Video Traffic Model for MPEG4 Encoded Video

Transcription

1 Video Traffic odel for EG Encoded Video. H. Liew. Kodikara.. Kondoz entre for ommunication Systems Research University of Surrey Guildford Surrey GU 7XH UK bstract To date video traffic models in the literature have mostly considered the autocorrelation modeling of empirical video traffic for fixed source quantization parameters. Existing models also ignore the inter-dependencies between - - and - frame types of EG coding which have great impact on empirical queuing performance prediction accuracy. We propose a new Video Traffic odel (VT) that is capable of generating output video traffic for wide range of quantization parameters in real time while at the same time capturing the inter-dependencies between different frame types. The VT performance is evaluated by means of packet loss rate prediction accuracy. Some existing models are implemented for performance benchmark. Simulation results show that the VT captures empirical video traffic characteristic accurately and outperforms the existing models. Simulation results also show that the models that ignore the inter-dependencies between frame types can greatly underestimate the empirical packet loss rate. Keywords-component; video traffic model mpeg wireless video communication. NTRODUTON ultimedia services are seen to be the revenue generating service for current and future networks. This can be observed from operator efforts to support services like video on demand and digital video broadcast (DV-H). n order to study and evaluate the performance of video applications over wired and wireless communication systems an accurate video traffic model is required. n this paper we propose a Video Traffic odel (VT) that is capable of synthetisizing EG video traffic in real time for a range of quantization parameters. This will enable use of VT for study of bandwidth adaptive video transmissions such as wireless video streaming where video source rate is periodically and adaptively matched to the time-varying channel bandwidth by varying the quantization parameters in real time. n contrast existing models e.g. [- ] have mostly modeled the video traffic characteristic for a fixed set of quatisation parameters (i.e. VR video) and do not provide the adaptive capability of the proposed VT. Secondly the existing video traffic models have ignored inter-dependencies between - - and - frame types of EG coding. gnoring the inter-dependencies between - - and - frame types leads to underestimation of empirical traffic burstiness and consequently the empirical packet loss rate prediction accuracy. n contrast the VT models the the inter-dependencies by using a ultinomial ethod () []. The inter-dependencies modeling using greatly improves the model accuracy in predicting the nter- Dependencies odeling utocorrelation odeling Frame Size odel Frame Size odel Frame Size odel UX Figure. The proposed Video Traffic odel (VT) empirical packet loss rate as will be shown later in this paper. Simulation has been performed to evaluate the proposed VT as well as comparing it against the existing models. Three existing models GO-method [] Nested-R [] FR [] have been implemented for performance comparisons. Nested-R FR and GO-method represent different cases of autocorrelation and interdependencies modeling. Simulation results show that the VT captures the video traffic burstiness accurately and is superior to the existing models. The proposed GVT has three major parts as shown in Fig.. The first part consists of - - and - frame size models. The second part models the inter-dependencies between different frame types using a ultinomial ethod (). The final part models the autocorrelation structure of video traffic. The rest of the paper is organized as follows. n Section we describe the frame size model. n Section and V we discuss the inter-dependencies and autocorrelation modeling techniques. The summary for video traffic generation is presented in Section V. Finally simulation results and discussions are presented in Section V FRE SZE ODEL n this section we propose a generalized frame size model that can synthesize - - and - frame sizes for a quantization parameter range Q [ ]. First we introduce a frame activity concept. Secondly we study the - - and - frame composition and its relative importance to the output frame size. We then establish a relationship between the frame activity and the output frame size for a given quantization parameter. Finally the overall frame size //$. EEE 8

2 model is presented.. Frame ctivity Frame activity can be used to measure frame visual complexity. n general an un-quantized and un-compressed picture frame with large frame activity would result in a large output frame size. n this paper we have adopted a DT method [] for frame activity calculation. Here we denote and as the frame activities for - - and - frames calculated before the quantization process in an EG encoder [].. Frame Size omposition The basic - - and - frame size compositions are shown in Fig.. The header bits contain information for frame decompression and are found to be almost constant from the empirical trace. The texture bits contain the compressed pixel information as a result of video coding. We jointly consider the header bits and texture bits as Texture Size T χ where χ { }. The motion vector bits contain information for temporal reference and are only present in - and - frames. The motion vector bits are denoted using χ. n general T χ is found to be dependent on the quantization parameter of an EG encoder while χ does not dependent on the quantization parameter. Header Tx T x (a) Texture Header Texture V (b) Figure. (a) frame size composition. (b) and frame size composition x. Relationship of Frame ctivity to Texture Size We have studied the relationship between the frame activity χ and the Texture size T χ for the quantization range Q χ [ ] where χ { }. Fig. portrays the plots of against T. For the sake of brevity we only plot curves for the quantization parameter and. We found that this family of curves can be closely fitted with a quadratic function. The corresponding quadratic curves are plotted in the same graph. Thus for a given quantization parameter Q χ the Texture size T χ is related to the frame activity χ using the following equation: T = + + () ( χ) ( χ) ( χ) χ χ χ where χ χ and ( χ ) are the least square fitted coefficients for a given Q χ. For frame the least square fitting is performed on the empirical trace for the quantization parameter Q χ [ ] and the calculated χ χ and χ are stored in an - frame table T. Same procedures are carried out for calculating the and frame tables T and D. arginal Distribution odeling T. The umulative Distribution Functions (DFs) of the frame activity and the motion vector size are respectively found to be closely fitted to the Gamma distribution. The plots of Gamma fits are shown in Fig.. E. The omplete Frame Size odel n this section we present the complete frame size model based on the previous discussions. The frame size S χ can be represented mathematically using: x x 7 x Q = Q = Q = Q = Q = Q = Q = Q = Q = Frame ctivity 7 Frame ctivity Frame ctivity (a) frame (b) frame (c) frame Figure. Graphs shows the plot of frame activity against texture size and its corresponding quadratic fitting curves for different quantiser. 8

3 .9.9 umulative Density Function frame activity frame activity frame activity Gamma fit umulative Density Function otion Vector Size otion Vector Size Gamma fit Frame ctivity () otion Vector Size (bit) (a) (b) Figure. (a) frame activity DF fitting. (b) - and - frame motion vector size DF fitting Generator to / motion vector size marginal distribution to frame activity marginal frame activity to frame size Generator distribution frame size to / frame activity marginal distribution Generator + / frame activity to / frame size / frame size Sχ = Tχ + χ χ () where χ { } and χ is an indicator function: χ { } χ = () χ { } Fig. shows the - - and - frame size models graphically. We use the - frame size generation as an example to explain the frame generation process. First a random variable X is generated. Then X is mapped to using a probability integral transform [7]: FΓ FG X = ( ) () where F Γ and F G are the inverse DF and the DF of and variable respectively. Given the quantization parameter Q is then mapped to T using () and ( ) ( ) ( ) are obtained from the - frame table T (explained in Section -). Finally the - frame size S can be calculated using (). - and - frame size S and (a) frame size model Quantiser Selection S can be calculated in a similar way. (b) and frame size model Figure. Frame size model for - - and - frame. ovariance atrix TLE. OVRNE TRX Σ Quantiser Selection NTER-DEENDENES ODELNG The - - and - frame sizes of EG encoded video are highly inter-dependent as shown in Fig.. t can be observed that the - - and - frame sizes follow a same trend. This represents a traffic burtiness to the network. This burstiness can be attributed to the inter-dependencies of underlying and. The interdependencies can also be observed from the normalized covariance matrix Σ shown in Table. t can be observed that the normalized covariance value can be as high as.9. n light of this observation we have used a ultinomial ethod () [] for modeling the inter-dependencies between and. We used to generate correlated variables 8

4 (a) frame (b) frame (c) frame Figure. The plots show the - - and - frame inter-frame dependency (i.e. following the same trend) X = ( X X X X X) which will respectively be mapped to and such that they are also correlated. The is described as follows: Let X = ( X X X X X) be a correlated vector calculated using X = LZ () where L is a weighting matrix and Z = ( Z Z Z Z Z) are zero mean and unity variance variables. t can be proved that X has zero mean T and LL covariance []. n order to generate X with the same cross correlation to the empirical trace we equate the T covariance matrix Σ from Table to LL and solve for matrix L using holesky Decomposition [8]. The L matrix is then used in () to generate the correlated vector X. V. UTOORRELTON ODELNG The autocorrelation structures of and frames are due to the autocorrelation structure of underlying and. We have used a Spatial Renewal rocess (SR) [9] to model the utoorrelation Functions (Fs) of and. We refer the reader to [9] for detail discussions of SR. The steps utilizing SR for F modeling are as follows:. alculate the background frame activity process for all the uncompressed raw video frames. Note that this calculation is performed on unprocessed video frames i.e. without motion compensation in the EG encoder chain [].. alculate the F of background frame activity process. Use SR to model the F.. Generate the sequence Z using SR. Z - Z are randomly generated as zero mean and unity variables.. Z = ( Z... Z) is mapped to X = ( X... X) using (). Note that in this case X derives its F directly from Z whereas X - X derive their Fs from Z by means of inter-dependencies modeling in ().. Finally X X X X and X are mapped to using (). The Fs modeling of are thus achieved. V and V. SURY OF VT TRFF GENERTON The traffic generation procedures are as follows: V. Decide current frame type χ in the GO. Decide the quantization parameter Q χ based on certain criteria e.g. output frame bit rate matching the channel bandwidth.. Given the frame type χ and Q χ initialize the coefficients χ χ and χ from the χ tables T χ as discussed in Section... Generate Z using SR. SR the models autocorrelation structure of background frame activity process (See Section V). Generate Z - Z randomly as zero mean unity variance variables.. ap ( Z... Z ) to ( X... X ) using (). () models the cross correlation structure in ( X... X ). Go to Step for frame Step for frame and Step 7 for frame. 87

5 log (acket Loss Ratio) Nested R FR GVT GO uffer Size (s) Figure 7. lot of empirical packet loss rate prediction of VT and existing models.. ap variable X to frame activity using (). Then map frame activity to frame size S using () () and coefficients ( ) ( ) ( ). See Fig. (a).. ap variables X and X to frame activity and motion vector size using (). Then map frame activity to texture T using () and coefficients. Sum texture T and motion vector size to obtain frame size S. See Fig. (b) 7. ap variables X and X to frame activity and motion vector size using (). Then map frame activity to texture size T using () and coefficients. Sum texture sizet and motion vector size to obtain frame size S. See Fig. (b). 8. Repeat -7 for the total number of required frames. V. SULTON RESULTS ND DSUSSON We validate the proposed VT by means of empirical packet loss rate prediction accuracy. We have implemented Nested-R [] FR [] GO-method [] for performance comparisons. Nested-R FR and GOmethod represent three different cases. Nester-R models the F of frame while ignoring the Fs of and frames. Nested-R also ignores the inter-dependencies between different frame types. FR models the Fs of - - and - frames but ignores the inter-dependencies between different frame types. GO-method models the Fs of and frames and also the inter-dependencies between frame types. Several video sequences have been tested. However we only include the result from film sequence Lord of the Rings: The Two Towers due to space limitation. We have set the quantization parameters of VT for the - - and - frames to fixed values of and so that it can be easily compared to the existing VR video traffic models. VT Nested-R FR and GO are respectively used to generate synthetic traffic to a FFO queue at different bandwidth utilizations. The packet loss rate for all the models are recorded and compared to the empirical packet loss rate. We present the results for the bandwidth utilization of % in Fig. 7. VT is shown to predict empirical packet loss rate accurately and is superior over the Nested-R and FR. The GO-based method has similar performance to VT but it lacks the flexibility of VT being able to reproduce video traffic for different quantization parameters in real time. Fig. 7 also shows that the model that ignores inter-dependencies between frame types e.g. Nested-R and FR can greatly underestimate the empirical packet loss rate. Similar results are found for the bandwidth utilizations % and 8%. V. ONLUSON n this paper a new Video Traffic odel (VT) that can generate video traffic for quantization parameter range [] is proposed. This is important for study of adaptive video transmission where rate adaptation is achieved by varying quantization parameters at the encoder. The VT also considers inter-frame dependencies between - - and - frames. Simulation results show that the VT accurately predicts empirical traffic characteristics and outperform the existing video traffic models that ignore the inter-frame dependencies. KNOWLEDGENT The authors would like to acknowledge the funding from EU F ST-NEWO project. REFERENES [] O. Rose "Statistical ropeties of EG Video Traffic and Their mpact on Traffic odelling in T Systems" Report No. Feb 99. [] D. Liu E.. Sara and W. Sun 哲 ested auto-regressive processes for mpeg-encoded video traffic modeling_ EEE Trans. ircuits Syst. vol. pp. 9 8 Febuary. [] N. nsari H. Liu Q. Shi and H. Zhao On modeling mpeg video traffics EEE Trans. roadcast. vol. 8 pp. 7 7 December. [] E.. Scheuer and Stoller On the generation of normal random vectors Technometrics vol. pp ay 9. [] W. J. Kim J. W. Yi and S. D. Kim bit allocation method based on picture activity for still image coding EEE Trans. mage rocessing vol. 8 pp July 999. []. Ghanbari Video oding: n ntroduction to Standard odecs st ed. EE 999. [7].. ood and. G. Graybill ntroduction to The Theory of Statistics rd ed. cgraw-hill 97. [8] R. K. ock W. Krischer The Data nalysis riefbook Version pril 998. [9] T. Taralp. Devetsikiotis and. Lambadaris Efficient fractional gaussian noise generation using the spatial renewal process in roc. of the EEE nternational onference on ommunications tlanta US June 998 pp.. 88