Research Article A Framework for QoE-Aware 3D Video ...

Research ArticleA Framework for QoE-Aware 3D Video Streaming Optimisationover Wireless Networks

Ilias Politis Asimakis Lykourgiotis and Tasos Dagiuklas

CONES Research Group Hellenic Open University 26335 Patras Greece

Correspondence should be addressed to Ilias Politis ilpolitiseapgr

Received 6 October 2015 Revised 5 February 2016 Accepted 29 February 2016

Academic Editor Ansgar Gerlicher

Copyright copy 2016 Ilias Politis et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

The delivery of three-dimensional immersive media to individual users remains a highly challenging problem due to the largeamount of data involved diverse network characteristics and user terminal requirements as well as userrsquos context This paperproposes a framework for quality of experience-aware delivering of three-dimensional video across heterogeneous wirelessnetworks The proposed architecture combines a Media-Aware Proxy (application layer filter) an enhanced version of IEEE80221 protocol for monitoring key performance parameters from different entities and multiple layers and a QoE controllerwith a machine learning-based decision engine capable of modelling the perceived video quality The proposed architecture isfully integrated with the Long Term Evolution Enhanced Packet Core networks The paper investigates machine learning-basedtechniques for producing an objective QoE model based on parameters from the physical the data link and the network layersExtensive test-bed experiments and statistical analysis indicate that the proposed framework is capable of modelling accurately theimpact of network impairments to the perceptual quality of three-dimensional video user

1 Introduction

Media-friendly Future Internet needs to cope with anincreased demand for streaming applications and three-dimensional (3D) media by modifying and optimising exist-ing protocols for streaming over a hybrid network environ-ment Accessing the Internet easily any time anywhere andwith any device has changed userrsquos Internet consumptionamount and behaviour especially onmedia consumption [1]Traditionally web surfing and file sharing were the dominantapplications whereas today real-time streaming and socialmedia applications (eg YouTube Facebook etc) constitutethe major portion of the Internet traffic The change in Inter-net usage patterns and the increasing needs of novel applica-tions (high performance availability security etc) togetherwith end usersrsquo increasing quality of experience (QoE) expec-tations open new challenges for the Future Internet due toincreasing popularity of real-time communication onlinesocial networks and heterogeneity of user devices [2ndash4]

It is evident that due to recent advances in video acquisi-tion techniques coding delivery and displaying 3D videohas sprouted in the consumer domain through a range of

applications and services which provide enhanced visualexperience for the viewers Recently the research on 3D videocoding and transportation intensified flared by enhance-ments and updates on coding strategies such as H264SVC(Scalable Video Coding) [5] H264MVC (Multiview VideoCoding) [6] and High Efficiency Video Coding (HEVC) [7]as well as the wide deployment of LTE (Long TermEvolution)systems for high-speed data delivery to both wireless andmobile users [8]

Based on the different representation formats for stereo-scopic and multiview videos different coding schemes haveevolved over time to compress such media efficiently Purecolour based (eg multiview stereo L-R) or depth based for-mats (eg colour-plus-depth or Layered Video plus Depth)are usually compressed using the legacy block based codingstandards such as MPEG-4 Part 10H264 Advance VideoCoding (AVC) or its extensions such as SVC [9] or MVC[10] AVC has also introduced new profiles to support codingof stereoscopic videos such as Stereoscopic High Profile [11]Both AVC derived codecs namely SVC and MVC get theirbase layer fully compliant with AVC specifications in order tobe compliant with legacy deploymentsThe purpose of SVC is

Hindawi Publishing CorporationMobile Information SystemsVolume 2016 Article ID 4913216 18 pageshttpdxdoiorg10115520164913216

2 Mobile Information Systems

to provide a universal media bit-stream that can be decodedby multiple decoders of different capacities to produce thereconstructed media at different states SVC also providesdynamic adaptation to a diversity of networks terminalsand formats This is extremely useful for simple adaptationof transmission and efficient storage and is beneficial fortransmission services with uncertain resolution channelconditions and device types The MVC standard on theother hand in order to impove the overall rate-distortionperformance exploits the inter-view redundancies throughthe Disparity Compensated Prediction (DCP) [12] Nonethe-less the multiview bit rate increases to unmanageable levelsas the number of source cameras increases Therefore itis impractical to encode and transport the entire set ofviews required to be displayed on most multiview displaysparticularly when dealing with heterogeneous wireless accessnetworks Yet it remains unclear how all these technologiesaffect the viewerrsquos perception of 3D video [13 14] Apparentlyit is still unrealistic to try and estimate the QoE of suchapplication without resorting to full subjective evaluations

Any 3D video coding and delivery system will ultimatelyneed to be measured in terms of userrsquos satisfaction andperceived experience QoE can be defined as the overallacceptability of an application or service strictly from theusers point of view It is a subjective measure of end-to-endservice performance from the user perspective and it is anindication of howwell the networkmeets the users needs [15]Encompassing many different aspects QoE is riveted on thetrue feelings of end users when they watch streaming videolisten to digitized music and browse the Internet through aplethora ofmethods and devices [16]There is a real challengein creating models which will accurately perform learning tomodel service behaviours by taking into account parameterssuch as arrival pattern request service time distributionsIO system behaviours and network usage [17] The aimof such attempts is to estimate QoE for both resource-centric (utilization availability reliability and incentive)and user-centric (response time and fairness) environmentsover certain predefined Quality of Service (QoS) and QoEthresholds Evidently there is still no concrete evidence toprove the correlation between QoS and QoE particularly for3D video Nevertheless there are several attempts to definethe relationship of human perception of video quality to theinherently objective network conditions and the majority ofthis work concludes that such a relationship cannot be linear[18 19] Specifically authors in [20] proved the importanceof the impact that network delivery speed and latency haveon the human satisfaction but also determined that in aheterogeneous networking environment neither bandwidthnor latency is sufficient as attributes that indicate the rangeof issues which render a service as nonappealing to the enduser

This paper describes a framework for QoE-aware deliv-ering of 3D video across heterogeneous wireless networksTowards this end a synergy is proposed between the LTEEPCarchitecture and IEEE 80221 protocol that will ensure real-time control and monitoring of key performance indicatorsand parameters relative to the perceived 3D QoE across dif-ferent layers and entities The proposed framework includes

a QoE controller module housing a machine learning- (ML-)based decision making engine which trains data collectedfrom the user and the network planes in order to modelas accurate as possible the perceptual 3D video quality Itneeds to be underlined that the proposed architecture doesnot require any alteration of the standardised LTE EPCinterfaces hence it could be smoothly integrated with thecurrent mobile operatorsrsquo deployments The paper investi-gates machine learning-based techniques for producing anobjective QoE model of network related impairments Theproposed QoE model is a function of parameters collectednot only from the application layer but also from theunderlying layers Previous efforts on modelling QoE for 3Dvideo have been based on measuring the end-to-end packetloss and determining the perceivedQoE a process that wouldrequire feedback from the receiving end which in the caseof wireless networks may never arrive or may arrive toolate for the decision engine This research investigates a newmethodology for monitoring network imposed impairmentsto the perceived video quality that will result inmore accurateQoEmodels Specifically a set of QoS parameters is collectedfrom different points in the delivery chain (ie bit errorrate from physical layer MAC layer load and networklayer delayjitter) which account for the overall numberof lost packets Three classification schemes with differentcharacteristics have been studied in order to identify themostappropriate method for modelling 3D video quality due tonetwork impairments

The rest of the paper is organised as follows Section 2briefly presents the concept of LTE EPC and its main entitiesthat are relevant to the purposes of this study Section 3provides a detailed description of the proposed media-awareframework and focuses on the synergy between the Media-Aware Proxy the IEEE 80221 protocol and the QoE con-troller which enables collectingQoE related key performanceindicators across multiple layers and network componentsthe training of these data and the modelling of the perceived3D video QoE Additionally in this section the three ML-based classification models are introduced and their maindifferences are explained The setup of the experimental test-bed is described in Section 4 Moreover the section definesthe physical MAC and network layer parameters whoseimpact on the perceptual quality of 3D video is under studyThe collected Mean Opinion Score (MOS) from the subjec-tive experiments along with a detailed statistical analysisis presented in Section 5 Based on the measured 3D MOSSection 6 compares the performance of the three machinelearning methods in terms of precision and accuracy inorder to determine the most suitable classification schemefor implementation as part of the QoE controller FinallySection 7 concludes the paper and highlights the future aimsof this research

2 3GPP EPS Architecture

The technical realization of next generation mobile net-works requires that complementary wireless technologiesare integrated in order to provide ubiquitous multimediaservices ldquoanywhere any time and on any devicerdquo In the

Mobile Information Systems 3

recent years 3GPPrsquos Evolved Packet System (EPS) [21] isincreasingly deployed by mobile operators in order to satisfythe requirement for uninterrupted high quality real-timemultimedia services across wireless access networks EPSconsists of the LTE wireless access which is forming theEvolved UTRAN (E-UTRAN) as the lower part of EPS andan IP connectivity control platform called Evolved PacketCore (EPC) as the upper part of EPS which enable wirelessaccess networks diversity (ie LTE UMTS WiMax WiFietc) EPS is relying on standardized routing and transportmechanisms of the underlying IP network In parallel EPSprovides capabilities for coexistence with legacy systems andmigration to the EPS while it supports functionalities foraccess selection based on operators policies user preferencesand networking conditions One of the most importantfeatures of EPS is its ability to support simultaneous activePacket Data Network (PDN) connections for the same userequipment whilst maintaining the negotiated QoS across thewhole system

3GPP has proposed the EPC in an effort to support higherdata rates lower latency packet optimisation and multipleradio access technologies [22] As part of EPS EPC is an all-IP architecture that fulfils those requirements and supportsthe characteristics of E-UTRAN The advantages of EPCas opposed to the legacy GPRS architecture are a moreclear depiction of control and data planes a more simplifiedarchitecture with a single core network and finally the fullassumption of the IETF protocols Therefore EPC allows fora truly converged packet core for trusted 3GPP networks (ieGPRSUMTS LTE etc) non-3GPPnetworks (ieWiMAX)and untrusted networks (ie WLAN) Additionally EPCmaintains seamless mobility and consistent and optimisedservices provisioning independent of the underlying accessnetwork

The proposed framework for QoE-aware delivery of 3Dvideo content over heterogeneous wireless networks relieson 3GPP EPC Therefore a description of the EPC modulesrelated to the study follows The functionalities of the dis-cussed modules are enhanced as part of the proposed frame-work in order to support seamless connectivity transparentreal-time adaptation of 3D video streaming and QoE-awaremonitoring and management

21 Key EPC Modules The relevance to the proposed frame-work EPC modules [22] is depicted in Figure 1 Specificallythese architecturalmodules and functions include the follow-ing

(i) Home Subscriber Server (HSS) is the main databaseof the EPC responsible for storing subscriberrsquos infor-mation that may include profile description severalidentification information and information regard-ing the subscriberrsquos established sessions The moduleinterfaces with the Mobility Management Entity andthe Authentication Authorization and Accountingserver (implemented as a Diameter server [23])

(ii) Mobility Management Entity (MME) is the modulethat performs the mobility management and bearermanagement and establishment It is also responsible

E‐UTRAN

IP services(IMS etc)

Internet

GSM

PSTN

UMTS

PDN gateway

Servinggateway

eNodeB

User equipment

PCRF

HSS

MME

Applicationfunctions

DataPolicycharging

Mobility mgtAuthentication

IP connectivity

Figure 1 Relevant EPC architectural elements

for performing mobility functions during handoversbetween 2G or 3G 3GPP access networksThemodulehas interfaces to the Serving-GW the eNodeB and theHSS

(iii) Serving Gateway (Serving-GW) acts as the gatewaywhere the interface towards E-UTRAN terminatesEach piece of user equipment (UE) associated withthe EPS is served by a single Serving-GW In terms ofimplementation both the Serving-GW and the PDN-GWmay be dwelling in a single physical component

(iv) Packet Data Network Gateway (PDN-GW) is thegateway which terminates the interface towards thepacket data networks (eg IMS [24] Internet) In casewhere the UE is connected with different IP servicesseveral PDN-GW are associated with this UE It per-forms UE IP allocation policy enforcement packetfiltering per user and packet screening MoreoverPDN-GW acts as the mobility anchor between 3GPPand non-3GPP access It has interfaces to the Serving-GW the Evolved PacketDataGateway (ePDG) whichis responsible for interworking between the EPC anduntrusted non-3GPP networks such as WiFi inter-face to the 3GPP AAA server the trusted non-3GPPaccess gateway and the Policy and Charging RulesFunction

(v) Policy and Charging Rules Function (PCRF) per-forms QoS control access control and chargingcontrol The function is policy based and autho-rizes bearer and session establishment and manage-ment A single PCRF is assigned to all the connec-tions of a subscriber It interfaces with Policy and


Charging Enforcement Function (PCEF) within thePDN-GW and the Bearer Binding and Event Report-ing Functions (BBERFs) within the trusteduntrusted3GPPnon-3GPP access gateways and the HSS

(vi) Evolved-UTRAN (E-UTRAN) as explained above isa simplified radio access network architecture com-prised of evolved NodeBs (eNodeBs) In terms offunctionality E-UTRAN supports radio resourcemanagement (RRM) and selection of MME and isresponsible for routing user data to the Serving-GWand for performing scheduling measurements andreporting

The described EPC modules have been implemented aspart of the OpenEPC testing platform [25] and they havebeen extended within the context of this work in order tosupport QoE-aware multimedia delivery as described in therest of the paper

3 Proposed Media-Aware Architecture

Themedia-aware architecture proposed in this paper aims toprovide seamless end-to-end services with ubiquitous use ofthe heterogeneous technologies In the centre of this archi-tecture lies a QoE controller which facilitates a ML-basedmodel of the quality of the perceived video experience Thelearning process is based on a collection of key performanceindicators from across the network architecture and fromdif-ferent layers This information is stored in a central databaseas described in the rest of this section Evidently the proposedframework is required to provide seamless uninterruptedand QoE-aware video services across heterogeneous linklayer interfaces and user devices Therefore the proposedmedia-aware framework integrates a novel entity namedMedia-Aware Proxy (MAP)with the IEEE 80221Media Inde-pendent Handover (MIH) protocol [26] In order to supportseamless handover and IP layer mobility the proposedframework supports also the Proxy Mobile IP however thisis beyond the scope of the current research

This framework is tightly coupled with the LTE EPCfunctions and modules [22] however it neither interruptsthe inherent process of EPC nor violates its protocols as itlies just before the core network The synergy of MAP andMIH along with a QoE controller which acts as the QoE-aware decision making function allows for real-time videostream adaptation (ie intelligent quality layer dropping orstream prioritization) based on multiple physical networkand application layer parameters collected from both themobile terminal and the access network side An overview ofthe proposed architecture is illustrated in Figure 2

31 Media-Aware Proxy MAP is defined as a transparentuser-space module responsible for low-delay adaptation andfiltering of scalable video streams [27] In conjunction withthe QoE controller which models QoE and informs MAPabout the predicted level of the perceived quality the latteris able to either drop or forward packets that carry specificlayers of a stream to the receiving video users In detail MAPis a network function based on the Media-Aware Network

Element (MANE) standard [28 29] Briefly MANE can beconsidered as either a middle box or an application layergetaway capable of aggregating or thinning RTP streams byselectively dropping packets that have the less significantimpact on the userrsquos video experience As such MANE hasbeen proposed as an intermediate system that is capable ofreceiving and depacketising RTP traffic in order to customisethe encapsulated network abstraction layer units accordingto clientrsquos and access networkrsquos requirementsWithin the con-text of the proposed media-aware architecture MAPrsquos role istwofold Firstly it acts as a central point of decision in orderto overcome networking limitations imposed by firewalls andNetwork Address Translation (NAT) protocol that are exten-sively used in real life networks Secondly it receives andparses RTP streams and customises the streaming accordingto the video clientrsquos requirements and network conditionsbased on the QoE-aware decision engine of the QoE con-troller

In particular MAP which is designed to run in Linuxkernel level in order to ensure minimum impact on the end-to-end delay acts as a transparent proxy of the mobile clientand it is responsible for parsing the packets that are destinedfor all mobile users over multiple ports Each received packetis forwarded in a queue and its header is parsed by anRTP parser process in order to identify the embedded videorelated information without changing the headerrsquos fieldsMoreover MAP has the responsibility to store this infor-mation in media independent information service (MIIS)database MAP is designed as a MIH-aware entity hence itcan directly store information required from the QoE modelto the MIIS database Specifically MAP regularly updates thedatabase with all client IDs (ie video service video view IDnumber of layers incoming data rate etc) thus it enablesthe QoE controller to train the collected data and predictthe perceptual video quality of experience The timely outputof the ML-based QoE model will be utilised by MAP inorder to maximise the perceived video QoE by selectingthe appropriate stream optimisationmethodThe abovemen-tioned algorithmic process is shown in Figure 3

32 IEEE 80221 Overview While the main purpose of IEEE80221 standard is to enhance the handover performance inheterogeneous networks [26] it also supports the importantaspect of link adaptation A resource-intensive multimediaapplication like video streaming requires link adaptation inorder for the network provider to offer the maximum levelof video experience As a result IEEE 80221 standard can beemployed to support cooperative use of information availableat both the mobile node and the network infrastructure Thisframework introduces a new entity called MIH Function(MIHF) which resides within the protocol stack of thenetwork elements and particularly between the link layer andthe upper layers Any protocol that uses the services providedby the MIHF is called a MIH user

As shown in Figure 4 MIHF provides three types ofservices the media independent event service (MIES) themedia independent command service (MICS) and themediaindependent information service (MIIS)TheMIESmonitorslink layer properties and initiates related events in order to


Media-Aware Proxy

DB functionsFiltering trafficRTP parserVideostreams

Media Independent Handover

MIIS

MIES(networkclient

side)DatabaseMICS

(client side)

DataSignalling

Enhance packet coreQoE controller

Figure 2 Proposed media-aware framework overview

MIIS database

(DB)

Receive IP packet

Parse headers(IP UDP RTP NAL)

Write to MIIS DB(refresh informationregarding layer_ID

bit rate etc)

Read MIIS DB(drop flag onoff)

Filtering process(drop packet)

Filtering process(forward packet)Off

Repeat process for each received packet

LTEEPC(tunnelingrouting)

On

Figure 3 Algorithmic representation of Media-Aware Proxy

inform both local and remote users TheMICS provides a setof commands for theMIHusers to control link properties andinterfaces Finally the MIIS provides information of staticnature for the candidate access networks such as frequencybands maximum data rate and link layer address

On the other hand there is the 3GGP approach whichdefines a new component called access network discovery

and selection function (ANDSF) [30] which in conjunctionwith the event reporting function (ERF) aims to facilitateinformation exchange for the network selection processHowever this approach is not as exhaustive and comprehen-sive as the MIES and can not support in depth monitoring ofthe link layer condition For this reason in the current workthe integration of MIH in the EPC architecture to support


L3 and above

MIH functions Remote MIHF

Link layer(80211 3G LTE)

MIE

SM

IES

MII

S

MIC

SM

ICS

MIESMICSMIIS

Figure 4 Illustration of overall MIH architecture and services

link layer monitoring is proposed and the related signallingis presented Furthermore the available set of monitoredcharacteristics is extended in order to include the applicationlayer and hence to encompass video related attributes

In order for theMIH framework to be effectively deployedin the EPC architecture the MIHF must be installed at bothnetwork elements and mobile nodes Additionally an MIISserver must be introduced that has a database with the staticattributes of the available access networks in the currentEPC domain We propose the extension of the MIIS serverrsquosfunctionalities so as to support the MIES and MICS servicesin order to store the reports of the involved entities andenforce policy rules applied by the QoE controller

A focal point of the described EPC architecture is thePDN-GW as it interconnects the various access routers suchas the Serving-GWfor LTE access and the evolved packet datagateway (ePDG) for non-3GPP access PGW also providesaccess to external packet data networks like the Internetthrough the SGi interface As a result it is reasonable topropose that this extended MIH server should be colocatedor at least should be at the vicinity of the PGW

33 MIH in EPC Architecture Within the context of thisresearch MIH is applied as a control mechanism respon-sible for monitoring and collecting QoE related parametersacross different layers and network entities in near realtime Towards this end MIH primitives including param-eters reporting can be utilised MIH supports differentoperating modes for parameters reporting For instancethe MIH Link Get Parameters command can be used toobtain the current value of a set of link parameters of a specificlink through the MIH Link Parameters Report primitiveAdditionally the same primitive can report the status of aset of parameters of a specific link at a predefined regularinterval determined by a user configurable timer Finallythe MIH Link Configure Thresholds command can beused to generate a report when a specified parameter crosses

Table 1 Summary of QoE related KPIs

Physical layer Network layer Application layerAccess technology Throughput Type of serviceSignal strength Delay Spatial index

Jitter Temporal indexPacket loss Quality layer

Quantization parameter

a configured threshold Subsequently the reporting can beperiodical event-driven or explicit depending on the desiredapproach or the available bandwidth for the monitoringfunctionality

The extended MIH server will be the destination of thesereports and will store themost recent values of themonitoredparameters to the MIIS database that can be shared with theQoE controller Both static attributes (eg the maximumavailable bandwidth) and dynamic attributes (eg the utilisedbandwidth) can be stored at the same database As the MAPis also aware of the characteristics of the video streamingsession the overall set of available monitored parameters willextend from the link layer to the application layer allowinga more efficient and precise modelling of the perceptual 3Dvideo quality Figure 5 depicts the basic signalling flow of theperiodic reporting scenarioThe available links are generatingperiodic reports of their status and inform their local MIHFthrough the Link Parameters ReportindicationThereafter the various MIHF report their set of monitoringparameters to MIIS through the MIH Link ParametersReportindication event Additionally a similar prim-itive is generated by the MAP in order to report video relatedparameters that are available to it

34 QoE Controller

341 QoE Related Key Performance Indicators There arethree basic categories of QoE parameters related to the phy-sical network and application layers that are monitored andcollected by MAP and MIH as shown in Table 1

(i) Physical layer related information includes the typeof the access technology the received signal strengthand oscillations of the signal strength Particularlyfor the WiFi this information can easily be collecteddirectly from the access point through queries orbroadcast messages

(ii) The network QoS parameters are also an importantsource of information for the QoE management andinclude the throughput the end-to-end IP layer delayjitter and packet loss MAP which is acting as anIP packet filtering function for the IP based videotraffic directed to the wireless users is able to parsethe packet header and retrieve the QoE related KPIs

(iii) The application parameters monitored involve thetype of application (video or audio although audio isout of the paperrsquos scope) the encoding characteristicsand the video content properties These are made


Mobile node

MIHF MAC-1

Access router

MIHF PoA

MAP

MIHF

PGW

MIH server

Link_Parameters_ReportindicationMIH_Link_Parameters_Reportindication

Link_Parameters_Reportindication

MIH_Link_Parameters_Reportindication

MIH_Video_Parameters_Reportindication

Periodic report of L2 parameters

Periodic report of video parameters

Figure 5 Signalling flow of periodic parameters report

available from the MAP by parsing the RTP headerof the received video traffic

In addition to the parameters of Table 1 a number ofadditional measurements need to be considered in order tomodel and monitor the perceived video QoE Three maingroups ofmeasurements are considered the radio quality thecontrol plane performance and the user plane QoS and QoEmeasurements [31 32]

The radio quality related measurements will be restrictedin just collecting the number of active user pieces of equipmentThe number of active user devices indicates how manysubscribers on average use the resources of the cell overa defined period of time and it can be used for traffic andradio resource planning Moreover a set of metrics relatedwith the control plane performance are collected includingthe number of connection drops where a suddenly occurringexception which may include loss of signal on the radiointerface can cause interruption of the connection betweenthe userrsquos equipment and the network In a scenario thatinvolves real-time video streaming a loss of an ongoingconnection with the content provider will cause extremedeterioration of the userrsquos experience In this case it ismandatory in addition to the connection drop ratio per cell toalso measure the connection drop ratio per service

Furthermore user plane QoE related performance indi-cators utilised by the QoE controller include the packet jitterdefined as the average of the deviations from the meannetwork latency This measurement is based on the arrivaltime stamp of successive received UDP packets at MAP Thethroughput is defined as the data volume of IP datagramstransmitted within a defined time period (usually this timeperiod is set to 1 second) In order to determine the IPthroughput of each client MAP parses the IP header andcollects and stores the source and destination address of theparticular packet The throughput will be measured at a high

sampling rate (ie every 333ms) for the duration of theongoing connection

342 ML-Based QoE Model The supervised ML is a learn-ing process based on instances that produce a generalisedhypothesis In turn this hypothesis could be applied as amean to forecast future instances [33] There are a numberof common steps distinct in all ML techniques (a) data setcollection (b) data prepossessing (c) features creation (d)algorithm selection and (e) learning and evaluation using testdataThe accuracy of themodel improves through the repeti-tion and the adjustment of any step that needs improvementSupervised ML algorithms could be categorized as follows[34]

(i) Logic-based in this category the most popular algo-rithms are decision trees where each node in thetree represents a feature of instances and each branchrepresents a value that the node can assume The dis-advantage of the algorithms in this category is thatthey cannot perform efficiently when numerical fea-tures are used

(ii) Perceptron-based artificial neural networks areincluded in this category Neural networks have beenapplied to a range of different real-world problemsand their accuracy is a function of the number ofneurons used as well as the processing cost Howeverneural networks may become inefficient when fedwith irrelevant features

(iii) Statistical the most well-known statistical learningalgorithms are the bit rate Bayesian network andthe k-nearest neighbour The first has the advantagethat it requires short computational time for trainingbut it is considered partial due to the fact that itassumes that it can distinguish between classes using a


single probability distribution The second algorithmis based on the principle that neighbouring instanceshave similar properties Although very simple to useas it requires only the number of nearest neighboursas input it is unreliable when applied on data sets withirrelevant features

(iv) Support VectorMachines (SVM) SVM learning algo-rithms have been proven to perform better whendealing withmultidimension and continuous featuresas well as when applied to inputs with a nonlinearrelationship between themAn example of inputswithnonlinear relationship is the video viewerrsquos QoE andthe qualitative metrics gathered from different pointsof the delivery chain

Within the context of this study a QoE controller whichis dwelling next to the wireless network providerrsquos coreimplements the proposed machine learning-based QoE pre-diction model The proposed model could be characterisedby fast response time in order to support seamless streamingadaptation low processing cost and efficiency In order toselect the best candidate for such a model three well-knownand widely used learning algorithms are investigated andcompared in terms of precision and efficiency The first isbit rate Bayesian classifier [35] the second is a decision treebased on the C45 algorithm [36] and the third is amultilayerperceptron network [37]

Naive Bayesian Classifier A simple method for acquiringknowledge and accurately predicting the class of an instancefrom a set of training instances which include the classinformation is the Bayesian classifiers Particularly the naiveBayesian classifier is a specialized form of Bayesian networkthat relies on two important simplifying assumptions Firstlyit assumes that given the class the predictive attributes areconditionally independent and secondly that the trainingdata set does not contain hidden attributes which potentiallycould affect the prediction process

These assumptions result in a very efficient classificationand learning algorithm In more detail assuming a randomvariable 119862 representing the class of an instance then 119888 is aparticular class label Similarly 119883 is the vector of randomvariables that denote the values of the attributes therefore119909 is a particular observed attribute value vector In order toclassify a test case 119909 the Bayes rule is applied to compute theprobability of each class given the vector of observed valuesfor the predictive attributes

119901 (119862 = 119888 | 119883 = 119909) =119901 (119862 = 119888) 119901 (119883 = 119909 | 119862 = 119888)

119901 (119883 = 119909) (1)

In (1) the denominator can be calculated given that theevent119883 = 119909 is a conjunction of assigned attribute values (ie1198831= 1199091and 1198832= 1199092and sdot sdot sdot and 119883

119896= 119909119896) and that these attributes

are assumed to be conditionally independent

119901 (119883 = 119909 | 119862 = 119888) = prod

119894

119901 (119883119894= 119909119894| 119862 = 119888) (2)

Evidently from (2) it is deduced that each numeric attri-bute is represented as a continuous probability distribution

over the range of the attributes values Commonly theseprobability distributions are assumed to be the normal dis-tribution represented by the mean and standard deviationvalues If the attributes are assumed to be continuous thenthe probability density function for a normal distribution is119901 (119883 = 119909 | 119862 = 119888) = 119866 (119909 120583

119888 120590119888) where

119866 (119909 120583119888 120590119888) =

119890minus(119909minus120583)

221205902

radic2120587120590

(3)

Hence given the class only the mean of the attribute andits standard deviation are required for the bit rate Bayes clas-sifier to classify the attributes For the purposes of this studythe naive Bayes classifier will be providing the baseline for theML-based QoE models in terms of accuracy and precisionThis is due to the fact that although it is very fast and robustto outliers and uses evidence from many attributes it is alsocharacterised by low performance ceiling on large databasesand assumes independence of attributes Moreover the naiveBayes classifier requires initial knowledge of many probabili-ties which results in a significant computational cost

Logical Decision Tree By definition a decision tree can beeither a leaf node labelled with a class or a structure contain-ing a test linked to two or more nodes (or subtrees) Hencean instance is classified by applying its attribute vector to thetree The tests are performed into these attributes reachingone or other leaf to complete the classification process Oneof the most widely used algorithms for constructing decisiontrees is C45 [36] Although C45 is not the most efficientalgorithm it has been selected due to the fact that the pro-duced results would be easily compared with similar researchworksThere are a number of assumptions commonly appliedfor C45 in order to increase its performance efficiency andclassification process

(i) When all cases belong to the same class then the treeis a leaf and is labelled with the particular class

(ii) For every attribute calculate the potential informa-tion provided by a test on the attribute according tothe probability of each case having a particular valuefor the attribute Additionally measure the informa-tion gain that results from a test on the attribute usingthe probabilities of each case with a particular valuefor the attribute being of a particular class

(iii) Depending on the current selection criterion findthe best attribute to create a new branch

In particular the selection (or splitting) criterion is thenormalized information gain Inherently C45 aims to iden-tify the attribute that possesses the highest information gainand create a splitting decision nodeThis algorithmic processcan be analytically represented with functions (4) assumingthat the entropy of the 119899-dimensional vector of attributes ofthe sample 119867() denotes the disorder on the data while theconditional entropy 119867(119895 | ) is derived from iterating overall possible values of

119867() = minus

119899

sum

119895=1

10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

log10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816


119867(119895 | ) =

10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

log10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

Gain ( 119895) = 119867 ( minus 119867 (119895 | )) (4)

The algorithm ultimately needs to perform pruning ofthe resulting tree in order to minimise the classificationerror caused by the outliers included in the training dataset However in the case of classifying video perceptualquality with the use of decision trees the training set containsspecialisations (eg a low number of very high or verylow MOS measurements) and hence the outlier detectionand cleansing of the training set needs to be performedbeforehand as described in Section 51

Multilayer PerceptronAmultilayer perceptron network (alsoknown as multilayer feed-forward network) has an inputlayer of neurons that is responsible for distributing the valuesin the vector of predictor variable values to the neurons ofthe hidden layers In addition to the predictor variables thereis a constant input of 10 called the bias that is fed to eachof the hidden layers The bias is multiplied by a weight andadded to the sum going into the neuron When in a neuronof the hidden layer the value from each input neuron ismultiplied by a weight and the resulting weighted values areadded together producing a weighted sum that in turn is fedto a transfer functionThe outputs from the transfer functionare distributed to the output layer Arriving at a neuron inthe output layer the value from each hidden layer neuronis multiplied by a weight and the resulting weighted valuesare added together producing a weighted sum which is fedinto the transfer function The output values of the transferfunction are the outputs of the network In principle thetransfer function could be any function and could also bedifferent for each node of the neural network In this studythe sigmoidal (s-shaped) function has been used in the formof 119891(119909) = 1(1 + exp(minus119909)) This is a commonly used functionthat also is mathematically convenient as it produces thefollowing derivative property 119889120590(119909)119889119909 = 120590(119909)(1 minus 120590(119909))

The training process of the multilayer perceptron is todetermine the set of weight values that will result in a closematch between the output from the neural network and theactual target values The algorithm precision depends on thenumber of neurons in the hidden layer If an inadequatenumber of neurons are used the network will be unable tomodel complex data and the resulting fit will be poor Iftoo many neurons are used the training time may becomeexcessively long and worse the networkmay overfit the dataWhen overfitting occurs the network will begin to modelrandom noise in the data In the context of this study severalvalidation experiments have been performed using differentnumber of neurons in the hidden layerThe best accuracy hasbeen achieved by using five neurons in one hidden layer in thenetwork as shown in Figure 6 The feed-forward multilayerperceptronminimises the error function but it may take timeto converge to a solution (ie the minimum value) whichmay be unpredictable due to the error that is added to theweight matrix in each iterationTherefore in order to control

Input layer Hidden layer (Wh) Output layer (WY)

X =

X1

X6

120590

120590

120590

Y

Wh56

Wh11

u5

u1

h5

h1

1

WYj5

WY11

Figure 6 Proposed multilayer perceptron model tested and vali-dated

the convergence rate the learning rate parameter was setto 03 and 350 iterations were performed until the systemconverged

4 Experimental Setup

41 Capturing and Processing of 3D Video Content The pro-posed QoE framework is validated through extensive exper-iments conducted over the test-bed platform of Figure 8For the purposes of this study four real-world capturedstereo video test sequences (ldquomartial artsrdquo ldquomusic concertrdquoldquopanel discussionrdquo and ldquoreportrdquo) with different spatial andtemporal indexes have been used in left-right 3D formatThe 3D video capturing was performed in accordance withthe specifications of [38] in order to provide 3D contentswhich can be simply classified (genres) for 3D video encodingperformances evaluation and comparison to provide multi-view contents with relevant 3D features from which one cancompute disparity maps and synthesise intermediate viewsto be rendered on different kinds of display and to allowadvances and exhaustive quality assessment benchmarksTowards this end an LGOptimus 3Dmobile phone was usedas stereoscopic capturing device during the shooting sessionThe cameras were configured as follows

(i) Synchronized stereo camera (f28 FullHD 30 fpsBayer format 65mm spaced)

(ii) Raw formats bayer format which is then convertedto obtain rgb and yuv sequences (+ color correctionnoise filtering mechanical misalignment and auto-convergence)

(iii) Encoding FullHD or lower side-by-side or top-bottom with subsampling + MPEG4 AVC

Snapshots of the test video sequences are shown inFigure 7 A detailed description of the capturing scenariosand content characteristics are described in [38] All sequen-ces are side-by-side with resolutions 640 times 720 pixels and960 times 1080 pixels per view at 25 frames per second TheH264SVC encoding was performed using the encoderprovided by Vanguard Software tool [39] configured to createtwo layers (one base layer and one enhancement layer) usingMGS quality scalability The SVC was the favourable choicefor the particular experiments as it allows 3D video contentto be delivered over heterogeneous wireless channels with amanageble bit rate The particular encoder inherently allows


(a) Martial arts (b) Music concert

(c) Panel discussion (d) Report

Figure 7 Test 3D video sequences used for MOS measurements

the use of a variety of pattent protected error concealmentschemes that can compansate for packet losses up to 5Each video frame was encapsulated in a single networkabstraction layer unit (NALU)with a size of 1400 bytes whichin turn was encapsulated into a single Real-Time TransportProtocol (RTP) packet with a size of 1400 bytes (payload)Thegenerated video packets are delivered through the networkusing UDPIP protocol stack An additional separate channelis responsible for the transmission of the Parameter Sets (PS)to the client through a TCPIP connection MoreoverNetEm[40] was used for emulating diverse networking conditionsvariable delay and loss in accordance with the describedexperimental scenarios Hence each experiment has beenrepeated 10 times in order to obtain valid statistical dataTable 2 summarises the encoding and streaming configura-tion parameters used in all experiments

Wireless Channel Error Model In order to model the impactof physical impairments on the QoE the Rayleigh fadingchannel of the simulated 80211g is represented by a two-stateMarkov model Each state is characterized by a different biterror rate (BER) in the physical layer that results from thestate and transitional probabilities of the Markov model asin Table 3Thewireless channel quality is characterized by theprobabilities of an error in the bad and good state 119875B and 119875Gand the probabilities 119875GG and 119875BB to remain at the good orthe bad state respectively

42MAC Layer LoadModel The load of the wireless channelwill result in time-outs whichwill then cause retransmissionsin the data link layer due to the contention-based accessthat is inherent in IEEE 80211 protocol Eventually since thepaper investigates real-time video streaming such a load inthe MAC layer will result in losses that in the application

Table 2 Encoding and streaming parameters

Video sequence Martialarts Munich Panel

discussion Report

Spatialtemporalindex 2117 258 3215 333

Number of frames 400 framesIntra period 8 framesQP used forbase-enhancementlayers

42ndash36 amp 36ndash30

Frame rate 25 frames per secondResolution perview 640 times 720amp 960 times 1080 pixels

Mediatransmission unit 1500 bytes

layer will be presented as artefacts and distortion of theperceptual video quality In order to emulate and control theload UDP traffic is generated and transmitted to both uplinkand downlink channels The constant-sized UDP packets aregenerated according to the Poisson distribution three timeswith mean values of 2Mbps 3Mbps and 4Mbps in eachdirection respectively Evidently such a setup will eventuallydouble the overall background traffic (ie load) over theWiFiaccess channel

43 IP Delay Variation Model Moreover for the purpose ofthis research and in order to model the delay variations inthe IP layer a constant plus gamma distribution is applied[41] similar to the one depicted in Figure 9 More preciselythe delay variations are modelled as a gamma distributionwith a constant scale factor of 120579 = 30ms and three different


MIHF

Media-aware filtering

SGi

Serving GW

S5S8

MIHF

S1-U

MIHF

ePDG

S2b

MIIS

PDN gateway

SWn

EPC platform MIHF

Video streamer

LTE access emulated over Ethernet

Emulated eNodeB

Dummynet

IEEE 8021

1n

QoE databaseQoE controller

Figure 8 Test-bed platform used during the experiments

00005

0010015

0020025

0030035

Gam

ma d

istrib

utio

n pd

f

50 100 150 200 250 300 350 400 450 5000Delay (ms)

k = 1 120579 = 30msk = 4 120579 = 30msk = 8 120579 = 30ms

Figure 9Delay variations in IP layer based on gammadistributions

shape factors in the range of 119896 = 1 4 and 8 Apparentlythis configuration will result in various mean delays andcorresponding variances During the experiments threesuch delay variations schemes were applied summarised inTable 4 in an effort to control the networking conditionsand produce comprehensible outcomes regarding the visualquality degradation of the decoded video

5 Subjective Experiments and Analysis

The subjective experiments were conducted in accordancewith the absolute category rating (ACR)method as defined in

Table 3 BER emulation parameters for the Rayleigh fading wirelesschannel

Bad channel Medium channel Good channel119875G 125119890

21291198902

1291198902

119875B 41311989014

1311989013

4111989012

119875GG 0996 0990 0987119875BB 0336 0690 0740

Table 4 IP layer delay variation parameters resulting from gammadistribution

Shape factor (119896) Mean (ms) Variance (ms)1 30 9004 120 36008 140 7200

ITU-TRecommendations [42ndash45]The perceived video qual-ity is rated in a scale from 0 [bad] to 5 [excellent] accordingto the standard In order to produce reliable and repeatableresults the subjective analysis of the video sequences hasbeen conducted in a controlled testing environment An LG32LM620S LED 3D screen with a resolution of 1920 times 1080pixels aspect ratio 16 9 peak display luminance 500 cdm2and contrast ratio of 5000000 1 along with passive glasses isused during the experiments to display the stereoscopic videosequencesThe viewing distance for the video evaluators is set


to 2m in accordance with the subjective assessment recom-mendationsThemeasured environmental luminance duringmeasurements is 200 lux and the wall behind the screenluminance is 20 lux as recommended by [42] The subjectiveevaluation was performed by twenty expert observers in agender balanced way who were asked to evaluate the overallvisual quality of the displayed video sequences A trainingsession was included in order for the observers to get familiarwith the rating procedure and understand how to recognizeartefacts and impairments on the video sequences

51 Statistical Measures Initially the collected MOSs wereanalysed for each subject across the different test conditionsusing the chi-square test [46] which verified the normality ofthe majority of MOS distributions Following the normalityverification of the MOS distributions the technique of out-liersrsquo detection and elimination was applied This techniqueis based on the detection and removal of scores in the caseswhere the difference between the mean subject vote and themean vote for this test case from all other subjects exceeds15 The applied outlier detection method determined amaximum of two outliers per test case

After removing the outliers the remaining MOSs werestatistically analysed in order to assess the perceptual qualityof the received video sequencesThe averageMOS of each set119878 of tested video sequence scenarios of size 119870 is denoted asMOS119896

MOS119896=sum119870

119896=1sum119873

119894=1MOS119894119896

119870119872 (5)

whereMOS119894119896is theMeanOpinion Score of the 119894th viewer for

the 119896th tested scenario of one of the video sequences and 119873is number of evaluators of the particular tested scenario Thestandard deviation of MOS

119896is defined by the square root of

the variance

1205902

119896=

sum119870

119896=1[sum119873

119894=1MOS119894119896119873 minusMOS

119896]

119870 minus 1 (6)

The confidence interval (CI) of the estimated mean MOSvalues which indicate the statistical relationship between themean MOS and the actual MOS is computed using Studentrsquos119905-distribution as follows

CI119895= 119905 times (1 minus

119886

2119873) times

120590119895

radic119873 (7)

where 119905times(1minus1205722119873) is the 119905-value that corresponds to a two-tailed 119905-Student distribution with 119873 minus 1 degrees of freedomand a desired significance level 120572 In this case 120572 = 005which corresponds to 95 significance and 120590

119895is the standard

deviation of theMOS of all subjects per test Furthermore thedegree of asymmetry of the collected MOS around the meanvalue of its distribution of its measured samples is measuredby skewness while the ldquopeakednessrdquo of the distribution ismeasured by its kurtosis as follows

120573 =1198983

11989832

2

120574 =1198984

1198982

2

(8)

1

2

3

4

5

Aver

age M

OS

2 4 6 8 100Delay variation shape factor k

Martial artMunich

Panel discussionReporting

Figure 10 Average MOS measured as a result of artefacts due todelay variations

where119898119895is the 119895th moment about the mean given by

119898119895=

sum119870

119896=1[sum119873

119894=1MOS119894119873 minusMOS

119896]2

119870

(9)

Finally the measured MOSs are subjected to an analysisof variance (ANOVA) The ANOVA tests the null hypothesisthat the means between two or more groups are equal underthe assumption that the sample populations are normallydistributed Although multifactor ANOVA would reveal thecomplex relationship of BER MAC layer load and delayvariations with the resulting QoE as in [47] in this study theaim is to determinewhether each of the three selected factorswhich can be considered as statistically independent doeshave a statistically significant impact on the resulting QoEIn this particular case the ANOVA is used in order to test thefollowing null hypothesis

(1) The mean MOSs due to the three delay variationgroups (ie for 119896 = 1 119896 = 4 and 119896 = 8) independentof the video sequence their spatial resolution andquantization step size are equal

(2) The mean MOSs due to the three MAC layer back-ground loads (ie 2Mbps 3Mbps and 4Mbps)independent of the video sequence their spatialresolution and quantization step size are equal

(3) The mean MOSs due to the three wireless channelBERs (ie Good Medium and Bad) independentof the video sequence their spatial resolution andquantization step size are equal

52 Analysis of MeasuredMOS The averagedMOSmeasure-ments of the four test video sequences under the abovemen-tioned experimental conditions are shown in Figures 10 11and 12 Each column of the figures represents the averageMOS scored by the viewers due to the impact of just one of thethree QoSmetrics (ie BERMAC load and delay variation)


1

2

3

4

5

1 2 3 4 5

Aver

age M

OS

MAC layer load (Mbs)

Martial artMunich


Figure 11 Average MOS measured as a result of artefacts due toMAC layer load

Good Medium BadAccess channel bit error rate characterisation

Martial artMunich


1

2

3

4

5

6

Aver

age M

OS

Figure 12 Average MOS measured as a result of artefacts due to biterror rate

while the other two metrics are assigned to their ldquobestrdquovalue (ie the value that results from the highest level ofvideo delivery QoS) The comparison of the averaged MOSsprovides an indication on the impact of the delay variationload and BER on the tested sequence It is evident that theQoS parameters have different effect depending on the videocharacteristics (ie spatial and temporal indexes) MoreoverTable 5 summarises the statistical parameters of MOS pertested video sequence

In order to get a better understanding of the MOS pertested video sequence the mean the standard deviation thevariance the skewness and the kurtosis are studiedThroughvariance the difficulty or ease of the evaluator to assess aparameter as well as the agreement between evaluators canbe addressed Hence a lower variance may indicate a higheragreement of the overall 3D MOS among viewers In Table 5the low variance of the MOS around the mean in all videosequences indicates that viewers agreed on the scores thatwere assigned to the overall 3D video quality in each testscenario A slightly higher variance of the ldquomartial artsrdquo

sequence may be due to the high temporal index of thesequence that caused a more erratic behaviour of losses thuscausing more discrepancies among the viewers

In the context of subjective evaluation skewness showsthe degree of asymmetry of the scores around the MOS valueof each distribution of samples for the given parameter (inthis case the video sequences) As a normal distribution hasa skewness of zero the results in Table 5 indicate a symmetryof the scores around theMOS for every tested video sequenceSimilarly to the study of the variance the ldquomartial artsrdquosequence has a slight asymmetry compared to the rest of thesequences due to the fact that the viewers of the sequenceapproached the maximum and minimum scores more fre-quently than in the other cases

The last column of Table 5 presents the kurtosis of theMOS of each video sequenceThe kurtosis is the indication ofhow outlier-prone a distribution is A normal distribution hasa default kurtosis value of 3 The lower the kurtosis the moreprone the distribution Nevertheless a kurtosis two timesbelow the standard error of the kurtosis which is calculatedas the square root of 24 divided by119873 (the number of MOSs)[48] could be considered as normal In this case as thereare 20 evaluators scoring 108 different scenarios per videosequence the kurtosis below 40 can be ignoredTherefore theconsistent low kurtosis of Table 5 for all video sequences canbe regarded as an indication of lack of outliers since all eval-uators assigned similar scores to the overall 3D video quality

Three one-way ANOVA methods have been performedin order to evaluate an equivalent number of null hypothesesregarding the three QoS related parameters under study(delay variations load and BER) The ANOVA results aresummarised in Tables 6 7 and 8 In more detail assumingthat the grand mean of a set of samples is the total of all thedata values divided by the total sample size then the totalvariation is comprised as the sum of the squares of the dif-ferences of each mean with the grand mean In ANOVA thebetween-group variation and the within-group variation aredefined By comparing the ratio of between-group varianceto within-group variance ANOVAdetermines if the variancecaused by the interaction between the samples is muchlarger when compared to the variance that appears withineach group then it is because the means are not the sameAdditionally if there are 119899 samples involved with one datavalue for each sample (the sample mean) then the between-groups variation has 119899 minus 1 degrees of freedom Similarly inthe case of within-group variation each sample has degreesof freedom equal to one less than their sample sizes and thereare 119899 samples the total degrees of freedom are 119899 less than thetotal sample size df = 119873119899 where 119873 is the total size of thesamplesThe variance due to the differences within individualsamples is denoted as MS (Mean Square) within groupsThis is the within-group variation divided by its degrees offreedomThe119865-test statistic is foundby dividing the between-group variance by the within-group variance The degrees offreedom for the numerator are the degrees of freedom for thebetween-group variance (119899minus1) and the degrees of freedom forthe denominator are the degrees of freedom for the within-group variance (119873 minus 119899) According to the 119865-test the decisionwill be to reject the null hypothesis if the 119865-test statistic from


Table 5 Statistical analysis of measured MOS per video test sequence

Video sequence Average MOS Standard dev Variance 95 CI Skewness KurtosisMartial arts 2390 0836 0698 0391 0284 2208Munich 2943 0695 0483 0325 0010 2288Panel discussion 3223 0570 0325 0266 0173 2416Reporting 3689 0451 0203 0211 minus0024 2278

Table 6 One-way ANOVA for the parameter delay variation

Source Sum of squares Degrees of freedom MS 119865 Probability of 119865Between 103542 2 517711 12695 491566119890

minus44

Within 174945 429 04078

Table 7 One-way ANOVA for the parameter MAC load

Source Sum of squares Degrees of freedom MS 119865 Probability of 119865Between 11697 2 58486 94 00001Within 26679 429 062189

Table 8 One-way ANOVA for the parameter BER


minus09

Within 254635 429 05936

Tables 6 7 and 8 is greater than the 119865-critical value with 119896minus1numerator and119873 minus 119896 denominator degrees of freedom

By studying Tables 6 7 and 8 it can be affirmed that allthree null hypotheses as defined previously can be rejectedhence there is significant statistical difference among thegroups with respect to delay variations (ie the differentshape factors 119896 of the gamma distribution have a significanteffect on the resulting 3D video MOS) MAC layer load (iethe three different mean values of the Poisson distributedbackground traffic affect significantly the 3D MOS) and theBER (ie the three different wireless channel conditions alterthe 3DMOS significantly)The above conclusions are derivedby the extremely low probability 119901 which is much lower thanthe significance level of 005 and the 119865-statistic which in turnindicates that one or two means significantly differ betweeneach other

Therefore one-way ANOVA affirms that the selection ofthe three QoS parameters (delay variation load and BER)as well as the values assigned to these parameters for thepurposes of the subjective evaluation of 3D MOS and sub-sequently the definition training and validation of amachinelearning QoE model based on these QoS parameters iscorrect

6 ML QoE Models Comparison

The training and the analysis of the results have beenperformed using Weka software tool [49] The output of thenaive Bayes classification is shown in Table 9 where for everyattribute of the data set the mean and standard deviationof the Gaussian distribution are shown Furthermore the

decision tree of the C45 ML algorithm that is implementedas J48 in Weka is illustrated in Figure 13 In accordance withthe results of the MOS comparison in the previous sectionthe jitter is the most important parameter hence it is locatedin the route of the tree (ie first splitting node)

All models are assessed in terms of predictive per-formance using the 10-fold cross-validation method Theconfusion matrix of each of the three models is shown inTable 10 It can be seen that the naive Bayesian classifier hasclassified wrongly 42 instances ofMOS class 2 43 instances ofMOS class 3 and all MOS class 5 instances This is expectedsince the naive Bayes classifier was selected to act as thebaseline of theML performance testsThe confusionmatricesof multilayer perceptron and C45 decision tree confirm thatbothmethods perform similarly and better than the Bayesianclassifier

Nevertheless the models accuracy (number of correctlyclassified instances) cannot be used for assessing the useful-ness of classification models built using unbalanced datasetsas in this case For this purpose the ldquoKappa statisticrdquo isused which is a generic term for several similar measuresof agreement used with categorical data Typically it is usedin assessing the degree to which two or more video viewersin this case examining the same video sequence agree whenit comes to assigning the MOSs Similarly to the correlationcoefficient its value is zero when there is no agreement onthe scores (purely random coincidences of rates) and is onewhen there is complete agreement The ldquoKappa statisticrdquo ofthe three models is shown in Table 11

Additionally the three ML algorithms considered in thiswork are compared with each other in terms of accuracy


Table 9 Output of naive Bayes classifier

Attributes Class (MOS)1 2 3 4 5

SI Mean 2021 2171 2228 2267 24Std dev 089 197 198 188 06667

TI Mean 1753 1416 1287 12 9Std dev 201 445 445 424 15

QP1

Mean 3915 3882 3955 3814 36Std dev 299 299 294 287 1

QP2

Mean 3315 3282 3355 3214 30Std dev 299 299 294 287 1

Resolution Mean 057 051 046 05 1Std dev 049 049 049 05 016

Jitter Mean 7 561 268 025 0Std dev 058 200 217 090 05833

Load Mean 342 313 298 264 2Std dev 067 082 081 071 016

Channel Mean 052 085 106 128 2Std dev 067 077 083 076 016

le4 gt4

le21 gt21

le2 gt2

gt1le1

le0 gt0

le3 gt3

Jitter

SI

Load

Channel2 (12020)

1 (16020) Resolution

1 (16020) Load

2 (20) 1 (20)

1 (16020)le1 gt1

le3 gt3Load

Channel

3 (8010) 2 (4010)

le1 gt1

gt21le21

le1 gt1

le2 gt2

le36 gt36

le0 gt0

Jitter

SI

Channel 3 (36060)

Load 3 (12010)

QP_1 2 (16010)

3 (40) Channel

2 (20) 3 (20)

le3 gt3

le1 gt1

le21 gt21

le36 gt36

le0 gt0

Load

Channel

SI3 (16030)

QP_1 4 (40)

Resolution 3 (20)

4 (10) 2 (10)

le21 gt21

le0 gt0

le2 gt2

SI

Channel

Load 4 (16010)

4 (40) 3 (4010)

le36 gt36

le0 gt0

le0 gt0

QP_1

Resolution4 (12020)

3 (60) Channel

3 (20) 4 (4010)

Figure 13 C45 based classification tree of the proposed model

precision (119901) and recall bit rate The precision of the algo-rithm which is also known as reproducibility or repeatabilitydenotes the degree to which repeated measurements (inthis case of MOS) under unchanged conditions show thesame results Recall is defined as the number of relevantinstances retrieved by a search divided by the total numberof existing relevant instances In addition to the two metricsthe algorithms are also evaluated in terms of 119865-measure (119865)which considers both the precision (119901) and the recall (119903) of thetest to compute the score Let 119894 isin [1 2 3 4 5] represent theclass which in this case is derived from the MOSs Then truepositive TP

119894is the number of the instances correctly classified

false positive FP119894is the number of the instances that belong to

the class bit rate but have not been classified there and falsenegative FN

119894is the number of the instances that do not belong

in class bit rate but have been classified there Thus (119901) (119903)and (119865) are derived as follows

119901 =TP119894

TP119894+ FP119894

119903 =TP119894

TP119894+ FN119894

119865 =2119901119903

119901 + 119903

(10)

Finally the accuracy of the models is also deduced bythe area under the Receiver Operating Characteristic (ROC)curve An area of one represents a perfect model while anarea of 05 means lack of any statistical dependencies Thesecomparison results are also shown in Table 11

It is evident that for the particular data set with themultiple classes and the nonlinear relationship among theattributes the best performance is achieved by the multilayerperceptron Nevertheless C45 is a very good alternative thatperforms MOS classification with acceptable accuracy andwith less processing effort Moreover since decision treesreturn results that have a physical significance thus theclassification can be easily interpreted and used as a decision


Table 10 Confusion matrices of the three ML models

119886 119887 119888 119889 119890 larr classified asNaive Bayesian confusion matrix

2 17 0 0 0 119886 = 1

2 39 39 1 0 119887 = 2

0 11 134 32 0 119888 = 3

0 0 40 109 0 119889 = 4

0 0 0 6 0 119890 = 5

Multilayer perceptron confusion matrix11 8 0 0 0 119886 = 1

6 63 12 0 0 119887 = 2

0 13 141 23 0 119888 = 3

0 0 21 128 0 119889 = 4

0 0 0 6 0 119890 = 5

C45 confusion matrix12 7 0 0 0 119886 = 1

6 63 12 0 0 119887 = 2

0 14 137 26 0 119888 = 3

0 1 25 121 2 119889 = 4

0 0 0 5 1 119890 = 5

Table 11 Performance comparison of ML algorithms

NaiveBayesian

Multilayerperceptron C45

Kappa statistic 0473 0693 0664Precision (119901) 0643 0782 0770Recall (119903) 0657 0794 0773119865-measure (119865) 0641 0788 0771Area under ROC 0851 0912 0886

rule C45 would be preferable as opposed to the neuralnetwork as a decision engine during real-time QoE control

7 Conclusion and Future Steps

Evidently a new wave of IP convergence with 3GPPrsquos LTEEPC in its heart is being fuelled by real-time multimediaapplications As new network architecture paradigms areconstantly evolving there is an ongoing effort by contentproviders and operators to provide and maintain premiumcontent access with optimised QoE This paper presents anovel media-aware framework for measuring and modellingthe 3D video user experience designed to be closely inte-grated with LTE EPC architecture The proposed solutioninvolves a synergy between aMedia-Aware Proxy networkingelement and the IEEE 80221 MIH protocol The former isresponsible for parsing RTP packets to collect QoE relatedKPIs and adapting the video streams to the current network-ing conditions in real timeThe latter is being enhanced to actas control mechanism for providing the necessary signallingto collect KPIs across different layers and network elementsMoreover MIH provides the central database which storesthe collected QoE related KPIs and allows a third component

named QoE controller to train this dataset and model theperceived 3D video quality Three machine learning classifi-cation algorithms for modelling QoE due to network relatedimpairments have been investigated Opposite to previousstudies the QoE model is a function of parameters collectedfrom the application the MAC and physical layers Henceinstead of determining the 3D QoE by measuring only theend-to-end packet loss this paper considers a set of qualityof service (QoS) parameters (ie bit error rate from physicallayer MAC layer load and network layer delayjitter) whichaccount for the overall number of lost packets A detailedstatistical analysis of the measured 3D MOS indicated thatthe predominant factor of QoE degradation is the IP layerjitter Therefore the proposed ML scheme aims to determinemore accurately the impact that different layer impairmentshave on the perceived 3D video experience A comparison ofthe three ML schemes under study in terms of precision andaccuracy indicated that themultilayer perceptron has the bestperformance closely followed by the C45 decision tree

This is an ongoing research Building upon the resultspresented in this paper the following steps involve the designand implementation of a QoE management mechanism thatwill optimise the perceived videoQoEby instructing differentnetwork elements of the EPC architecture and the MAPto either adapt the delivered video traffic to the networkconditions or enforce with the aid of MIH a content-awarehandover to a less loaded neighbouring channel or accessnetwork Moreover a closer integration with the currentand future LTE EPC implementations is currently understudy The aim is to reconfigure the proposed framework toincorporate the LTE policy and charging control (PCC) tech-niques in an effort to upgrade it into a content-aware trafficmechanism extended over all network operator deployments

Competing Interests

The authors declare that they have no competing interests

Acknowledgments

This work was supported by the SIEMENS ldquoExcellencerdquoProgram by the State Scholarship Foundation (IKY) Greece

References

[1] A Lykourgiotis K Birkos T Dagiuklas et al ldquoHybrid broad-cast and broadband networks convergence for immersive TVapplicationsrdquo IEEE Wireless Communications vol 21 no 3 pp62ndash69 2014

[2] G-M Su Y-C Lai A Kwasinski andHWang ldquo3D video com-munications challenges and opportunitiesrdquo International Jour-nal of Communication Systems vol 24 no 10 pp 1261ndash12812011

[3] C Tselios I Politis K Birkos T Dagiuklas and S KotsopoulosldquoCloud for multimedia applications and services over hetero-geneous networks ensuring QoErdquo in Proceedings of the IEEE18th International Workshop on Computer Aided Modeling andDesign of Communication Links andNetworks (CAMAD rsquo13) pp94ndash98 Berlin Germany September 2013


[4] K Birkos I Politis A Lykourgiotis et al ldquoTowards 3D videodelivery over heterogeneous networks the ROMEO approachrdquoin Proceedings of the International Conference on Telecommuni-cations and Multimedia (TEMU rsquo12) pp 60ndash65 IEEE ChaniaGreece August 2012

[5] ITU-T and ISOIEC ldquoITU-T recommendation H264 (2010)advanced video coding for generic audiovisual servicesrdquo Infor-mation Technology Coding of Audiovisual Objects Part 10Advanced Video Coding ISOIEC 14496-102010 2010

[6] A Vetro TWiegand and G J Sullivan ldquoOverview of the stereoand multiview video coding extensions of the H264MPEG-4AVC standardrdquo Proceedings of the IEEE vol 99 no 4 pp 626ndash642 2011

[7] G J Sullivan J-R OhmW-J Han and TWiegand ldquoOverviewof the high efficiency video coding (HEVC) standardrdquo IEEETransactions on Circuits and Systems for Video Technology vol22 no 12 pp 1649ndash1668 2012

[8] 3GPP ldquoTechnical specification group services and systemaspectsrdquo 3GPP TS 23002 V1330 (2015-09) Network Architec-ture 2015

[9] H Schwarz D Marpe and T Wiegand ldquoOverview of the scal-able video coding extension of the H264AVC standardrdquo IEEETransactions on Circuits and Systems for Video Technology vol17 no 9 pp 1103ndash1120 2007

[10] A Vetro P Pandit H Kimata A Smolic andY-KWang ldquoJointdraft 8 of multi-view video codingrdquo Tech Rep JVT-AB204Joint Video Team (JVT) Hannover Germany 2008

[11] G J Sullivan AM Tourapis T Yamakage andC S Lim ldquoDraftAVC amendment text to specify constrained baseline profilestereo high profile and frame packing SEImessagerdquo Joint VideoTeam (JVT) Document JVT-AE204 Joint Video Team (JVT)London UK 2009

[12] P Merkle A Smolic K Muller and T Wiegand ldquoEfficient pre-diction structures for multiview video codingrdquo IEEE Transac-tions on Circuits and Systems for Video Technology vol 17 no 11pp 1461ndash1473 2007

[13] I Politisa L Dounis C Tselios A Kordelas T Dagiuklas andA Papadakis ldquoAmodel of network relatedQoE for 3D videordquo inProceedings of the IEEE Globecom Workshops (GC Wkshps rsquo12)pp 1335ndash1340 Anaheim Calif USA December 2012

[14] C Tselios C Mysirlidis I Politis T Dagiuklas and S Kot-sopoulos ldquoOn quality evaluation of 3D video using Colour-plus-Depth and MDCrdquo in Proceedings of the IEEE 19th Inter-national Workshop on Computer Aided Modeling and Design ofCommunication Links and Networks (CAMAD rsquo14) pp 66ndash69IEEE Athens Greece December 2014

[15] Nokia-Siemens Networks ldquoQoS-enabled IP networks forassured multiplay QoErdquo Technical White Paper 2011

[16] K-T Chen C-J Chang C-C Wu Y-C Chang and C-LLei ldquoQuadrant of euphoria a crowdsourcing platform for QoEassessmentrdquo IEEE Network vol 24 no 2 pp 28ndash35 2010

[17] A Balachandran V Sekar A Akella S Seshan I Stoica andHZhang ldquoDeveloping a predictive model of quality of experiencefor internet videordquoACMSIGCOMMComputer CommunicationReview vol 43 no 4 pp 339ndash350 2013

[18] K U R Laghari and K Connelly ldquoToward total quality of expe-rience a QoE model in a communication ecosystemrdquo IEEECommunications Magazine vol 50 no 4 pp 58ndash65 2012

[19] A K Moorthy L K Choi A C Bovik and G De VecianaldquoVideo quality assessment onmobile devices subjective behav-ioral and objective studiesrdquo IEEE Journal on Selected Topics inSignal Processing vol 6 no 6 pp 652ndash671 2012

[20] V De Silva H K Arachchi E Ekmekcioglu et al ldquoPsycho-physical limits of interocular blur suppression and its applica-tion to asymmetric stereoscopic video deliveryrdquo in Proceedingsof the 19th International Packet Video Workshop (PV rsquo12) pp184ndash189 IEEE Munich Germany May 2012

[21] 3GPP ldquoGeneral Packet Radio Service (GPRS) enhancementsfor Evolved Universal Terrestrial Radio Access Network (E-UTRAN) accessrdquo 3GPP TS 23401 2016

[22] 3GPP ldquo3rd Generation Partnership Project Technical Specifi-cation Group Services and System Aspects Network Architec-turerdquo Tech Rep 3GPP TS 23002 V1250 2014

[23] IETF ldquoDiameter base protocolrdquo Tech Rep RFC 6733 IETF2012

[24] ETSI TS 123 228 V1290 ldquoIP Multimedia Subsystem (IMS)stage 2rdquo 3GPP TS 23228 version 1290 Release 12 2015

[25] Fraunhofer FOKUS OpenEPC Toolkit httpwwwopenepcnet

[26] IEEE 80221 Working Group ldquoIEEE standard for local andmetropolitan area networks part 21 media independent han-doverrdquo IEEE Standard 802 2008

[27] A Kordelas I Politis A Lykourgiotis T Dagiuklas and SKotsopoulos ldquoAn media aware platform for real-time stereo-scopic video streaming adaptationrdquo in Proceedings of the IEEEInternational Conference on Communications Workshops (ICCrsquo13) pp 687ndash691 Budapest Hungary June 2013

[28] IETF ldquoRTP payload format for H264 videordquo IETF Draft RFC3984 IETF 2010 draft-ietf-avt-rtp-rfc3984bis-12

[29] IETF Draft AudioVideo Transport WG ldquoRTP Payload Formatfor SVC Videordquo draft-ietf-avt-rtp-svc-27 2011

[30] 3GPP ldquoArchitecture enhancements for non-3GPP accessesrdquo3GPP TS 23402 2015

[31] H Ekstrom ldquoQoS control in the 3GPP evolved packet systemrdquoIEEE Communications Magazine vol 47 no 2 pp 76ndash83 2009

[32] H Luo S Ci DWu JWu andH Tang ldquoQuality-driven cross-layer optimized video delivery over LTErdquo IEEE Communica-tions Magazine vol 48 no 2 pp 102ndash109 2010

[33] V Menkovski A Oredope A Liotta and A Cuadra SanchezldquoPredicting quality of experience in multimedia streamingrdquo inProceedings of the 7th International Conference on Advances inMobile Computing and Multimedia (MoMM rsquo09) pp 52ndash59Kuala Lumpur Malaysia December 2009

[34] S B Kotsiantis ldquoSupervised machine learning a review ofclassification techniquesrdquo Informatica vol 31 no 3 pp 249ndash268 2007

[35] G H John and P Langley ldquoEstimating continuous distributionsin Bayesian classifiersrdquo in Proceedings of the 11th Conferenceon Uncertainty in Artificial Intelligence (UAI rsquo95) pp 338ndash345Montreal Canada August 1995

[36] J R Quinlan C45 Programs for Machine Learning vol 1Morgan Kaufmann Boston Mass USA 1993

[37] M W Gardner and S R Dorling ldquoArtificial neural networks(the multilayer perceptron)mdasha review of applications in theatmospheric sciencesrdquoAtmospheric Environment vol 32 no 14-15 pp 2627ndash2636 1998

[38] EU-FP7-ICT-ROMEO ldquoD31 Report on 3D media capture andcontent preparationrdquo May 2012 httpwwwict-romeoeu

[39] Vanguard Software Solutions H264SVC SDK httpwwwvanguardswcom

[40] S Hemminger ldquoNetwork emulation with NetEm Linux confaurdquo in Proceedings of the Australiarsquos National Linux Conference(linuxconfau rsquo05) Canberra Australia April 2005


[41] J C Bolot ldquoEnd-to-end packet delay and loss behavior in theInternetrdquo in Proceedings of the Conference on CommunicationsArchitectures Protocols and Applications (SIGCOMM rsquo93) pp289ndash298 1993

[42] ITU Radio Communication Sector ldquoMethodology for thesubjective assessment of the quality of television picturesrdquo TechRep ITU-R BT500-11 2002

[43] ITU Radio Communication Sector ldquoSubjective video qualityassessment methods for multimedia applicationsrdquo Tech RepITU-T Rec P910 1999

[44] ITU ldquoSubjective audiovisual quality assessment methods formultimedia applicationsrdquo ITU-T Recommendation P911 Inter-national TelecommunicationUnionGeneva Switzerland 1998

[45] ITU-T BT2021 Subjective Methods for the Assessment of Stereo-scopic 3DTV Systems International Telecommunication UnionGeneva Switzerland 2012

[46] DTICChi-Square Tests Defense Technical Information Center1976

[47] A Khan L Sun and E Ifeachor ldquoQoE predictionmodel and itsapplication in video quality adaptation over UMTS networksrdquoIEEE Transactions on Multimedia vol 14 no 2 pp 431ndash4422012

[48] B G Tabachnick L S Fidell and S J Osterlind Using Multi-variate Statistics 2001

[49] M Hall E Frank G Holmes B Pfahringer P Reutemann andI H Witten ldquoThe WEKA data mining software an updaterdquoACM SIGKDD Explorations Newsletter vol 11 no 1 pp 10ndash182009

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Distributed Sensor Networks


Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014


ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014


Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications


Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia


Biomedical Imaging


ArtificialNeural Systems

Advances in


RoboticsJournal of



Computational Intelligence and Neuroscience

Industrial EngineeringJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in



to provide a universal media bit-stream that can be decodedby multiple decoders of different capacities to produce thereconstructed media at different states SVC also providesdynamic adaptation to a diversity of networks terminalsand formats This is extremely useful for simple adaptationof transmission and efficient storage and is beneficial fortransmission services with uncertain resolution channelconditions and device types The MVC standard on theother hand in order to impove the overall rate-distortionperformance exploits the inter-view redundancies throughthe Disparity Compensated Prediction (DCP) [12] Nonethe-less the multiview bit rate increases to unmanageable levelsas the number of source cameras increases Therefore itis impractical to encode and transport the entire set ofviews required to be displayed on most multiview displaysparticularly when dealing with heterogeneous wireless accessnetworks Yet it remains unclear how all these technologiesaffect the viewerrsquos perception of 3D video [13 14] Apparentlyit is still unrealistic to try and estimate the QoE of suchapplication without resorting to full subjective evaluations

Any 3D video coding and delivery system will ultimatelyneed to be measured in terms of userrsquos satisfaction andperceived experience QoE can be defined as the overallacceptability of an application or service strictly from theusers point of view It is a subjective measure of end-to-endservice performance from the user perspective and it is anindication of howwell the networkmeets the users needs [15]Encompassing many different aspects QoE is riveted on thetrue feelings of end users when they watch streaming videolisten to digitized music and browse the Internet through aplethora ofmethods and devices [16]There is a real challengein creating models which will accurately perform learning tomodel service behaviours by taking into account parameterssuch as arrival pattern request service time distributionsIO system behaviours and network usage [17] The aimof such attempts is to estimate QoE for both resource-centric (utilization availability reliability and incentive)and user-centric (response time and fairness) environmentsover certain predefined Quality of Service (QoS) and QoEthresholds Evidently there is still no concrete evidence toprove the correlation between QoS and QoE particularly for3D video Nevertheless there are several attempts to definethe relationship of human perception of video quality to theinherently objective network conditions and the majority ofthis work concludes that such a relationship cannot be linear[18 19] Specifically authors in [20] proved the importanceof the impact that network delivery speed and latency haveon the human satisfaction but also determined that in aheterogeneous networking environment neither bandwidthnor latency is sufficient as attributes that indicate the rangeof issues which render a service as nonappealing to the enduser

This paper describes a framework for QoE-aware deliv-ering of 3D video across heterogeneous wireless networksTowards this end a synergy is proposed between the LTEEPCarchitecture and IEEE 80221 protocol that will ensure real-time control and monitoring of key performance indicatorsand parameters relative to the perceived 3D QoE across dif-ferent layers and entities The proposed framework includes

a QoE controller module housing a machine learning- (ML-)based decision making engine which trains data collectedfrom the user and the network planes in order to modelas accurate as possible the perceptual 3D video quality Itneeds to be underlined that the proposed architecture doesnot require any alteration of the standardised LTE EPCinterfaces hence it could be smoothly integrated with thecurrent mobile operatorsrsquo deployments The paper investi-gates machine learning-based techniques for producing anobjective QoE model of network related impairments Theproposed QoE model is a function of parameters collectednot only from the application layer but also from theunderlying layers Previous efforts on modelling QoE for 3Dvideo have been based on measuring the end-to-end packetloss and determining the perceivedQoE a process that wouldrequire feedback from the receiving end which in the caseof wireless networks may never arrive or may arrive toolate for the decision engine This research investigates a newmethodology for monitoring network imposed impairmentsto the perceived video quality that will result inmore accurateQoEmodels Specifically a set of QoS parameters is collectedfrom different points in the delivery chain (ie bit errorrate from physical layer MAC layer load and networklayer delayjitter) which account for the overall numberof lost packets Three classification schemes with differentcharacteristics have been studied in order to identify themostappropriate method for modelling 3D video quality due tonetwork impairments

The rest of the paper is organised as follows Section 2briefly presents the concept of LTE EPC and its main entitiesthat are relevant to the purposes of this study Section 3provides a detailed description of the proposed media-awareframework and focuses on the synergy between the Media-Aware Proxy the IEEE 80221 protocol and the QoE con-troller which enables collectingQoE related key performanceindicators across multiple layers and network componentsthe training of these data and the modelling of the perceived3D video QoE Additionally in this section the three ML-based classification models are introduced and their maindifferences are explained The setup of the experimental test-bed is described in Section 4 Moreover the section definesthe physical MAC and network layer parameters whoseimpact on the perceptual quality of 3D video is under studyThe collected Mean Opinion Score (MOS) from the subjec-tive experiments along with a detailed statistical analysisis presented in Section 5 Based on the measured 3D MOSSection 6 compares the performance of the three machinelearning methods in terms of precision and accuracy inorder to determine the most suitable classification schemefor implementation as part of the QoE controller FinallySection 7 concludes the paper and highlights the future aimsof this research

2 3GPP EPS Architecture

The technical realization of next generation mobile net-works requires that complementary wireless technologiesare integrated in order to provide ubiquitous multimediaservices ldquoanywhere any time and on any devicerdquo In the








E‐UTRAN


Internet

GSM

PSTN

UMTS

PDN gateway

Servinggateway

eNodeB

User equipment

PCRF

HSS

MME


DataPolicycharging


IP connectivity



















Media-Aware Proxy



MIIS

MIES(networkclient

side)DatabaseMICS

(client side)

DataSignalling



MIIS database

(DB)

Receive IP packet



bit rate etc)






On






L3 and above



MIE

SM

IES

MII

S

MIC

SM

ICS

MIESMICSMIIS












34 QoE Controller






Mobile node

MIHF MAC-1

Access router

MIHF PoA

MAP

MIHF

PGW

MIH server























119901 (119862 = 119888 | 119883 = 119909) =119901 (119862 = 119888) 119901 (119883 = 119909 | 119862 = 119888)

119901 (119883 = 119909) (1)




119901 (119883 = 119909 | 119862 = 119888) = prod

119894

119901 (119883119894= 119909119894| 119862 = 119888) (2)



119888 120590119888) where

119866 (119909 120583119888 120590119888) =

119890minus(119909minus120583)

221205902

radic2120587120590

(3)







119867() = minus

119899

sum

119895=1

10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

log10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816


119867(119895 | ) =

10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

log10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

Gain ( 119895) = 119867 ( minus 119867 (119895 | )) (4)





X =

X1

X6

120590

120590

120590

Y

Wh56

Wh11

u5

u1

h5

h1

1

WYj5

WY11


















discussion Report









MIHF


SGi

Serving GW

S5S8

MIHF

S1-U

MIHF

ePDG

S2b

MIIS

PDN gateway

SWn

EPC platform MIHF

Video streamer


Emulated eNodeB

Dummynet

IEEE 8021

1n



00005

0010015

0020025

0030035

Gam

ma d

istrib

utio

n pd

f

50 100 150 200 250 300 350 400 450 5000Delay (ms)

k = 1 120579 = 30msk = 4 120579 = 30msk = 8 120579 = 30ms







21291198902

1291198902

119875B 41311989014

1311989013

4111989012

119875GG 0996 0990 0987119875BB 0336 0690 0740








MOS119896=sum119870

119896=1sum119873

119894=1MOS119894119896

119870119872 (5)




the variance

1205902

119896=

sum119870

119896=1[sum119873

119894=1MOS119894119896119873 minusMOS

119896]

119870 minus 1 (6)


CI119895= 119905 times (1 minus

119886

2119873) times

120590119895

radic119873 (7)




120573 =1198983

11989832

2

120574 =1198984

1198982

2

(8)

1

2

3

4

5

Aver

age M

OS


Martial artMunich




119898119895=

sum119870

119896=1[sum119873

119894=1MOS119894119873 minusMOS

119896]2

119870

(9)







1

2

3

4

5

1 2 3 4 5

Aver

age M

OS


Martial artMunich




Martial artMunich


1

2

3

4

5

6

Aver

age M

OS













minus44

Within 174945 429 04078





minus09

Within 254635 429 05936













SI Mean 2021 2171 2228 2267 24Std dev 089 197 198 188 06667

TI Mean 1753 1416 1287 12 9Std dev 201 445 445 424 15

QP1

Mean 3915 3882 3955 3814 36Std dev 299 299 294 287 1

QP2

Mean 3315 3282 3355 3214 30Std dev 299 299 294 287 1



Load Mean 342 313 298 264 2Std dev 067 082 081 071 016


le4 gt4

le21 gt21

le2 gt2

gt1le1

le0 gt0

le3 gt3

Jitter

SI

Load

Channel2 (12020)


1 (16020) Load

2 (20) 1 (20)

1 (16020)le1 gt1

le3 gt3Load

Channel

3 (8010) 2 (4010)

le1 gt1

gt21le21

le1 gt1

le2 gt2

le36 gt36

le0 gt0

Jitter

SI

Channel 3 (36060)

Load 3 (12010)

QP_1 2 (16010)

3 (40) Channel

2 (20) 3 (20)

le3 gt3

le1 gt1

le21 gt21

le36 gt36

le0 gt0

Load

Channel

SI3 (16030)

QP_1 4 (40)

Resolution 3 (20)

4 (10) 2 (10)

le21 gt21

le0 gt0

le2 gt2

SI

Channel

Load 4 (16010)

4 (40) 3 (4010)

le36 gt36

le0 gt0

le0 gt0

QP_1

Resolution4 (12020)

3 (60) Channel

3 (20) 4 (4010)








119901 =TP119894

TP119894+ FP119894

119903 =TP119894

TP119894+ FN119894

119865 =2119901119903

119901 + 119903

(10)






2 17 0 0 0 119886 = 1

2 39 39 1 0 119887 = 2

0 11 134 32 0 119888 = 3

0 0 40 109 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 13 141 23 0 119888 = 3

0 0 21 128 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 14 137 26 0 119888 = 3

0 1 25 121 2 119889 = 4

0 0 0 5 1 119890 = 5


NaiveBayesian








Competing Interests


Acknowledgments


References



























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in










E‐UTRAN


Internet

GSM

PSTN

UMTS

PDN gateway

Servinggateway

eNodeB

User equipment

PCRF

HSS

MME


DataPolicycharging


IP connectivity



















Media-Aware Proxy



MIIS

MIES(networkclient

side)DatabaseMICS

(client side)

DataSignalling



MIIS database

(DB)

Receive IP packet



bit rate etc)






On






L3 and above



MIE

SM

IES

MII

S

MIC

SM

ICS

MIESMICSMIIS












34 QoE Controller






Mobile node

MIHF MAC-1

Access router

MIHF PoA

MAP

MIHF

PGW

MIH server























119901 (119862 = 119888 | 119883 = 119909) =119901 (119862 = 119888) 119901 (119883 = 119909 | 119862 = 119888)

119901 (119883 = 119909) (1)




119901 (119883 = 119909 | 119862 = 119888) = prod

119894

119901 (119883119894= 119909119894| 119862 = 119888) (2)



119888 120590119888) where

119866 (119909 120583119888 120590119888) =

119890minus(119909minus120583)

221205902

radic2120587120590

(3)







119867() = minus

119899

sum

119895=1

10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

log10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816


119867(119895 | ) =

10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

log10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

Gain ( 119895) = 119867 ( minus 119867 (119895 | )) (4)





X =

X1

X6

120590

120590

120590

Y

Wh56

Wh11

u5

u1

h5

h1

1

WYj5

WY11


















discussion Report









MIHF


SGi

Serving GW

S5S8

MIHF

S1-U

MIHF

ePDG

S2b

MIIS

PDN gateway

SWn

EPC platform MIHF

Video streamer


Emulated eNodeB

Dummynet

IEEE 8021

1n



00005

0010015

0020025

0030035

Gam

ma d

istrib

utio

n pd

f

50 100 150 200 250 300 350 400 450 5000Delay (ms)

k = 1 120579 = 30msk = 4 120579 = 30msk = 8 120579 = 30ms







21291198902

1291198902

119875B 41311989014

1311989013

4111989012

119875GG 0996 0990 0987119875BB 0336 0690 0740








MOS119896=sum119870

119896=1sum119873

119894=1MOS119894119896

119870119872 (5)




the variance

1205902

119896=

sum119870

119896=1[sum119873

119894=1MOS119894119896119873 minusMOS

119896]

119870 minus 1 (6)


CI119895= 119905 times (1 minus

119886

2119873) times

120590119895

radic119873 (7)




120573 =1198983

11989832

2

120574 =1198984

1198982

2

(8)

1

2

3

4

5

Aver

age M

OS


Martial artMunich




119898119895=

sum119870

119896=1[sum119873

119894=1MOS119894119873 minusMOS

119896]2

119870

(9)







1

2

3

4

5

1 2 3 4 5

Aver

age M

OS


Martial artMunich




Martial artMunich


1

2

3

4

5

6

Aver

age M

OS













minus44

Within 174945 429 04078





minus09

Within 254635 429 05936













SI Mean 2021 2171 2228 2267 24Std dev 089 197 198 188 06667

TI Mean 1753 1416 1287 12 9Std dev 201 445 445 424 15

QP1

Mean 3915 3882 3955 3814 36Std dev 299 299 294 287 1

QP2

Mean 3315 3282 3355 3214 30Std dev 299 299 294 287 1



Load Mean 342 313 298 264 2Std dev 067 082 081 071 016


le4 gt4

le21 gt21

le2 gt2

gt1le1

le0 gt0

le3 gt3

Jitter

SI

Load

Channel2 (12020)


1 (16020) Load

2 (20) 1 (20)

1 (16020)le1 gt1

le3 gt3Load

Channel

3 (8010) 2 (4010)

le1 gt1

gt21le21

le1 gt1

le2 gt2

le36 gt36

le0 gt0

Jitter

SI

Channel 3 (36060)

Load 3 (12010)

QP_1 2 (16010)

3 (40) Channel

2 (20) 3 (20)

le3 gt3

le1 gt1

le21 gt21

le36 gt36

le0 gt0

Load

Channel

SI3 (16030)

QP_1 4 (40)

Resolution 3 (20)

4 (10) 2 (10)

le21 gt21

le0 gt0

le2 gt2

SI

Channel

Load 4 (16010)

4 (40) 3 (4010)

le36 gt36

le0 gt0

le0 gt0

QP_1

Resolution4 (12020)

3 (60) Channel

3 (20) 4 (4010)








119901 =TP119894

TP119894+ FP119894

119903 =TP119894

TP119894+ FN119894

119865 =2119901119903

119901 + 119903

(10)






2 17 0 0 0 119886 = 1

2 39 39 1 0 119887 = 2

0 11 134 32 0 119888 = 3

0 0 40 109 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 13 141 23 0 119888 = 3

0 0 21 128 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 14 137 26 0 119888 = 3

0 1 25 121 2 119889 = 4

0 0 0 5 1 119890 = 5


NaiveBayesian








Competing Interests


Acknowledgments


References



























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in
















Media-Aware Proxy



MIIS

MIES(networkclient

side)DatabaseMICS

(client side)

DataSignalling



MIIS database

(DB)

Receive IP packet



bit rate etc)






On






L3 and above



MIE

SM

IES

MII

S

MIC

SM

ICS

MIESMICSMIIS












34 QoE Controller






Mobile node

MIHF MAC-1

Access router

MIHF PoA

MAP

MIHF

PGW

MIH server























119901 (119862 = 119888 | 119883 = 119909) =119901 (119862 = 119888) 119901 (119883 = 119909 | 119862 = 119888)

119901 (119883 = 119909) (1)




119901 (119883 = 119909 | 119862 = 119888) = prod

119894

119901 (119883119894= 119909119894| 119862 = 119888) (2)



119888 120590119888) where

119866 (119909 120583119888 120590119888) =

119890minus(119909minus120583)

221205902

radic2120587120590

(3)







119867() = minus

119899

sum

119895=1

10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

log10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816


119867(119895 | ) =

10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

log10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

Gain ( 119895) = 119867 ( minus 119867 (119895 | )) (4)





X =

X1

X6

120590

120590

120590

Y

Wh56

Wh11

u5

u1

h5

h1

1

WYj5

WY11


















discussion Report









MIHF


SGi

Serving GW

S5S8

MIHF

S1-U

MIHF

ePDG

S2b

MIIS

PDN gateway

SWn

EPC platform MIHF

Video streamer


Emulated eNodeB

Dummynet

IEEE 8021

1n



00005

0010015

0020025

0030035

Gam

ma d

istrib

utio

n pd

f

50 100 150 200 250 300 350 400 450 5000Delay (ms)

k = 1 120579 = 30msk = 4 120579 = 30msk = 8 120579 = 30ms







21291198902

1291198902

119875B 41311989014

1311989013

4111989012

119875GG 0996 0990 0987119875BB 0336 0690 0740








MOS119896=sum119870

119896=1sum119873

119894=1MOS119894119896

119870119872 (5)




the variance

1205902

119896=

sum119870

119896=1[sum119873

119894=1MOS119894119896119873 minusMOS

119896]

119870 minus 1 (6)


CI119895= 119905 times (1 minus

119886

2119873) times

120590119895

radic119873 (7)




120573 =1198983

11989832

2

120574 =1198984

1198982

2

(8)

1

2

3

4

5

Aver

age M

OS


Martial artMunich




119898119895=

sum119870

119896=1[sum119873

119894=1MOS119894119873 minusMOS

119896]2

119870

(9)







1

2

3

4

5

1 2 3 4 5

Aver

age M

OS


Martial artMunich




Martial artMunich


1

2

3

4

5

6

Aver

age M

OS













minus44

Within 174945 429 04078





minus09

Within 254635 429 05936













SI Mean 2021 2171 2228 2267 24Std dev 089 197 198 188 06667

TI Mean 1753 1416 1287 12 9Std dev 201 445 445 424 15

QP1

Mean 3915 3882 3955 3814 36Std dev 299 299 294 287 1

QP2

Mean 3315 3282 3355 3214 30Std dev 299 299 294 287 1



Load Mean 342 313 298 264 2Std dev 067 082 081 071 016


le4 gt4

le21 gt21

le2 gt2

gt1le1

le0 gt0

le3 gt3

Jitter

SI

Load

Channel2 (12020)


1 (16020) Load

2 (20) 1 (20)

1 (16020)le1 gt1

le3 gt3Load

Channel

3 (8010) 2 (4010)

le1 gt1

gt21le21

le1 gt1

le2 gt2

le36 gt36

le0 gt0

Jitter

SI

Channel 3 (36060)

Load 3 (12010)

QP_1 2 (16010)

3 (40) Channel

2 (20) 3 (20)

le3 gt3

le1 gt1

le21 gt21

le36 gt36

le0 gt0

Load

Channel

SI3 (16030)

QP_1 4 (40)

Resolution 3 (20)

4 (10) 2 (10)

le21 gt21

le0 gt0

le2 gt2

SI

Channel

Load 4 (16010)

4 (40) 3 (4010)

le36 gt36

le0 gt0

le0 gt0

QP_1

Resolution4 (12020)

3 (60) Channel

3 (20) 4 (4010)








119901 =TP119894

TP119894+ FP119894

119903 =TP119894

TP119894+ FN119894

119865 =2119901119903

119901 + 119903

(10)






2 17 0 0 0 119886 = 1

2 39 39 1 0 119887 = 2

0 11 134 32 0 119888 = 3

0 0 40 109 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 13 141 23 0 119888 = 3

0 0 21 128 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 14 137 26 0 119888 = 3

0 1 25 121 2 119889 = 4

0 0 0 5 1 119890 = 5


NaiveBayesian








Competing Interests


Acknowledgments


References



























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




L3 and above



MIE

SM

IES

MII

S

MIC

SM

ICS

MIESMICSMIIS












34 QoE Controller






Mobile node

MIHF MAC-1

Access router

MIHF PoA

MAP

MIHF

PGW

MIH server























119901 (119862 = 119888 | 119883 = 119909) =119901 (119862 = 119888) 119901 (119883 = 119909 | 119862 = 119888)

119901 (119883 = 119909) (1)




119901 (119883 = 119909 | 119862 = 119888) = prod

119894

119901 (119883119894= 119909119894| 119862 = 119888) (2)



119888 120590119888) where

119866 (119909 120583119888 120590119888) =

119890minus(119909minus120583)

221205902

radic2120587120590

(3)







119867() = minus

119899

sum

119895=1

10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

log10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816


119867(119895 | ) =

10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

log10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

Gain ( 119895) = 119867 ( minus 119867 (119895 | )) (4)





X =

X1

X6

120590

120590

120590

Y

Wh56

Wh11

u5

u1

h5

h1

1

WYj5

WY11


















discussion Report









MIHF


SGi

Serving GW

S5S8

MIHF

S1-U

MIHF

ePDG

S2b

MIIS

PDN gateway

SWn

EPC platform MIHF

Video streamer


Emulated eNodeB

Dummynet

IEEE 8021

1n



00005

0010015

0020025

0030035

Gam

ma d

istrib

utio

n pd

f

50 100 150 200 250 300 350 400 450 5000Delay (ms)

k = 1 120579 = 30msk = 4 120579 = 30msk = 8 120579 = 30ms







21291198902

1291198902

119875B 41311989014

1311989013

4111989012

119875GG 0996 0990 0987119875BB 0336 0690 0740








MOS119896=sum119870

119896=1sum119873

119894=1MOS119894119896

119870119872 (5)




the variance

1205902

119896=

sum119870

119896=1[sum119873

119894=1MOS119894119896119873 minusMOS

119896]

119870 minus 1 (6)


CI119895= 119905 times (1 minus

119886

2119873) times

120590119895

radic119873 (7)




120573 =1198983

11989832

2

120574 =1198984

1198982

2

(8)

1

2

3

4

5

Aver

age M

OS


Martial artMunich




119898119895=

sum119870

119896=1[sum119873

119894=1MOS119894119873 minusMOS

119896]2

119870

(9)







1

2

3

4

5

1 2 3 4 5

Aver

age M

OS


Martial artMunich




Martial artMunich


1

2

3

4

5

6

Aver

age M

OS













minus44

Within 174945 429 04078





minus09

Within 254635 429 05936













SI Mean 2021 2171 2228 2267 24Std dev 089 197 198 188 06667

TI Mean 1753 1416 1287 12 9Std dev 201 445 445 424 15

QP1

Mean 3915 3882 3955 3814 36Std dev 299 299 294 287 1

QP2

Mean 3315 3282 3355 3214 30Std dev 299 299 294 287 1



Load Mean 342 313 298 264 2Std dev 067 082 081 071 016


le4 gt4

le21 gt21

le2 gt2

gt1le1

le0 gt0

le3 gt3

Jitter

SI

Load

Channel2 (12020)


1 (16020) Load

2 (20) 1 (20)

1 (16020)le1 gt1

le3 gt3Load

Channel

3 (8010) 2 (4010)

le1 gt1

gt21le21

le1 gt1

le2 gt2

le36 gt36

le0 gt0

Jitter

SI

Channel 3 (36060)

Load 3 (12010)

QP_1 2 (16010)

3 (40) Channel

2 (20) 3 (20)

le3 gt3

le1 gt1

le21 gt21

le36 gt36

le0 gt0

Load

Channel

SI3 (16030)

QP_1 4 (40)

Resolution 3 (20)

4 (10) 2 (10)

le21 gt21

le0 gt0

le2 gt2

SI

Channel

Load 4 (16010)

4 (40) 3 (4010)

le36 gt36

le0 gt0

le0 gt0

QP_1

Resolution4 (12020)

3 (60) Channel

3 (20) 4 (4010)








119901 =TP119894

TP119894+ FP119894

119903 =TP119894

TP119894+ FN119894

119865 =2119901119903

119901 + 119903

(10)






2 17 0 0 0 119886 = 1

2 39 39 1 0 119887 = 2

0 11 134 32 0 119888 = 3

0 0 40 109 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 13 141 23 0 119888 = 3

0 0 21 128 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 14 137 26 0 119888 = 3

0 1 25 121 2 119889 = 4

0 0 0 5 1 119890 = 5


NaiveBayesian








Competing Interests


Acknowledgments


References



























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




Mobile node

MIHF MAC-1

Access router

MIHF PoA

MAP

MIHF

PGW

MIH server























119901 (119862 = 119888 | 119883 = 119909) =119901 (119862 = 119888) 119901 (119883 = 119909 | 119862 = 119888)

119901 (119883 = 119909) (1)




119901 (119883 = 119909 | 119862 = 119888) = prod

119894

119901 (119883119894= 119909119894| 119862 = 119888) (2)



119888 120590119888) where

119866 (119909 120583119888 120590119888) =

119890minus(119909minus120583)

221205902

radic2120587120590

(3)







119867() = minus

119899

sum

119895=1

10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

log10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816


119867(119895 | ) =

10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

log10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

Gain ( 119895) = 119867 ( minus 119867 (119895 | )) (4)





X =

X1

X6

120590

120590

120590

Y

Wh56

Wh11

u5

u1

h5

h1

1

WYj5

WY11


















discussion Report









MIHF


SGi

Serving GW

S5S8

MIHF

S1-U

MIHF

ePDG

S2b

MIIS

PDN gateway

SWn

EPC platform MIHF

Video streamer


Emulated eNodeB

Dummynet

IEEE 8021

1n



00005

0010015

0020025

0030035

Gam

ma d

istrib

utio

n pd

f

50 100 150 200 250 300 350 400 450 5000Delay (ms)

k = 1 120579 = 30msk = 4 120579 = 30msk = 8 120579 = 30ms







21291198902

1291198902

119875B 41311989014

1311989013

4111989012

119875GG 0996 0990 0987119875BB 0336 0690 0740








MOS119896=sum119870

119896=1sum119873

119894=1MOS119894119896

119870119872 (5)




the variance

1205902

119896=

sum119870

119896=1[sum119873

119894=1MOS119894119896119873 minusMOS

119896]

119870 minus 1 (6)


CI119895= 119905 times (1 minus

119886

2119873) times

120590119895

radic119873 (7)




120573 =1198983

11989832

2

120574 =1198984

1198982

2

(8)

1

2

3

4

5

Aver

age M

OS


Martial artMunich




119898119895=

sum119870

119896=1[sum119873

119894=1MOS119894119873 minusMOS

119896]2

119870

(9)







1

2

3

4

5

1 2 3 4 5

Aver

age M

OS


Martial artMunich




Martial artMunich


1

2

3

4

5

6

Aver

age M

OS













minus44

Within 174945 429 04078





minus09

Within 254635 429 05936













SI Mean 2021 2171 2228 2267 24Std dev 089 197 198 188 06667

TI Mean 1753 1416 1287 12 9Std dev 201 445 445 424 15

QP1

Mean 3915 3882 3955 3814 36Std dev 299 299 294 287 1

QP2

Mean 3315 3282 3355 3214 30Std dev 299 299 294 287 1



Load Mean 342 313 298 264 2Std dev 067 082 081 071 016


le4 gt4

le21 gt21

le2 gt2

gt1le1

le0 gt0

le3 gt3

Jitter

SI

Load

Channel2 (12020)


1 (16020) Load

2 (20) 1 (20)

1 (16020)le1 gt1

le3 gt3Load

Channel

3 (8010) 2 (4010)

le1 gt1

gt21le21

le1 gt1

le2 gt2

le36 gt36

le0 gt0

Jitter

SI

Channel 3 (36060)

Load 3 (12010)

QP_1 2 (16010)

3 (40) Channel

2 (20) 3 (20)

le3 gt3

le1 gt1

le21 gt21

le36 gt36

le0 gt0

Load

Channel

SI3 (16030)

QP_1 4 (40)

Resolution 3 (20)

4 (10) 2 (10)

le21 gt21

le0 gt0

le2 gt2

SI

Channel

Load 4 (16010)

4 (40) 3 (4010)

le36 gt36

le0 gt0

le0 gt0

QP_1

Resolution4 (12020)

3 (60) Channel

3 (20) 4 (4010)








119901 =TP119894

TP119894+ FP119894

119903 =TP119894

TP119894+ FN119894

119865 =2119901119903

119901 + 119903

(10)






2 17 0 0 0 119886 = 1

2 39 39 1 0 119887 = 2

0 11 134 32 0 119888 = 3

0 0 40 109 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 13 141 23 0 119888 = 3

0 0 21 128 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 14 137 26 0 119888 = 3

0 1 25 121 2 119889 = 4

0 0 0 5 1 119890 = 5


NaiveBayesian








Competing Interests


Acknowledgments


References



























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in









119901 (119862 = 119888 | 119883 = 119909) =119901 (119862 = 119888) 119901 (119883 = 119909 | 119862 = 119888)

119901 (119883 = 119909) (1)




119901 (119883 = 119909 | 119862 = 119888) = prod

119894

119901 (119883119894= 119909119894| 119862 = 119888) (2)



119888 120590119888) where

119866 (119909 120583119888 120590119888) =

119890minus(119909minus120583)

221205902

radic2120587120590

(3)







119867() = minus

119899

sum

119895=1

10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

log10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816


119867(119895 | ) =

10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

log10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

Gain ( 119895) = 119867 ( minus 119867 (119895 | )) (4)





X =

X1

X6

120590

120590

120590

Y

Wh56

Wh11

u5

u1

h5

h1

1

WYj5

WY11


















discussion Report









MIHF


SGi

Serving GW

S5S8

MIHF

S1-U

MIHF

ePDG

S2b

MIIS

PDN gateway

SWn

EPC platform MIHF

Video streamer


Emulated eNodeB

Dummynet

IEEE 8021

1n



00005

0010015

0020025

0030035

Gam

ma d

istrib

utio

n pd

f

50 100 150 200 250 300 350 400 450 5000Delay (ms)

k = 1 120579 = 30msk = 4 120579 = 30msk = 8 120579 = 30ms







21291198902

1291198902

119875B 41311989014

1311989013

4111989012

119875GG 0996 0990 0987119875BB 0336 0690 0740








MOS119896=sum119870

119896=1sum119873

119894=1MOS119894119896

119870119872 (5)




the variance

1205902

119896=

sum119870

119896=1[sum119873

119894=1MOS119894119896119873 minusMOS

119896]

119870 minus 1 (6)


CI119895= 119905 times (1 minus

119886

2119873) times

120590119895

radic119873 (7)




120573 =1198983

11989832

2

120574 =1198984

1198982

2

(8)

1

2

3

4

5

Aver

age M

OS


Martial artMunich




119898119895=

sum119870

119896=1[sum119873

119894=1MOS119894119873 minusMOS

119896]2

119870

(9)







1

2

3

4

5

1 2 3 4 5

Aver

age M

OS


Martial artMunich




Martial artMunich


1

2

3

4

5

6

Aver

age M

OS













minus44

Within 174945 429 04078





minus09

Within 254635 429 05936













SI Mean 2021 2171 2228 2267 24Std dev 089 197 198 188 06667

TI Mean 1753 1416 1287 12 9Std dev 201 445 445 424 15

QP1

Mean 3915 3882 3955 3814 36Std dev 299 299 294 287 1

QP2

Mean 3315 3282 3355 3214 30Std dev 299 299 294 287 1



Load Mean 342 313 298 264 2Std dev 067 082 081 071 016


le4 gt4

le21 gt21

le2 gt2

gt1le1

le0 gt0

le3 gt3

Jitter

SI

Load

Channel2 (12020)


1 (16020) Load

2 (20) 1 (20)

1 (16020)le1 gt1

le3 gt3Load

Channel

3 (8010) 2 (4010)

le1 gt1

gt21le21

le1 gt1

le2 gt2

le36 gt36

le0 gt0

Jitter

SI

Channel 3 (36060)

Load 3 (12010)

QP_1 2 (16010)

3 (40) Channel

2 (20) 3 (20)

le3 gt3

le1 gt1

le21 gt21

le36 gt36

le0 gt0

Load

Channel

SI3 (16030)

QP_1 4 (40)

Resolution 3 (20)

4 (10) 2 (10)

le21 gt21

le0 gt0

le2 gt2

SI

Channel

Load 4 (16010)

4 (40) 3 (4010)

le36 gt36

le0 gt0

le0 gt0

QP_1

Resolution4 (12020)

3 (60) Channel

3 (20) 4 (4010)








119901 =TP119894

TP119894+ FP119894

119903 =TP119894

TP119894+ FN119894

119865 =2119901119903

119901 + 119903

(10)






2 17 0 0 0 119886 = 1

2 39 39 1 0 119887 = 2

0 11 134 32 0 119888 = 3

0 0 40 109 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 13 141 23 0 119888 = 3

0 0 21 128 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 14 137 26 0 119888 = 3

0 1 25 121 2 119889 = 4

0 0 0 5 1 119890 = 5


NaiveBayesian








Competing Interests


Acknowledgments


References



























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




119867(119895 | ) =

10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

log10038161003816100381610038161003816119910119895

1003816100381610038161003816100381610038161003816100381610038161003816100381610038161003816

Gain ( 119895) = 119867 ( minus 119867 (119895 | )) (4)





X =

X1

X6

120590

120590

120590

Y

Wh56

Wh11

u5

u1

h5

h1

1

WYj5

WY11


















discussion Report









MIHF


SGi

Serving GW

S5S8

MIHF

S1-U

MIHF

ePDG

S2b

MIIS

PDN gateway

SWn

EPC platform MIHF

Video streamer


Emulated eNodeB

Dummynet

IEEE 8021

1n



00005

0010015

0020025

0030035

Gam

ma d

istrib

utio

n pd

f

50 100 150 200 250 300 350 400 450 5000Delay (ms)

k = 1 120579 = 30msk = 4 120579 = 30msk = 8 120579 = 30ms







21291198902

1291198902

119875B 41311989014

1311989013

4111989012

119875GG 0996 0990 0987119875BB 0336 0690 0740








MOS119896=sum119870

119896=1sum119873

119894=1MOS119894119896

119870119872 (5)




the variance

1205902

119896=

sum119870

119896=1[sum119873

119894=1MOS119894119896119873 minusMOS

119896]

119870 minus 1 (6)


CI119895= 119905 times (1 minus

119886

2119873) times

120590119895

radic119873 (7)




120573 =1198983

11989832

2

120574 =1198984

1198982

2

(8)

1

2

3

4

5

Aver

age M

OS


Martial artMunich




119898119895=

sum119870

119896=1[sum119873

119894=1MOS119894119873 minusMOS

119896]2

119870

(9)







1

2

3

4

5

1 2 3 4 5

Aver

age M

OS


Martial artMunich




Martial artMunich


1

2

3

4

5

6

Aver

age M

OS













minus44

Within 174945 429 04078





minus09

Within 254635 429 05936













SI Mean 2021 2171 2228 2267 24Std dev 089 197 198 188 06667

TI Mean 1753 1416 1287 12 9Std dev 201 445 445 424 15

QP1

Mean 3915 3882 3955 3814 36Std dev 299 299 294 287 1

QP2

Mean 3315 3282 3355 3214 30Std dev 299 299 294 287 1



Load Mean 342 313 298 264 2Std dev 067 082 081 071 016


le4 gt4

le21 gt21

le2 gt2

gt1le1

le0 gt0

le3 gt3

Jitter

SI

Load

Channel2 (12020)


1 (16020) Load

2 (20) 1 (20)

1 (16020)le1 gt1

le3 gt3Load

Channel

3 (8010) 2 (4010)

le1 gt1

gt21le21

le1 gt1

le2 gt2

le36 gt36

le0 gt0

Jitter

SI

Channel 3 (36060)

Load 3 (12010)

QP_1 2 (16010)

3 (40) Channel

2 (20) 3 (20)

le3 gt3

le1 gt1

le21 gt21

le36 gt36

le0 gt0

Load

Channel

SI3 (16030)

QP_1 4 (40)

Resolution 3 (20)

4 (10) 2 (10)

le21 gt21

le0 gt0

le2 gt2

SI

Channel

Load 4 (16010)

4 (40) 3 (4010)

le36 gt36

le0 gt0

le0 gt0

QP_1

Resolution4 (12020)

3 (60) Channel

3 (20) 4 (4010)








119901 =TP119894

TP119894+ FP119894

119903 =TP119894

TP119894+ FN119894

119865 =2119901119903

119901 + 119903

(10)






2 17 0 0 0 119886 = 1

2 39 39 1 0 119887 = 2

0 11 134 32 0 119888 = 3

0 0 40 109 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 13 141 23 0 119888 = 3

0 0 21 128 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 14 137 26 0 119888 = 3

0 1 25 121 2 119889 = 4

0 0 0 5 1 119890 = 5


NaiveBayesian








Competing Interests


Acknowledgments


References



























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in












discussion Report









MIHF


SGi

Serving GW

S5S8

MIHF

S1-U

MIHF

ePDG

S2b

MIIS

PDN gateway

SWn

EPC platform MIHF

Video streamer


Emulated eNodeB

Dummynet

IEEE 8021

1n



00005

0010015

0020025

0030035

Gam

ma d

istrib

utio

n pd

f

50 100 150 200 250 300 350 400 450 5000Delay (ms)

k = 1 120579 = 30msk = 4 120579 = 30msk = 8 120579 = 30ms







21291198902

1291198902

119875B 41311989014

1311989013

4111989012

119875GG 0996 0990 0987119875BB 0336 0690 0740








MOS119896=sum119870

119896=1sum119873

119894=1MOS119894119896

119870119872 (5)




the variance

1205902

119896=

sum119870

119896=1[sum119873

119894=1MOS119894119896119873 minusMOS

119896]

119870 minus 1 (6)


CI119895= 119905 times (1 minus

119886

2119873) times

120590119895

radic119873 (7)




120573 =1198983

11989832

2

120574 =1198984

1198982

2

(8)

1

2

3

4

5

Aver

age M

OS


Martial artMunich




119898119895=

sum119870

119896=1[sum119873

119894=1MOS119894119873 minusMOS

119896]2

119870

(9)







1

2

3

4

5

1 2 3 4 5

Aver

age M

OS


Martial artMunich




Martial artMunich


1

2

3

4

5

6

Aver

age M

OS













minus44

Within 174945 429 04078





minus09

Within 254635 429 05936













SI Mean 2021 2171 2228 2267 24Std dev 089 197 198 188 06667

TI Mean 1753 1416 1287 12 9Std dev 201 445 445 424 15

QP1

Mean 3915 3882 3955 3814 36Std dev 299 299 294 287 1

QP2

Mean 3315 3282 3355 3214 30Std dev 299 299 294 287 1



Load Mean 342 313 298 264 2Std dev 067 082 081 071 016


le4 gt4

le21 gt21

le2 gt2

gt1le1

le0 gt0

le3 gt3

Jitter

SI

Load

Channel2 (12020)


1 (16020) Load

2 (20) 1 (20)

1 (16020)le1 gt1

le3 gt3Load

Channel

3 (8010) 2 (4010)

le1 gt1

gt21le21

le1 gt1

le2 gt2

le36 gt36

le0 gt0

Jitter

SI

Channel 3 (36060)

Load 3 (12010)

QP_1 2 (16010)

3 (40) Channel

2 (20) 3 (20)

le3 gt3

le1 gt1

le21 gt21

le36 gt36

le0 gt0

Load

Channel

SI3 (16030)

QP_1 4 (40)

Resolution 3 (20)

4 (10) 2 (10)

le21 gt21

le0 gt0

le2 gt2

SI

Channel

Load 4 (16010)

4 (40) 3 (4010)

le36 gt36

le0 gt0

le0 gt0

QP_1

Resolution4 (12020)

3 (60) Channel

3 (20) 4 (4010)








119901 =TP119894

TP119894+ FP119894

119903 =TP119894

TP119894+ FN119894

119865 =2119901119903

119901 + 119903

(10)






2 17 0 0 0 119886 = 1

2 39 39 1 0 119887 = 2

0 11 134 32 0 119888 = 3

0 0 40 109 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 13 141 23 0 119888 = 3

0 0 21 128 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 14 137 26 0 119888 = 3

0 1 25 121 2 119889 = 4

0 0 0 5 1 119890 = 5


NaiveBayesian








Competing Interests


Acknowledgments


References



























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




MIHF


SGi

Serving GW

S5S8

MIHF

S1-U

MIHF

ePDG

S2b

MIIS

PDN gateway

SWn

EPC platform MIHF

Video streamer


Emulated eNodeB

Dummynet

IEEE 8021

1n



00005

0010015

0020025

0030035

Gam

ma d

istrib

utio

n pd

f

50 100 150 200 250 300 350 400 450 5000Delay (ms)

k = 1 120579 = 30msk = 4 120579 = 30msk = 8 120579 = 30ms







21291198902

1291198902

119875B 41311989014

1311989013

4111989012

119875GG 0996 0990 0987119875BB 0336 0690 0740








MOS119896=sum119870

119896=1sum119873

119894=1MOS119894119896

119870119872 (5)




the variance

1205902

119896=

sum119870

119896=1[sum119873

119894=1MOS119894119896119873 minusMOS

119896]

119870 minus 1 (6)


CI119895= 119905 times (1 minus

119886

2119873) times

120590119895

radic119873 (7)




120573 =1198983

11989832

2

120574 =1198984

1198982

2

(8)

1

2

3

4

5

Aver

age M

OS


Martial artMunich




119898119895=

sum119870

119896=1[sum119873

119894=1MOS119894119873 minusMOS

119896]2

119870

(9)







1

2

3

4

5

1 2 3 4 5

Aver

age M

OS


Martial artMunich




Martial artMunich


1

2

3

4

5

6

Aver

age M

OS













minus44

Within 174945 429 04078





minus09

Within 254635 429 05936













SI Mean 2021 2171 2228 2267 24Std dev 089 197 198 188 06667

TI Mean 1753 1416 1287 12 9Std dev 201 445 445 424 15

QP1

Mean 3915 3882 3955 3814 36Std dev 299 299 294 287 1

QP2

Mean 3315 3282 3355 3214 30Std dev 299 299 294 287 1



Load Mean 342 313 298 264 2Std dev 067 082 081 071 016


le4 gt4

le21 gt21

le2 gt2

gt1le1

le0 gt0

le3 gt3

Jitter

SI

Load

Channel2 (12020)


1 (16020) Load

2 (20) 1 (20)

1 (16020)le1 gt1

le3 gt3Load

Channel

3 (8010) 2 (4010)

le1 gt1

gt21le21

le1 gt1

le2 gt2

le36 gt36

le0 gt0

Jitter

SI

Channel 3 (36060)

Load 3 (12010)

QP_1 2 (16010)

3 (40) Channel

2 (20) 3 (20)

le3 gt3

le1 gt1

le21 gt21

le36 gt36

le0 gt0

Load

Channel

SI3 (16030)

QP_1 4 (40)

Resolution 3 (20)

4 (10) 2 (10)

le21 gt21

le0 gt0

le2 gt2

SI

Channel

Load 4 (16010)

4 (40) 3 (4010)

le36 gt36

le0 gt0

le0 gt0

QP_1

Resolution4 (12020)

3 (60) Channel

3 (20) 4 (4010)








119901 =TP119894

TP119894+ FP119894

119903 =TP119894

TP119894+ FN119894

119865 =2119901119903

119901 + 119903

(10)






2 17 0 0 0 119886 = 1

2 39 39 1 0 119887 = 2

0 11 134 32 0 119888 = 3

0 0 40 109 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 13 141 23 0 119888 = 3

0 0 21 128 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 14 137 26 0 119888 = 3

0 1 25 121 2 119889 = 4

0 0 0 5 1 119890 = 5


NaiveBayesian








Competing Interests


Acknowledgments


References



























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in







MOS119896=sum119870

119896=1sum119873

119894=1MOS119894119896

119870119872 (5)




the variance

1205902

119896=

sum119870

119896=1[sum119873

119894=1MOS119894119896119873 minusMOS

119896]

119870 minus 1 (6)


CI119895= 119905 times (1 minus

119886

2119873) times

120590119895

radic119873 (7)




120573 =1198983

11989832

2

120574 =1198984

1198982

2

(8)

1

2

3

4

5

Aver

age M

OS


Martial artMunich




119898119895=

sum119870

119896=1[sum119873

119894=1MOS119894119873 minusMOS

119896]2

119870

(9)







1

2

3

4

5

1 2 3 4 5

Aver

age M

OS


Martial artMunich




Martial artMunich


1

2

3

4

5

6

Aver

age M

OS













minus44

Within 174945 429 04078





minus09

Within 254635 429 05936













SI Mean 2021 2171 2228 2267 24Std dev 089 197 198 188 06667

TI Mean 1753 1416 1287 12 9Std dev 201 445 445 424 15

QP1

Mean 3915 3882 3955 3814 36Std dev 299 299 294 287 1

QP2

Mean 3315 3282 3355 3214 30Std dev 299 299 294 287 1



Load Mean 342 313 298 264 2Std dev 067 082 081 071 016


le4 gt4

le21 gt21

le2 gt2

gt1le1

le0 gt0

le3 gt3

Jitter

SI

Load

Channel2 (12020)


1 (16020) Load

2 (20) 1 (20)

1 (16020)le1 gt1

le3 gt3Load

Channel

3 (8010) 2 (4010)

le1 gt1

gt21le21

le1 gt1

le2 gt2

le36 gt36

le0 gt0

Jitter

SI

Channel 3 (36060)

Load 3 (12010)

QP_1 2 (16010)

3 (40) Channel

2 (20) 3 (20)

le3 gt3

le1 gt1

le21 gt21

le36 gt36

le0 gt0

Load

Channel

SI3 (16030)

QP_1 4 (40)

Resolution 3 (20)

4 (10) 2 (10)

le21 gt21

le0 gt0

le2 gt2

SI

Channel

Load 4 (16010)

4 (40) 3 (4010)

le36 gt36

le0 gt0

le0 gt0

QP_1

Resolution4 (12020)

3 (60) Channel

3 (20) 4 (4010)








119901 =TP119894

TP119894+ FP119894

119903 =TP119894

TP119894+ FN119894

119865 =2119901119903

119901 + 119903

(10)






2 17 0 0 0 119886 = 1

2 39 39 1 0 119887 = 2

0 11 134 32 0 119888 = 3

0 0 40 109 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 13 141 23 0 119888 = 3

0 0 21 128 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 14 137 26 0 119888 = 3

0 1 25 121 2 119889 = 4

0 0 0 5 1 119890 = 5


NaiveBayesian








Competing Interests


Acknowledgments


References



























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




1

2

3

4

5

1 2 3 4 5

Aver

age M

OS


Martial artMunich




Martial artMunich


1

2

3

4

5

6

Aver

age M

OS













minus44

Within 174945 429 04078





minus09

Within 254635 429 05936













SI Mean 2021 2171 2228 2267 24Std dev 089 197 198 188 06667

TI Mean 1753 1416 1287 12 9Std dev 201 445 445 424 15

QP1

Mean 3915 3882 3955 3814 36Std dev 299 299 294 287 1

QP2

Mean 3315 3282 3355 3214 30Std dev 299 299 294 287 1



Load Mean 342 313 298 264 2Std dev 067 082 081 071 016


le4 gt4

le21 gt21

le2 gt2

gt1le1

le0 gt0

le3 gt3

Jitter

SI

Load

Channel2 (12020)


1 (16020) Load

2 (20) 1 (20)

1 (16020)le1 gt1

le3 gt3Load

Channel

3 (8010) 2 (4010)

le1 gt1

gt21le21

le1 gt1

le2 gt2

le36 gt36

le0 gt0

Jitter

SI

Channel 3 (36060)

Load 3 (12010)

QP_1 2 (16010)

3 (40) Channel

2 (20) 3 (20)

le3 gt3

le1 gt1

le21 gt21

le36 gt36

le0 gt0

Load

Channel

SI3 (16030)

QP_1 4 (40)

Resolution 3 (20)

4 (10) 2 (10)

le21 gt21

le0 gt0

le2 gt2

SI

Channel

Load 4 (16010)

4 (40) 3 (4010)

le36 gt36

le0 gt0

le0 gt0

QP_1

Resolution4 (12020)

3 (60) Channel

3 (20) 4 (4010)








119901 =TP119894

TP119894+ FP119894

119903 =TP119894

TP119894+ FN119894

119865 =2119901119903

119901 + 119903

(10)






2 17 0 0 0 119886 = 1

2 39 39 1 0 119887 = 2

0 11 134 32 0 119888 = 3

0 0 40 109 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 13 141 23 0 119888 = 3

0 0 21 128 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 14 137 26 0 119888 = 3

0 1 25 121 2 119889 = 4

0 0 0 5 1 119890 = 5


NaiveBayesian








Competing Interests


Acknowledgments


References



























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in








minus44

Within 174945 429 04078





minus09

Within 254635 429 05936













SI Mean 2021 2171 2228 2267 24Std dev 089 197 198 188 06667

TI Mean 1753 1416 1287 12 9Std dev 201 445 445 424 15

QP1

Mean 3915 3882 3955 3814 36Std dev 299 299 294 287 1

QP2

Mean 3315 3282 3355 3214 30Std dev 299 299 294 287 1



Load Mean 342 313 298 264 2Std dev 067 082 081 071 016


le4 gt4

le21 gt21

le2 gt2

gt1le1

le0 gt0

le3 gt3

Jitter

SI

Load

Channel2 (12020)


1 (16020) Load

2 (20) 1 (20)

1 (16020)le1 gt1

le3 gt3Load

Channel

3 (8010) 2 (4010)

le1 gt1

gt21le21

le1 gt1

le2 gt2

le36 gt36

le0 gt0

Jitter

SI

Channel 3 (36060)

Load 3 (12010)

QP_1 2 (16010)

3 (40) Channel

2 (20) 3 (20)

le3 gt3

le1 gt1

le21 gt21

le36 gt36

le0 gt0

Load

Channel

SI3 (16030)

QP_1 4 (40)

Resolution 3 (20)

4 (10) 2 (10)

le21 gt21

le0 gt0

le2 gt2

SI

Channel

Load 4 (16010)

4 (40) 3 (4010)

le36 gt36

le0 gt0

le0 gt0

QP_1

Resolution4 (12020)

3 (60) Channel

3 (20) 4 (4010)








119901 =TP119894

TP119894+ FP119894

119903 =TP119894

TP119894+ FN119894

119865 =2119901119903

119901 + 119903

(10)






2 17 0 0 0 119886 = 1

2 39 39 1 0 119887 = 2

0 11 134 32 0 119888 = 3

0 0 40 109 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 13 141 23 0 119888 = 3

0 0 21 128 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 14 137 26 0 119888 = 3

0 1 25 121 2 119889 = 4

0 0 0 5 1 119890 = 5


NaiveBayesian








Competing Interests


Acknowledgments


References



























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






SI Mean 2021 2171 2228 2267 24Std dev 089 197 198 188 06667

TI Mean 1753 1416 1287 12 9Std dev 201 445 445 424 15

QP1

Mean 3915 3882 3955 3814 36Std dev 299 299 294 287 1

QP2

Mean 3315 3282 3355 3214 30Std dev 299 299 294 287 1



Load Mean 342 313 298 264 2Std dev 067 082 081 071 016


le4 gt4

le21 gt21

le2 gt2

gt1le1

le0 gt0

le3 gt3

Jitter

SI

Load

Channel2 (12020)


1 (16020) Load

2 (20) 1 (20)

1 (16020)le1 gt1

le3 gt3Load

Channel

3 (8010) 2 (4010)

le1 gt1

gt21le21

le1 gt1

le2 gt2

le36 gt36

le0 gt0

Jitter

SI

Channel 3 (36060)

Load 3 (12010)

QP_1 2 (16010)

3 (40) Channel

2 (20) 3 (20)

le3 gt3

le1 gt1

le21 gt21

le36 gt36

le0 gt0

Load

Channel

SI3 (16030)

QP_1 4 (40)

Resolution 3 (20)

4 (10) 2 (10)

le21 gt21

le0 gt0

le2 gt2

SI

Channel

Load 4 (16010)

4 (40) 3 (4010)

le36 gt36

le0 gt0

le0 gt0

QP_1

Resolution4 (12020)

3 (60) Channel

3 (20) 4 (4010)








119901 =TP119894

TP119894+ FP119894

119903 =TP119894

TP119894+ FN119894

119865 =2119901119903

119901 + 119903

(10)






2 17 0 0 0 119886 = 1

2 39 39 1 0 119887 = 2

0 11 134 32 0 119888 = 3

0 0 40 109 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 13 141 23 0 119888 = 3

0 0 21 128 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 14 137 26 0 119888 = 3

0 1 25 121 2 119889 = 4

0 0 0 5 1 119890 = 5


NaiveBayesian








Competing Interests


Acknowledgments


References



























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






2 17 0 0 0 119886 = 1

2 39 39 1 0 119887 = 2

0 11 134 32 0 119888 = 3

0 0 40 109 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 13 141 23 0 119888 = 3

0 0 21 128 0 119889 = 4

0 0 0 6 0 119890 = 5


6 63 12 0 0 119887 = 2

0 14 137 26 0 119888 = 3

0 1 25 121 2 119889 = 4

0 0 0 5 1 119890 = 5


NaiveBayesian








Competing Interests


Acknowledgments


References



























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in



Research Article A Framework for QoE-Aware 3D Video ...

Documents

Transcript of Research Article A Framework for QoE-Aware 3D Video ...