road network. The congeston contrbuton m j of edge e j n CCM s m j = C V 1 3 (D 3)+C V 2 2 (D 2). After they pass edge e j,theentrym j of the CCM s updated by subtractng the congeston contrbuton value C V 1 3 (D 3) and C V 2 2 (D 2). In the next secton, we ntroduce the congeston contrbuton model-based Delay-constraned Shortest Path (DSP) Algorthm. 2) Delay Constraned Shortest Path Algorthm: To explan DSP algorthm, we frst ntroduce the α-ncrease travel path [18]. The α-ncrease travel path for a vehcle s a path such that the E2E delay of ths path s wthn (1 + α)d, where D s the shortest path travel tme on the current traffc condton.

7 1044 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 19, NO. 4, APRIL 2018 Algorthm 1 Delay-Constraned Shortest Path Algorthm 1: functon DSP(G, u,v,α) 2: P P wll contan the lst of ntersectons for a selected α-ncrease shortest path. 3: D uv Compute-Djkstra-Path-E2E-Delay(G, u,v) 4: ˆD uv (1 + α) D uv 5: K Compute-k-Smallest-Congeston-Increase- Paths(G, u, v) 6: k K 7: for 1tok do 8: D Compute-E2E-Delay(K, ) 9: f D ˆD uv then 10: P Get-Path(K, ) 11: return P 12: end f 13: end for 14: P Compute-Djkstra-Path(G, u,v) 15: return P 16: end functon The objectve of the DSP algorthm s to fnd an α-ncrease travel path based on CCM and Yen s k-shortestpath algorthm [28]. As shown n Algorthm 1, the nput of the algorthm s a graph G, source ntersecton u, destnaton ntersecton v, and the ncrease percent of travel delay α. Frst, the E2E delay of a tme-wse shortest path s computed based on Djkstra s shortest path algorthm. Then we obtan the α-ncrease E2E delay ˆD uv. Based on CCM, the k smallest congeston ncrease paths are stored n K and the sze of K s k. For each path n K, the E2E delay s calculated and put nto D. IfD s smaller than the E2E delay ˆD uv of an α-ncrease travel path, the correspondng path for the par of (K, ) s selected. Otherwse, contnue to check the next path. If all k smallest congeston ncrease paths cannot fulfll the condton, a tme-wse Djkstra shortest path s selected. The dea behnd the DSP algorthm s that at the current moment, a constraned detour route s selected to mnmze the future congeston n the target road network. Comparng ths to the greedy algorthm (e.g., Djkstra algorthm), DSP ntentonally plans a tme-wse suboptmal route for vehcles, but the overall travel delay can be sgnfcantly reduced. The results n Secton VII confrm ths dea. Through ths algorthm, a vehcle can fnd a global optmzed path to mnmze future possble congeston. In the next secton, we wll descrbe the desgn of the new navgaton scheme for emergency servce delvery. V. THE DESIGN OF SAINT+ EMERGENCY NAVIGATION Ths secton explans the desgn of SAINT+ navgaton for the delvery of road emergency servces from the followng perspectves: () congeston contrbuton adapton for an accdent road segment and emergency vehcles, () zone-based accdent area protecton, and () traffc flow-based dynamc congeston contrbuton adjustment for the protecton zones. As explaned n the prevous secton, the congeston contrbuton model provdes a predcton of the future congeston so Fg. 5. Accdent Road Segment Protecton and Emergency Navgaton. that each vehcle can be assgned an optmal route that not only s the constraned detour path but also dstrbutes unformly the traffc flows of all the vehcles on all possble road segments n a target road network. Based on ths model the goal of the emergency navgaton of SAINT+ s to optmze a route for each vehcle when any accdent happens n the road network, and allow emergency vehcles to reach the accdent ste as fast as possble whle reducng the traffc congeston nfluence on other vehcles as much as possble. For the explanaton of SAINT+, we frst gve several defntons as follows: Defnton 2 (Accdent Road Segment): Let an Accdent Road Segment be an edge e uv E(G) where a vehcle accdent occurs. Defnton 3 (Contour Vertces): Let the Contour Vertces be a set S CV V (G) that ncludes all adjacent vertces of the two vertces u and v of the Accdent Road Segment e uv, wth u / S CV, v/ S CV. Defnton 4 (Protecton Zone): Let Protecton Zone be a set S PZ E(G) such that all edges of S PZ are drectly jonted to the two vertces u and v of the Accdent Road Segment e uv or the Contour Vertces of another Protecton Zone. A. DSP-Based Reroute Algorthm for an Accdent Road Segment and Emergency Vehcle Congeston contrbuton model gves a good predcton for future congeston based on a decreasng step functon that utlzes the percentage of travel tme on the whole travel course. That s, the farther road segment on the vehcle s trajectory has the lower congeston contrbuton accordng to the travel delay of the road segment. Moreover, the CCM mantaned by TCC s perodcally updated through the nteracton wth vehcles va V2I and I2V, or a 4G-LTE lnk, and provdes a global optmzed route for each vehcle. 1) Emergency Servce Delvery Optmzaton: Now we consder a specal case where a vehcle v has an accdent (e.g., vehcles collson or crash) on a road segment e uv of a target road network, as shown n Fg. 5. Ths accdent vehcle drectly (by tself when the DSRC devce works well, even after

8 SHEN et al.: SAINT+ FOR EMERGENCY SERVICE DELIVERY OPTIMIZATION 1045 a serous damage by the accdent) or ndrectly (by other vehcles or RSU when the DSRC devce n the accdent vehcle does not functon) communcates wth a nearby RSU to report the accdent, and then RSU forwards ths nformaton to TCC n vehcular cloud. In Fg. 5, the green rectangles on road segments represent the congeston contrbuton values n CCM. A thcker rectangle ndcates a hgher congeston n the near future. Frst, TCC delvers the accdent nformaton to EC, where an an emergency vehcle s prepared (e.g., ambulance, polce car, and fre engne) for ths accdent. Second, n order to protect the accdent road segment e uv, the correspondng entry m u,v n CCM s set to an artfcal congeston ncrease value γ (e.g., 5000) as follows: m u,v = m u,v + γ, for m u,v M and γ>n, (10) where N s the total number of vehcles n the current road network. In Fg. 5, the thck yellow rectangle shows the protecton of the accdent road segment. In addton, because TCC has the route nformaton of all SAINT+ clents, by the updated matrx and DSP algorthm, TCC recalculates the new routes for the vehcles that wll pass through the accdent edge n ther current routes, as shown n Algorthm 2. Then, TCC ntates the messages for the reroute request, ncludng the new route, to each vehcle va the I2V protocol. Subsequently, the affected vehcles wll follow the new routes from the requests. As shown n Fg. 5, the blue dashed lnes are possble detour paths. Any newly joned vehcle wll also receve an accdent-protected route. After a short perod, an EV v E wll be ready to head to the accdent road segment desgnated by EC. Once the EV moves to the road network, t requests TCC to calculate a fastest route from the current poston to the accdent spot. TCC reples to t wth a calculated route. Also, the correspondng entres m, j n CCM, ncludng all road segments n the EV s route R EV = v 1,v 2,...,v n for v V (G), aresettoanartfcal congeston ncrease value δ (e.g., 5000) as follows: m v k,v k+1 = m vk,v k+1 + δ, for m vk,v k+1 M and δ>n, (11) where N s the total number of vehcles n the current road network. TCC collects the nformaton for all vehcles that wll pass through any road segment of the EV s route, and then calculates and sends the new routes to these vehcles, as shown n Algorthm 2. These vehcles wll avod the route of the EV so that the EV can get a clear way toward the accdent road segment. The blue dashed lnes n Fg. 5, near the red path of the EV are the possble detour routes for the vehcles that wll pass through the red lne. The am of Algorthm 2 s to check every vehcle s route as to whether t s overlapped wth the accdent edge or the route of EV. If so, a new route P replaces the current route R n n lne 8 of Algorthm 2. The nputs of Algorthm 2 are the road graph G, vehcle set N, the route of EV R EV, the accdent edge e ACC, and the detour factor α for the DSP algorthm. Once the accdent s handled and the road segment s cleaned, the EV nforms TCC that the accdent has been handled. Then TCC wll recover the CCM to ts orgnal level Algorthm 2 SAINT+ Reroute Algorthm 1: procedure REROUTE(G, N, R EV, e ACC,α) 2: for all n N do 3: R n Get-Route(n) 4: f any e R n = e ACC or any e R EV then 5: s Get-Current-Locaton(n) 6: d Get-Destnaton(n) 7: P DSP(G, s, d,α) 8: R n P 9: end f 10: end for 11: end procedure by subtractng the γ and δ, compute the new route for each vehcle and dstrbute t to each of them. Thereafter, all the vehcles wll reroute based on the updated routes for effcent navgaton. 2) Zone-Based Accdent Area Protecton: Snce the surroundng areas of the accdent road segment are usually affected by the accdent, many vehcles may experence congeston, so they slowdown ther movng speed, whch causes further congeston; ths n turn degrades the navgaton effcency. For example, because of a collson, a vehcle drvng n the vcnty wll take a longer tme than n normal traffc condtons. Ths s the reason we propose zone-based accdent area protecton to lessen the mpact of the accdent on the traffc flow near the accdent area. In a road network graph G, the edge e uv s an accdent road segment, and the 1st Protecton Zone, Zone 1(Z 1 ), conssts of all edges jonted to vertces u and v except e uv. Formally, the set of edges S Z1 for the zone Z 1 s defned as follows: S Z1 ={e : e N(e uv )}, (12) where N(e uv ) s the neghbor edges of the accdent edge e uv. Smlarly, the 2nd Protecton Zone, Zone 2(Z 2 ), ncludes all edges jonted to the Contour Vertces, whch can be represented as: S Z2 ={e : e N(S Z1 ) and e / S Z1 and e = e uv }, (13) where N(S Z1 ) s the neghbor edges of all edges n S Z1. For nstance, Z 1 and Z 2 are formed as llustrated n Fg. 6. As a drvng rule n Z 1 and Z 2, all the vehcles try to evacuate from the both areas. The outsde vehcles try a detour to avod movng nto the both areas, that s, a detour around these areas. B. Dynamc Traffc Flow Control for the Protecton Zones In ths secton, we wll explan the desgn of dynamc congeston contrbuton adjustment for the protecton zones based on traffc flow. Even though the protecton zones provde strong protecton for the accdent road segment and fast delvery of emergency servces, other vehcles may have to use long detour paths to avod usng the road segments n protecton zones. Especally, n a road network domnated by one-way road segments, the detoured vehcles may be congested due to lmted optons of detour road segments. Through the observaton of traffc flows n protecton zones, t

9 1046 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 19, NO. 4, APRIL 2018 Fg. 6. Fg. 7. Zone2. Zone Protecton for the Accdent. Inflow and Outflow Rates for the Road Segments n Zone1 and s reasonable to permt some vehcles to use the road segments n the zones when the traffc n ths area s relatvely lght. Accordng to ths ratonale, we propose a traffc flow control scheme for the zone areas. Wthout loss of generalty, we consder a road network graph G based on an accdent road segment e uv as shown n Fg. 7, where the vertces u and v are the two ends of the edge e uv. Referrng to the Defnton 4, Zone1(Z 1 )andzone2(z 2 ) are formed to protect the accdent spot. We assume that the vehcle set S V (T ) ncludes each vehcle durng the th sample tme T, and the nflow rate λ(t ) of a road segment durng the th sample duraton T can be expressed as: λ(t ) = S V (T ) T for T > 0. (14) In order to measure the traffc status nsde zones, we defne the nflow rate as λ x, x {Z 1, Z 2 }.Then λ x = j S x λ j (T ) for x {Z 1, Z 2 }. (15) Outflow rate μ x has a smlar calculaton process to the nflow rate λ x for x {Z 1, Z 2 }. Then we defne a traffc ndcator I, whch s the rato of nflow rate to outflow rate, for each zone as λ x μ x, λx > 0, μ x > 0 I x = λ x, λ x > 0, μ x = 0 0, λ x = 0, μ x > 0 1, λ x = 0, μ x = 0 for x {Z 1, Z 2 }. (16) Intutvely, f I s larger than 1, ths means that the nflow rate s larger than the outflow rate,.e., more vehcles move nto the area. On the contrary, f t s smaller than 1, ths means that more vehcles leave the area, and f t s close to 1, the correspondng zone has a balanced traffc flow. For example, as shown n Fg. 7, the nflow rates, λ Z 1 1, λz 1 2,, λz 1 n,andthe outflow rates, μ Z 1 1, μz 1 2,, μz 1 n, (here we omt the sequence of the sample tme T ) of all the edges n Z 1 are measured. Therefore, for Z 1 wth n edges and Z 2 wth m edges, the total nflow rate and the total outflow rates are λ Z 1 = n =1 λ Z 1, μ Z 1 = n =1 μ Z 1,andλ Z 2 = m k=1 λ Z 2 k, μz 2 = m k=1 μ Z 2 k, respectvely. Then, we obtan the correspondng I Z1 and I Z2. Recall CCM n equaton(9); we know that the CC value m e on an edge e n CCM s the summaton of all CC values from dfferent vehcles accumulated on ths edge, whch can be descrbed as: m e = C k j (D j ) for S V ={V 1, V 2,...,V a }, (17) k S V where S V s the set of vehcles that wll pass through the edge e, andc k j (D j ) s the correspondng value of CCSF for edge e contrbuted by a vehcle (denoted by k) nthesets V,and j s the ndex of the edge e n the travel path of k. SotheCC value me x of each edge e for x {Z 1, Z 2 } can be updated va the followng model: m x e = me x + C I x, x {Z 1, Z 2 }, (18) where C s the mean CC value of edges not belongng to Z 1 or Z 2,andme x s the CC value of each edge e for x {Z 1, Z 2 }. The ratonale for ths model reles on the observaton of vehcle behavors changng before and after the constructon of zone areas. Before the constructon of zones, vehcles could go through zones regardless of the speed and traffc affected by the accdent, whereas after the constructon of them, the detoured paths for the vehcles avod the edges of zones by checkng CCM,.e., comparng the CC values of the edges nsde the zones to those of the edge outsde the zones. Ths model updates the CC values for edges nsde each zone wth the mean CC value C of all the edges outsde the zones and the traffc ndcator I x.wheni x s larger than 1, me x s enlarged by I x tmes C,.e., me x ncreases dramatcally to the sum of mx e and C I x. On the other hand, when I x s smaller than 1 but larger than 0, the augment s the fracton of C, whch enables path plannng to assgn the zone-nsde edges to the vehcles wth less probablty than the zone-outsde edges for the zone protecton. So far, we have llustrated the desgn of SAINT+ for emergency servce delvery and accdent ste protecton. In the next secton, we show the overall navgaton procedure of SAINT+.

10 SHEN et al.: SAINT+ FOR EMERGENCY SERVICE DELIVERY OPTIMIZATION 1047 VI. SAINT+ NAVIGATION PROCEDURE Ths secton explans the navgaton procedure of emergency servce delvery and accdent protecton along wth practcal consderatons n ths procedure. A. Navgaton Procedure The navgaton procedure s as follows: 1) A vehcle functonng as a SAINT+ clent follows the same procedure of our prevous work SAINT [18] n regular traffc condton (.e., the non-accdent case). By V2I [8] and I2V [9] data delvery schemes or a 4G-LTE lnk [4], the vehcle sends ts navgaton request, ncludng ts source and destnaton, to a nearby RSU, and then the RSU forwards the request to the TCC. The TCC calculates a global optmzed route for the vehcle based on the congeston contrbuton matrx, and reples to the vehcle va an RSU. If the vehcle devates from the notfed route, the same procedure can be repeated to obtan a new optmzed route from the TCC. 2) Whenever a vehcle usng the SAINT+ servce has an accdent wth other vehcles, the accdent nformaton wll be forwarded to a nearby RSU (or enodeb). The RSU nforms the TCC, and then the congeston contrbuton of the accdent road segment wll be updated wth a very large value, and the orgnal CC value wll be recorded n TCC for future CC recovery. 3) Meanwhle, TCC vrtual protecton areas (.e., protecton zones) for the accdent road segment are constructed around the accdent road segment. The congeston contrbuton of the protecton zones wll be dynamcally adjusted based on the traffc flow control model descrbed n Secton V-B. 4) After that, TCC mmedately broadcasts a reroute request to all vehcles va RSUs, but only those that wll pass through the accdent road segment and protecton zones n the planned routes wll be asked to reroute,.e., detour from accdent road segment and protecton zones. Note that the detour route stll follows the SAINT [18] procedure to guarantee the global optmzed route. 5) After an accdent happens, an EV wll be dspatched to the accdent road segment as soon as possble. Wth the SAINT+ servce, the EV follows the same procedure as SAINT [18] to obtan a global optmzed route, but the dfference s that the congeston contrbutons of the road segments along ts route to the accdent road segment wll be set to a very large value to dsallow other vehcles to use the road segments along ts route, as descrbed n V-A. The orgnal CC values wll be recorded n TCC for future CC recovery after the delvery of the EV. Smlarly, TCC wll broadcast a detour request to all vehcles and only the vehcles that wll pass through the road segments of EV wll reroute accordng to the method of SAINT [18]. 6) When arrvng at the accdent road segment and after fnshng the rescue and medcal care, the EV wll ntate a msson accomplshed message to TCC va an RSU. TCC wll recover the congeston contrbuton of the accdent road segment and the road segments of the protecton zones. When startng to move back to the EC from the accdent road segment, the EV can use SAINT+ servce agan to obtan a global optmzed route so that the travel tme can be reduced. 7) All SAINT+ clents navgate accordng to the normal SAINT navgaton procedure [18]. Note that a vehcle wth the orgn or the destnaton or the both that are n the protecton zones can follow ts current planned path to reach the destnaton. SAINT+ does not stop a GV movng, and schedules an optmal path for t. If the destnaton of a GV s n the accdent road segment, the GV can move to the destnaton, but wll experence a serous delay due to the nfluence of the accdent. B. Practcal Consderatons To use SAINT+ n real traffc envronment, a navgator n a vehcle shall connect to an On-Board Unt (OBU). The navgator can receve a scheduled route for the current travel. The drver frst nputs the destnaton for the travel, and the current poston and the destnaton wll be transmtted to TCC va OBU. TCC calculates a global optmal travel path and sends t back to the navgator. The vehcle can start to move. When there s an emergency event and the vehcle wll pass the road segment of the emergency event, TCC wll multcast the reroute request to the vehcle. The vehcle automatcally reports current poston wthout an nterventon of the drver. TCC uncasts the new scheduled route to the vehcle. To allow SAINT+ to work well n the real world, there are practcal consderatons: Durng an emergency servce delvery, there are two reroute procedure called n SAINT+. The frst s the moment when the accdent happened, and the second s the moment when the EV s dspatched. Consderng the tme between the accdent occurrence and the dspatch of an EV, f t s a reasonably short tme, the frst reroute can be lmted to vehcles n the vcnty of the accdent road, or the two calls of the reroute procedure can be combned as one call after the EV s dspatched. Ths reduces the reroute frequency of SAINT+ clents snce the frequent reroute may cause an uncomfortable drvng experence. Ths smplfcaton shall be consdered on a case-by-case bass, e.g., emergency response tme and nfrastructure level. Normally, to deal wth an accdent event, several EVs would be dspatched at the same tme or dfferent tme perod. These EVs may nclude polce cars, EMS, frefght team, etc. The proposed scheme n SAINT+ may show an unsatsfactory performance n the delvery of several EVs. Bascally each EV s route wll be protected, so every new entered EV needs to follow the planned path based on the current CCM, and ths results n serous detours for the latest joned EVs. One practcal soluton s mantanng two CCMs n the TCC such that one s a normal CCM used by the vehcles except EVs, and the other s talored for EVs by not addng extra CC values for the routes of

11 1048 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 19, NO. 4, APRIL 2018 TABLE I DESCRIPTION OF THE ROAD NETWORK TABLE II SIMULATION CONFIGURATIONS Fg. 8. Road Network n the Downtown Area of Mnneapols for Smulaton. the EVs. In the next secton, we evaluate the scenaro of multple EVs to dsallow the other vehcles to use the road segments along the routes of the EVs. So far, we have explaned the desgn and procedure of SAINT+, as well as practcal ssues. In the next secton, we wll show performance evaluaton of SAINT+ wth other state-of-the-art schemes. VII. PERFORMANCE EVALUATION We demonstrate the performance through smulatons based on SUMO (Smulaton of Urban MOblty) [44]. We use a realworld map obtaned from OpenStreetMap [45], whch s the downtown area of Mnneapols, MN, US, as shown n Fg. 8. The nformaton of the map s shown n Table I. We edted the map by Java OpenStreetMap Edtor (JOSM) [46] and NETCONVERT [47] to remove dead ends and modfy redundant streets. The emergency center s located at an end of the road on the left bottom of the map. The accdent spot s set at a road segment of the center of the map. Smulaton settngs n SUMO were as follows: the vehcle length s 5 m, the mnmal gap between vehcles s 2.5 m, and the deceleraton s 6 m/s 2. The conducted smulatons nclude two scenaros: Consecutve accdents are reached by one emergency vehcle (EV). To measure the travel delay of an EV and general vehcles (GVs), every new accdent vehcle s njected after 30 s when the prevous accdent s handled by an EV. 20 vehcles are selected to contnuously travel between two fxed ponts, L1 andl2, n order to make road traffc exst around the accdent spot, as shown n Fg. 8. One accdent s reached by ten EVs. To measure the delvery rato of the ten EVs, an accdent vehcle s only njected once at a fxed smulaton tme pont, and ten EVs are set to be dspatched by the EC. At the begnnng of a smulaton, vehcles are sequentally placed on the target road network and the destnaton of each vehcle s randomly chosen. Once a vehcle reaches ts destnaton, t wll dsappear, and then a new vehcle wll be placed at a random road segment to start a new travel. The emergency response tme (.e., the delay from the moment when the accdent s reported to the moment when the EV s dspatched) s set to 60 s. Note that we only measure the performance that EV heads to the accdent road segment, the return of EV to EC s not measured n the current smulaton. All vehcles conform to the car followng model (Krauss et al. [49]). All traffc lghts n the target road network follow a statc traffc lght scheduler [50]. All road segments are nstalled wth two loop detectors at ther entry and ext. GVs do not gve way to an EV so that the EV can move fast n the road segment havng those GVs n our smulatons. Other evaluaton settngs are as follows: Performance Metrcs: () the mean E2E travel delay of EV and GVs, () the mean lnk delay of all vehcles, () the mean number of accdents handled, (v) the successful delvery rato of EVs, and(v)the tradeoff of mean E2E delay between EV and GVs. Baselne: () SAINT [18], () Djkstra [13] [17], [19] [21], and () RTP [29]. Parameters: In the performance evaluaton, we nvestgate the mpact of the followng parameters: () Vehcular traffc densty N (.e., the number of vehcles), () Maxmum vehcle speed v max (.e., speed lmt), and () Vehcle acceleraton a v. The smulaton tme s set to two hours (.e., 7200 s) and the smulatons are repeated ten tmes wth dfferent seeds. The performance results nclude an error bar to show the 90% confdence ntervals. The default values of the smulaton are specfed n Table II. The CC ncrease-values γ and δ mentoned n Secton V-A are set to a value larger than the upper bound of the CC value n the matrx CCM. In our smulatons, they are all set to The EMS scene tme n Table II s based on real-world data [48], whch s the average tme for an EV s crew to deal wth an emergency event. Note that the general urban speed lmt vares from

12 SHEN et al.: SAINT+ FOR EMERGENCY SERVICE DELIVERY OPTIMIZATION 1049 Fg. 9. The Cumulatve Dstrbuton Functons of Navgaton Delays. (a) The CDF of EV E2E Delay. (b) The CDF of GVs E2E Delay. (c) The CDF of Mean Lnk Delay. Fg. 10. The Navgaton Delays. (a) E2E Delay of EV. (b) E2E Delay of GVs. (c) Mean Lnk Delay. 50 km/h to 80 km/h (.e., meter/s to 22.5 meter/s or mle/h to 50 mle/h), and the actual vehcle speed s affected by many factors (e.g., traffc lght and congeston level). The 80 km/h speed lmt for the road segments n the smulaton shown n Table II can reflect a general scenaro, and we also studed the mpact of the speed lmt n Secton VII-C. A. The Comparson of Navgaton Behavor The purpose of SAINT+ s to optmze the delvery tme of EVs. To show the overall performance comparson, we collected all the data of the smulatons for the CDFs of SAINT+, SAINT, Djkstra, and RTP llustrated n Fg. 9. From Fg. 9(a), we can see that more than 95% of mean emergency delvery tmes of SAINT+ are less than 160 s, whch sgnfcantly outperforms the Djkstra algorthm. Even for SAINT and RTP, SAINT+ mantans ts good performance all the tme, and when the traffc grows heavy, the advantage ncreases. For the mean E2E delay of GVs, as shown n Fg. 9(b), the CDF of SAINT+ s slghtly below that of SAINT and stll mantans a large gap compared to Djkstra, and a huge gap compared to RTP. Snce the prorty of SAINT+ s to guarantee fast delvery of EVs, t sacrfces to some extent the E2E delay of GVs, but ths slght sacrfce s worthwhle because the tme saved for the EV s related to savng a human lfe. Fg. 9(c) shows the mean lnk delays (.e., mean travel tme of road segments) of all the schemes. The results n the fgure suggest that SAINT+ grows faster than other schemes, whch confrms the mprovement n the E2E delay of the EV. Also, we notced an nterestng curve shape n the fgure. As shown n Fg. 9(c), when the mean lnk delay (x-axs) s smaller than 10 s, the shapes of all the curves have a clear step style, and as the mean lnk delay ncreases, the step style becomes vague. The lower lnk delay (e.g., smaller than 10 s) represents the case of the lght traffc condton (.e., low vehcle densty), whch s manly decded by the length of each road segment and traffc lghts. When the traffc grows heavy (.e., hgh vehcles densty), the traffc congeston ncreases the travel tme for each vehcle on each road segment, whch causes a longer lnk delay for each road segment. In addton, the step style shape s softened, shown as part of larger mean lnk delay (.e., larger than 10 s at x-axs) n Fg. 9(c). B. The Impact of Vehcle Number The results from Fg. 10 show that the E2E delays of the EV and GVs, and the lnk delays n all the schemes ncrease accordng to an ncrease n the number of vehcle runnng n the smulaton. In Fg. 10(a), the E2E delay of the EV for the other schemes ncreases dramatcally as the the number of runnng vehcles ncreases, but SAINT+ keeps t low as a constant-lke value almost all of the tme. Especally when vehcle densty s the hghest (e.g., the data 1300 at x-axs), n comparson wth SAINT, SAINT+ reduces the mean E2E delay of EV from s (SAINT) to s (SAINT+), about a 42.2% reducton, and has very small varance (.e., t s stable). If we consder the worst case at the hghest densty, compared to SAINT and Djkstra, the mean E2E delay of EV n SAINT+ s reduced by s (SAINT) and s (Djkstra), respectvely. Meanwhle, the E2E delay of GVs n SAINT+ gans very small degradaton, as shown n Fg. 10(b). Consderng the sgnfcant reducton n E2E delay of the EV n SAINT+, the tradeoff here s acceptable (we wll show ths tradeoff later n the secton). That s, agan analyzng the data at the hghest vehcle densty (.e., vehcle number of 1300), the E2E delay

13 1050 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 19, NO. 4, APRIL 2018 overall decreases accordng to ncreases n the speed lmt, such that as the vehcle speed lmt ncreases, the travel tme spent on each lnk decreases. Thus, the relatve lower mean lnk delay of SAINT+ verfes ts good performance n Fg. 12(a) and Fg. 12(b). Fg. 11. The Number of Accdents Handled Wthn Lmted Tme. for the GVs for SAINT+ ncreased from s to s, only about 30 s added (.e., 11.2% downgrade), compared wth SAINT. Fg. 10(c) llustrates the change of the mean lnk delay accordng to the ncrease of runnng vehcles. We found that the mean lnk delays for all schemes go up when the vehcle densty ncreases. However, when the number of vehcles reaches 700 or above, the dfference of the mean lnk delay among Djkstra, SAINT, RTP, and SAINT+ s enlarged, whch confrms the mprovement of SAINT+ n both E2E delay of EV and GVs to Djkstra n the hgh densty stuaton. From the observaton n the smulaton, the mean lnk delay of SAINT+ s smaller than that of SAINT, and ths small gap on each lnk accounts for the overall reducton of E2E delay for EV. Fg. 11 shows the number of accdents handled from each scheme wthn the smulaton tme. We can see that SAINT+ mantans good performance even though the vehcle densty s hgh, whereas such a number from the other schemes gradually decreases. The mprovement from SAINT+ s due to the optmzed EV delvery scheme n whch the EV s route can be protected, and so the delvery speed s mproved. Therefore, wthn lmted tme, more accdents can be handled. C. The Impact of Vehcle Speed Lmt To assess the mpact of the vehcle speed lmt, we measured the metrcs by varyng the vehcle s speed lmt from 20 km/h to 80 km/ h wth a step of 5 km/ h (.e., from 5.56 m/s to m/s wth a step of about 1.39 m/s). Usually, 80 km/h s the hghest speed lmt n an urban area, whch s why t s selected as the upper bound. As dsplayed n Fg. 12(a), the E2E delay of the EV s constantly below that of the other schemes on the whole x-axs, whch shows a promsng performance. Comparng to SAINT wth a speed lmt of m/s, the E2E delay of SAINT+ was mproved by 13.5%. Also, though Djkstra has a hgh varaton, SAINT+ and SAINT exhbt a very stable E2E delay regardless of the speed change. In Fg. 12(b), the E2E delay of GVs decreases as the speed lmt of vehcles ncreases. Although mprovement n the E2E delay of the EV s acheved, the E2E delay of GVs of SAINT+ are smlar to that of SAINT, and much better than that of Djkstra. For the mean lnk delay n Fg. 12(c), t D. The Impact of Vehcle Acceleraton We also examned the mpact of vehcle acceleraton. Fg. 13 shows the metrcs comparson among the schemes under the change of acceleraton. Smlarly, the E2E delay of EV of SAINT+ s constantly better than those of SAINT and Djkstra for all test ponts. There s about a 15% mprovement for SAINT+. The E2E delay of the GVs for SAINT+ s smlar to that of SAINT when the acceleraton s lower than 2.5 m/s 2, and slghtly worse than that of SAINT when the acceleraton ncreases. The mean lnk delays of SAINT+ shown n Fg. 13(c) show a smlar pattern wth those of the speed lmt n Fg. 12(c). SAINT+ n all cases has a short mean lnk delay than other schemes. E. The Tradeoff of E2E Delay In order to show how much SAINT+ mproves the mean E2E delay of the EV n comparson wth SAINT, as well as the tradeoff of E2E delay between the EV and GVs, we nvestgated the E2E delay reducton of SAINT+ aganst vehcle densty, speed lmt, and acceleraton llustrated n Fg. 14. From Fg. 14(a) we can see that the E2E delay reducton of SAINT+ ncreases as the number of vehcles ncreases, whereas the E2E delay of the GVs ncreases moderately. For example, when the number of vehcle s 1200, SAINT+ reduces the mean E2E delay of EV by 30%, but that of the GVs ncreases by less than 5%. On the other hand, Fgs. 14(b) and 14(c) demonstrate the mpact of the vehcle speed lmt and acceleraton on the tradeoff of the E2E delay of EV. Fg. 14(b) shows that the E2E delay reducton decreases as the speed lmt ncreases. When the speed lmt s larger than 12.5 m/s, the E2E delay reducton shows a stable mprovement of about 14%. Meanwhle, the E2E delay on the GV sde shows a modest downgrade. Fg. 14(c) shows that as the acceleraton vares, the E2E delay reducton of EV fluctuates between 10% to 17%, whereas the GVs E2E delay does not show a dstngushable pattern. F. The Successful Delvery Rato of Multple EVs To show the robustness of SAINT+ n real-world emergency stuatons, we conducted smulatons wth another scenaro (the 2nd scenaro mentoned at the begnnng of ths secton) to calbrate the successful delvery rato of EVs by dspatchng several EVs to an accdent ste. As shown n Fg. 15, compared wth the other schemes, from a low densty to a hgh densty of vehcles, SAINT+ always obtans a 100% successful delvery rato of EVs. However, SAINT fals to fully delver all EVs to the accdent ste when the vehcle densty s hgh. Ths s because some EVs are jammed on the way to the accdent ste. Djkstra and RTP lack a mechansm for a

14 SHEN et al.: SAINT+ FOR EMERGENCY SERVICE DELIVERY OPTIMIZATION 1051 Fg. 12. The Impact of Vehcle Speed Lmt. (a) E2E Delay vs. Vehcle Speed Lmt of EV. (b) E2E Delay vs. Vehcle Speed Lmt of GVs. (c) Mean Lnk Delay vs. Vehcle Speed Lmt. Fg. 13. The Impact of Vehcle Acceleraton. (a) E2E Delay vs. Vehcle Acceleraton of EV. (b) E2E Delay vs. Vehcle Acceleraton of GVs. (c) Mean Lnk Delay vs. Vehcle Acceleraton. Fg. 14. The Tradeoff Between EV and GVs n Terms of E2E Delay of SAINT+ Compared wth SAINT. (a) E2E Delay Reducton of EV and GVs vs. the Number of Vehcles. (b) E2E Delay Reducton of EV and GVs vs. Vehcle Speed Lmt. (c) E2E Delay Reducton of EV and GVs vs. Vehcle Acceleraton. Fg. 15. The Performance of Multple EVs Delvery. (a) The Comparson of the Successful Delvery Rato of EVs. (b) E2E Delay of Multple EVs Delvery. dedcated reroutes for the EV and accdent protecton, so the success rate s low when the densty of vehcles grows. On the other hand, Fg. 15(b) shows the E2E delay of multple EVs for the schemes. The performance of SAINT+ s much better than SAINT when the vehcle densty ncreases. For the Djkstra and RTP, the performance s acceptable when the traffc s lght. However, when the densty ncreases, ther performance seems to mprove because of the lack of data due to the plummetng of successful delvery ratos, as seen n Fg. 15(a). However, the E2E delay can be computed only by the several ntal EVs, so the actual performance becomes worse.

15 1052 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 19, NO. 4, APRIL 2018 VIII. CONCLUSION Ths paper proposes an evolved Self-Adaptve Interactve Navgaton Tool (SAINT+) for emergency servce delvery and accdent area protecton. Based on our prevous work, SAINT, SAINT+ optmzes the emergency servce delvery by ntentonally ncreasng the congeston level of an accdent area and the path of an emergency vehcle n an urban dstrct. Other vehcles are detoured around the accdent area and the path of the emergency vehcle. Further, the formaton of accdent protecton zones s suggested to protect the accdent area and mprove the delvery effcency of the emergency servce. In addton, SAINT+ reduces the mpact of the accdent by leveragng a dynamc traffc flow control model for the accdent protecton zones. Through the extensve and realstc smulatons, the results demonstrate that SAINT+ outperforms other state-of-the-art schemes for the travel delay of the emergency vehcle(s). Wth the SAINT+ servce, the effcency of the emergency servce delvery can be mproved, and the mpact of the accdent upon the other vehcles can be reduced. For future work, we wll study a more accurate watng tme model of the traffc lghts, snce t s related to the accuracy of the congeston contrbuton model. We are also nterested n an approprate traffc lght control scheme n order to globally schedule traffc lght for the emergency servce delvery. Moreover, we wll nvestgate a traffc-lght-free ntersecton passng scheme based on the SAINT+ navgaton system as another nterestng drecton for enhancng effcent traffc flow at ntersectons. ACKNOWLEDGMENT The authors thank Dr. Junghyun (Peter) Jun for hs valuable comments. REFERENCES [1] Natonal Center for Statstcs and Analyss, Traffc safety facts 2014: A complaton of motor vehcle crash data from the fatalty analyss reportng system and the general estmates system, Nat. Hghway Traffc Safety Admn, Washngton, DC, USA, Tech. Rep. DOT HS , Mar [2] E. Uhlemann, Introducng connected vehcles [connected vehcles], IEEE Veh. Technol. Mag., vol. 10, no. 1, pp , Mar [3] A. Abdrabou and W. Zhuang, Probablstc delay control and road sde unt placement for vehcular ad hoc networks wth dsrupted connectvty, IEEE J. Sel. Areas Commun., vol. 29, no. 1, pp , Jan [4] R. Kru, Accessng cloud computng resources over 4G LTE, Dept. Inf. Technol., Helsnk Metropola Unv. Appl. Sc., Helsnk, Fnland, Tech. Rep , May [5] C. D. Monfred, The LTE network archtecture A comprehensve tutoral, Alcatel-Lucent, Boulogne-Bllancourt, France, Tech. Rep., [Onlne]. Avalable: rdantu/fall_2013_ WIRELESS_NETWORKS/LTE_Alcatel_Whte_Paper.pdf [6] Dstrct 6-0 Traffc Management Center, accessed [Onlne]. Avalable: [7] J. Zhao and G. Cao, VADD: Vehcle-asssted data delvery n vehcular ad hoc networks, IEEE Trans. Veh. Technol., vol. 57, no. 3, pp , May [8] J. Jeong, S. Guo, Y. Gu, T. He, and D. H. C. Du, Trajectory-based data forwardng for lght-traffc vehcular ad-hoc networks, IEEE Trans. Parallel Dstrb. Syst., vol. 22, no. 5, pp , May [9] J. Jeong, S. Guo, Y. Gu, T. He, and D. H. C. Du, Trajectorybased statstcal forwardng for multhop nfrastructure-to-vehcle data delvery, IEEE Trans. Moble Comput., vol. 11, no. 10, pp , Oct [10] J. P. Jeong, T. He, and D. H. C. Du, TMA: Trajectory-based multanycast for multcast data delvery n vehcular networks, Comput. Netw., vol. 57, no. 13, pp , Sep [11] J. P. Jeong et al., TPD: Travel predcton-based data forwardng for lght-traffc vehcular networks, Comput. Netw., vol. 93, no. P1, pp , Dec [12] NHTSA. U.S. DOT Advances Deployment of Connected Vehcle Technology to Prevent Hundreds of Thousands of Crashes, accessed [Onlne]. Avalable: [13] Garmn. Dedcated Navgator, accessed [Onlne]. Avalable: [14] TomTom. Dedcated Navgator, accessed [Onlne]. Avalable: [15] Waze. Smartphone App for Navgator, accessed [Onlne]. Avalable: [16] Navfree. Free GPS Navgaton for Androd Smartphone, accessed [Onlne]. Avalable: [17] E. W. Djkstra, A note on two problems n connexon wth graphs, Numer. Math., vol. 1, no. 1, pp , Dec [18] J. Jeong, H. Jeong, E. Lee, T. Oh, and D. H. C. Du, SAINT: Selfadaptve nteractve navgaton tool for cloud-based vehcular traffc optmzaton, IEEE Trans. Veh. Technol., vol. 65, no. 6, pp , Jun [19] NAVI. NAVI Navgaton System, accessed [Onlne]. Avalable: [20] Skobbler. Smartphone App for GPS Navgaton and Maps, accessed on [Onlne]. Avalable: androd [21] Tmap. SKT Smartphone Navgator, accessed [Onlne]. Avalable: [22] S. Wang, S. Djahel, Z. Zhang, and J. McMans, Next road reroutng: A multagent system for mtgatng unexpected urban traffc congeston, IEEE Trans. Intell. Transp. Syst., vol. 17, no. 10, pp , Oct [23] J. Pan, I. S. Popa, and C. Borcea, DIVERT: A dstrbuted vehcular traffc re-routng system for congeston avodance, IEEE Trans. Moble Comput., vol. 16, no. 1, pp , Jan [24] G. Km, Y. S. Ong, T. Cheong, and P. S. Tan, Solvng the dynamc vehcle routng problem under traffc congeston, IEEE Trans. Intell. Transp. Syst., vol. 17, no. 8, pp , Aug [25] L.-W. Chen and C.-C. Chang, Cooperatve traffc control wth green wave coordnaton for multple ntersectons based on the Internet of vehcles, IEEE Trans. Syst., Man, Cybern., Syst. [26] J. Pan, I. S. Popa, K. Zetoun, and C. Borcea, Proactve vehcular traffc reroutng for lower travel tme, IEEE Trans. Veh. Technol., vol. 62, no. 8, pp , Oct [27] P. E. Hart, N. J. Nlsson, and B. Raphael, A formal bass for the heurstc determnaton of mnmum cost paths, IEEE Trans. Syst. Sc., vol. 4, no. 2, pp , Jul [28] J. Y. Yen, Fndng the k shortest loopless paths n a network, Manage. Sc., vol. 17, no. 11, pp , Jul [29] M. Wang, H. Shan, R. Lu, R. Zhang, X. Shen, and F. Ba, Realtme path plannng based on hybrd-vanet-enhanced transportaton system, IEEE Trans. Veh. Technol., vol. 64, no. 5, pp , May [30] C.-Y. Chen, P.-Y. Chen, and W.-T. Chen, A novel emergency vehcle dspatchng system, n Proc. IEEE 77th Veh. Technol. Conf. (VTC Sprng), Jun. 2013, pp [31] G. Tan, Y. Ren, and M. Zhou, Dual-objectve schedulng of rescue vehcles to dstngush forest fres va dfferental evoluton and partcle swarm optmzaton combned algorthm, IEEE Trans. Intell. Transp. Syst., vol. 17, no. 11, pp , Nov [32] G. Tan, M. Zhou, P. L, C. Zhang, and H. Ja, Multobjectve optmzaton models for locatng vehcle nspecton statons subject to stochastc demand, varyng velocty and regonal constrants, IEEE Trans. Intell. Transp. Syst., vol. 17, no. 7, pp , Jul [33] N. Malvya, S. Madden, and A. Bhattacharya, A contnuous query system for dynamc route plannng, n Proc. IEEE 27th Int. Conf. Data Eng., Apr. 2011, pp [34] J. Xu, Y. Gao, C. Lu, L. Zhao, and Z. Dng, Effcent route search on herarchcal dynamc road networks, Dstrb. Parallel Databases, vol. 33, no. 2, pp , [Onlne]. Avalable: /s x [35] Y. L. Morgan, Notes on DSRC & WAVE standards sute: Its archtecture, desgn, and characterstcs, IEEE Commun. Surveys Tuts., vol. 12, no. 4, pp , 4th Quart., 2010.

16 SHEN et al.: SAINT+ FOR EMERGENCY SERVICE DELIVERY OPTIMIZATION 1053 [36] ETSI. DSRC Standardzaton, accessed [Onlne]. Avalable: [37] Y. Yao, L. Rao, X. Lu, and X. Zhou, Delay analyss and study of IEEE p based DSRC safety communcaton n a hghway envronment, n Proc. IEEE INFOCOM, Apr. 2013, pp [38] Y. Naga, L. Zhang, T. Okamawar, and T. Fuj, Delay performance analyss of LTE n varous traffc patterns and rado propagaton envronments, n Proc. IEEE 77th Veh. Technol. Conf. (VTC Sprng), Jun. 2013, pp [39] J. Jeong and E. Lee, VCPS: Vehcular cyber-physcal systems for smart road servces, J. Korean Inst. Commun. Sc., vol. 31, no. 3, pp , Feb [40] A. Polus, A study of travel tme and relablty on arteral routes, Transportaton, vol. 8, no. 2, pp , Jun [41] M. DeGroot and M. Schervsh, Probablty and Statstcs, 3rd ed. Readng, MA, USA: Addson-Wesley, [42] G. Dmtrakopoulos and P. Demestchas, Intellgent transportaton systems, IEEE Veh. Technol. Mag., vol. 5, no. 1, pp , Mar [43] Research and Innovatve Technology Admnstraton (RITA). IntellDrve: Safer, Smarter and Greener, accessed [Onlne]. Avalable: [44] D. Krajzewcz, J. Erdmann, M. Behrsch, and L. Beker, Recent development and applcatons of SUMO Smulaton of urban moblty, Int. J. Adv. Syst. Meas., vol. 5, nos. 3 4, pp , Dec [45] M. Haklay and P. Weber, OpenStreetMap: User-generated street maps, IEEE Pervasve Comput., vol. 7, no. 4, pp , Oct [46] D. S. I. Scholz. (2016). Java OpenStreetMap Edtor. [Onlne]. Avalable: [47] DLR and Contrbutors. (2016). NETCONVERT. [Onlne]. Avalable: [48] The NEMSIS techncal assstance center. (2016). Natonal EMS Informaton System. [Onlne]. Avalable: reportngtools/reports/natonalreports/createarep%ort.html [49] S. Krauss, P. Wagner, and C. Gawron, Metastable states n a mcroscopc model of traffc flow, Phys.Rev.E,Stat.Phys.PlasmasFluds Relat. Interdscp. Top., vol. 55, pp , May [Onlne]. Avalable: [50] SUMO. Smulaton of Urban Moblty, accessed [Onlne]. Avalable: Ywen (Chrs) Shen receved the B.S. degree from the Department of Communcaton Engneerng and the M.S. degree from the Department of Mechatroncs Engneerng, North Unversty of Chna, n 2009 and 2013, respectvely. He s currently workng toward the Ph.D. degree wth the Department of Computer Scence and Engneerng, Sungkyunkwan Unversty, South Korea. Hs research areas nclude ntellgent transportaton systems, wreless networks, and VANET. He receved the Chnese Government Scholarshp from the Chna Scholarshp Councl. Hohyeon Jeong receved the B.S. and M.S. degrees from the School of Electrcal and Computer Engneerng, Ajou Unversty, South Korea, n 2012 and 2014, respectvely. He s currently workng toward the Ph.D. degree wth the Department of Computer Scence and Engneerng, Sungkyunkwan Unversty, South Korea. Hs research area s self-adaptve software systems. Jaehoon (Paul) Jeong (M 14) receved the B.S. degree from the Department of Informaton Engneerng, Sungkyunkwan Unversty, South Korea, n 1999; the M.S. degree from the School of Computer Scence and Engneerng, Seoul Natonal Unversty, South Korea, n 2001; and the Ph.D. degree from the Department of Computer Scence and Engneerng, Unversty of Mnnesota, n He s currently an Assstant Professor wth the Department of Software, Sungkyunkwan Unversty. Hs research areas are vehcular networks, wreless sensor networks, and moble ad hoc networks. He s a member of the ACM and the IEEE Computer Socety. Hs two data-forwardng schemes (called TBD and TSF) for vehcular networks were selected as spotlght papers n IEEE TRANS- ACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS n 2011 and IEEE TRANSACTIONS ON MOBILE COMPUTING n 2012, respectvely. Eunseok Lee receved the B.S. degree from the Department of Electronc Engneerng, Sungkyunkwan Unversty, South Korea, n 1985, and the M.S. and Ph.D. degrees from the Department of Informaton Engneerng, Tohoku Unversty, Japan, n 1991 and 1988, respectvely. He was an Assstant Professor wth Tohoku Unversty, Japan. He was a Research Scentst wth Mtsubsh Electrc Corporaton. He s currently a Professor wth the Department of Computer Engneerng, Sungkyunkwan Unversty. Hs research areas are self-adaptve software systems, software testng, and autonomc computng. Jnho Lee receved the B.S. degree from the Department of Mathematcs, Sungkyunkwan Unversty, South Korea, n 2014, where he s currently workng toward the M.S. degree wth the Department of Computer Scence and Engneerng. Hs research areas nclude ntellgent transportaton systems and VANET. DavdH.C.Dureceved the B.S. degree n mathematcs from Natonal Tsng Hua Unversty, Tawan, n 1974, and the M.S. and Ph.D. degrees n computer scence from Unversty of Washngton, Seattle, n 1980 and 1981, respectvely. He s the Qwest Char Professor wth the Computer Scence and Engneerng Department, Unversty of Mnnesota, Mnneapols. Hs research nterests nclude cyber securty, sensor networks, multmeda computng, storage systems, and hgh-speed networkng.