Analysis of Power Consumption of H.264/AVC-based Video Sensor Networks through Modeling the Encoding Complexity and Bitrate

Transcription

1 Analysis o Power Consumption o H.264/AVC-based Video Sensor Networks through Modeling the Encoding Complexity and Bitrate Bambang A.B. Sari, Panos Nasiopoulos and Victor C.M. eung Department o Electrical and Computer Engineering, University o British Columbia Vancouver, Canada {bambangs, panosn, vleung}@ece.ubc.ca Mahsa T. Pourazad Department o Electrical and Computer Engineering, University o British Columbia TEUS Communications Inc. Vancouver, Canada pourazad@ece.ubc.ca Abstract The H.264/AVC video encoding standard has many advanced eatures that can be tailored to suit a wide range o applications. In order to obtain optimal coding perormance in video sensor networks (VSNs, it is essential to ind the right setting parameters or the encoder. There is a trade-o between required energy or encoding and transmission o video content in VSNs that can be exploited to minimize total power consumption. In this study, we model the complexity and bitrate in H.264/AVC codec. By using the model, the trade-o between encoding and transmission energy consumption is urther exploited. Our experiments show that the complexity modeling error is less than or equal to 3.45%. However, the bitrate modeling error that we obtain is less than or equal to 11.6%. Keywords-H.264/AVC; complexity and bitrate modeling; energy consumption model; and video sensor network I. INTRODUCTION With the increasing concern about security in homes or public spaces, the demands or monitoring and surveillance systems is growing. In this regard, video sensor networks (VSNs oer an alternative to several existing monitoring technologies [1], [2]. However, unlike the traditional sensor networks which require negligible power to process captured data in the sensor nodes, VSNs need signiicant processing power to encode and transmit the captured videos. With the limitations o energy resources in VSNs, maximizing the power eiciency o coding and transmission operations becomes very important. In general, there is a tradeo between encoding complexity and compression perormance in the sense that to obtain higher compression perormance (i.e., lower bit rate, more complex and computationally expensive encoding scheme is required. On the other hand, transmission o lower bitrate content requires less amount o energy. Fig. 1 illustrates the relationship between coding complexity, compression perormance and the required power or encoding and transmission o the content. It can be observed that, to minimize the overall VSN power consumption, encoding process needs to be handled careully. Among the existing video coding standards, H.264/AVC is the most widely used standard in the consumer market [3], [4]. Some o the existing studies on the perormance o H.264/AVC codec look into maximizing the coding perormance without considering the total power Figure 1. Relation between encoding and transmission power consumption consumption o the coding process [3], [5]. J.J. Ahmad et al. [6] studied the required energy or encoding and transmission o video content in the case o using H.264/AVC codec. Unortunately, the number o coniguration settings considered or the encoder in that study is limited. To address this issue, we extended the study in [6] by including more encoder coniguration settings in our previous work [7]. We proposed a guideline table or encoder coniguration setting which include dierent combinations o coding complexity and coding eiciency in terms o bitrate that produces compressed videos with similar quality in terms o peak signal to noise ratio (PSNR. Our study shows that the energy consumption o a VSN can be reduced by careully selecting the encoder settings at each VSN node based on the proposed table. This paper is an extension to our previous work [7] where the relationship between coding complexity and coding eiciency (in terms o bitrate o H.264/AVC codec is modeled. By using this model, the trade-o between encoder complexity and bitrate can be urther elaborated, unrestrained with the encoder setting parameters. The rest o the paper is organized as ollows. Section II describes the H.264/AVC encoding complexity and bitrate modeling. The encoding and transmission power consumption model is then discussed in Section III. Conclusions are drawn in Section IV.

2 II. H.264/AVC COMPEXITY AND BITRATE MODEING H.264/AVC is a block-based hybrid video coding standard utilizing intra-rame and inter-rame prediction. While inter-rame prediction is more involved than intrarame prediction, it results in lower bitrate. By increasing the number o inter-rames coded picture within a successive video stream, i.e., group o picture (GOP size, the bitrate o the coded video is reduced at the cost o higher encoding complexity. In the case o inter-rame prediction, the complexity and bitrate can be controlled by adjusting the search range (SR in motion estimation process. The SR determines the size o searching area in the reerence rame to ind the best match to be used or inter prediction. Increasing the SR size may result in better compression perormance at the cost o increased complexity. However this observation is quite content dependant and there are cases where increasing the value o SR does not provide signiicant beneit in terms o compression perormance [7]. The other actor that controls the complexity and the perormance o the H.264/AVC codec is the number o block sizes used in the inter prediction process. Increasing the number o used block sizes results in better prediction and consequently higher compression perormance at the expense o increased complexity. The complexity o motion estimation (ME can be classiied into dierent level o complexity, depending on the number o block size candidates used. In general, there are seven block sizes deined or inter-prediction in H.264/AVC. In this paper, we analyze the eect o dierent coding parameters on the coding complexity using a set o training videos and propose a model or the relationship between coding coniguration and coding complexity, and later this model is tested on a set o unseen test video set. The ollowing subsections provide more details on our experiment settings and the proposed model. A. Experiment Settings In VSN applications, due to the limitations in energy and processing resources, less complex encoder conigurations are used. To this end, we used baseline proile o H.264/AVC that is suitable or low complexity applications and uses only I and P rames (no B-rames in our study. The other encoding parameters in our experiments include using context-adaptive variable-length coding (CAVC entropy coding and one reerence rame, setting SR equal to 8, and disabling the rate distortion optimization (RDO, rate control, deblocking ilter and Intra coding or P rames options. Furthermore, to have an objective measure or the encoding complexity, we use the instruction level proiler ipro [8], which provides us with the total number o instruction counts. The H.264/AVC reerence sotware, JM version 18.2 is used in our experiments. Five representative videos rom [9] are used in our study (BQMall, Traic, TABE I ME COMPEXITY EVE (M Fig. 2. Normalized C P or dierent M or BQMall video Race Horse, PeopleOnStreet and Vidyo1. To mimic a common VSN data, these sequences are downsampled to the common intermediate ormat (CIF resolution (352x288 pixels and also their rame rate was reduced to 15 rames per second (ps. The BQMall and Traic video sequences are used as the training set or the model and the rest o videos as the test set. B. Complexity Modelling The coding process complexity o a video sequence (C S is ormulated as ollows: C S = C I n I + C P n P (1 where C I is the complexity to encode an I-rame, C P is the complexity to encode a P-rame, n I is the number o I- rames in the sequence and n P is the number o P-rames in the sequence. For a video sequence with no scene change, the value o C I can be considered almost constant. On the other hand, C P depends on the complexity level o the ME process. In our study, the complexity level o ME process (called M is classiied based on the used block-size candidates in the encoding process as shown in Table I. As illustrated in Fig. 2, the GOP size does not aect the normalized coding complexity o P rames at each M. Note that the complexity o coding P-rame (C P is normalized with respect to C p when M is equal to one. Furthermore, as it can be seen rom Fig. 3, the plot o normalized C P or dierent training videos has the same slope but scaled by a constant. It can be seen rom this igure that the normalized C P or the Traic video ranges rom 1 to 1.485, which also

3 TABE II ME COMPEXITY EVE (M AND δ CP M δ CP Fig. 5. Bitrate o a P-rame or dierent M o BQMall video ω 1 = (2 Cp M =1 Using ω 1, the complexity to encode a P-rame is ormulated as: Cp ( 1+ δ ( ω M = i = Cp M = 1 CP i 1 (3 Fig. 3. Normalized C P or GOP=2 o the training videos Considering that n I =N/GOP, where N is total number o rames and n P =N N/GOP, then the average complexity per rame is computed as ollows: C = ( CI + CpM (1 CP ( GOP 1 / GOP = 1 + δ ω1 (4 C. Bitrate Modelling The bitrate o the encoded video is modeled as R=R F r, where R is the average bitrate o a rame and F r is the rame rate. The total size o the encoded sequence (in bit is then modeled as: Fig. 4. Fractional increase o normalized C P or the training videos R = R n + R n S I I P P (5 means that the normalized C P range or this video is On the other hand, the normalized C P range or the BQMall video is equal to Scaling the range o the normalized C P to one, we can plot the ractional increase o normalized C P as shown in Fig. 4. It is interesting to see that the increase o normalized C P with respect to M is almost similar or both videos. We deine δ CP as the amount o increase normalized C P at dierent M. δ CP is calculated by averaging the values obtained in Fig. 4, as shown in Table II. Another interesting observation is that, the value o range o normalized C P shown in Fig. 2 is proportional to the value o. Thereore, using the values obtained rom Cp M =1 the training videos, the range o normalized C P values or a speciic video sequence is calculated as: where R I is the average size o an I-rame and R P is the average size o a P-rame. The value o R P depends on the M and GOP used by the encoder. Fig. 5 shows that, the value o R P decreases as M increases. Thereore, or a certain GOP value, the R P is modeled as: R PGOP = ω = i Ri ( M where ω Ri is the bitrate o a P-rame when GOP=i and M =1, and (M is a decay unction with respect to M, which is modeled using the generalized logistic unction. The logistic unction is a widely used sigmoid unction or growth/decay modeling where the growth/decay is exponential at irst, but eventually slower and then levels o. This matches the way R P is reduced with the increase o (6

4 TABE III COMPEXITY MODEING ERROR TABE IV BITRATE MODEING ERROR Fig. 6. The normalized bitrate ( BQMall, GOP=2 and the logistic unction M (see Fig. 5. The logistic unction (M used in our study is as ollows: M b a = a (7 1+ e ( + c( x d where a and b indicate the minimum and maximum asymptote o the plot respectively, c is the growth rate, while d signiy the time or maximum growth (see Fig. 6. Furthermore, Fig. 5 also shows that the slope o the R P plot or dierent GOP sizes is the same. Thereore, R P is modeled equal to: where ω Rp is the bitrate o P-rame when GOP=2 and M =1, and ω 2 is the weight or (GOP. To obtain the parameters or the (M, we applied least mean square approach using the normalized R P o training video sequences when GOP=2. Also to estimate (GOP, we applied curve itting approach on the Rp values o training video sequences at dierent GOP size settings, and ound that ω 2 ln(gop provides a good estimate or (GOP. The value o ω 2 is estimated using least square regression rom the training sequences. Assuming that the average bitrate o an I-rame is equal to R I the average bitrate o a rame (R is estimated as: R R P = ω ( M + ω2 ( GOP (8 R I 0.08 ( GOP 1 + ωrp CP GOP ( (1 e δ. (9 + GOP ( GOP 1 + ω2 ln( GOP GOP = D. Implementation o the Proposed Model Rp To implement the proposed model, we need to obtain several variables rom each video sequence. To this end, we encode the irst two rames o each video sequence. Assuming that there is no scene change in the video sequence, the bitrate o each I-rame will be almost similar. Thereore, R I is assumed to be equal to the bitrate o the encoded irst rame while ω Rp is equal to the bitrate o the second rame. For the complexity modeling, the ipro tool will provide us with the complexity o encoding the irst two rames o the video sequence, i.e., C 2-rames =C I +. Since we already have the value o R I rom encoding the irst two rames o each test sequence, we can estimate the value o C I o these sequences. The value o can then be calculated using C 2-rames C I. Consequently, the value o ω 1 is calculated using (2. To estimate the modeling error, the average percentage o complexity and bitrate error or GOP={1, 2, 4, 8, 16, 32, 64} and M ={1, 2, 3, 4, 5, 6, 7} is calculated. As Table III shows, the average error or complexity modeling is less than or equal to 3.45% or the test video sequences, while the average error o bitrate modeling is less than or equal to 11.6% as reported in Table IV. III. ENCODING AND TRANSMISSION POWER CONSUMPTION MODE The total power dissipation at a sensor node consists o the power consumption or encoding (P e, transmission (P t and reception (P r. P e can be calculated as ollows: P = C F CPI E (10 e where CPI is the number o CPU cycles to perorm one basic instruction and E c is the energy depletion per cycle. The transmission power consumption is calculated as: P t r η ( + β d c Cp M =1 Cp M =1 = α R (11 where α is a constant coeicient related to coding and modulation, β is the ampliier energy coeicient, d is the transmission distance, η is path loss exponent and R is the bitrate. The reception power consumption is calculated as: P r = λ R (12

5 TABE V. PARAMETERS USED. Parameters Description value Energy cost or transmitting 1 α 0.5 J/Mb bit β Transmit ampliier coeicient J/Mb/m4 λ Energy cost or receiving 1 bit 0.5 J/Mb η Path loss exponent 4 XScale average cycle per CPI 1.78 instruction [10] Energy depleted per cycle or E c nj imote2 [6] where λ is a constant coeicient representing energy cost or receiving 1 bit. Table V shows the parameters used or our experiments. In this paper, we analyze a simple topology consisting o one video node and the sink. The total power consumption o a video node or dierent transmission distances or PeopleOnStreet video sequence is shown in Fig. 7. In this igure, we analyze two scenarios: a the GOP size is ixed while the M varies, and b the M is ixed while the GOP size changes. In Fig. 7a, the GOP size is set equal to eight and M changes. It is observed that or transmission distance less than 200m, the use o bigger M results in higher total power consumption. This result shows that varying M values do not signiicantly aect the trade-o between computation and communication. This trend is also seen in other test video sequences. Fig. 7b shows the plot o total power consumption when M is equal to our and the GOP size changes. The igure shows that when the transmission distance is small, the coniguration that leads to low power consumption is the one using smaller GOP. It means that the low encoding power consumption (due to the use o smaller GOP is compensating the higher transmission power consumption (due to higher bitrate. However, when the transmission distance is large, the energy cost to transmit the data increased signiicantly. Thereore, we need to use the coniguration with better compression perormance, i.e., larger GOP size, to reduce the transmission energy consumption. The trade-o between computation and communication can be clearly seen when the transmission distance is less than 100m as shown in Fig. 8. However, it can be seen that the transmission distance at which the use o bigger GOP minimizes power consumption is content dependent. For example, in the case o PeopleOnStreet video sequence, using GOP equal to one will minimize the total power consumption when the transmission distance is less than 63m (see Fig. 8a. However, or the RaceHorses video sequence, the use o GOP equal to one will minimize total power consumption when the transmission distance is less than 88m (see Fig. 8b. (a (b Fig. 7. Total power consumption or dierent transmission distance: (a GOP=8 and varying M (b M =4 and varying GOP sizes (a (b Fig. 8. Total power consumption or transmission distance less than 100m (M =4 and varying GOP sizes: (a PeopleOnStreet sequence(b RaceHorses sequence

6 IV. CONCUSION In this paper, we propose the encoding complexity and bitrate model o H.264-based video sensor networks. The experimental results show that the proposed complexity model provides a very small prediction error (less than or equal to 3.45%, while the bitrate modeling error is rom 8.57% to 11.6% or the video sequences tested. The proposed model is used to show the trade-o between encoding and communication that can be exploited to minimize the total power consumption o VSNs. ACKNOWEDGMENT This work was supported by the NPRP grant # NPRP rom the Qatar National Research Fund (a member o the Qatar Foundation. The statements made herein are solely the responsibility o the authors. REFERENCES [1] I. F. Akyildiz, T. Melodia, and K. R. Chowdhury, A survey on wireless multimedia sensor networks, Computer Networks: The International Journal o Computer and Telecommunications Networking, vol. 51, no. 4, pp , Mar [2] T. D. R aty, Survey on Contemporary Remote Surveillance Systems or Public Saety, Systems, Man, and Cybernetics, Part C: Applications and Reviews, IEEE Transactions on, vol. 40, no. 5, pp , Sep [3] I. E. Richardson, The H.264 Advanced Video Compression Standard, Second Edition. John Wiley & Sons, td, [4] T. Wiegand, G. J. Sullivan, G. Bjontegaard, and A. uthra, Overview o the H.264/AVC video coding standard, IEEE Transactions on Circuits and Systems or Video Technology, vol. 13, no. 7, pp , Jul [5] H. K. Zrida, A. C. Ammari, M. Abid, and A. Jemai, Complexity/Perormance Analysis o a H.264/AVC Video Encoder, in Recent Advances on Video Coding, InTech. [6] J. J. Ahmad, H. A. Khan, and S. A. Khayam, Energy eicient video compression or wireless sensor networks, presented at the Inormation Sciences and Systems, CISS rd Annual Conerence on, Baltimore, MD, 2009, pp [7] B. A. B. Sari, M. T. Pourazad, P. Nasiopoulos, and V. C. M. eung, Encoding and communication energy consumption trade-o in H.264/AVC based video sensor network, presented at the Accepted or publication IEEE World o Wireless, Mobile and Multimedia Networks, WoWMoM 13, 2013, pp [8] P. M. Kuhn, A Complexity Analysis Tool: ipro (version 0.41, ISO/IEC JTC1/SC29/WG11/M3551, Jul [9] ISO/IEC JTC1/SC29/WG11, Joint Call or Proposals on Video Com-pression Technology. Jan [10] D. Chinnery and K. Keutzer, Closing the Power Gap between ASIC & Custom: Tools and Techniques or ow Power Design, 1st edition. Springer, 2007.