AS an important part of intelligent transportation system

Transcription

1 734 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO. 1, JANUARY 2017 Chael Allocatio for Adaptive Video Streamig i Vehicular Networks Log Su, Haggua Sha, Member, IEEE, Aipig Huag, Seior Member, IEEE, Li Cai, Seior Member, IEEE, ad Hogli He Abstract Video services i vehicular etworks play a importat role i future itelliget trasportatio systems ad vehicular ifotaimet systems. Yet, at the presece of other services with high priorities, the remaiig radio resources for video services are highly dyamic. To support video service of multiple vehicles i vehicular etworks, we propose a joit chael allocatio ad adaptive video streamig algorithm that makes the vehicles compete for chael access opportuities ad to request video data with a proper visual quality accordig to their utilities. A vehicle s request is determied by takig several key factors ito cosideratio, icludig the locatio ad the velocity of the vehicle, the activity of the high-priority services, the itesity of the competitio amog multiple vehicles, ad the smoothess requiremet of visual quality. Simulatio results show that the proposed algorithm is superior to the existig algorithms i both iterruptio ratio ad visual quality. Idex Terms Auctio, chael allocatio, scalable video codig (SVC), video streamig. I. INTRODUCTION AS a importat part of itelliget trasportatio system (ITS), vehicular etworks allow commuicatios betwee vehicles ad roadside uits (RSUs) ad support various applicatios [1], [2] that improve the safety ad efficiecy of public trasportatio [3] [5]. Video services over vehicular etworks play a key role i driver assistace, emergecy iformatio delivery, ad ifotaimet distributio. For example, cotetrich video streams ca be used i road coditio moitorig for driver assistace ad traffic safety, the real-time moitorig of Mauscript received July 20, 2015; revised December 19, 2015; accepted Jauary 29, Date of publicatio February 29, 2016; date of curret versio Jauary 13, This work was supported i part by the Natioal Key Basic Research Program of Chia uder Grat 2012CB316104; by the Natioal Sciece ad Techology Specific Major Projects uder Grat 2015ZX ; by the Natioal Hi-Tech R&D Program of Chia uder Grat 2014AA01A702; by the Natioal Natural Sciece Foudatio of Chia uder Grat ; by the Fudametal Research Fuds for the Cetral Uiversities uder Grat 2015XZZX001-02; ad by a research grat from the Natural Scieces ad Egieerig Research Coucil of Caada. The review of this paper was coordiated by Prof. H. Nishiyama. L. Su, H. Sha, A. Huag, ad H. He are with the Istitute of Iformatio ad Commuicatio Egieerig, Zhejiag Uiversity, Hagzhou , Chia, ad also with Zhejiag Provicial Key Laboratory of Iformatio Processig, Commuicatio ad Networkig, Hagzhou , Chia ( hades@zju.edu.c; hsha@zju.edu.c; aipig.huag@zju.edu.c; @zju.edu.c). L. Cai is with the Departmet of Electrical ad Computer Egieerig, Uiversity of Victoria, Victoria, BC V8W 3P6, Caada ( cai@ece.uvic. ca). Color versios of oe or more of the figures i this paper are available olie at Digital Object Idetifier /TVT the iside ad the surroudigs of a cash truck or a jewelry truck, the remote medical guidace for emergecy treatmet i a ambulace, ad the real-time televisio recordig ad broadcastig of a public evet or a atural disaster [6]. Providig video service over vehicular etworks also raises ew challeges to commuicatio ad etworkig techologies. Geerally, video service demads stable ad sufficiet badwidth due to its large data volume, log playback duratio, ad the user expectatio of smooth playback ad high visual quality [7]. Yet, the radio resources for video services are limited ad highly dyamic, due to the presece of other services with higher priority (e.g., safety-related applicatios). The radio resources are further straied whe multiple vehicles are demadig video services. Further, the playback of video services should be uiterrupted, but RSU deploymet might ot be dese eough, ad the data trasmissio (DT) will be iterrupted whe a vehicle drives out of the coverage areas of RSUs. To address all the give challeges, a efficiet chael allocatio ad video streamig algorithm is idispesable for multiple vehicles i a vehicular etwork with highly dyamic radio resources. Several video streamig algorithms have bee proposed to support smooth video playback i vehicular etworks, but most of them simplified or eve igored the cotetio amog multiple vehicles. To alleviate the shortage of radio resources, some researchers adopted spectrum-sharig techologies [8], [9], which, however, may ot be directly applicable to vehicular etworks where the vehicles move fast ad the topology chages quickly. Researches o chael allocatio amog multiple vehicles i vehicular etworks have bee doe to improve the etwork performace, but most of them did ot cosider the characteristics of video services. I this paper, we study chael allocatio ad video streamig i a vehicular etwork with highly dyamic radio resources, i which both the umber of available chaels ad that of vehicles i the coverage area of a RSU chage with time. The RSU deploymet is sporadic; thus, vehicles face highly dyamic competitio to access chaels. We propose a auctio-based chael allocatio ad adaptive video streamig algorithm. To esure smooth video playback outside the coverage areas of RSUs, we adopt a fiite-legth buffer i each vehicle to store video data. To provide better video services uder dyamic chael coditios, we employ the scalable video codig (SVC) extesio of the H.264 stadard [10] as the ecodig scheme ad ecode the video trace ito several layers. Therefore, the video data with more layers ad, thus, higher visual quality ca be trasmitted uder better chael IEEE. Persoal use is permitted, but republicatio/redistributio requires IEEE permissio. See for more iformatio.

2 SUN et al.: CHANNEL ALLOCATION FOR ADAPTIVE VIDEO STREAMING IN VEHICULAR NETWORKS 735 coditio ad/or less itesive competitio, whereas video data with a fewer umber of layers ca be trasmitted to avoid playback iterruptio uder worse chael coditios ad/or itesive competitio. To support multiple vehicles with limited radio resources, we employ a auctio mechaism [11] that makes the vehicles bid accordig to their utility values so that RSUs ca allocate the available chaels efficietly. We allow the vehicles to request a proper umber of video layers to support smooth playback ad high visual quality accordig to the trasmissio rate, the umber of data uits i the buffer, the umber of data uits eeded for video playback, ad the possibility to access the chael i the future. To describe the characteristics of vehicles drive-thru ad video playback, we use a stochastic game to model the states ad state trasitio probabilities of the vehicles. To deal with the competitio amog vehicles, we adopt a statistic method to model the eviromet states with the help of historical records. To calculate the optimal request of a vehicle, we adopt low-complexity dyamic programmig. The mai cotributios of this paper are fourfold. The auctio mechaism is adopted i the chael allocatio problem i vehicular etworks. As each vehicle bids accordig to its utility value, the RSU ca allocate the chaels reasoably. Vehicles calculate their ow bids; therefore, the computatio is distributed, ad the solutio eables real-time services. The problem of multivehicle chael access is modeled as a stochastic game. Several key factors are cosidered i the model, icludig the umber of available chaels, which is affected by the activity of high-priority services, the umber of vehicles i the coverage area of a RSU, ad the vehicle states. The vehicle state cotais the iformatio o the locatio of the vehicle, the amout of data uits i the buffer, ad the requested umber of video layers i the last successful chael access. A utility fuctio is proposed to reflect the experiece of vehicle ad, thus, to guide each vehicle to determie its bid ad the required umber of video layers. The utility fuctio is desiged to reflect two factors, amely, the video iterruptio ratio ad visual quality. A method is proposed for vehicles to bid reasoably by modelig the eviromet states ad estimatig the bids of the others through historical records. The remaider of this paper is orgaized as follows. I Sectio II, the related works are itroduced. Sectio III describes the system model. Sectio IV presets the chael allocatio ad adaptive video streamig algorithm, ad Sectio V discusses the problem solvig. Performace evaluatio is preseted i Sectio VI, followed by the coclusios i Sectio VII. without cosiderig the effect of chagig the frame rate o the user experiece. The bouds of iterruptio probability was ivestigated i [14]. To achieve high visual quality, a adaptive video streamig algorithm was preseted i [15], which cosiders the relatioship betwee the total data volume to be obtaied whe drivig i the coverage area of a RSU ad the data volume eeded to play video with a certai umber of layers. I [16], a approach of trasmittig the basemet layer of the video via a vehicular etwork ad trasmittig the ehacemet layers through LTE etwork was proposed. I [17], a scheme was proposed to trasmit the basemet layer of the video by multihop, ad to obtai high peak-sigal-to-oise ratio (PSNR) for as may vehicles as possible by adjustig the modulatio mode of the ehacemet layers of the video trace. To achieve both low iterruptio ratio ad high visual quality, a thresholdbased adaptive video streamig algorithm utilizig the iformatio of the queue legth ad curret trasmissio rate was proposed i [18]. A predictio-widow-based video streamig algorithm was preseted i [19], which adjusts the requested umber of video layers accordig to the relatioship betwee the data volume that the vehicle eeds for video playback ad that it expects to obtai. Most of the existig works igored or simplified the competitio agaist other vehicles ad thus are difficult to be used i the scearios where a ever-chagig umber of vehicles compete for chaels. There have bee extesive research works aimig at chael allocatio for vehicle-to-rsu commuicatios. To improve overall system throughput, a chael allocatio algorithm was preseted i [20], which takes the curret traffic loads ad bit error rates of differet service chaels ito cosideratio. I [21], a desity-adaptive medium-access-cotrol protocol that predicts the vehicle-desity dyamics was proposed. A schedulig algorithm called MV-MAX, which opportuistically grats wireless access to the vehicles with the highest trasmissio rate, was proposed i [22]. I [23], the throughput capacity of vehicular ad-hoc etworks exploitig mobility diversity was aalyzed. For vehicle-to-rsu osafety service, a schedulig algorithm utilizig the iformatio of the queue legths ad chael qualities to trasmit more data was preseted i [24]. A game-based allocatio algorithm was proposed i [25], i which the vehicles optimize their requests to miimize their costs uder the costraits of the required quality of services. To decrease the packet loss rate, a chael allocatio protocol was proposed i [26] based o vehicles locatios ad their remaiig time slots i a RSU. However, for video service, high etwork throughput, low cost, ad low packet loss rate do ot mea smooth video playback ad high visual quality. This motivates us to desig a joit chael allocatio ad video streamig algorithm cosiderig user experiece. II. RELATED WORK Video streamig i vehicular etworks has bee ivestigated i the literature, cosiderig how to esure uiterrupted playback ad high visual quality. To reduce the iterruptio ratio, a algorithm for vehicular etworks that adjusts the retrasmissio times i the medium access cotrol protocol was proposed i [12]. A frame-rate adaptatio algorithm was proposed i [13], III. SYSTEM MODEL A oe-way highway sceario as show i Fig. 1 is cosidered. It cosists of RSUs, vehicles, ad a video server. The video server is coected to RSUs via a backhaul with sufficiet badwidth, ad RSUs commuicate with vehicles via air. RSUs are deployed alog the road, ad the distace betwee two adjacet RSUs is X R2R. It is assumed that the

3 736 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO. 1, JANUARY 2017 Fig. 2. Buffer storage. Fig. 1. Highway sceario ad chael model. iterarrival distace of vehicles follows a Poisso distributio with the mea of λ [12], [27]. Each vehicle has a uique ID. A GPS device is istalled o each vehicle to obtai the realtime positio ad velocity. It is also assumed that a vehicle ca sychroize with a RSU whe it eters the RSU s coverage area ad receives the RSU s beaco message. Therefore, the vehicle works i a sychroized maer with the RSU whe drivig i the RSU s coverage area. The radius of a RSU s coverage area is deoted by R, where R<X R2R /2. Therefore, some parts of the road are outside the RSUs coverage area. It is assumed that, usig appropriate adaptive modulatio ad codig, as log as the distace betwee a vehicle ad a RSU is less tha R, the vehicle ca receive RSU s sigal without error. It is assumed that the badwidths of all data chaels are equal, ad the trasmissio rate is oly affected by path loss ad depeds o the distace betwee the vehicle ad the RSU. The same assumptio ad settig ca also be foud i [28] ad [29]. Therefore, the road segmet i the coverage area of a RSU is divided ito K regios, as show i Fig. 1. The trasmissio rate of ay data chael betwee ay vehicle i regio k ad the RSU remais costat ad is deoted r k, k = 1, 2,...,K. The farther a vehicle is away from the RSU, the lower the rate r k is, e.g., r 1 = r K <r 2 = r K 1. There is a dedicated chael called commo cotrol chael (CCC), whose badwidth is much arrower tha that of the data chaels. I additio, there are a total of M data chaels for the RSU to trasmit video data to a vehicle whe they are ot occupied by ay higher priority service. The state of a data chael, occupied or ot occupied by higher priority services, is modeled as a two-state Markov chai. It is assumed that the states of the M data chaels are idepedet of each other. At ay time, oe chael is allocated to oe vehicle oly. The video trace stored i the server is divided ito data uits with equal duratio of 1/μ secods, amely μ data uits are played per secod. Each data uit is ecoded ito L layers usig SVC techology, i.e., oe basemet layer ad L 1 ehacemet layers. The source rate of the data uit with l layers (l = 1, 2,...,L) is deoted τ l, ad τ i <τ j if i<j. The umber of video layers is called quality level for short. The average PSNR of the data uits with quality level l is deoted Q l, ad Q i <Q j if i<j. The two factors affectig the quality of user experiece most are the playback iterruptio ratio ad visual quality. The visual quality is usually objectively measured by PSNR, ad a higher value of PSNR meas a higher visual quality. For the Fig. 3. Slot structure. SVC ecoded data uits, more video layers result i a higher PSNR. The buffer i each vehicle is divided ito uits each with the size of B max, as show i Fig. 2. A buffer uit (black block) is large eough to store a data uit with L layers ad a duratio of B max /μ secods. A data uit (blue block) with l layers ad a duratio of B max /μ secods occupies a buffer uit, ad its volume icreases as l icreases, where l = 1, 2,...,L. Such a buffer divisio ad applicatio ca also be foud i [30] to represet the vehicle states. It is assumed that the vehicular etwork is a time-slotted system, ad the slot legth is deoted Δ. Both the umber of vehicles withi the coverage area of a RSU ad the umber of available chaels remai costat durig a slot. The slot structure of vehicular etwork is show i Fig. 3. A slot cotais a resource auctio phase of duratio Δ auc secods ad a DT phase of duratio Δ data secods. The resource auctio phase cosists of four subphases. The first is beaco broadcast (BB) subphase, i which a RSU broadcasts a beaco massage to make the vehicles withi its coverage area sychroized. Secod is the iformatio broadcast (IB) subphase, i which a RSU broadcasts commo iformatio, icludig the umber of vehicles withi its coverage area, the umber of available chaels, ad the check list of ID ad miislot correspodece, to facilitate chael cotetio i the curret slot. Third is the request upload (RU) subphase, which cosists of N max miislots [25]. Here, each miislot is used by a sigle vehicle to sed its request cotaiig both bid ad quality level, ad N max stads for the maximal umber of vehicles i a RSU s coverage area. A RSU ca be aware of the umber of vehicles withi its coverage area because the vehicle eterig its coverage area must iform the RSU its ID whe CCC is idle. Fourth is the feedback (FB) subphase, i which a RSU broadcasts chael allocatio results ad the average bid of all the vehicles that wi the auctio. For better readability, a otatio list of the key symbols used i this paper is provided i Table I.

4 SUN et al.: CHANNEL ALLOCATION FOR ADAPTIVE VIDEO STREAMING IN VEHICULAR NETWORKS 737 TABLE I NOTATION LIST IV. CHANNEL ALLOCATION AND ADAPTIVE VIDEO STREAMING ALGORITHM Whe multiple vehicles eed to access chael to obtai video data, limited ad ever-chagig radio resources of a RSU may oly meet the demads of part of the vehicles. Cosiderig the relative positios betwee the vehicles ad the RSU, the chaels should be reasoably allocated to 1) the vehicles that have already ecoutered or are about to ecouter playback iterruptio ad 2) the vehicles that are closer to the RSU ad, thus, have a higher trasmissio rate tha those far away from the RSU. Allocatig to the former ca decrease the playback iterruptio ratio, whereas allocatig to the latter ca icrease the overall visual quality. With these two cosideratios, we propose a auctio-based chael allocatio mechaism i which the vehicles bid accordig to their ow utility values ad the RSU decides the optimal chael allocatio based o the bids. Auctio-based chael allocatio eables each vehicle to calculate its ow decisio; hece, the computatio complexity per vehicle ca be much lower as compared with a cetralized algorithm. Moreover, the oly iformatio that vehicles eed to upload are their requests, which decreases the overhead ad protects privacy. The auctio is coducted oce per slot. A adaptive video streamig algorithm is developed for a vehicle to request a proper quality level, to get data uits esurig smooth playback ad high visual quality. The request of a vehicle is made based o the curret data volume i its buffer, the trasmissio rate, the data volume eeded for smooth playback, ad the possibility to access chael i the future. A. Chael Allocatio ad Video Streamig Here, the iformatio exchage procedure betwee RSU ad vehicles is described, followed by the discussio o two key factors affectig the vehicles decisios, amely, the budget ad the quality level. Fig. 4. Iformatio exchage procedure i each slot. The iformatio exchage procedure is coducted i each slot betwee the RSU ad the vehicles i its coverage area, as show i Fig. 4. After broadcastig a beaco message i the BB subphase, the RSU broadcasts a vector b 1 =[M,N,W 1,..., W Nmax ] i the IB subphase, where M ( M) deotes the umber of available chaels, N ( N max ) is the umber of the vehicles withi the coverage area of the RSU, ad W is the ID of the vehicle who will upload its request i the th miislot, with = 1,...,N max. I the RU subphase, vehicle uploads its ecrypted request A i the miislot assiged to it. The request A icludes both the bid C ad the quality level l.the RSU puts requests from all vehicles ito a request matrix [A ]. If N<N max, for some elemets A s, we have C =0, ad l =0. The RSU allocates the M available chaels to the M vehicles who ted the highest bids ad prepares data uits of the quality levels specified i the requests. I the FB subphase, the RSU broadcasts a vector b 2 =[ C,O 1,O 2,...,O Nmax ], where O =m ( {1, 2,...,M}) meas that vehicle ca access chael m, ad O =0 represets that vehicle obtais o chael. C deotes the average bid of all the vehicles who wi the auctio ad ca also be called the average trasactio price. C reflects the competitio itesity; thus, it is provided to the vehicles as a referece for their future bids. I the DT phase, the RSU trasmits data uits o the M available chaels.

5 738 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO. 1, JANUARY 2017 Budget is a importat basis for the vehicles to bid. Whe drivig ito the coverage area of a RSU, a vehicle is offered some iitial budget of amout F max, which is o less tha the umber of the time slots that it drives withi the RSU s coverage. Whe a vehicle gets the opportuity to access a chael, its bid amout is deducted from its remaiig budget. The vehicle has to give up auctio if it has used up its budget util it drives ito the coverage of the ext RSU. The quality level affects the playback iterruptio ad visual quality. Requestig a high quality level leads to a small umber of data uits beig dowloaded, i.e., the visual quality is high at the cost of a shorter playable duratio, ad vice versa. Moreover, the quality level ad the bid affect each other. If the vehicle requests a high quality level, the buffer will obtai fewer data uits ad the estimated iterruptio probability will icrease, which leads to a higher bid i the ext slot. I additio, the quality level of the vehicle should also be adjusted accordig to its remaiig budget, which limits the chaces to the access chael i the future. Aother factor that affects the quality level is the positio of the vehicle, amely, the distace ad the trasmissio rate betwee the vehicle ad the RSU. Eve with the same quality level, the umbers of data uits received i a slot at differet regios are differet; thus, the correspodig playable duratios are differet. Thus, the bid ad quality level should be joitly decided. Givig a reasoable bid ad quality level at the appropriate time slot is the key to esure user experiece. B. Vehicle State ad Multiuser Game For each vehicle, the bid ad quality level are ot oly related to its state but affected by the other vehicles bids as well as the results of a multiuser game. The, we employ a stochastic game to model the chael allocatio ad video streamig problem. The stochastic game is defied as (< S, A, O, Q, U > N =1, M, N ) (1) where M = {0, 1,...,M} is the set of the umber of available chaels, N =(0, 1,...,N max ) deotes the set of the umber of possible vehicles, S represets the set of vehicle s states, A stads for the set of vehicle s requests, O = {0, 1,..., M} is the set of the possible chael allocatio results to vehicle, Q stads for the set of trasitio probabilities of vehicle s state, ad U is the set of utility values of vehicle. The utility fuctio will be detailed i Sectio IV-C. A elemet S i set S i (1) represets the state of vehicle ad is expressed as a quaterio, i.e., S = ( B,T,L last ),F (2) where B ( [0,B max ]) deotes the umber of data uits i the buffer at the begiig of the curret slot. T deotes the umber of slots left before vehicle drives out of the RSU s coverage area. T > 0 if vehicle drives withi a RSU s coverage area; otherwise, T = 0. L last represets the quality level vehicle obtaied i the last chael access. F stads for the remaiig budget of vehicle. A elemet A i set A i (1) represets the request of vehicle ad is deoted by a tuple A =(C,l ) (3) where l ( [0,L]) stads for the quality level vehicle requests i the curret slot ad C (0 C mi (F T,F limit )) represets the bid of vehicle, with F T represetig the left budget after guarateeig at least 1 uit of budget for each of the rest time slots, which guaratees that a vehicle must have some budget, as log as it drives withi the coverage of a RSU. l = 0 whe vehicle is outside of ay RSU s coverage area ad thus does ot request video data. 1 l L whe vehicle is withi the coverage area of a RSU. To smooth the variatios of visual quality, we let 1 (l L last ) 1 betwee successive data uits [31]. C = 0 oly happes whe vehicle is outside the RSUs coverage areas, idicatig that the vehicle caot bid. The bid rage of vehicle i RSUs coverage areas is 1 C mi (F T,F limit ). The lower boud of bid is desiged as 1 to icrease the chael utilizatio, amely, to avoid the situatio that the umber of competitio vehicles is less tha the umber of available chaels. F limit stads for the bid upper boud, which ca limit the competitio itesity. A elemet q (S S,O ) i set Q i (1) deotes the trasitio probability of vehicle from state S to state S give the chael allocatio result O ( O ). S is vehicle s state i the ext slot S =(B,T,L last,f ). Whe the chael allocatio result is kow, q (S S,O ) is determiistic with a value of either 0 or 1. The expressios of the umber of data uits i buffer B, the umber of slots left before drivig out the RSU s coverage area T, the quality level of the last successful access L last, the remaiig budget F of vehicle at the begiig of the ext slot, ad the correspodig variables i the curret slot are as follows: { [ ] mi μ B B max, (B + G τ = l D), if O 0 (4a) max(0,b D), if O = 0 { T T 1, if T > 0 = (4b) 0, if T = 0 { L last l, if O 0 = L last (4c), if O = 0 { F F C, if O 0 = (4d) F, if O = 0 where G = i+δ data y=i r ky dy deotes the data volume trasmitted to vehicle i the curret slot, with i stadig for the start time istat of the DT phase i the curret slot ad r ky represetig the trasmissio rate i road regio k y i time istat y; D deotes the umber of data uits eeded to smoothly play video i a slot ad is a costat for the give duratio of a slot ad the duratio of a data uit; mi[ ] i (4a) guaratees that the buffer will ot overflow; ad G μ/τ l stads for the received umber of data uits.

6 SUN et al.: CHANNEL ALLOCATION FOR ADAPTIVE VIDEO STREAMING IN VEHICULAR NETWORKS 739 C. Utility Fuctio The effect of both playback iterruptio ad visual quality o the user experiece should be cosidered whe a vehicle decides its bid ad the quality level. However, smooth playback ad high visual quality are cotradictory requiremets for give chael access opportuities ad trasmissio rates. To make a good tradeoff, a utility fuctio is proposed for vehicle to decide its bid ad quality level. The utility fuctio of vehicle is deoted U ( U ). Accordig to the vehicle state S ad request A, U is expressed as U (S,S,A )=α (B D) +(1 α) Ql Q L p(o 0 S,C ) (5) where S deotes the eviromet state to be defied i Sectio V-D, icludig the umber of available chaels i the curret slot, the umber of vehicles withi the coverage area of the RSU, ad the competitio itesity for a vehicle to access chael. The set composed of all eviromet states is deoted as S. Coefficiet α ( (0, 1)) is a weight to balace playback iterruptio ratio ad visual quality. It is usually set close to 1 because the effect of playback iterruptio o user experiece is much higher tha that of visual quality. Operatio (x) = mi{0,x}. p(o 0 S,C ) stads for the probability of vehicle obtaiig a chael give eviromet state S ad bid C. The first item i (5) is the utility decremet caused by playback iterruptio. It is egative if the umber of data uits B i buffer at the begiig of the curret slot is smaller tha the umber of data uits D eeded for smooth playback, ad is 0 otherwise. The secod item i (5) is the utility icremet expectatio whe quality level l is ordered ad is oegative. The larger l is, the larger Q l ad the utility value will be. Therefore, the utility value of (5) may be positive or egative. D. Request for Maximal Utility O the road segmets outside RSUs coverage areas, the vehicle keeps playig video usig the data uits already stored i its buffer. Therefore, whe decidig its bid ad quality level, vehicle should take both the storage ad the probability of gettig chael access opportuities ito cosideratio, ad maximize the mea of the total utility value accumulated (MTUA) from the curret slot (e.g., slot t) to the slot before drivig ito the coverage area of the ext RSU. The optimal request of vehicle at slot t ca be modeled as,t =arg max E A,t A A opt [ t ext i=t U (S,i,S,i,A,i ) } {{ } MTUA s.t. 1 C,i mi(f,i T,i,F limit ) l,t L last,t 1 ] (6a) (6b) (6c) Fig. 5. Descriptio of time poits ad time duratios. where the subscript i stads for slot idex, ad i = t, t + 1,...,t ext, with t ext deotig the idex of the slot before vehicle drivig ito the coverage area of the ext RSU, ad A opt,t represets the request that maximizes the MTUA of vehicle i slot t. Equatio (6b) implies that the bid i each slot should ot exceed either the remaiig budget or the bid upper boud F limit ad is o less tha oe. Equatio (6c) meas that the divergece of the curret quality level ad the quality level of the last successful chael access should be o larger tha oe. To explai the temporal relatioships i this paper, the time istats ad duratios i a vehicle s drive-thru are show i Fig. 5. The duratio from a vehicle drivig ito a RSU s coverage area to that of the ext RSU is called a period, ad a period cotais T RSU slots. t boud is the idex of the slot whe a vehicle just drives out of a RSU s coverage area. T out is the umber of slots whe the vehicle drives through the road outside the RSUs coverage areas, i.e., T out =t ext t boud +1. T i stads for the total umber of the slots of a vehicle i a RSU s coverage area, i.e., T i = T RSU T out. T is the umber of the remaiig slots i curret RSU s coverage area, i.e., T = t boud t, ad t is the idex of the curret slot. V. P ROBLEM SOLVING Here, we discuss the approach to solve the optimizatio problem (6). First, the mathematical expressio of the optimal request A opt,t is give. The, the item with Bellma equatio form i A opt,t is aalyzed, ad a fiite-horizo dyamic programmig method is proposed to obtai the optimal request. Fially, a method to obtai the eviromet states, ad their trasitio probabilities is give. A. Mathematical Expressio of A opt,t The optimal request should maximize the MTUA from slot t to slot t ext. Therefore, vehicle should predict its ow states ad the eviromet states of the followig slots from t + 1 to t ext ad calculate the weighted sum of the total utility value accumulated give A,t ad S,t with respect to the probability to the correspodig states. Here, the predictio should be doe based o vehicle s request to be determied

7 740 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO. 1, JANUARY 2017 ad its curret state, as well as eviromet state trasitio probability, ad S,t deotes the possible eviromet state i slot t. To this ed, (6a) is expaded as (7), show at the bottom of the page. Whe calculatig its request, vehicle has bee iformed by the RSU s broadcast of the umber of available chaels M ad the umber of vehicles withi the coverage area of the same RSU i the curret slot, but it does ot kow the other vehicles bids. Thus, vehicle should calculate the MTUA give S,t ad A,t with respect to the probability q ( S,t S,t 1,O,t 1 ). Because the eviromet state is partially kow to vehicle, the possible eviromet state S,t comes from a subset S of eviromet state set S, which is show as the first summatio i (7). As the eviromet state is completely ukow i slot t + 1, the possible eviromet state ca be ay elemet i eviromet state set S. Therefore, the symbol of eviromet from t to t ext stateislott + 1isS,t+1 i the secod summatio istead of S,t+1. Item (a) i (7) deotes the curret utility obtaied by adoptig request A,t. Item (b) is the probability weighted sum of item (c). Item (c) represets the maximal MTUA from slot t + 1toslott ext, whose probability is proportioal to the product of three probabilities: the probability p(o,t S,t,C,t ) that the chael allocatio result is O,t give eviromet state S,t ad vehicle s bid C,t,the probability q (S,t+1 S,t,O,t ) that vehicle s state trasforms from S,t to S,t+1 give the chael allocatio result O,t, ad the probability q (S,t+1 S,t,O,t ) that the eviromet state trasfers to S,t+1 from S,t give the chael allocatio result O,t. Item (c) ca be expaded as (8), show at the bottom of the page. B. Maximal MTUA With the form of Bellma equatio [32], (8) ca be rewritte as (9), show at the bottom of the page, where V (S,S ) deotes the maximal MTUA from slot t(< t boud ) uder the vehicle state S ad the eviromet state S to slot t ext. S ad S are the simplified forms of S,i ad S,i+1, respectively, where i is the idex of slot. The idex of slot is omitted i (9) A opt,t =arg max q ( S,t S,t 1,O,t 1 ) U (S,t, S ),t,a,t A,t A S,t S }{{} (a) (b) {}} { + p(o,t S,t,C,t )q (S,t+1 S,t,O,t )q (S,t+1 S t ext,t,o,t ) max U(S,j,S,j,A,j ) S,t +1 S S,t +1 S O,t O A,j A E j=t+1 }{{} (c) (7) max E A,j A t ext j=t+1 U (S,j,S,j,A,j ) = max A,t+1 A U(S,t+1,S,t+1,A,t+1 )+ q (S,t+2 S,t+1,O,t+1 )q (S,t+2 S,t+1,O,t+1 ) max A,g A E S,t+2 S S,t+2 S O,t+1 O [ t ext g=t+2 p(o,t+1 S,t+1,C,t+1 ) ] U (S,g,S,g,A,g ) (8) V (S,S )= max U(S,S,A )+ A A S S S S O O ( ) p (O S,C ) q (S S,O ) q S S,O ( ) V S,S (9)

8 SUN et al.: CHANNEL ALLOCATION FOR ADAPTIVE VIDEO STREAMING IN VEHICULAR NETWORKS 741 ad (11), show below, for the followig reaso. Whe calculatig V by (9), oly vehicle s states ad the eviromet states i two successive slots are ivolved, ad the calculatio is idepedet of specific slot; thus, the subscripts i ad i + 1 should ot appear. Likewise, whe updatig V usig (11), the subscripts i ad i+1 should ot appear either. Equatios (7) (9) show that, to decide the optimal request, all the slots before drivig ito the coverage area of the ext RSU are ivolved, ad a iteratio should be performed from slot t ext to slot t. Notice that from slot t boud to slot t ext, vehicle caot participate i auctio, its buffer caot be filled, ad the umber of data uits for playback is kow ad decreases i a fixed rate as slot idex icreases. Moreover, to obtai high visual quality, budget should ot be left after slot t boud 1. Thus, the maximal MTUA i this duratio ca be calculated accordig to vehicle s state i slot t boud as ( ) b,t V boud (S,t boud,s,t boud)= D T out α δ F,t boud,t boud t t ext (10) where S,t boud ad S,t boud stad for vehicle s state ad eviromet state i slot t boud, respectively; F,t boud deotes the remaiig budget of vehicle i slot t boud, i.e., wasted budget; b,t boud represets the umber of data uits i the buffer of vehicle i slot t boud ; δ represets puishmet coefficiet ad is a positive costat; b,t boud/d i (10) deotes the umber of slots that the data uits i buffer ca play without iterruptio; ad (b,t boud/d T out ) represets the umber of iterruptio slots whe vehicle drives o the road segmet outside RSUs coverage areas. Therefore, the first item i (10) is of egative value, ad more iterruptio slots will result i a larger utility decremet. The secod item i (10) is a puishmet item, meaig that more wasted budget produces a larger utility decremet. Therefore, this item ecourages ratioal budget allocatio as well as makig full use of the budget. From the two items i (10), it ca be judged that the value of (10) is less tha zero. The maximal MTUA from slot t boud to slot t ext ca be calculated by usig (10), ad the optimal request i the curret slot ca simply be calculated iteratively from slot t boud to slot t. C. Fiite-Horizotal Dyamic Programmig If (7) is cotiuously expaded accordig to (8) ad (9), computatio may be heavy. Therefore, we employ fiite-horizo dyamic programmig to calculate (9) ad the solve (7) with low computatio complexity. Dyamic programmig divides the target problem ito several subproblems ad solves it by solvig subproblems, as well as iteratig or cumulatig results of subproblems. The subproblems of our target problem V (S,S ) are the correspodig values V (S,S ) of the states i the ext slot. Therefore, we set up a maximal MTUA table, i which each item is a combiatio of vehicle state ad eviromet state (S,S ) ad the correspodig maximal MTUA. The dimesio of the table is Z v = Z S Z S, where Z S = T i L(B max + 1)(F max + 1) deotes the umber of vehicle s states, ad Z S represets the umber of eviromet states ad will be detailed i Sectio V-D. Whe calculatig (9), a vehicle just eeds to search V (S,S ) correspodig to (S,S ) i its maximal MTUA table. Whe the maximal MTUA table is iitialized, all the maximal MTUA values of state combiatios (S,S ), except for the oes defied i (10) are 0; thus, the table eeds to be updated durig the auctios. After receivig the broadcast vector b 2, vehicle cofirms the eviromet state i the curret slot, ad it ca calculate V (S,S ) ad update the related item i the maximal MTUA table usig V (S,S ) (1 γ) V old (S,S ) = +γ V ew (S,S ), if (S,t,S,t )=(S,S ) V old (S,S ), otherwise (11) where γ is the learig rate factor, V old (S,S ) deotes the origial value i the table, ad V ew (S,S ) stads for the result of (9). D. Eviromet State ad the Estimatio of Eviromet State Trasitio Probability To implemet the aforemetioed algorithm, we eed the eviromet state S ad its trasitio probability. The eviromet state cotais the private iformatio of the other vehicles ad caot be obtaied directly. The statistical eviromet state trasitio probability ca be estimated by each vehicle from the historical records, i.e., the series of broadcast vectors b 1 ad b 2 i the past slots. The RSU ca also collect the statistical eviromet state trasitio probability from the vehicles i its coverage area ad ca sed the itegrated eviromet state trasitio probability to the ewly arrivig vehicles as a referece. The statistical eviromet state trasitio probability of the vehicle is more importat because it is estimated accordig to the vehicle s states ad the competitio itese. The itegrated eviromet state trasitio probability may ot reflect the idividual differeces amog vehicles, but it ca be offered to the vehicles that have o iitial iformatio ad speed the covergece of the trasitio probability estimatio. Here, we propose a method for a vehicle to estimate the statistical eviromet state trasitio probability accordig to the accumulated historical records. First of all, the defiitio of eviromet state S used i (7) (9) is give. The, provided that the chael allocatio result is O i (7) (9), the expressio of the eviromet state trasitio probability is give. Fially, the formula of the probability of the chael allocatio result O is preseted whe the eviromet state ad the bid of vehicle are give. 1) Eviromet State: The eviromet state is deoted a tuple, i.e., S =(N,M,v), where v ( {1, 2,...,V}) is the price rage idex of the average trasactio price C. The V price rages are [1, Γ 1 ), [Γ 1, Γ 2 ),..., ad [Γ V 1,F limit ],

9 742 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO. 1, JANUARY 2017 TABLE II ENVIRONMENT STATE TRANSITION COUNTING TABLE TABLE III AUCTION RESULT TABLE respectively; thus, v describes the competitio itesity. The total umber of eviromet states is Z S = N max (M + 1)V. Here, N ad M ca be obtaied from the broadcast vector b 1, ad v ca be derived from C i b 2. 2) Eviromet State Trasitio Probability: The eviromet state trasitio probability i (7) (9) ca be writte i a geeral form q (S # S,O ), where oly O 0 ad O = 0 are differetiated. It is the trasitio probability from eviromet state S to S # whe the chael allocatio result for vehicle is O ad ca be expressed as q ( S # S,O ) = S # f S,S #,O f S,S #,O,O = 0or O 0 (12) where the umerator f S,S #,O ad the deomiator stad for the times of eviromet state trasitios from S to S # ad the total times from S to all the possible eviromet states, respectively, both give the chael allocatio result O. q (S S #,O ) ca be obtaied by lookig up Table II ad calculatig usig (12). It is a geeral form of q (S,t+1 S,t,O,t ) ad q (S,t S,t 1,O,t 1 ) i Sectio V-A so that the values of the latter two ca be easily obtaied from Table II ad (12). The value of f S S #,O i (12) ca be obtaied by coutig. Let vehicle maitai a eviromet state trasitio coutig table (see Table II) to record the times of the eviromet state trasitios. The dimesio of Table II is 2Z S Z S. The factor 2 exists because all the chael allocatio results are icluded for both O = 0 ad O 0. Whe Table II is iitially set up, the diagoal elemets with O 0 are set to 1, ad the others are set to 0, which eables the calculatios of (12) ad the request i the first auctio. Durig the auctios, vehicle updates the correspodig item by addig 1 accordig to the eviromet state i the last slot after it receives the broadcast vector b 2 ; thus, the eviromet state becomes kow. Alog with the accumulatio of historical records, the estimated eviromet state trasitio probability obtaied usig Table II ad (12) becomes accurate. 3) Probability to Access a Chael: p(o 0 S,C ) i (5) is the probability of the chael allocatio result O 0 give eviromet state S ad bid C ad satisfies p(o 0 S,C )= f suc S,C f suc S,C + f fail S,C (13) where fs suc,c deotes the times of successful chael access attempts whe the eviromet state is S ad the bid is C, ad fs fail,c stads for the times of failed chael access whe the eviromet state is S ad the bid is C.BothfS suc,c ad fs fail,c ca be obtaied by coutig. The probability of failed chael access give eviromet state S ad bid C is deoted p(o = 0 S,C ). Let each vehicle maitai a auctio result table (see Table III). Table III records the times of successful or failed chael access attempts uder differet eviromet states. The dimesio of Table III is Z S F limit. Whe Table III is iitially set up, all the successful times of diagoal items are set to 1, ad the failed times are set to 0 to eable the calculatio of (13) i the first auctio. Whe receivig the broadcast vector b 2, vehicle kows the result of whether it ca access a chael ad the adds 1 to the item of correspodig successful (failed) times. Notice that, if vehicle bids C ad successfully obtais (but fails to obtai) a chael, ay bid Ĉ >C (Ĉ <C ) will also wi (lose) the auctio. E. Algorithm We propose a algorithm to perform chael allocatio ad adaptive video streamig proposed i Sectio IV for vehicle. The steps are summarized i Algorithm I. The operatios i Step 5 make ecessary preparatio for calculatig the bid ad the quality level i the ext slot.

10 SUN et al.: CHANNEL ALLOCATION FOR ADAPTIVE VIDEO STREAMING IN VEHICULAR NETWORKS 743 Algorithm I Chael allocatio ad video streamig algorithm for vehicle 1. Receive broadcast vector b 1 from the RSU; obtai the umber of vehicles ad that of available chaels as well as the correspodece betwee the vehicle IDs ad the mii-slot idexes; cofirm the allocated mii-slot for itself. 2. Obtai q (S S,O ) via Table II ad (12); solve p(o S,C ) usig Table III ad (13); calculate q (S S, O ) accordig to the vehicle state of the curret slot ad the possible states of ext slots; calculate the optimal request A opt =(C opt,l opt ) by substitutig the values of q (S S, O ), V (S,S ), q (S S,O ), ad p(o S,C ) ito (7); sed A opt to the RSU i the allocated miislot. 3. Receive the broadcast vector b 2, ad obtai O ad the average trasactio price C. 4. If O = m, receive video data through the allocated chael m. 5. Cofirm the eviromet sate S, ad update Table II; update Table III accordig to O ; compute (9), ad obtai V (S,S ) by substitutig the result ito (11) as V ew (S,S ); update vehicle state accordig to (4). The computatio ad space complexities of the proposed algorithm are discussed i the followig. To allocate the chaels, the RSU should sort the requests from vehicles i a decreasig order accordig to their bids. If a quick sort algorithm is applied, the time complexity of the resource allocatio at RSU is O(N max log 2 (N max )) [33]. To calculate the optimal request accordig to (7), each vehicle should compare the results of all possible requests ad states. Specifically, a vehicle should cosider all possible requests with dimesio F limit L, all the possible eviromet states i the curret slot with dimesio V,the possible chael allocatio results with dimesio 2, the possible vehicle states with dimesio Z S (seesectio V-C), ad the eviromet states with dimesio Z S (see Sectio V-D1), respectively. The computatio complexity is proportioal to the product of the aforemetioed dimesios ad ca be expressed as O(2T i F limit N max V 2 L 2 (B max + 1)(F max + 1)(M + 1)). It is oted that the actual computatio complexity is much lower tha that aforemetioed sice ot all requests, vehicle states, ad eviromet states will be reached i practice. The space complexity of each vehicle depeds o the sizes of the tables. The dimesio of the maximal MTUA table is Z S Z S = T i LN max V (M + 1)(F max + 1)(B max + 1), the dimesio of eviromet state trasitio coutig table (see Table II) is 2Z S Z S = 2V 2 Nmax(M 2 + 1) 2, ad the dimesio of the auctio result table (see Table III) is 2Z S F limit = 2VN max F limit (M +1). It will be a importat further research issue to desig a joit chael allocatio scheme for heterogeeous traffic, icludig both data ad video traffic. A possible approach is to defie multiple utility fuctios for heterogeeous traffic, ad the objective of chael allocatio is to maximize the total utility. Aother further issue is that cellular etworks ca play as a alterative to provide access with RSU. As the RSUs resource is cheap but sporadic distributed, ad the cellular resource TABLE IV REGIONS AND CORRESPONDING TRANSMISSION RATES TABLE V PARAMETERS OF THE VIDEO TRACE is expesive but ca cover the whole road segmet with a certai trasmissio rate, the system should make decisios o allocatig the chael resources to vehicles properly. Moreover, device-to-device commuicatios [34], [35] ad multihop relayig [36], [37] ca also be able to facilitate efficiet vehicular commuicatios. VI. PERFORMANCE EVALUATION To validate the effectiveess of the proposed chael allocatio ad video streamig algorithm, simulatios are coducted to evaluate ad compare the performace of the proposed algorithm ad two baselie algorithms. A. Simulatio Settig The umber of data chaels M is 3. The probability that the umber of available chaels chages from i to j is deoted p ij, where i, j {1, 2, 3}. Wesetp 11 = p 12 = p 22 = 0.4, p 13 = p 33 = 0.2, p 21 = p 23 = p 31 = 0.3, ad p 32 = 0.5. The radius of a RSU s coverage area is 75 m, ad the distace betwee two adjacet RSUs X R2R is 225 m. The coverage area of a RSU is divided ito three regios (K = 3). The relatioship betwee the trasmissio rate r k ad the distace from a vehicle to a RSU X V2R is show i Table IV. To describe the relatioship betwee the vehicle velocity ad vehicle desity, the free-flow model obtaied from field observatios [28] is adopted, i.e., v = v f (1 ρ/ρ jam ). Here, ρ is the vehicle desity, i.e., the average umber of vehicles per kilometer; ρ jam = 250 veh/km is the vehicle desity i traffic jams; ad v f = 140 km/h deotes the free-flow velocity. The relatioship betwee itervehicle distace λ ad the vehicle velocity v is expressed as λ = ρv, ad the maximal umber of vehicles i the coverage area of a RSU is N max = 2Rρ jam, where deotes floor fuctio. Other settigs iclude the learig rate factor γ = 0.4, bid upper boud F limit = 5, ad puishmet coefficiet δ = 0.1. The slot duratio Δ is 0.53 s, ad the duratio of resource auctio phase Δ auc is 0.05 s. The video trace Forema [39] with resolutio is used i the simulatios. The frame rate is 30 fps ad the group of picture (GOP) is 16 frames. Oe GOP that ca be played i a slot of 0.53 s is cosidered a data uit. The video trace is ecoded ito a basemet layer ad two ehacemet layers [40]. The parameters of the video trace are listed i Table V. I the simulatio, the legth of the road equals the distace betwee two adjacet RSUs X R2R. For each vehicle desity, te rus are carried out to get the average. For each ru, vehicles

11 744 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO. 1, JANUARY 2017 were radomly placed o the road with certai vehicle desity. The vehicles drive ito the RSU s coverage area at differet time istats; thus, i ay slot, they may drive o differet road regios with differet trasmissio rates. Whe drivig to the ed of the road, a vehicle will be loop back to the begiig of the road with a radom itervehicle distace. I this way, 6500 periods are simulated per ru to observe the covergece of the proposed algorithm. The received umber of video layers ad that of iterrupted slots of each vehicle i a period are recorded, ad the average iterruptio ratio (AIR) ad average umber of received layers (ARL) of the 6500 periods are the calculated. AIR is defied as (1/(V um 6500)) 6500 Vum h=1 i=1 J i,h/t RSUh, where N um is the total umber of vehicles, T RSUh is the total umber of slots i the hth period, ad J i,h is the umber of iterrupted slots of vehicle i i the hth period. ARL is defied as (1/(N um 6500)) 6500 Num h=1 i=1 Y i,h /T RSUh, with Y i,h beig the total umber of video layers received by vehicle i i the hth period. Fig. 6. Effect of the weight coefficiet α. B. Baselie Algorithms Two baselie algorithms are adopted i this paper. To the best of our kowledge, for multivehicle commuicatio sceario, there is o previous work that joitly cosiders chael allocatio ad video streamig. Therefore, we desiged the two baselie algorithms by combiig the video streamig algorithm (THRESHOLD for short) proposed i [18] ad differet chael allocatio schemes that appeared i [22], [41], ad [42]. The first algorithm is called MAX+THRESHOLD. A RSU allocates the available chaels to the closest vehicles withi its coverage area, amely, the vehicles with the highest DT rates. Each vehicle that is allocated a chael requests quality level accordig to its buffer level B ad the two thresholds defied i [18]. The dow threshold is ϑx R2R /v/δ, where v is the vehicle velocity, ad coefficiet ϑ is set to be 1 i the simulatio. The up threshold is twice the dow threshold. The requested quality level l is decreased by 1 if B is less tha the dow threshold ad is icreased by 1 whe B is betwee the two thresholds; otherwise, l = 3. The secod algorithm is called RD+THRESHOLD. A RSU radomly allocates the available chaels to the vehicles withi its coverage area, ad each vehicle that is allocated a chael decides its request based o its buffer level B ad the two aforemetioed thresholds. I both algorithms, each vehicle ca be allocated at most oe chael. C. Impact of Parameters i Proposed Algorithm The impact of the parameters α ad F max o the performace of the proposed algorithm are ivestigated. Fig. 6 shows the relatioship betwee AIR (ARL) ad weight coefficiet α defied i (5) uder ρ = 20 veh/km. It ca be observed from Fig. 6 that a larger α leads to a lower AIR, which is preferred, whereas a smaller α leads to a larger ARL, which is desired. This is because a larger α results i a higher weight for the first item related to iterruptio ratio i (5); thus, playback iterruptio is preferetially preveted ad vice versa. The impact of iitial budget amout F max o the performace of the proposed algorithm is show i Fig. 7 uder Fig. 7. Effect of iitial budget. ρ = 20 veh/km. The horizotal axis is the ratio of F max ad T i, where T i is the umber of slots that a vehicle i the coverage area of a RSU, as show i Fig. 5. It ca be observed that as F max /T i icreases, AIR decreases at first, ad the icreases. This ca be explaied as follows. Whe the ratio of F max ad T i is 1, all the vehicles just eterig a RSU s coverage area ca oly bid 1 because of the lower boud C 1 ad caot icrease bid to improve the chael access opportuity; thus, the iterruptio ratio is high. Yet, whe the iitial budget is abudat (F max /T > 1.5), the vehicle will bid with a high price to avoid puishmet caused by wasted budget. Due to the bid upper boud F limit, the bids of the vehicles with high utility value ad the vehicles with excessive budget may be equal, which leads to ureasoable chael allocatio ad a icrease i AIR because the RSU caot idetify the vehicles with a really high utility value. Therefore, it is proper to set F max /T i the rage of 1.1 to 1.5. It is also observed that the ARL is basically ot affected by F max. This is because, by havig the same iitial budget, all the vehicles adopt similar pricig strategy ad, thus, obtai similar umber of access chaces. Coclusio ca be draw from Figs. 6 ad 7 that a higher α ad a lower F max should be set i the proposed auctio-based chael allocatio mechaism. Therefore, i the followig simulatios, we set α = ad F max = 1.1 T i, where x calculates the miimum iteger larger tha or equal to x.

12 SUN et al.: CHANNEL ALLOCATION FOR ADAPTIVE VIDEO STREAMING IN VEHICULAR NETWORKS 745 Fig. 8. Performace compariso whe RSUs are sporadically deployed. (a) AIR versus vehicle desity. (b) ARL versus vehicle desity. D. Average Iterruptio Ratio ad Average Number of Received Layers Fig. 8 shows the AIRs ad ARLs of the proposed algorithm ad the two baselie algorithms uder differet vehicle desities whe RSUs are sporadically deployed. It ca be see that all the algorithms have the same tred: The AIR icreases ad the ARL decreases as the vehicle desity icreases. This is because higher vehicle desity results i more vehicles competig i a RSU. Moreover, oe ca otice that both the AIR ad the ARL of the proposed algorithm are better tha those of the baselie algorithms. This is because the proposed algorithm cosiders the chael allocatio ad video streamig joitly; thus, the RSUs ad vehicles ca make proper decisios, whereas the baselie algorithms do ot. Fig. 9 shows the AIRs ad ARLs of differet algorithms whe the road is fully covered by RSUs, amely vehicles ca receive the RSUs sigal all the time. We simple set the distace of adjacet RSUs equal to the diameter of the RSU s coverage i the simulatios. This sceario is a special ad easy case compared with whe the RSUs are sporadically deployed. It ca be observed from the figure that AIRs ad ARLs of all the algorithms perform well uder low vehicle desities. Yet, as the vehicle desity icreases, the performace deteriorates. Compared with the baselie algorithms, the proposed algorithm ca achieve lower iterruptio ratio ad higher umber of received layers except whe vehicle desity equals 25 veh/km because Fig. 9. Performace compariso whe road is fully covered. (a) AIR versus vehicle desity. (b) ARL versus vehicle desity. the proposed algorithm teds to keep as low iterruptio ratio as possible ad thus sacrifices the performace of ARL. It ca be observed that the performaces i Fig. 9 are better tha those i Fig. 8. It is because the road is fully covered by RSUs; thus, the vehicles are i the coverage areas of RSUs all the time ad have more chaces to assess chaels. E. Performace Improvemet by Accumulatio of Historical Records A method to estimate the eviromet state trasitio probability based o the accumulated historical records is proposed i Sectio V-D. The effect of this method is verified by simulatio, ad the results are preseted here. The system average utility value (SAUV) is adopted as a metric ad is the average of all vehicles utilities i a time widow aimig to smooth the fluctuatio. It ca be defied as Ū(y) =(1/H) y Nmax =1 U,g. Here, H represets the g=y H duratio of the widow i period is set to 200, ad U,g is the total utility of vehicle obtaied i the gth period ad is defied as U,g = α (B,g D,g ) +(1 α) Ql Q L. (14) Notice that as compared with (5), (14) is ot weighted by the probability of accessig a chael because the results of chael allocatio ad the umber of received video layers have already bee kow.

13 746 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO. 1, JANUARY 2017 Fig. 10. Coverget tedecy versus the accumulatio of past RSUs. (a) SAUV. (b) AIR. (c) ARL. The subfigures i Fig. 10 show the coverget tedecy as the accumulatio of past RSUs. Fig. 10(a) shows the SAUVs obtaied uder differet vehicle desities. As vehicle desity icreases, the SAUV icreases rapidly at first ad the teds to be flat, meaig that the SAUV ca be improved ad achieve covergece with the help of accumulated historical records. Fig. 10(b) ad (c) show the obtaied AIRs ad ARLs alog with the accumulatio of historical records. It ca be observed that, as the vehicle desity icreases, the AIRs i Fig. 10(b) icrease, ad ARLs i Fig. 10(c) decrease. That is because if there are more vehicles i the coverage area of a RSU, the fewer opportuities for each vehicle to access a chael, ad the worse the quality of video playback will be. It ca also be observed i Fig. 10(b) that, for a give vehicle desity, AIR decreases first ad the teds to be flat with the icrease i the accumulated umber of RSUs. This is because with the help of accumulated historical records, the vehicles ca bid ad request layers more reasoably; therefore, the iterruptio ratio decreases, ad visual quality is maitaied as high as possible. Fig. 10(c) shows that ARLs keep almost costat as the RSU idex icreases, revealig that more historical records eable more reasoable bids. It is show that the AIR i Fig. 10(b) ad the SAUV i Fig. 10(a) get flatteed after about 2000 RSUs, although the vehicles have o prior iformatio about the eviromet state. This implies that the ratioality of the vehicle bid will be greatly improved ad the covergece process will be speeded up if RSUs provide the iformatio of eviromet states to their ewly arrived vehicles. VII. CONCLUSION I this paper, we have proposed a chael allocatio ad adaptive streamig algorithm for supportig video services i vehicular etworks. The cotributios of this paper are as follows. By employig the auctio-based chael allocatio mechaism, the vehicles ca bid reasoably accordig to their utility values, which help the RSU to allocate chaels to the vehicles with the most urget eed for DT ad the vehicles with high trasmissio rates ad, therefore, guaratees the vehicles smooth playback ad visual quality. The vehicles ca make the tradeoff betwee visual quality ad oiterruptio playback by chagig the required layers. The proposed algorithm ca coverge quickly with the help of historical records ad achieve lower iterruptio ratio ad higher visual quality tha the baselie algorithms. The proposed algorithm is also effective whe the umber of data chaels for video service is costat. I future work, we will cosider a more geeral chael model ad mobility model to make the proposed algorithm more practical. REFERENCES [1] N. Lu ad X. She, Capacity Aalysis of Vehicular Commuicatio Networks. New York, NY, USA: Spriger-Verlag, [2] T. Lua, X. She, F. Bai, ad L. Su, Feel bored? Joi verse! Egieerig vehicular proximity social etworks, IEEE Tras. Veh. Techol., vol. 64, o. 3, pp , Mar [3] E. Uhlema, Itroducig coected vehicles, IEEE Veh. Techol. Mag., vol. 10, o. 1, pp , Mar [4] H. T. Cheg, H. Sha, ad W. Zhuag, Ifotaimet ad road safety service support i vehicular etworkig: From a commuicatio perspective, Mech. Syst. Sigal Process., vol. 25, o. 6, pp , Aug [5] M. Wag et al., Real-time path plaig based o hybrid-vanetehaced trasportatio system, IEEE Tras. Veh. Techol., vol. 64, o. 5, pp , May [6] M. Guo, M. Ammar, ad E. Zegura, V3: A vehicle-to-vehicle live video streamig architecture, i Proc. IEEE PerCom, 2005, pp [7] T. Lua, L. X. Cai, ad X. She, Impact of etwork dyamics o user s video quality: Aalytical framework ad QoS provisio, IEEE Tras. Multimedia, vol. 12, o. 1, pp , Ja [8] J. Peha, Approaches to spectrum sharig, IEEE Commu. Mag.,vol. 43, o. 2, pp , Feb [9] Q. Zhao ad B. Sadler, A survey of dyamic spectrum access: Sigal processig, etworkig, ad regulatory policy, IEEE Sigal Process. Mag., vol. 55, o. 5, pp , Apr [10] H. Schwarz, D. Marpe, ad T. Wiegad, Overview of the scalable video codig extesio of the H.264/AVC stadard, IEEE Tras. Circuits Syst. Video Techol., vol. 17, o. 9, pp , Sep [11] Y. Zhag, C. Lee, D. Niyato, ad P. Wag, Auctio approaches for resource allocatio i wireless systems: A survey, IEEE Commu. Surveys Tuts., vol. 15, o. 3, pp , 3rd Quart [12] M. Asefi, J. W. Mark, ad X. She, A mobility-aware ad qualitydrive retrasmissio limit adaptatio scheme for video streamig over VANETs, IEEE Tras. Wireless Commu.,vol.11,o.5,pp , May [13] M. Kalma, E. Steibach, ad B. Girod, Adaptive media playout for lowdelay video streamig over error-proe chaels, IEEE Tras. Circuits ad Syst. Video Techol., vol. 14, o. 6, pp , Ju [14] H. Lu ad C. Che, Playback iterruptio probability aalysis for roadside-to-vehicle media streamig, i Proc. IEEE WoWMoM, 2011, pp [15] R. A, Z. Liu, ad Y. Ji, Video streamig for highway VANET usig scalable video codig, i Proc. IEEE VTC Fall, 2014, pp [16] E. Yaacoub, F. Filali, ad A. Abu-Dayya, SVC video streamig over cooperative LTE/802.11p vehicle-to-ifrastructure commuicatios, i Proc. IEEE CIT, 2013, pp [17] M. Hu, Z. Zhog, ad C.-Y. Chag, A multicast schedulig approach for layered video service i vehicular ad hoc etworks, i Proc. IEEE IMIS, 2013, pp [18] M. Xig ad L. Cai, Adaptive video streamig with iter-vehicle relay for highway VANET sceario, i Proc. IEEE ICC,2012,pp [19] L. Su, A. Huag, H. Sha, M. Xig, ad L. Cai, Quality-drive adaptive video streamig for cogitive VANETs, i Proc. IEEE VTC Fall, 2014, pp. 1 6.

14 SUN et al.: CHANNEL ALLOCATION FOR ADAPTIVE VIDEO STREAMING IN VEHICULAR NETWORKS 747 [20] S. Park, Y. Chag, F. Kha, ad J. Copelad, Dyamic service-chaels allocatio (DSCA) i vehicular ad-hoc etworks, i Proc. IEEE CCNC, 2013, pp [21] M. Wag, Q. She, R. Zhag, H. Liag, ad X. She, Vehicle-desitybased adaptive MAC for high throughput i drive-thru etworks, IEEE Iteret Thigs J., vol. 1, o. 6, pp , Dec [22] D. Hadaller, S. Keshav, ad T. Brecht, MV-MAX: Improvig wireless ifrastructure access for multi-vehicular commuicatio, i Proc. ACM Challeged Netw., 2006, pp [23] M. Wag et al., Asymptotic throughput capacity aalysis of VANETs exploitig mobility diversity, IEEE Tras. Veh. Techol., vol. 64, o. 9, pp , Sep [24] J. Alcaraz, J. Vales-Aloso, ad J. Garcia-Haro, Cotrol-based schedulig with QoS support for vehicle to ifrastructure commuicatios, IEEE Wireless Commu., vol. 16, o. 6, pp , Dec [25] D. Niyato, E. Hossai, ad P. Wag, Competitive wireless access for data streamig over vehicle-to-roadside commuicatios, i Proc. IEEE GLOBECOM, 2009, pp [26] R. Tomar ad S. Verma, RSU cetric chael allocatio i vehicular ad-hoc etworks, i Proc. IEEE WCSN, 2010, pp [27] J. Wu, Coectivity aalysis of a mobile vehicular ad hoc etwork with dyamic ode populatio, i Proc. IEEE GLOBECOM, 2008, pp [28] J. D. Fricker ad R. K. Whitford, Fudametals of Trasportatio Egieerig: A Multimodal Systems Approach. Upper Saddle River, NJ, USA: Pretice-Hall, [29] L. Cheg, B. Hety, D. Stacil, F. Bai, ad P. Mudalige, Mobile vehicleto-vehicle arrow-bad chael measuremet ad characterizatio of the 5.9 GHz dedicated short rage commuicatio (DSRC) frequecy bad, IEEE J. Sel. Areas Commu., vol. 25, o. 8, pp , Oct [30] A. Bokai, M. Hassa, ad S. Kahere, HTTP-based adaptive streamig for mobile cliets usig Markov decisio process, i Proc. IEEE PV, 2013, pp [31] M. Xig, S. Xiag, ad L. Cai, A real-time adaptive algorithm for video streamig over multiple wireless access etworks, IEEE J. Sel. Areas Commu., vol. 32, o. 4, pp , Apr [32] M. L. Puterma, Markov Decisio Processes: Discrete Stochastic Dyamic Programmig. Hoboke, NJ, USA: Wiley, [33] W. Xiag, Aalysis of the time complexity of quick sort algorithm, i Proc. IEEE ICIII, 2011, pp [34] J. Liu ad N. Kato, Device-to-device commuicatio overlayig two-hop multi-chael uplik cellular etworks, i Proc. ACM MobiHoc, 2015, pp [35] J. Liu, S. Zhag, H. Nishiyama, N. Kato, ad J. Guo, A stochastic geometry aalysis of D2D overlayig multi-chael dowlik cellular etworks, i Proc. IEEE INFOCOM, 2015, pp [36] Y. Zhag, S. He, ad J. Che, Data gatherig optimizatio by dyamic sesig ad routig i rechargeable sesor etworks, i Proc. IEEE SECON, 2013, pp [37] Z. Su, Q. Xu, H. Zhu, ad Y. Wag, A ovel desig for cotet delivery over software defied mobile social etworks, IEEE Tras. Netw., vol. 29, o. 4, pp , Jul./Aug [38] Cisco, Maximum Throughput Calculatios for b WLAN. [Olie]. Available: [39] Video Test Media. [Olie]. Available: [40] Joit Scalable Video Model Referece Software JSVM [Olie]. Available: [41] K. Ota, M. Dog, S. Chag, ad H. Zhu, MMCD: Max-throughput ad mi-delay cooperative dowloadig for drive-thru iteret systems, i Proc. IEEE ICC, 2014, pp [42] S. Yag, C. K. Yeo, ad B. S. Lee, MaxCD: Efficiet multi-flow schedulig ad cooperative dowloadig for improved highway drive-thru iteret systems, Comput. Netw., vol. 57, o. 8, pp , Ju Log Su received the B.Sc. degree i iformatio egieerig i 2012 from Zhejiag Uiversity, Hagzhou, Chia where he is curretly workig toward the Ph.D. degree with the Istitute of Iformatio ad Commuicatio Egieerig. His curret research iterests iclude video streamig i vehicular ad hoc etworks ad qualityof-service provisioig i cogitive radio etworks. Haggua Sha (M 10) received the B.Sc. degree i electrical egieerig from Zhejiag Uiversity, Hagzhou, Chia, i 2004 ad the Ph.D. degree i electrical egieerig from Fuda Uiversity, Shaghai, Chia, i From 2009 to 2010, he was a Postdoctoral Research Fellow with the Uiversity of Waterloo, Waterloo, ON, Caada. Sice February 2011, he has bee with the College of Iformatio Sciece ad Electroic Egieerig, Zhejiag Uiversity, where he is curretly a Associate Professor. His curret research iterests iclude cross-layer protocol desig, resource allocatio, ad quality-of-service provisioig i wireless etworks. Dr. Sha has served as a Techical Program Committee member for various iteratioal cofereces, icludig the IEEE Global Commuicatios Coferece, the IEEE Iteratioal Coferece o Commuicatios, the IEEE Wireless Commuicatios ad Networkig Coferece (WCNC), ad the IEEE Vehicular Techology Coferece. He also served as the Publicity Cochair for the third ad fourth IEEE Iteratioal Workshops o Wireless Sesor, Actuator, ad Robot Networks ad the fifth Iteratioal Coferece o Wireless Commuicatios ad Sigal Processig. He coreceived the Best Idustry Paper Award from the 2011 IEEE WCNC held i Quitaa Roo, Mexico. Aipig Huag (SM 08) received the B.S. degree from Najig Istitute of Post ad Telecommuicatios, Najig, Chia, i 1977; the M.S. degree from Najig Istitute of Techology (Southeast Uiversity), Najig, i 1982; ad the Licetiate of Tech. degree from Helsiki Uiversity of Techology (HUT), Espoo, Filad, i From 1977 to 1980, she was a Egieer with the Desig ad Research Istitute of the Chiese Miistry of Post ad Telecommuicatios. From 1982 to 1994, she was with Zhejiag Uiversity (ZJU), Hagzhou, Chia, as a Assistat Professor ad the as a Associate Professor. From 1994 to 1998, she was a Visitig Scholar ad a Research Scietist with HUT (Aalto Uiversity). Sice 1998, she has bee a Full Professor with ZJU. She is the author of oe book ad more tha 160 papers i refereed jourals ad cofereces o commuicatios ad etworks ad sigal processig. Her curret research iterests iclude heterogeeous etworks, performace aalysis ad cross-layer desig, ad plaig ad optimizatio of cellular mobile commuicatio etworks. Ms. Huag serves as the Vice Chair of IEEE Commuicatios Society Najig Chapter. Li Cai (S 00 M 06 SM 10) received the M.A.Sc. ad Ph.D. degrees i electrical ad computer egieerig from the Uiversity of Waterloo, Waterloo, ON, Caada, i 2002 ad 2005, respectively. Sice 2005, she has bee with the Departmet of Electrical ad Computer Egieerig, Uiversity of Victoria, Victoria, BC, Caada, where she is curretly a Professor. Her research iterests iclude several areas i commuicatios ad etworkig, with a focus o etwork protocol ad architecture desig supportig emergig multimedia traffic over wireless, mobile, ad hoc, ad sesor etworks. Dr. Cai served as a Techical Program Committee Symposium Cochair for the IEEE Global Commuicatios Coferece i 2010 ad She served as a Associate Editor for the IEEE TRANSACTIONS ON WIRELESS COMMUNICATION, the IEEE TRANSACTIONS ON VEHICULAR TECHNOL- OGY, theeurasip Joural o Wireless Commuicatios ad Networkig, the Iteratioal Joural of Sesor Networks, adthejoural of Commuicatios ad Networks. She received the NSERC Discovery Accelerator Supplemet Grats i 2010 ad i 2015 ad the Best Paper Awards at the IEEE Iteratioal Coferece o Commuicatios i 2008 ad the IEEE Wireless Commuicatios ad Networkig Coferece i Hogli He received the B.Sc. degree i iformatio egieerig i 2014 from Zhejiag Uiversity, Hagzhou, Chia, where he is curretly workig toward the Ph.D. degree with the Istitute of Iformatio ad Commuicatio Egieerig. His curret research iterests iclude video streamig i vehicular ad hoc etworks ad Log- Term Evolutio over ulicesed spectrum.