RECEIVER DRIVEN LAYERED MULTICAST WITH
LAYER-AWARE FORWARD ERROR CORRECTION
Cornelius Hellge1, Thomas Schierl1, and Thomas Wiegand1,2
{hellge|schierl|wiegand}@hhi.fraunhofer.de
1
Fraunhofer Institute for Telecommunications
Heinrich-Hertz-Institut (HHI)
Einsteinufer 37, D-10587 Berlin, Germany
ABSTRACT
A wide range of different capabilities and connection qualities
typically characterizes receivers of mobile television services.
Receiver driven layered multicast (RDLM) offers an efficient way
for providing different capabilities over such a broadcast channel.
Scalable video coding (SVC) allows for the transmission of
multiple video qualities within one media stream. Using SVC
generates a video bit stream with various inter layer dependencies
due to references between the layers. This work proposes a layeraware forward error correction (L-FEC) approach in combination
with SVC. L-FEC increases robustness of the more important
layers by generating protection across layers following existing
dependencies of the media stream. The L-FEC is integrated as an
extension of a Raptor FEC implementation in a DVB-H broadcast
system. It is shown by experimental results that L-FEC
outperforms traditional UEP protection schemes.
Index Terms— DVB-H, SVC, Raptor, RDLM, Layered
Transmission
1. INTRODUCTION
Digital video broadcasting for handhelds (DVB-H) [1] seems to
become a popular solution for mobile broadcast. Due to the variety
of different device capabilities, e.g. different display resolutions or
computational power, transmitting only one video quality could be
problematic due to extra computation like downscaling or
transcoding at the battery-powered mobile devices.
An approach similar to receiver driven layered multicast
(RDLM) [2] in combination with the recently adopted SVC
extension of H.264/AVC [3] offers an efficient way to provide
multiple video signals over a broadcast channel as shown in
previous work [4]. In such a mobile RDLM scenario, a client
joining the broadcast service only requests the scalable layers,
which provide either a signal that the device is capable or chooses
to process. The transmission of multiple video signals using SVC
is much more efficient in terms of bit-rate compared to simulcast
transmission [5].
A mobile broadcast channel typically suffers from burst
errors. With the additional delay constraints of the streaming
service, the reliable transmission is still a big challenge in the
mobile broadcast scenario. Since broadcast services only provide a
unidirectional downlink channel, a possible solution to increase
coverage is to use an additional forward error correction (FEC) at
the link or application layer. FEC is applied in DVB-H at the link
layer with the optional multi protocol encapsulation FEC (MPE-
2
Image Communications Group
Technical University of Berlin
Einsteinufer 17, D-10587 Berlin, Germany
FEC) and using Raptor coding [7] as application layer FEC.
Although MPE-FEC is intended to be used for streaming services
and application layer FEC for file download, the application layer
FEC offers more flexibility for media aware protection. Possible
approaches are unequal error protection (UEP) [8], priority
encoding transmission (PET) [9] or dependency aware UEP (DAUEP) as proposed in [10]. The aforementioned approaches do not
take the existence of layers in the video stream and their multiple
dimensions of dependencies into account. E.g., SVC allows up to
three different scalability dimensions within one bit stream.
Scalability in SVC can be applied to temporal, spatial and quality
dimension. The proposed layer-aware FEC (L-FEC) approach
generates redundancy symbols following existing dependencies in
the scalable dimensions, i.e., redundancy symbols of layers of
lower priority can be used to correct symbols of layers of higher
priority. The proposed transmission scheme uses RDLM in
combination with SVC to serve different device capabilities.
Furthermore, we use the L-FEC as extension of the Raptor code
defined in DVB-H as proposed in [14]. Simulation results show an
increase in reliability of an SVC transmission in a mobile RDLM
scenario.
The rest of the paper is organized as follows. Section 2 gives a
very brief overview of the SVC standard. Section 3 introduces the
Raptor L-FEC defined in DVB-H and in section 4 we outline the
proposed layer-aware FEC extension and show a simple example.
In section 5 we apply the L-FEC to the Raptor defined in DVB-H
and in 6 we show selected simulation results and conclude in
section 7 with a summary.
2. SCALABLE VIDEO CODING
The SVC design, which is an extension of the H.264/AVC video
coding standard, can be classified as a layered video codec. An
SVC bit-stream can be structured so that devices with different
capabilities can decode parts of it that have a quality very similar
to the case when the bit-stream for each device would be a singlelayer H.264/AVC bit-stream. In SVC, the hybrid video coding
approach of motion-compensated transform coding of H.264/AVC
is extended in a way that a wide range of spatio-temporal and
quality scalability is achieved. The base layer (BL) is an
H.264/AVC compliant bit-stream that ensures backwardcompatibility for existing receivers. The temporal scaling
functionality of SVC for high delay configurations is typically
based on a temporal decomposition using hierarchical bi-predictive
pictures. The spatial scalability is achieved by different encoder
loops with an over-sampled pyramid for each resolution. For
details of SVC, see [3][6].
Using layered multicast, typically the redundancy symbols are
generated separately for each layer. Due to the dependencies
within the SVC bit-stream, lower priority layers and the associated
redundancy symbols cannot be used without successfully decoded
higher priority layers. I.e., if a higher priority layer is lost, even if it
is fully received, all dependent layers including related redundancy
become useless.
The idea of L-FEC is, to follow media coding dependencies in
the media stream for the generation of across layer protection.
Using the proposed approach, protection of less important layers
can be jointly used with protection of more important layers for
recovering the SSs of all participating layers.
A dependency path (DP) contains all referenced layers for
decoding a particular frame in order of importance. Using the LFEC, all redundancy symbols in the same DP can be jointly used
for error correction. Figure 1 sketches the L-FEC approach for one
scalable dimension, e.g., temporal, spatial or SNR scalability.
FEC L-1
FEC L-1
FEC L-1
...
...
...
...
...
...
...
Dependency
Layer 2
FEC 2
Dependency
Layer 1
FEC 1
FEC 1
FEC 1
Dependency
Layer 0
FEC 0
FEC 0
FEC 0
Time = 0
FEC 2
Time = t
FEC 2
...
...
D3 1
D2 1
...
D1 1
FEC
Base layer 010
FEC
110
3
...
...
Base layer
...
D2 1
D2 1
n
sio
en 3)
m
D i (D
...
...
...
Base layer
...
D1 2
D1 1
D1 1
FEC
Base layer 100
FEC
200
Base layer
...
Dimension 1 (D1)
Fig.2: L-FEC generation over multiple dimensions
In each dimension, several layers lDi are present where the
arrows represent the dependencies. In this example, each layer
depends on all layers of higher priority of the same dimension and
partially on the layers of other dimensions. All redundancy
symbols FEC lD1lD2lD3 are generated over all depending layers. The
redundancy symbols within a particular dependency path can be
jointly used for correcting all source symbols of that path. The BL,
e.g., is included in all FEC symbols. Hence, there are multiple
paths where redundancy symbols can be jointly used for correcting
errors in the BL. The L-FEC approach allows for the joint use of
symbols within multiple dimensions. I.e. for a single dimension,
FEC 200, FEC 100, and FEC 000 can be jointly used for decoding
source symbols in dimension D1 up to the maximum layer used for
FEC generation. In the multidimensional case, FEC 111, FEC 110,
FEC 010, FEC 100, and FEC 000 can be jointly used for correcting
errors up to the maximum layer used for L-FEC generation.
In Fig.3 we present an encoding example of the L-FEC using a
very simple systematic FEC code, where redundancy symbols are
generated by simple XOR combinations of the source symbols.
Dependency
Layer B
L-FEC Extension
Layer A
iA0 iA1 iA2 iB0 iB1 iB2
0 0 1
1 1 0
IB=(110)
XOR
1 à CB=(11011)
pB1
1
pB0
Time = 2t
iA0 iA1 iA2
Fig.1: Layer Aware FEC generation over one dimension
The depicted SVC media bit-stream consists of L layers where
the arrows show the dependencies between the layers. The
redundancy symbols FEC 0 of the highest priority layer l=0 are
typically generated given by the FEC coding technique T.
Redundancy symbols of the enhancement layer l=x are calculated
incorporating SSs of all layers l ≤ x. I.e., FEC 1 symbols are
generated over SSs of layer l=0 and layer l=1. Furthermore, FEC 2
symbols are generated over SSs of layer l=0, layer l=1 and layer
l=2 and so on up to FEC L-1, which is generated over SSs of itself
and all layers of higher priority. Using the L-FEC approach the
redundancy symbols of different layers, but same DP, can be
FEC
111
D1 1
D2 2
FEC
D2 1
020
Base layer
FEC
Base layer
000
4. LAYER-AWARE FEC
Dependency
Layer L-1
...
Raptor codes as first introduced in [11] belong to the category of
Rateless or Fountain codes. Such a type of a FEC code can produce
a theoretically infinite number of encoding symbols (ESs) from a
limited number of source symbols (SSs) with a linear complexity.
The receiver can recover the original data by an inverse encoding
process after receiving an amount of ESs only slightly larger than
the number of SSs.
The Raptor code adopted by DVB-H [7] is a systematic code
based on the concatenation of a Luby-Transform (LT-) Code [12],
and an additional pre-code. The pre-code produces intermediate
symbols, which are used as input symbols of the LT-Code. For
systematic behavior, the pre-code is designed so that the output of
the LT-Code contains the SSs. The ESs are generated by XORing
randomly selected intermediate symbols following a given
distribution. More details about the systematic Raptor design can
be found in [7][12][14].
jointly used for error correction. Note that the number of
redundancy symbols remains constant.
This approach can also be extended to multiple dimensions of
dependencies as present in SVC. Fig.2 depicts the L-FEC
generation for three scalable dimensions D1…D3 corresponding to
the temporal, spatial or SNR scalability in SVC.
Dimension 2 (D2)
3. RAPTOR CODE IN DVB-H
Dependency
Layer A
0
0
IA=(001)
1
XOR
0
pA0
1 à
pA1
CA=(00101)
Fig.3: Exemplary encoding using L-FEC
We assume two layers A and B with SSs IA and IB. Layer B
depends on layer A. Each layer has k SSs and p=n-k redundancy
symbols where k=3 and p=2 with symbol size t=1 Bit. Layer A is
protected by typical FEC technique. I.e. the parity bits pA0 and pA1
are calculated by XOR combinations of the SSs. The code word CA
xor
RB=(11011)
0
Dependency
Layer B
Dependency
Layer A
0
1
1
1
0
iB0 iB1 iB2
iA0 iA1 iA2
IA=(001)
IB=(110)
rA0 rA1 rA2 rA3 rA4
? ? ? 0 1
RA=(???01)
In this section, we present selected results for transmission of
QVGA and VGA resolution using SVC over a DVB-H channel. A
Gilbert-Elliot (GE) model is used as statistical model for
simulation of burst losses on the DVB-H channel as used in [13].
The transmission blocks (TB) of the wireless channel are of size
186 bytes. The mean error burst length is about 100 TBs.
We simulated two different sequences. The CIRCLE sequence
has 25 fps and a length of 1297 frames and the SOCCER sequence
has 30 fps and a length of 1794 frames. For encoding we used the
SVC reference software JSVM8.8 with a H.264 BL at QVGA
resolution and a spatial enhancement layer (EL) at VGA, a groupof-picture (GOP) size of 16 and random access point at each
second GOP. In case, a VGA receiver does not receive the spatial
EL, we calculated the PSNR value of an up-scaled VGA resolution
and we use freeze frame error concealment in case the BL is lost.
The BL is protected with a standard Raptor, whereas the spatial EL
protection uses the L-FEC approach. Fig.5 sketches the different
protection types for exemplary two layers, i.e. the L-FEC of the EL
also incorporates the source symbols of the BL.
?
?
?
IA=(???)
iA0 iA1 iA2
Fig.4: Exemplary decoding using L-FEC
Note, if layer A can be corrected by itself, the additional
connections introduced by the L-FEC to layer A can be removed by
the use of the SSs of layer A. In such a case, layer B can be
corrected following standard FEC using only the redundancy
symbols of layer B. We define the code rate (CR) as the number of
source symbols k over the number of related output symbols n
including all redundancy symbols p. Using the L-FEC, the CR of
layer B remains constant with CRB=kB/nB. Contrary to that, the CR
of layer A decreases due to the additional protection by the L-FEC
in layer B. If there is no error in layer B, the minimum CRmin of
layer A using the L-FEC decreases to CRmin = ka/(na+pb), because
all redundancy symbols of layer B can be used for layer A.
Therefore, using the L-FEC, the CR of layer A is between ka/na ≥
CRa ≥ ka/(na+pb) depending on the number of received symbols in
layer A and B. I.e., for a layered transmission, the CR can never
become higher or the protection can never become lower by using
the L-FEC than with standard FEC.
2 Layer
xor
1.Enh. Layer
FEC
1
BaseLayer
FEC
0
L-FEC approach
1.Enh. Layer
BaseLayer
L-FEC
1
FEC
0
Fig.5: Standard FEC generation and L-FEC approach
The ESs are computed over the amount of data between the
random access points. Tab.1 and Tab.2 show the CR of the
standard FEC as well as the CRmin of the L-FEC and the operation
points of the SVC stream. Since CRmin is estimated by adding the
redundant bit-rate of the EL to the redundant bit-rate of the BL.
Tab.1: Code-rate (CR) and operation points for SOCCER
H.264 Base
layer (QVGA)
Spatial Enh.
Layer (VGA)
Source
Bit-rate
352 kbps
685 kbps
PSNR
VGA
31.70 dB
up-scaled
34.85 dB
CR
FEC/UEP
0.62/0.50
≈CRmin
FEC/UEP
0.36/0.35
0.62/0.70
0.62/0.70
Tab.2: Code-rate (CR) and operation points for CIRCLE
5. L-FEC WITH SYSTEMATIC RAPTOR CODES
In order to apply the idea of L-FEC to the systematic Raptor FEC,
the encoding process has to be modified for dependency layers l >
0 following the L-FEC procedure shown in section 4. The standard
LT encoding only covers the SSs of the actual layer. To extend it
following the L-FEC approach and to keep the code rate constant,
the encoding process has to be extended to all referenced layers.
Standard FEC
2 Layer
rB0 rB1 rB2 rB3 rB4
1 1 0 1 1
6. SIMULATION RESULTS
Spatial Scalability
L-FEC Extension
Layer A
rA0 rA1 rA2 rA3 rA4
? ? ? 0 1
I.e. the XORing process covers all layers of higher importance. To
keep the systematic behavior of the extended code, the pre-code
has to be modified as well. Further details about the required
modification can be found in [14]. In the simulations in section 6,
we use such an extended Raptor code, which provides systematic
behavior and layer-aware protection, similar to the example given
in section 4. The resulting ESs of lower priority layers are
generated incorporating the SSs of higher priority layers, and the
first symbols correspond to the original SSs of the encoded layer.
Spatial Scalability
is generated by concatenating the SSs with the redundancy
symbols. Layer B is protected by the L-FEC, i.e. the redundancy
symbols pB0 and pB1 of layer B are generated by additionally
XORing the source symbols of layer A. Finally, the redundancy bits
are concatenated with the SSs of layer B to the code word CB. With
the L-FEC, the redundancy symbols of layer B now also protect
layer A. It is obvious, that the bit-rate of the encoded stream
remains equal to the standard FEC generation. Furthermore, each
codeword is transmitted over an error prone channel. A client
receives the code words RA and RB as depicted in Fig.4. We
assume, that the code word of layer A is affected by three errors,
which overburdens the error correction capabilities of the applied
protection of layer A. I.e. using standard FEC, layer A could not be
decoded. Layer B is received without any error. Therefore, its SSs
can be recovered. But the video stream cannot be decoded due to
the missing references of layer A. Using the L-FEC, the received
redundancy bits of layer B can be used in combination with the
received symbols of layer A for a combined error correction. Since
there are enough received symbols in both layers, the SSs of both
layers can be successfully recovered.
H.264 Base
layer (QVGA)
Spatial Enh.
Layer (VGA)
Source
Bit-rate
222 kbps
275 kbps
PSNR
VGA
33.26 dB
up-scaled
36.84 dB
CR
FEC/UEP
0.60/0.51
≈CRmin
FEC/UEP
0.40/0.40
0.60/0.70
0.60/0.70
Fig.6 and Fig.7 show the results of the simulations, where the
Y-axis shows the mean received video quality in terms of PSNR at
different TB loss rates. We compare two settings using different
CR distributions FEC, UEP as shown in Tab.1 and Tab.2. All
settings have the same transmission bit-rate of 840kbps for
CIRCLE and 1833kbps for SOCCER sequence incorporating
header overhead. Each setting is simulated using standard FEC
encoding (Normal) and the L-FEC (Extended) in the spatial EL.
We focus on the VGA receiver, due to a QVGA client only
receiving the BL would show similar performance compared to
standard FEC. However, the results in [14] show, that with
additional reception of the VGA stream, even a QVGA receiver
would profit by the use of the L-FEC.
SOCCER VGA Receiver 1833 kbps - Burst length: 100 TB
35
7. CONCLUSION AND SUMMARY
In this work, we propose a layer-aware forward error correction (LFEC) approach. L-FEC generates redundancy symbols
incorporating layered structures in modern media codes like SVC.
SVC is used to transmit two different resolutions QVGA and VGA
at the same time. We applied the L-FEC approach to a Raptor
code. The L-FEC approach enhances the protection capability of
the spatial enhancement layer without increasing the bit-rate.
Simulation results in an RDLM-like scenario (DVB-H) show that
the proposed approach outperforms a standard UEP scheme. In
such a scenario, the L-FEC can never show a weaker performance
than standard FEC. Therefore, we recommend the use of the LFEC for the transmission of layered media, e.g. SVC.
34
8. REFERENCES
33
Mean PSNR [dB]
32
31
30
29
FECNormal
FECExtended
28
UEPNormal
27
26
UEPExtended
0
0.05
0.1
0.15
0.2
0.25
Mean TB loss rate
0.3
0.35
0.4
Fig.6: SOCCER sequence using a VGA receiver
CIRCLE VGA Receiver 840 kbps - Burst length: 100 TB
37
36
35
Mean PSNR [dB]
34
33
32
31
30
FECNormal
29
FECExtended
UEPNormal
28
27
UEPExtended
0
0.05
0.1
0.15
0.2
0.25
Mean TB loss rate
0.3
0.35
0.4
Fig.7: CIRCLE sequence using a VGA receiver
Using a VGA receiver, the L-FEC approach shows a gain in
PSNR for all settings and sequences. I.e. the additional protection
of the L-FEC reduces losses in the base layer. In contrast to
standard FEC, using the L-FEC, the FEC and UEP scheme show a
similar performance at lower loss rates. First at higher loss rates,
the settings with the highest protection for the BL show the best
performance. The quality difference between the Normal and
Extended approaches decreases with the increase in difference
between CR of the layers, this is due to the gain is caused by the
redundancy in the EL. If there is less redundancy in the EL, the
observed gain decreases as can also be observed by the difference
between the CRmin value and the CR value as shown in Tab.1 and
Tab.2.
[1] ETSI, “Digital video broadcasting (DVB); transmission
system for handheld terminals (DVB-H),” European Standard
EN 302 304, V1.1.1, November 2004.
[2] S. McCanne, V. Jacobson, and M. Vetterli, “Receiver-driven
layered multicast,” ACM SIGCOMM’96, Aug. 1996.
[3] ITU-T and ISO/IEC JTC 1, “Advanced video coding for
generic audiovisual services, ITU-T Recommendation H.264
and ISO/IEC 14496-10 (AVC),” Version 8, July 2007.
[4] T. Schierl,
C. Hellge,
S. Mirta,
K. Grüneberg,
and
T. Wiegand, “Using H.264/AVC-based Scalable Video
Coding (SVC) for Real Time Streaming in Wireless IP
Networks,” IEEE ISCAS, May 2007.
[5] H. Schwarz, and T. Wiegand, “Further Results for SVC
coding: QVGA to VGA spatial scalability in Mobile TV,”
DVB, TM-AVC, TM-AVC0351, Dec. 2007.
[6] H. Schwarz, D. Marpe, and T. Wiegand, “Overview of the
Scalable Video Coding Extension of the H.264/AVC
Standard,” IEEE Trans. on Circuits and Systems for Video
Technology, Special Issue on Scalable Video Coding, Sept.
2007.
[7] ETSI, “Digital Video Broadcasting (DVB); IP Datacast over
DVB-H: Content Delivery Protocols,” TS 102 472 V1.2.1,
Dec. 2006.
[8] B. Girod, U. Horn, and B. Belzer, “Scalable Video Coding
With Multiscale Motion Compensation And Unequal Error
Protection,” in Proc. International Symposium on Multimedia
Communications and Video Coding, New York, Oct. 1995.
[9] A. Albanese, J. Blomer, J. Edmonds, M. Luby, and M. Sudan,
”Priority encoding transmission,” IEEE Trans. on Information
Theory, 1996.
[10] A. Bouabdallah, and J. Lacan, “Dependency-aware unequal
erasure protection codes,” Journal of Zhejiang University Science A, Volume 7 (Suppl. 1): 27-33, April 2006.
[11] A. Shokrollahi, ”Raptor codes,” IEEE Trans. on Information
Theory, Vol. 52, No. 6, June 2006.
[12] M. Luby, T. Gasiba, , T. Stockhammer, and M. Watson,
”Reliable Multimedia Download Delivery in Cellular
Broadcast Networks,” IEEE Trans. on Broadcasting, March
2007.
[13] M. Hannuksela,
V. Vadakital
and
S. Jumislo-Pyykk,
”Comparison of Error Protection Methods for Audio-Video
Broadcast over DVB-H”, EURASIP Journal on Advances in
Signal Processing, 2007.
[14] C. Hellge, T. Schierl, and T. Wiegand, “Mobile TV using
scalable video coding and layer-aware forward error
correction,” IEEE ICME, June 2008.