Transmission of video telephony images over wireless channels

Transmission of video telephony images overwireless channels

k

HangLiu andMagdaEl Zarki

VideoProcessing andTelecommunicationsLaboratory,Department of Electrical Engineering,University of Pennsylvania,

Philadelphia, PA 19104,USA

Abstract. In this paper, the effects of digital transmission errors on H.263 codecs are analyzed and the transmission of H.263

coded video over a TDMA radio link is investigated. The impact of channel coding and interleaving on video transmission quality is

simulated for different channel conditions. Fading on radio channels causes significant transmission errors and H.263 coded bit

streams are very vulnerable to errors. Powerful Forward Error Correction (FEC) codes are therefore necessary to protect the data

so that it can be successfully transmitted at acceptable signal power levels. FEC, however, imposes a high bandwidth overhead. In

order tomake best use of the available channel bandwidth and to alleviate the overall impact of errors on the video sequence, a two-

layer data partitioning and unequal error protection scheme based onH.263 is also studied. The scheme can toleratemore transmis-

sion errors and leads to more graceful degradation in quality when the channel SNR decreases. In lossy environments, it can

improve the video transmission quality at no extra bandwidth cost.

1. Introduction

The advances in low bitrate video coding technology

have led to the possibility of delivering video services to

users through band limited wireless networks. Both

ITU-T/SG15 and ISO-MPEG4 are working to set stan-

dards for very low bit rate video coding. Recently, ITU-

T/SG15 finished the first draft recommendation of

H.263 which targets the transmission of video telephony

through the Public Switched Telephone Network

(PSTN) at data rates less than 64 kbit/s [1]. The experts'

group is starting to adaptH.263 for wireless applications

because the low bit rate makes it best suitable for band

limited wireless networks. A couple of proposals are cur-

rently being evaluated [2].

H.263 coded bit streams are very vulnerable to errors

and require high channel reliability. Radio channels on

the other hand are error prone. Fading on radio channels

causes significant transmission errors. Therefore,

powerful Forward Error Correction (FEC) codes are

necessary to protect the data so that it can be successfully

transmitted at acceptable signal power levels. However,

the extra bandwidth for FEC overhead is also critical in

wireless networks because the bandwidth in the wireless

domain is much more limited than in wireline networks.

The relationship between error control and video trans-

mission quality must therefore be investigated in order

to get an optimal trade-off.

Layered (or classified) source coding and unequal

error protection have received a lot of attention in recent

years [2^5]. They are two techniques that one can imple-

ment to alleviate the impact of errors on the video

sequence and make best use of the channel bandwidth.

There are two main issues related to layered source cod-

ing and unequal error protection.One is how to partition

the data into different priority layers. The other is how

to choose the appropriate error protection schemes for

the different priority layers. In order to obtain optimal

transmission quality, these two issues have to be jointly

considered in conjunction with the channel characteris-

tics.

In this paper, the effects of digital transmission errors

on H.263 codecs are analyzed, the transport of H.263

video over a TDMA radio link is systematically investi-

gated, and the impact of FEC and interleaving on video

quality is evaluated. Numerical results for the perfor-

mance of the video transmission under different channel

coding and interleaving strategies are presented for var-

ious channel conditions. In order to make best use of the

available channel bandwidth and to alleviate the overall

impact of errors on the video sequence, a two-layer

source coding and unequal error protection scheme

based on H.263 is also studied. The scheme can tolerate

more transmission errors and leads to more graceful

degradation in quality when the channel SNR decreases.

In lossy environments, it improves video transmission

quality at no extra bandwidth cost.

This paper is organized as follows: Section 2 outlines

the H.263 coding structure and analyzes the effect of

transmission errors on the H.263 video. Based on this

analysis, a two-layer data partitioning and unequal error

protection scheme is presented. In section 3, we briefly

describe the FEC and channel model used in this work.

In section 4, the transmission of single priority layer and

two priority layer H.263 video with different FEC, inter-

leaving and data partitioning strategies has been simu-

lated for various channel SNR, and the results are

reported. Finally, section 5 concludes ourwork.

WirelessNetworks 2 (1996) 219^228219

kPart of this paper was presented at IS&T/SPIE Symposium on

Electronic Imaging, San Jose, CA,USA, January 1996.

Ä J.C. Baltzer AG, Science Publishers

2.H.263 video coder and effect of transmission

errors

2.1.Outline ofH.263 video coding structure

H.263 coded data is arranged in a hierarchical struc-

ture with four syntax layers as shown in Fig. 1 [1]. From

top to bottom they are: Picture, Group of Blocks (GOB),

Macroblock (MB) and Block. Data for each picture is

composed of a picture header followed by data for

GOBs, eventually followed by an optional end-of-

sequence code (EOS) and stuffing bits (STUF). The pic-

ture header contains a picture start code (PSC), temporal

reference (TR), type information (PTYPE), quantizer

information (PQUANT), continuous presence multi-

point (CPM), picture logical channel indicator (PLCI) if

CPM mode is indicated by external means, extra inser-

tion information (PEI), and spare information

(PSPARE). Temporal reference for B-frames (TRB) and

quantization information for B-pictures (DBQUANT)

are also included in the picture headers if the optional

PB-framesmode is used.

Each picture is divided into GOBs. Data for each

GOB consists of a GOB header followed by data for

macroblocks. Each GOB comprises one macroblock

row for sub-QCIF, QCIF, and CIF format pictures, and

two macroblock rows for 4CIF and four macroblock

rows for 16CIF. For the first GOB, the GOB header is

empty and therefore only macroblock data is present.

For all otherGOBs, it is up to the encoder whether or not

the GOB header will be left empty or not. If the GOB

header is not empty, it will include GOB start code

(GBSC), group number (GN), GOB logical channel

indicator (GLCI) if CPM mode is indicated by external

means, GOB frame ID (GFID), and quantizer informa-

tion (GQUANT).

A macroblock contains four 8 by 8 blocks of lumi-

nance data (Y) and the two corresponding 8 by 8 blocks

of chrominance data (one of each of the blue chromi-

nance Cb and red chrominance Cr). Data for each

macroblock consists of a macroblock header followed

by data for the blocks. The Macroblock header includes

coded macroblock indication (COD), macroblock type

& coded block pattern for chrominance (MCBPC),

macroblock mode for B-blocks (MODB), coded block

pattern for B-blocks (CBPB), coded block pattern

for luminance (CBPY), quantizer information

(DQUANT), motion vector data (MVD), motion vector

data for optional advanced prediction mode (MVD2-4),

and motion vector data for B-macroblocks (MVDB).

Some fields may not be present in the MB header

depending on the picture header and the other MB

header fields. A block comprises an 8-row by 8-column

matrix of luminance or chrominance data samples (Y,

Cb or Cr). The DC coefficient for INTRA block

(INTRADC) is present for every block of the intra coded

macroblocks. Transform coefficients (TCOFFs) repre-

sent the other DCT coefficients coded by run-length and

variable length coding (VLC) that are present if indi-

cated byMCBPC,CBPB, orCBPY.

2.2. Effect of transmission errors and data partitioning

The impact of a single bit in error will depend on

which bit is hit. A single bit error in the DCT coefficients

will at least damage one block as the DCT coefficients of

each block are coded using run-length and variable

length coding. If loss of synchronization occurs, all the

subsequent blocks in the GOB may be destroyed. As

motion vectors are differentially encoded for the inter-

coded macroblocks of the same GOB, then a bit error in

a motion vector may result in the corruption of this

macroblock and the following predicted macroblocks.

In addition, because VLC is also employed in the macro-

block headers, synchronization will probably be lost

when a bit error occurs. Therefore, the worst case of a bit

error in the macroblock headers is the loss of the com-

pleteGOB. The start codes in the picture andGOBhead-

ers provide the synchronization in the spatial domain

and stop error propagation. The effect of transmission

errors is confined to a GOB alone, i.e., one or more bit

errors in aGOB does not affect otherGOBs. Only if a bit

error occurs in the control information symbols of the

picture headers,may it seriously impair the total frame.

In the temporal domain, predictive coded pictures

(P-pictures) are coded using motion compensated pre-

diction from a past intra or predictive coded picture and

are generally used as a reference for further prediction.

The error will propagate among the P-pictures until the

next intra coded picture (I-picture) which is coded using

information only from itself, thereby providing error

resilience in the temporal domain. Furthermore, as the

higher frequencies are visually less important than the

lower ones, a bit error occurring in the higher frequency

DCT coefficients will have very little impact on video

quality.

Wireless channels are much less reliable than wire-Fig. 1. Simplified syntax diagram for theH.263 video bitstream.

H.Liu,M. ElZarki / Transmission of video telphony images220

links; both random and bursty errors generated by noise

and fading exist when signals are transmitted. Powerful

FEC is required in order to combat the errors. However,

FEC adds a lot of overhead to the system which at worst

could render the delivery of acceptable quality video

impossible because the required data rate exceeds the

available channel capacity. A natural approach is to

rearrange the coded video information such that the

important information is better protected and more

likely to be received correctly. The unimportant infor-

mation is less protected in order to reduce bandwidth

overhead.

Based upon the above observations, a simple

approach is to partition the H.263 coded data into two

priority layers (classes) [5]. The high priority layer

(partition 0 or base layer) consists of the important data,

which contains the control information, motion vectors

and maybe lower order DCT coefficients, while the low

priority layer (partition 1 or enhancement layer) con-

tains the higher frequency DCT coefficients. The GOB

start codes andGOB numbers may or may not be redun-

dantly copied in the low priority layer, depending on the

ratio of additional overhead to significance of synchro-

nization and error recovery in the low priority layer. In

general, it is better to copy them in the low priority layer

if a lot of data is partitioned into this layer. A priority

breakpoint in the picture headers indicates what ele-

ments are to be included in the high priority layer. The

remainder of the bitstream is to be placed in the low

priority layer. This is similar to the MPEG-2 data parti-

tioning syntax [6] except that one more bit is used in the

picture header to indicate whether or not the GOB start

codes and GOB numbers are copied in the low priority

layer. Fig. 2 shows an example of how the decoder

switches between the partitions when a bit stream with

two partitions is decoded. The controller can change the

priority breakpoint so that the I-pictures havemore data

in the high priority layer than the P-pictures.

The two priority layers from the video source coder

employ different FEC codes, reflecting the importance

of the information. After channel coding, a sequential

multiplexing scheme can be used to multiplex the

encoded data from the different priority layers [2]. The

two layers of data are serially interleavedwith each other

in the multiplexed stream. Each layer is preceded by a

layer header in order to allow demultiplexing at the

receiver side. For each frame, the data of the high prior-

ity layer with the prefixed layer header is transmitted

first, followed by the low priority layer header and data.

Interleaving can then be used on the multiplexed stream

to randomize the bursty errors. Finally, the coded data is

placed in the transmitter buffer for transmission.

If block FEC codes are employed and the priority

layer length from a coded frame is not exactly a multiple

of the FEC block code length (i.e., when the coded data

from one priority layer of a frame is divided into blocks,

the last block does not have enough bits), the last block

can include some bits from the same priority layer of the

next transmitted frame. In our simulation, block codes

with127bits longareemployed.Thepriority layerheader

is of fixed length and includes three fields. The first 1 bit

field indicates the layer type.The secondfield is 7bits long

and indicates how many bits in the last block come from

the next frame. The third field is 8 bits long and indicates

the length of the data field following the header and

belonging to this priority layer in units of 127 bits. The

layer length is variable depending on the coded frames.

Generally, 8 bits are enough to represent the layer length

of encoded QCIF format video sequences. In case the

layer length is longer thanwhat the length field can repre-

sent, the field is extendedbyanother8bits.Then thevalue

of the first 8 bit length field is 2

8

ÿ 1 and the value of the

last 8 bit length field is the difference of the layer length

and 2

8

ÿ 1. If convolution codes are used, the layer

header only needs two fields, one is for the layer type and

the other for the layer length in bits because the data for

the layer canexactly come fromone frame [2].

Note that the picture start codes are not necessary

after channel coding because the layer header informa-

tion has indicated where a picture will start. Therefore,

the picture start code will be removed before channel

coding at the transmitter and added after channel decod-

ing at the receiver. The layer headers are well protected

for transmission so that the errors in the layer headers

are negligible. The bit stream is corrupted during the

transmission over the wireless channel. At the receiver,

the corrupted bit stream is deinterleaved, and the two

layers are reconstituted by the demultiplexer, then chan-

nel decoding and source decoding are performed. Good

error tolerance can be achieved if the high priority data

is well protected. This scheme has the advantage ofmini-

mum complexity and a bit rate efficiency very close to

the single layer encoder. Furthermore, the high priority

layer can carry enough information to produce an accep-

table visual quality image even if the low priority data is

in error.

Fig. 2. A segment from a bit stream with two partitions. The priority

breakpoint is set so that one (last, run, level) coefficient event is

included in the high priority layer. The arrows indicate how the decoder

switches between two partitions.

H.Liu,M. ElZarki / Transmission of video telphony images 221

3. Channel error protection

3.1. Forward error correction codes

Real-time video services require high reliability, lim-

ited time delays, and reasonably high transmission rates.

The selection of FEC codes needs to take into considera-

tion of several factors in order to meet the video trans-

mission requirements: (1) The capability of the FEC: the

objective is to improve the end-to-end bit error rate as

much as possible. It should be noted however that the

capability of the FEC codes strongly depends on the

channel characteristics and error patterns. (2) The over-

head or code rate: the FEC codes should add as little as

possible overhead and maximize the code rate. (3) The

block size for each code: most codes becomemore robust

with larger blocks because more redundancy is available

for a given code rate. However, the increase in block size

causes additional delay. (4) The complexity: it should be

simple so that the design/implementation cost can be

minimized. Interleaving spreads the bursty errors due to

Rayleigh fading into random errors required by most

FEC codes. It should be noted that a single bit error

could have the same impact on the reconstructed pic-

tures as if all the bits of the GOB are in error because of

error propagation in a GOB in the compressed video bit

stream. The spreading of errors, via interleaving, in a

compressed video bit stream, may damage more GOBs

thereby having a negative effect on the overall perfor-

mance unless the used channel coding is strong enough

to correct those erroneous bits. The same observation

can be made for MPEG coded video [7]. Furthermore,

interleaving results in delay. Some of the above criteria

in the selection of FEC codes and interleaving degrees

are contradictory. The overriding issue is that the design

of an effective error control scheme should consider all

of the above factors to achieve the best engineering

trade-off.

BCH codes provide a good trade-off in terms of error

correction capability versus complexity. They can effec-

tively correct random errors and have also been success-

fully applied in bursty error environments when

combined with interleaving. We use BCH codes in our

simulations. Fig. 3 depicts the BER performance versus

channel SNR over a Rayleigh fading channel for

BCH(127, 120, 1), BCH(127, 113, 2), BCH(127, 106, 3),

BCH(127, 99, 4), BCH(127, 92, 5), BCH(127, 85, 6),

BCH(127, 78, 7) and BCH(127, 71, 8) codes. We assume

here that themodulation scheme is�=4-QPSK and inter-

leaving is used to randomize the error bursts. The asso-

ciated overheads are 5.5%, 11%, 16.5%, 22%, 27.5%,

33%, 38.5%, and 44%, respectively.

3.2.Channelmodel

The channel and transmissionmodel used in the simu-

lation is based on the Personal Access Communications

Services (PACS) system [8,9], which is one of the indus-

try standard proposals for emerging Personal

Communications Services (PCS) in the United States.

PACS uses TDMA and FDD, with 8 slots per carrier

and 2.5 ms frame duration. �=4-QPSK is chosen as the

Fig. 3. BERperformance of BCH codes on interleavedRayleigh fading channels.


modulation format because of its high spectrum and

power efficiency. The channel bandwith is 300 kHz and

the radio link rate is 384 kb/swith 32 kb/s basic data rate

for each user. One time slot carries 120 bits, including 80

bits of user information. A base station can support mul-

tiple handsets simultaneously with a transmitter and a

receiver. Furthermore, it is able to allocate several slots

to a single call to provide higher data rates.

A fading simulator based on the above channel and

transmissionmodel is used to generate the error patterns

of the channel [10]. We assume that the channel coher-

ence bandwidth ismuch larger than the signal bandwidth

and delay spread is not a serious problem. Otherwise,

anti-intersymbol interference (ISI) measures such as

adaptive equalizers should be used to alleviate ISI. The

carrier frequency is 1.9 GHz which is in the frequency

band of PCS.

4. Simulation results

4.1. Transmission of standard single layerH.263 video

Fig. 4 shows the block diagram of the wireless video

transmission system under investigation. Source coding

is performed first, then various FEC and interleaving

error control schemes are used. To evaluate the effect of

channel errors, error patterns generated by the fading

simulator are added modulo-2 to the binary bit stream

representing the coded output from the interleaver. The

coded bit stream with errors is then reconstructed into a

video sequence.

Simulation was carried out on several QCIF video

sequences. The results of the ``Mother and Daughter''

sequence are reported in this paper, which contains typi-

cal video telephony-like images. The original YUV

4 : 1 : 1 video sequence is encoded with 15 frames per

second, and an I frame is used every second (i.e. every 15

frames). Video encoding and decoding are performed

with the modified Telenor R&DH.263 software and the

optional PB-frames mode, unrestricted motion vector

mode and advanced predictionmode are employed.

Peak Signal-to-Noise Ratio (PSNR) as a function of

the average channel SNR is used as an objectivemeasure

of video quality for a given error control scheme. The

overall bit rate including the video and FEC overhead is

always 32 kb/s for all simulation scenarios. 200 encoded

frames with different FEC and interleaving error protec-

tion schemes are transmitted at various average channel

SNR conditions. In order to reduce the sensitivity of the

encoded video to the temporal and spacial location of an

error, each transmission was run 20 times using a differ-

ent starting time. The average value over all the runs for

each transmission is presented. For the two layer coded

sequences, the sum of the video data in the high and low

priority layers combined with the FEC used for each

layer and the overhead due to data partitioning deter-

mines the overall data rate.

Fig. 5 depicts the average PSNR as a function of the

average channel SNR using BCH codes with 127 bits

block length and different error correction capabilities.

The standard single layer H.263 coded bit streams are

used and interleaving (INV) is performed over 20 127-

bit-blocks. The vehicle speed is 2.5 mi/h. For each FEC

case, the video quality rapidly degrades as the average

channel SNR decreases below a threshold. This is

because there is a dramatic drop in video quality once

errors occur in the headers and motion vectors. The

threshold of the curve occurs at a lower channel SNR for

more powerful FEC codes. However, the PSNR is lower

for stronger FEC codes at high channel SNRwhen there

is no error, because higher overhead required by stron-

ger FEC codes reduces the video source rate.

Fig. 6 shows the impact of interleaving on video qual-

ity. BCH(127, 113, 2) is used as error protection with dif-

ferent interleaving degrees and the vehicle speed is 2.5

mi/h. When the degree of interleaving increases, the

error bursts are better randomized so that FEC codes,

which are ideally suited to correct uncorrelated errors,

can handle them better. On the other hand, as discussed

in section 3, the spreadingof errors in a compressed video

bit streammaydamagemoreGOBs if the channel coding

can not correct these errors. This is why higher degrees of

interleaving result in better PSNRperformancewhen the

average channel SNR is high and error rate is low.

However, no interleaving is the best when the average

channel SNR is very low and error rate is very high.

Interleaving also causes additional delay. As an

example, we assume that the processing delay of the

transmitter and the receiver is 50 ms which is the total

time required to perform video encoding, channel encod-

ing, modulation, demodulation, channel decoding, and

video decoding, etc., and each protected PB frame is 2.54

kbit including FEC overhead. The transmission delay

for each PB frame is 79ms for a 32 kb/s channel. If inter-

leaving is performed over a PB frame (20 127-bit-long

blocks), the receiver can start processing the data only

after all the interleaved bits are received. Thismeans that

a 79 ms transmission delay is added. The video frame

interval is 67 ms for a 15 frames/s video connection, andFig. 4. Diagramof video transmission system.


the video encoder can start encoding only after both the

P and B frames enter its buffer so that 67 ms additional

delaymust be counted for the PB codingmode. Thus, the

total delay for a single PB frame interleaving is 196 ms

(the propagation delay is assumed to be negligible).

If interleaving is performed over two PB frames (40

127-bit-long blocks), the first PB frame suffers an addi-

tional delay of 2� 67ms at the transmitter because inter-

leaving can be performed only after the second PB frame

is processed. The second frame suffers a delay of

2� 67 ms at the receiver while the first frame is being

processed. In addition, the receiver needs to wait one

more 79 ms transmission delay before it can start dein-

terleaving and decoding. The additional delay is 213 ms

Fig. 5. Average PSNRvs. average channel SNR for different FEC codes with a vehicle speed of 2.5mi/h and interleaving over 20 codes.

Fig. 6. Average PSNRvs. average channel SNR for different interleaving degrees with a vehicle speed of 2.5mi/h and BCH(127, 113, 2).


due to the two frame interleaving. Thus, the total delay

for two frame interleaving is 409ms.

The size of I frames is much lager than that of P and

PB frames and its transmission delay is longer. For

example, if the protected I frame is 14 kbits, its transmis-

sion delay is 438 ms. The total delay for an I frame is the

sum of the transmission delay and processing delay (488

ms if the processing delay is still 50ms and interleaving is

over the I frame itself). One way to reduce I frame delay

is to reduce the size of I frames. The total delay for the

system is determined by the larger of the I frame delay

and the interleaving delay.

Fig. 7 shows the average PNSR performance for dif-

ferent vehicle speeds, where BCH(127, 113, 2) and inter-

leaving over 10 127-bit-blocks are used. Fades are longer

for a slower vehicle. When the degree of interleaving is

limited, it may result in higher residual error rates after

channel decoding than if the vehicle speed were higher.

The higher vehicle speed on the other hand means that

the transmitted radio signals experience larger Doppler

effect which increases the channel error rate [11]. When

the channel SNR is high (the error rate is low), the per-

formance of the error control scheme dominates the

video transmission quality. Then higher speed gives bet-

ter average PSNR since the error control is more effec-

tive for short error bursts than for long error bursts with

a fixed degree of interleaving. However, when the chan-

nel SNR is very low (the error rate is very high), lower

speed gives better average PSNR since the error control

performance is not important for this case (the channel

errors are beyond the error correction capability of the

FEC code no matter what the vehicle speed is) and the

channel error rate for a lower speed vehicle is smaller.

The curves in Fig. 7 are similar to those shown in Fig. 6,

where for a fixed vehicle speed, the degree of interleaving

affects the performance.

4.2. Transmission of two layerH.263 video

In Fig. 8, the average PSNR is depicted for the com-

parison of the single layer standardH.263 bit stream and

two layer unequally protected bit streams. The five sce-

narios under consideration are: (1) single layer H.263

coded bit stream with BCH(127, 99, 4) error protection,

(2) two layer H.263 coded bit stream with BCH(127, 92,

5) for high priority layer and BCH(127, 113, 2) for low

priority layer protection, (3) two layer bit stream with

BCH(127, 92, 5) for high priority layer and BCH(127,

120, 1) for low priority layer protection, (4) two layer bit

stream with BCH(127, 92, 5) for high priority layer and

BCH(127, 127, 0) for low priority layer protection, (5)

two layer bit streamwith BCH(127, 85, 6) for high prior-

ity layer and BCH(127, 127, 0) for low priority layer pro-

tection. For the two layer scenarios, the data

partitioning point is adjusted so that the total overhead

due to channel coding and data partition is the same as

the channel coding overhead for the single layer scenario

(1), i.e. all the scenarios have the same source data rate.

Therefore, when there is no error, the PSNR of the two

layer schemes are equal to that of the single layer scheme

protected with BCH(172, 113, 4). This value is greater

than those values for the single layer scheme using

BCH(127, 92, 5) or BCH(127, 85, 6) error protection

because the source rate for the two layer coder is higher

(only the high priority layer uses such high overhead pro-

tection, not the low priority layer).

Fig. 7. Average PSNRvs. average channel SNR for different vehicle speedswith BCH(127, 113, 2) and interleaving over 10 codes.


For the two layer case, errors in the low priority layer

first exceed the error correction capability of its FEC

code, this results in a slight degradation of the video

quality. However the important data is correctly

received for a wider range of SNR conditions since the

high priority layer is protected with more powerful FEC

Fig. 8. Average PSNR vs. average channel SNR for 2 layer unequal error protection with a vehicle speed of 2.5 mi/h and interleaving over 40

codes.

Fig. 9. The PSNR of each frame for the 180 decoded frames of the ``Mother and Daughter'' sequence after transmitted over the wireless channel

at the average channel SNR 22.5 dB under the equal and the unequal error protection as well as the ideal case with no error and no error

protection.


codes. Therefore, in general, two layer scenarios can tol-

eratemore transmission errors and lead tomore graceful


This is an advantage of the two layer scheme. Wireless

channels are time varying and it may be impossible to

determine a single optimal FEC, we therefore prefer a

scheme that can yield a constant video quality over a

wide range of channel conditions. In addition, layered

source coding and unequal error protection provides

more freedom to achieve an optimal overall video qual-

ity. As shown in Fig. 8, an optimum (scenario (2) here)

exists, providing an acceptable video quality for a wide

range of channel SNR and exhibiting much better per-

formance at low SNR conditions than the single layer

scheme with the same source coding rate. It is obvious

that careful choice of the partition point andmatching of

the error protection levels to the source bits is necessary

to obtain the optimal performance. For example, we

notice that there are too many unprotected bits for sce-

nario (5) so that the PSNR starts to drop at high SNR. It

is better to use a less powerful code on a larger number of

video bits and leave fewer bits unprotected.

In Fig. 9, the PSNR of every frame for the 180

decoded frames of the ``Mother andDaughter'' sequence

is depicted for the above scenarios (1) and (2) after trans-

mission over the wireless channel at an average channel

SNR of 22.5 dB. The PSNR of the decoded frames for

the same video sequence encoded by the single layer

standard H.263 at 32 kb/s with no error (and no error

protection) is also shown for the purpose of comparison.

The average PSNR values of scenario (1) (33.2 dB) and

scenario (2) (33.5 dB) are approximately equal here. For

Fig. 10. The fifteenth frame of the decode ``Mother and Daughter''

sequence after transmitted over the wireless channel at the average

channel SNR 22.5 dB, (a) equal error protection, (b) unequal error

protection, (c) ideal case with no error and no error protection.

Fig. 11. The eighty-sixth frame of the decode ``Mother and Daughter''

sequence after transmitted over the wireless channel at the average

channel SNR 22.5 dB, (a) equal error protection, (b) unequal error pro-

tection, (c) ideal case with no error and no error protection.


the single layer case (scenario (1)), all the data is equally

protected. The important data such as headers and

motion vectors are more easily in errors as they are

equally likely to be hit. We notice that some extreme

drops in PSNR are possible as shown in Fig. 9. This is

because an error in the header or themotion vector could

cause major damage to the frame. This heavily degrades

the video quality of some frames. The fluctuation in

video quality is much less in the case of the unequal error

protection (scenario (2)) as the important data are better

protected. This increases the overall video quality of

unequally protected sequence. Figs. 10(a)^(c) display

the fifteenth frame of the decoded ``Mother and

Daughter'' sequence under the scenarios (1) and (2), and

the single layer error free case, respectively. Figs. 11(a)^

(c) show the eighty-sixth frame. The fifteenth frame and

the eighty-sixth frame correspond to the poorest quality

frames (out of 180 frames total) for the equally protected

sequence (scenario (1)) and for the unequally protected

sequence (scenario (2)), respectively. Subjective evalua-

tion tests indicate that the overall visual quality of the

unequally protected sequence is much better than that

with the single layer source coding and equal error

protection.

5. Conclusions

In this paper, the transmission of H.263 coded video

over a TDMA radio link is investigated, and the impact

of FEC and interleaving on video quality is evaluated.

We obtain the results for the average PSNRas a function

of the average channel SNR under various channel cod-

ing and interleaving strategies and at different vehicle

speeds. Fading on radio channels causes significant

transmission errors and H.263 coded bit streams are

very vulnerable to these errors. Therefore, powerful

FEC codes are necessary to protect the data so that it can

be successfully transmitted at acceptable signal power

levels. However, the high bandwidth overhead of FEC is

also crucial in wireless networks. In order to make best

use of the available channel bandwidth and to alleviate

the overall impact of errors on the video sequence, a two-

layer source coding and unequal error protection scheme

based on H.263 is also studied. The scheme can tolerate

more transmission errors and leads to more graceful


In lossy environments, this simple scheme, combining

two layer source coding with adequate unequal error

protection, can improve the video transmission quality

at no extra bandwidth cost.

References

[1] ITU-T/SG15, Draft Recommendation H.263, Video coding for

low bitrate communication (1995).

[2] Robert Bosch GmbH and IENT, RWTH Aachen, Proposal for

extension of recommendation H.26P for mobile application:

H.26P/M, ITU-T/SG15Proposal, LBC-95-003 (1995).

[3] R. Stedman, H. Gharavi, L. Hanzo and R. Steele, Transmission

of subband-coded images via mobile channels, IEEE Trans.

Circuits and Syst. forVideo Tech. 3 (1993) 15^26.

[4] Y. Zhang, Y. Liu and R. Pickholtz, Layered image transmission

over cellular radio channels, IEEE Trans. Vehicular Tech. 43

(1994) 786^794.

[5] H. Liu and M. El Zarki, Data partitioning and unbalanced error

protection for H.263 video transmission over wireless channels,

Int. Symp. on Multimedia Communications & Video Coding, New

York,NY (1995).

[6] ISO/IEC JTC1/SC29/WG11 (MPEG-2), Generic coding of

moving pictures and associated audio: video (1994).

[7] Y. Zhang and X. Lee, Performance of MPEG codecs in the

presence of errors, J. Visual Comm. and Image Representation 5

(1994) 379^387.

[8] J. Padgett, C. Gunther and T. Hattori, Overview of wireless

personal communications, IEEE Commun. Mag. 33 (Jan. 1995)

28^41.

[9] D.C. Cox, Wireless network access for personal

communications, IEEECommun.Mag. 30 (Dec. 1992) 96^115.

[10] M. Jeruchim, P. Balaban and K. Shanmugan, Simulation of

Communication Systems (Plenum Press, New York, NY, 1992)

chap. 4.

[11] M. Yacoub, Foundations of Mobile Radio Engineering (CRC

Press, BocaRaton, FL, 1993) chap. 4.

Hang Liu received the B.S. from Tianjin

University, China, in 1985, and the M.S. from

the University of New Orleans, New Orleans,

LA, USA, in 1992. He is currently a Ph.D. can-

didate in the Department of Electrical

Engineering, University of Pennsylvania. His

research interests include video compression

and communications, wireless networking,

ATMbased networks and digital signal proces-

sing.

Magda El Zarki received the B.E.E. degree

from Cairo University, Egypt, in 1979, and the

M.S. and Ph.D. degrees in electrical engineer-

ing from Columbia University, New York,

NY, in 1981 and 1987, respectively. Sheworked

from1981 to 1983 as a communication network

planner in the Department of International

Telecommunications at Citibank in NewYork.

She joined Columbia University in 1983 as a

research assistant in the Computer

Communications Research Laboratory where she was involved in the

design and development of an integrated local area network testbed

called MAGNET. In 1988 she joined the faculty of the Department of

Electrical Engineering of the University of Pennsylvania, teaching

courses and conducting research in the field of telecommunications,

where she currently serves as an associate professor. She also holds a

secondary appointment in the Department of Computer and

Information Sciences. In January 1993, she was appointed as a part-

time professor of Telecommunication Networks in the Faculty of

Electrical Engineering at Delft University of Technology, in Delft, The

Netherlands. Dr. El Zarki is a member of the ACM, IEEE and Sigma

Xi. She is actively involved in many of their sponsored conferences and

journals. She was the Technical Program Chair of IEEE

INFOCOMM'94.


Transmission of video telephony images over wireless channels

Documents

Transcript of Transmission of video telephony images over wireless channels