UNEQUAL ERROR PROTECTION RATELESS CODING FOR EFFICIENT MPEG VIDEO TRANSMISSION Ali Talari and Nazanin Rahnavard School of Electrical and Computer Engineering Oklahoma State University Stillwater, OK 74078 Emails: {ali.talari, nazanin.rahnavard}@okstate.edu

I. ABSTRACT This paper proposes a rateless coding for efficient MPEG video transmission over loss prone networks. An MPEG video stream is comprised of various frame types, which contribute differently to the displayed video quality. We employ rateless codes with unequal error protection (UEP) property to have more protection on frames with higher influence on the quality of the displayed video. We define protection levels for different frame types and find their optimum values. Next, we evaluate our scheme using simulations and show that MPEG video coding with UEP rateless codes has a higher performance compared to video coding with EEP rateless codes, while new properties of low coding/decoding complexity, ratelessness, and adaptability to varying loss rates are kept intact. II. INTRODUCTION The growing demand for video streaming over error prone wireless channels requires designing new efficient video coding schemes. In this paper, we propose an efficient and simple forward-error-correction (FEC) coding scheme for transmission of MPEG videos over wireless channels. MPEG video encoders generate three types of frames I , P , and B , from the source raw video stream. The encoded frames are then grouped into batches of frames, called group of pictures (GOP), which contain one I -frame and several P - and B -frames. These frames are not equally important, since decoding of P - and B -frames depends on the availability of the preceding I - and P -frames. Therefore, I -frames are more important than P -frames, and P -frames are more important than B -frames. Existing work [1], [2] propose to use fixed-rate codes to provide unequal error protection (UEP) for MPEG video transmission. However, the proposed schemes have high coding/decoding complexity and their implementation might not be feasible in practice. In contrast, rateless codes have low coding/decoding complexity and can adapt to any channel condition with varying loss rate since they do not have fixed coding rates. For instance, compensating the burst errors of wireless channels with fixed-rate codes

requires a large overhead for all blocks of coded packets, which would result in an inefficient data transmission scheme. In contrast, rateless codes do not impose a large fixed overhead for all blocks of coded packets, and the coded source packets can be decoded as long as a sufficient number of encoded packets is delivered to receiver. Authors in [1] propose a comprehensive unequal error protection coding for a layered MPEG video. Besides considering different protection levels of frames, they have also considered the layers’ importance as well. They employ fixed-rate Reed-Solomon codes per layer per frame, and optimize the amount of redundancy assigned to each frame using genetic algorithms. Although the proposed UEP scheme has a comprehensive approach, it might not be feasible to implement in practice since encoder and decoder have to deal with multiple complex Reed-Solomon coding and decoding on the fly. Computational power available on wireless devices is usually expensive and limited. Paper [2] proposes to use maximum distance separable (MDS) codes to provide unequal error protection for I -, P -, and B -frames. MDS codes are complex, and they have fixed rates. The proposed scheme has both drawbacks of using fixed-rate and complex codes, which might make this scheme hard to implement. Authors in [3] propose an MPEG video transmission scheme which provides more protection for I -frames only by transmitting multiple copies of I -frames instead of using UEP-FEC codes. This video transmission scheme is obviously suboptimal since a large amount of redundant packets are transmitted from I -frames. On the other hand, importance of P -frames compared to B -frames is not considered. In this paper, we design a video coding scheme which inherits the low coding/decoding complexity of rateless codes while it maintains or improves the high performance of UEP video codings. We employ the recently introduced FEC codes, called UEP-rateless codes [4], [5], to implement our efficient and flexible proposed protocol. In this paper, we assume that the receiver can eliminate probable video delay and jitter by buffering a few GOPs before the video playout.

1 of 7

This paper mainly concentrates on designing a new efficient coding for MPEG video transmission over lossy networks rather than video playback issues. The paper is organized as follows. Section III provides an insight into MPEG video format and UEP-rateless codes. In Section IV, we formulate the decodable frame rate for the conventional case where EEP-rateless codes are employed to transmit MPEG video frames. Our proposed scheme is introduced in Section V. Section VI reports the performance of the proposed scheme. Finally, Section VII concludes the paper. III. BACKGROUND A. MPEG VIDEO FORMAT An MPEG video stream includes several GOPs, which contain three frame types, I , P , and B [1]. The structure of a GOP is often referred to by two numbers, M and N . M refers to the distance between P -frames, which is also the distance between an I -frame and the first P -frame, and N defines the distance between two I -frames, which is also the length of the GOP. For instance, the structure of a GOP defined by M = 3, N = 12 would be IBBPBBPBBPBB. When decoding an MPEG video, first the I -frame is decoded independently, and next, P -frames are recovered using the previous I -frame and P -frames. Finally, the B frames are constructed using preceding and succeeding I or P -frames. The last (M − 1) B -frames in each GOP are decoded using the previous I -frame and P -frames and also the next GOP’s I -frame. The structure of a GOP and frames dependencies are depicted in Figure 1.

B

B

I

B

GOP n

B

P

B

B

P

capacity performance. They can potentially encode n source data packets into unlimited number of outgoing (encoded) packets. Luby proposed LT codes which are the first practical implementation of rateless codes [7]. In LT encoding, first a packet degree, d, is chosen from a degree distribution {Ω1 , Ω2 , . . . , Ωm }, where Ωi is the probability that d = i, and m ≤ n is the largest packet degree. We can also denote this probability distribution by its generator polynomial P i Ω(x) = m i=1 Ωi x . Next, d packets are uniformly chosen at random among the source packets, and are XORed together to form an encoded packet. This procedure is repeated until γn packets are collected at the receiver. Any γn packets are sufficient for a successful decoding where γ ≥ 1 is called coding overhead. The decoding of rateless codes is performed in an iterative fashion. First, the decoder finds received coded packets of degree-one and recovers one source packet from each one of these encoded packets. Next, by removing the recovered source packets from higher degree encoded packets more degree-one coded packets emerge. This iterative decoding continues until either no more coded packets can be reduced to degree one or all source packets are decoded. Initially, all studied rateless codes provided equal error protection of the entire source data. In [4], [5] authors proposed generalized rateless codes that provide unequal error protection property. The idea used in the generalized rateless codes is to modify the source packets selection from uniform to nonuniform. In generalized rateless codes, n source packets are partitioned into r sets: s1 , s2 , . . . , sr of sizes α1 n, α2 n, . . . , αr n P such that rj=1 αj = 1. Let pj be the probability that a packet from set sj is chosen to build an encoded packet, and let yl,j be the probability that an original packet in sj is not recovered after l decoding iterations at receiver. In [4], [5] it is shown that for j = 1, . . . , r we have

GOP n+1

yl,j = δj (1 − β(1 −

Fig. 1. Illustration of different frame types in GOPs of an MPEG video stream. Arrows show which frames are used to reconstruct each dependent frame.

One can see that if an error occurs in an I -frame it propagates throughout the GOP, and if a P -frame is defected the error propagates until the next I -frame. On the other hand, a defected B -frame, in the worst case, causes only one frame drop. Therefore, the I -frames have the highest level of importance, and in contrast, B -frames have the lowest importance.

pk αk nyl−1,k )),

l≥1

(1)

k=1

in which y0,j = 1, β(x) = Ω′ (x)/Ω′ (1), and δj (x) = enpj µγ(x−1) with µ = Ω′ (1). It can be shown that sequence {yl,j }l converges to a fixed point yj [5]. This fixed point is the final packet error rate of set j , j ∈ {1, 2, . . . r}, with a rateless code with the parameters Ω(x), γ , α1 , α2 , . . . , αr , and p1 , p2 , . . . , pr . Later, we employ the UEP-rateless codes to encode MPEG video frames, and we employ the results provided here to optimize the parameters of these FEC codes. IV. DECODABLE FRAME RATE FOR MPEG VIDEO WITH EEP-RATELESS CODING

B. UEP-RATELESS CODES Rateless codes [6]–[8] are a class of modern FEC codes with low computational complexity and close to channel

r X

The quality of a video streams can be measured with various quality metrics, peak signal-to-noise ratio (PSNR),

2 of 7

mean opinion score (MOS), packet delivery ratio, decodable frame rate, etc. From an error control coding view, we select decodable frame rate, Q, to measure the quality of the decoded video since it can be mathematically expressed, and it closely reflects the behavior of PSNR [9]–[15]. The value of Q varies between 0 and 1, where a larger value shows a higher frame recovery rate. Q is defined as

can be decoded if all belonging packets to this I -frame are delivered. As a result, the probability that an I -frame is decodable is derived next.

E[Number of decoded frames] . (2) Total number of transmitted frames When the frames are encoded using rateless codes, each frame is broken into smaller transmittable packets according to the maximum possible packet size of the wireless network. A frame can be decoded at the receiver if at least a fraction of packets, called the decodable threshold (DT), is delivered. For instance, DT = 0.8 means that the decoder can tolerate 20% packet loss. To simplify our analysis w.l.o.g., we assume DT = 1. In practice, MPEG video decoders employ error concealment algorithms that makes the decoder error resilience, thus the decoder can recover MPEG video frames even if a portion of frames are distorted. When a video stream is encoded with EEP-rateless codes, at the receiver side all packets of different frame types are recovered with equal packet error rate, P EREEP . P EREEP can be found by (1) for a special case with p1 = p2 = . . . = pr = n1 , where r here is equal to N , and n is the total number of packets of all frames in a GOP. This provides equal protection of frames which results in the conventional EEP-rateless codes. In the EEP-rateless coding case, we have y1 = y2 = . . . = yr = P EREEP , which is the final packet error rate after l decoding iterations. Let us define the adopted notation as given in Table I, for our future reference.

NdecI = (1 − P EREEP )CI ∗ NGOP .

P rdec (I) = (1 − P EREEP )CI .

Consequently, the expected number of correctly decodable I -frames for the entire video is

Q=

TABLE I N OTATIONS ADOPTED IN THIS PAPER Ntotal NdecI NdecP NdecB NGOP CI , CP , CB

(4)

The P -frames are decodable if the preceding I and P frames are recovered and all the packets belonging to the corresponding frame are received with no defects. There are N NP = M − 1 P -frames in each GOP, and the probability that each P -frames is decodable is given by P rdec (P1 ) = (1 − P EREEP )CI +CP , P rdec (P2 ) = (1 − P EREEP )CI +2CP , .. . P rdec (PNP ) = (1 − P EREEP )CI +NP CP .

Therefore, the expected number of correctly decodable P frames, NdecP , for the entire video is given by NdecP =

NP X

(1 − P EREEP )CI +jCP ∗ NGOP .

(5)

j=1

Finally, each (M − 1) B -frames, enveloped between two consecutive P -frames, have the same successful decoding probability. Let Bj denote the j th group of B -frames, which N groups of (M − includes (M − 1) B -frames. There are M 1) consecutive B -frames in total. All groups of B -frames except the last group are decodable if the preceding P - and I -frames are decodable and all packets belonging to the B frames are correctly received. The last (M − 1) B -frames are decoded if the I -frame of the succeeding GOP is also decodable. Each group of (M − 1) B -frames has the same decoding probability given by

The total number of transmitted frames Expected number of decodable frames of each type for the entire video at the receiver side Total number of GOPs in the video stream The number of packets in each type of frame

P rdec (B1 ) = (1 − P EREEP )CI +CP +CB , P rdec (B2 ) = (1 − P EREEP )CI +2CP +CB , .. . N

P rdec (B N −1 ) = (1 − P EREEP )CI +( M −1)CP +CB , M N P rdec (B N ) = (1 − P EREEP )2CI +( M −1)CP +CB . M

Similar to [12], [15] the decodable frame rate in (2) is given by Q=

NdecI + NdecP + NdecB Ntotal

Consequently, NdecB is given by NdecB = [(1 − P EREEP )CI +NP CP +

(3)

where Ntotal = NGOP ∗ N , and NdecI , NdecP , and NdecB are as derived next. An I -frame can be decoded independently, regardless of decoding of other frames. Since we set DT = 1, an I -frame

NP X (1 − P EREEP )jCP ] j=1

∗(M − 1) ∗ (1 − P EREEP )CI +CB ∗ NGOP . (6)

Now, by substituting (4), (5), and (6) into (3) we have an expression for the value of Q when EEP-rateless codes are employed.

3 of 7

V. PROPOSED EFFICIENT VIDEO TRANSMISSION CODING BASED ON UEP-RATELESS CODES This section introduces our new efficient video broadcasting scheme. To increase the video transmission efficiency, i.e. increasing Q and consequently the PSNR of the video, we protect frames unequally according to their importance using generalized UEP-rateless codes. We encode each GOP with applying one independent UEP-rateless code over the entire GOP. There is one I -frame in each GOP with the highest level of importance. Let the importance level of an I -frame be kI = pI n, where pI shows the probability of choosing a source packet from the I -frame to generate an outgoing encoded packet, and n be the total number of packets of all frames in the GOP. Furthermore, we have several B -frames with equal and the lowest level of importance among all frame types. Let us show the importance level of B -frames by kB = pB n. Finally, according to the length of the GOP, different types of P -frames exist. We denote different P N frames with Pj , j ∈ {1, 2, . . . , M − 1} and their importance level by kPj = pPj n. For the sake of simplicity in illustration, assume that the I -frame is transmitted first followed by series of all P -frames, and afterwards B -frames are transmitted. Let the series αI n, αP1 n, αP2 n, . . . , αP N −1 n, αB n, where P N −1

P rdec (I) = (1 − P ERI )CI .

Therefore, the expected number of decodable I -frames for all GOPs is NdecI = (1 − P ERI )CI ∗ NGOP .

(7)

M

M αI + αB + i=1 αPi = 1, denote the segmentation sizes as shown in Figure 2. The corresponding source packet selection probabilities for different frames are also depicted in Figure 2. Naturally, I -frames have the largest size due to their lowest compression level and independency, and B -frames have the smallest frame size.

P 1 P2

PN/M-1

BB BB

The P -frames are decoded if the preceding I - and P -frames are recovered and all packets belonging to the corresponding P -frame are received with no defects. Each P -frame has N a different packet loss rate, P ERPj , j ∈ {1, 2, . . . , M − 1}, according to its protection level, and a different probability of successful decoding, P rdec (Pj ), given by P rdec (P1 ) = (1 − P ERI )CI ∗ (1 − P ERP1 )CP , P rdec (P2 ) = (1−P ERI )CI ∗(1−P ERP1 )CP ∗(1−P ERP2 )CP , .. . QNP P rdec (PNP ) = (1 − P ERI )CI ∗ j=1 (1 − P ERPj )CP .

One GOP

I

tection levels on each GOP. The encoded packets are transmitted over a loss prone wireless channel and recovered at the receiver side with different packet error rates. Let N P ERI , P ERB , and P ERPj , j ∈ {1, 2, . . . , M − 1} be the packet error rates of I -, B -, and Pj -frames, respectively. P ERI , P ERB , and P ERPj can be found in a similar way to P EREEP from (1) by setting the generalized rateless coding parameters as αi = {αI , αPj , αP2 , . . . , αP N −1 , αB } M and pi = {pI , pP1 , pP2 , . . . , pP N −1 , pB }. M After l decoding iterations of generalized rateless codes, the fixed points yj show the final packet error rates of different sets after l decoding iterations. In other words, we have P ERI = y1 , P ERP1 = y2 ,. . .,P ERP N −1 = y N , M M P ERB = y N +1 . Now we can derive a formula for Q with M our new coding scheme. In a similar way to EEP-rateless codes, Q is given by (3), and NdecI , NdecP , and NdecB are formulated below. The I -frame can be decoded independently if all packets belonging to the I -frames are correctly received, thus the probability that an I -frame is decodable is

BB

Selection Probability

Consequently, the expected number of decodable P -frames is given by 1

pI

pP1 p P2

pN/M-1

NdecP = (1 − P ERI )CI ∗

pB

NP Y k X

(1 − P ERPj )Cp ∗ NGOP .

k=1 j=1

(8)

0 ĮI.n

ĮP1.n ĮP2.n

ĮP(N/M-1).n

Finally, B -frames are decodable if the preceding P - and I frames are decodable. According to frames dependencies, the probabilities that B -frames are decodable are given by

ĮB.n

Fig. 2. Illustration of order of frames for transmitting a GOP, probability of source packet selection for frames, and relative sizes assigned to different parts of a GOP.

The UEP-rateless coding is performed with various pro-

P rdec (B1 ) = (1−P ERI )CI ∗(1−P ERP1 )CP ∗(1−P ERB )CB , P rdec (B2 ) = (1 − P ERI )CI ∗ (1 − P ERP1 )CP ∗ (1 − P ERP2 )CP ∗ (1 − P ERB )CB , .. .

4 of 7

P rdec (B N −1 ) = M N QM −1 (1 − P ERI ) ∗ (1 − P ERB )CB ∗ j=1 (1 − P ERPj )CP , P rdec (B N ) = M N QM −1 2CI (1 − P ERI ) ∗ (1 − P ERB )CB ∗ j=1 (1 − P ERPj )CP . CI

Consequently, the expected number of decodable B -frames is given by NdecB = [(1 − P ERI )CI ∗

NP Y

(1 − P ERPj )CP

our objective function is Q, which we try to maximize by finding optimum protection levels. By searching the whole decision space, we find the global optimum values of the protection levels as reported in Table III for MPEG-I and in Table IV for MPEG-II for different values of received overhead, γ . TABLE III O PTIMUM VALUES OF MPEG-I VIDEO STREAM PROTECTION LEVELS FOR DIFFERENT VALUES OF γ.

j=1

+

NP Y k X

(1 − P ERPj )CP ]

γ 1.1 1.2 1.3 1.4 1.5 1.6 1.7

(9)

k=1 j=1

∗(M − 1) ∗ (1 − P ERI )CI ∗ (1 − P ERB )CB ∗ NGOP .

Now, we substitute (7), (8), and (9) into (3) to find the final expression for the packet recovery rate with UEP-rateless video coding. The values assigned to protection levels affect the packet loss rates, and consequently change the value of the resulting Q. We can find the maximum Q by optimizing the values assigned to protection levels. Since the number of P -frames and their protection levels change by varying the length of the GOP, the optimum protection level values depend on values of M and N . Here, we provide two optimization examples for a GOP with M = 3, N = 15 for two cases of MPEG-I and MPEGII. This GOP format is the most common GOP size used in practice, and is shown by IBBPBBPBBPBBPBB. Table II summarizes the typical number of packets in each frame type, i.e. CI , CP , and CB , based on MPEG-I and MPEG-II video streams frame size. TABLE II NUMBER OF PACKETS IN EACH FRAME TYPE FOR MPEG-II VIDEO STREAMS

Frame type MPEG-I MPEG-II

CI 150 400

CP 50 200

MPEG-I AND

CB 20 80

Throughout our simulation, we consider the following degree distribution for rateless encoding Ω(x) derived in [6] Ω(x) = 0.007969x + 0.493570x2 + 0.166220x3 + 0.072646x4

A GOP with M = 3 and N = 15 has six protection levels shown by kI , kB , kP1 , kP2 , kP3 , and kP4 . According P to [4], [5] we have kI αI + kB αB + 4j=1 kPj αPj = 1, which shows that one of the importance level values is dependant to other protection levels; therefore, we have only five independent protection levels to optimize. We note that

kP1 1.17 1.15 1.14 1.13 1.12 1.11 1.11

kP2 1.13 1.12 1.11 1.1 1.1 1.09 1.08

kP3 1.08 1.07 1.07 1.06 1.06 1.05 1.05

kP4 1.01 1.01 1 1.01 1.01 1 1

kB 0.75 0.77 0.79 0.81 0.82 0.83 0.84

TABLE IV O PTIMUM VALUES OF MPEG-II VIDEO STREAM PROTECTION LEVELS FOR DIFFERENT VALUES OF γ.

γ 1.1 1.2 1.3 1.4 1.5 1.6 1.7

kI 1.24 1.21 1.19 1.17 1.16 1.15 1.14

kP1 1.21 1.18 1.16 1.15 1.14 1.13 1.12

kP2 1.16 1.14 1.13 1.12 1.11 1.1 1.1

kP3 1.09 1.09 1.08 1.08 1.07 1.07 1.06

kP4 0.99 1.01 1.01 1.02 1.02 1.02 1.02

kB 0.76 0.79 0.81 0.82 0.83 0.84 0.85

As we expected and it is confirmed by Tables III and IV, the highest protection levels are assigned to I -frames, and the lowest protection levels are assigned to B -frames. P frames protection levels are set to decreasing values, which is larger than B -frames protection levels and lower than I -frames protection levels, according to their position in the GOP. Based on the desired final packet error rate, one should choose appropriate overhead and optimum protection level values from Tables III and IV to acquire the highest performance. In the next section, we evaluate our proposed scheme’s performance.

+ 0.082558x5 + 0.056058x8 + 0.037229x9 + 0.055590x19 + 0.025023x65 + 0.003135x66 .

kI 1.2 1.18 1.17 1.15 1.14 1.14 1.13

VI. PERFORMANCE EVALUATION We set the protection levels to the optimum values of Table III and IV and find Q based on the derived formulas in this paper as our performance metric. We compare the performance of video coding employing UEP rateless codes with the case where video is coded using UEP rateless codes. The simulation results are shown in Figure 3(a) for MPEG-I and in Figure 3(b) for MPEG-II.

5 of 7

Q by utilizing UEP-rateless codes instead of EEP-rateless codes in MPEG-I and MPEG-II video streams transmission.

0.9

60 0.8

Percentage of improvement

Decodable frame rate, Q

1

UEP−rateless coding EEP−rateless coding

0.7 0.6 0.5

1.1

1.2

1.3 1.4 1.5 Received overhead, γ

1.6

1.7

(a) Q values for MPEG-I video stream with UEP and EEP-rateless encoding.

MPEG−II MPEG−I

40 30 20 10 0

1 Decodable frame rate, Q

50

1.1

1.2

1.3 1.4 1.5 Received overhead, γ

1.6

1.7

Fig. 4. Percentage of improvement made in Q versus received overhead, γ, when employing optimized UEP-rateless codes instead of EEP-rateless codes.

0.8 0.6 UEP−rateless coding EEP−rateless coding

0.4 0.2 1.1

1.2

1.3 1.4 1.5 Received overhead, γ

1.6

1.7

(b) Q values for MPEG-II video stream with UEP and EEP-rateless encoding. Fig. 3.

Obtained Q values for MPEG-I and MPEG-II video streams.

These graphs can be described in two ways. First, let us assume we have a fixed amount of data and channel bandwidth. We can see that by using UEP-rateless codes there is an increase in the number of decoded frames at the receiver, or equivalently, the receiver will perceive the video with a higher PSNR. For instance, at γ = 1.3, for MPEG-II video the number of decoded frames increases from 58% to 74% employing UEP-rateless codes. Second, for a fixed decodable frame rate Q, UEP-rateless coding requires a smaller overhead, γ , than EEP-rateless coding. For example, instead of using EEP-rateless codes with overhead γ = 1.5 for 95% frame recovery in MPEG-I video, we can encode the video with UEP-rateless codes and transmit only an overhead of γ = 1.4. Since we set DT = 1, frames of MPEG-II video have higher probability to be dropped due to their larger size. This is the rationale behind the lower Q values for MPEGII video compared to MPEG-I video. Figure 4 shows the improvement percentage made in

In this paper, we employed the rateless codes from [6] which originally have two-layer, inner and outer (precoder), coding algorithm. The precoder can be a conventional fixedrate code, and the inner code is a rateless code similar to the codes proposed in [7]. To keep the coding scheme simple, we have implemented this coding scheme with the rateless coding layer only and no precoding was considered. With such a setup, we need to have larger received overhead for a successful decoding, thus Figures 3 indicate a γ larger than 1. However, If a two layer coding scheme is implemented, the decoding succeeds when γ is slightly larger than one. Therefore, we consider γ = 1.1 as the point where we can compare our scheme with [1]. We can see from [1, Fig. 8] that for Foreman video stream the PSNR has increased on average by 1.5dB by using UEP scheme. This increase in PSNR is approximately equal to an increase in Q by 43% (from 0.55 to 0.79) based on [11, Fig. 5]. In our proposed protocol, we can observe 57% improvement in Q for MPEG-II from Figure 4, which shows our proposed protocol can outperform the protocol proposed in [1]. Moreover, an important advantage of our proposed protocol is having much lower complexity compared to previous coding schemes. Rateless codes [6] have linear time coding and decoding complexity of O(n), while Reed-Solomon codes [16] as employed in [1], have coding complexity of O(n2 ). As a result, besides having higher efficiency, our proposed scheme suites for wireless applications where the computational power is limited. Furthermore, as mentioned earlier fixed-rate codes such as Reed-Solomon codes cannot adapt to varying loss rates, thus they may not be employed on wireless links which have

6 of 7

dynamic characteristics. However, rateless codes are universal erasure channel error correction codes that can adapt to any loss rate and can still exhibit low coding/decoding complexity. On the other hand, rateless codes are the perfect coding choice for multicast video content delivery over wireless channels, since the number of output packets can be limitless in contrast to fixed-rate codes. Our simulation results are based on asymptotic formulas for decoding of generalized rateless codes [4], [5], i.e., Equation (1), which assumes that there are a large number of source packets, while in practice the number of packets is limited. To compare the performance of asymptotic rateless decoding results with rateless decoding with finite number of source packets, we evaluate the value of Q for an MPEGII video stream with frame sizes as given in Table II, and with GOPs with N = 15 and M = 3 for both cases. According to the number of packets in different frame types, each GOP would have n = 2000 source packets in total. Figure 5 compares resulting Q for both cases.

Decodable frame rate, Q

1 0.8

Decoding with n=2000 Asymptotic decoding with n=∞

0.6 0.4 0.2 0 1.05

1.1

1.15

1.2 1.25 1.3 1.35 Received overhead, γ

1.4

Fig. 5. Comparison of Q for rateless decoding with n = 2000 and rateless asymptotic decoding with n = ∞ for an MPEG-II video stream.

Figure 5 shows and confirms that the asymptotic rateless decoding performs close to rateless decoding with moderate number of packets. VII. CONCLUSION In this paper, we proposed a new efficient video broadcasting scheme using low complexity UEP-rateless coding. In UEP-rateless coding, we can have different protection levels for different frames of a video according to their importance. We used the metric decodable frame rate, Q, to evaluate the performance of our proposed protocol. Simulation results showed that rateless codes have the same performance compared to fixed-rate codes while they have the properties of low complexity and flexibility on wireless

channels. We also showed that asymptotic results are close to the case where the number of packets is limited. For future extends to this paper, we propose to apply our protocol to real MPEG and H.264 video streams and evaluate the performance employing other quality metrics such as PSNR and MOS. R EFERENCES [1] T. Fang and L.-P. Chau, “GOP-based channel rate allocation using genetic algorithm for scalable video streaming over errorprone networks,” IEEE Transactions on Image Processing, vol. 15, pp. 1323–1330, June 2006. [2] A. Bouabdallah and J. Laca, “Dependency-aware unequal erasure protection codes,” Journal of Zhejiang University - Science A, vol. 7, pp. 27–33, Jan 2006. [3] A. Abd El Al, T. Saadawi, and M. Lee, “Unequal error protection for real-time video in mobile ad hoc networks via multi-path transport,” Comput. Commun., vol. 30, no. 17, pp. 3293–3306, 2007. [4] N. Rahnavard, B. Vellambi, and F. Fekri, “Rateless codes with unequal error protection property,” IEEE Transactions on Information Theory, vol. 53, pp. 1521–1532, April 2007. [5] N. Rahnavard and F. Fekri, “Generalization of rateless codes for unequal error protection and recovery time: Asymptotic analysis,” IEEE International Symposium on Information Theory, 2006, pp. 523–527, July 2006. [6] A. Shokrollahi, “Raptor codes,” IEEE Transactions on Information Theory, vol. 52, pp. 2551–2567, June 2006. [7] M. Luby, “LT codes,” The 43rd Annual IEEE Symposium on Foundations of Computer Science, 2002. Proceedings., pp. 271– 280, 2002. [8] P. Maymounkov, “Online codes,” NYU Technical Report TR2003883, 2002. [9] P. Acelas, P. Arce, and J. Guerri, “Effect of the Multiple Description Coding over a Hybrid Fixed-AdHoc Video Distribution Network,” in Future Multimedia Networking: Second International Workshop, FMN 2009, Coimbra, Portugal, June 22-23, 2009, Proceedings, p. 176, Springer, 2009. [10] A. Ziviani, B. E. Wolfinger, J. F. Rezende, O. C. Duarte, and S. Fdida, “Joint adoption of QoS schemes for MPEG streams,” Multimedia Tools Appl., vol. 26, no. 1, pp. 59–80, 2005. [11] C.-Y. Yu, C.-H. Ke, C.-K. Shieh, and N. Chilamkurti, “MyEvalvidNT - a simulation tool-set for video transmission and quality evaluation,” IEEE Region 10 Conference TENCON 2006., pp. 1–4, Nov. 2006. [12] C.-H. Lin, C.-H. Ke, C.-K. Shieh, and N. Chilamkurti, “The packet loss effect on MPEG video transmission in wireless networks,” 20th International Conference on Advanced Information Networking and Applications, 2006. AINA 2006., vol. 1, pp. 565–572, April 2006. [13] A. Khan, L. Sun, and E. Ifeachor, “Impact of video content on video quality for video over wireless networks,” in Fifth International Conference on Autonomic and Autonomous Systems, 2009. ICAS ’09., pp. 277–282, April 2009. [14] A. Khan, L. Sun, and E. Ifeachor, “An ANFIS-based hybrid video quality prediction model for video streaming over wireless networks,” in The Second International Conference on Next Generation Mobile Applications, Services and Technologies, 2008. NGMAST ’08., pp. 357–362, Sept. 2008. [15] C. L. Ke and C. Shieh., “Evaluation of streaming MPEG video over wireless channels,” vol. 3, no. 1, pp. 047–064, 2007. [16] I. S. Reed and G. Solomon, “Polynomial codes over certain finite fields,” Journal of the Society for Industrial and Applied Mathematics, vol. 8, no. 2, pp. 300–304, 1960.

7 of 7