Frame corruption estimation from route messages for video coding over mobile ad hoc networks

Abstract

Recently we proposed a cross-layer design to support video communication over error-prone mobile ad-hoc networks. The idea is to utilize routing messages and network parameters to estimate the corrupted frames, and to guide the reference frame selection at the video encoder to mitigate error propagation. In this paper, we focus on the frame corruption estimation method used in the design. We build a packet loss probability model from the MAC layer mechanism and network parameters; then we utilize the model along with the routing messages received at the network layer to estimate the possible corrupted frames. We study our estimation method under different network settings and demonstrate its effectiveness and robustness. We further show the video quality gains achieved by adapting reference frame selection to the estimated results.

Full text

Page 1

Frame Corruption Estimation from Route Messages
for Video Coding over Mobile Ad Hoc Networks
Yiting Liao and Jerry D. Gibson
Department of Electrical and Computer Engineering
University of California, Santa Barbara, CA, 93106
Email: {yiting, gibson}@ece.ucsb.edu
Abstract—Recently we proposed a cross-layer design to sup-
port video communication over error-prone mobile ad-hoc net-
works. The idea is to utilize routing messages and network
parameters to estimate the corrupted frames, and to guide the
reference frame selection at the video encoder to mitigate error
propagation. In this paper, we focus on the frame corruption
estimation method used in the design. We build a packet loss
probability model from the MAC layer mechanism and network
parameters; then we utilize the model along with the routing
messages received at the network layer to estimate the possi-
ble corrupted frames. We study our estimation method under
different network settings and demonstrate its effectiveness and
robustness. We further show the video quality gains achieved by
adapting reference frame selection to the estimated results.
I. INTRODUCTION
Mobile ad-hoc networks have drawn increasing attention
in recent years due to their wide deployment in military,
homeland defense, and disaster recovery applications. Sup-
porting real-time services such as voice and video over such
networks is essential but not easily accomplished due to the
dynamic nature of the network. Node mobility and the lack of
infrastructure in the network leads to frequent link failures
and route changes while the effects of fading, noise and
interference in the wireless medium may affect the link quality.
For video transmitted over such lossy networks, a major
problem is that error propagation introduced in the motion-
compensated prediction loop may greatly degrade the deliv-
ered video quality. There are many research efforts to ad-
dress this problem, including intra refresh techniques [1], [2],
redundant picture coding [3], and feedback-based reference
picture selection (RPS) [4]. Intra refresh techniques insert
intra macroblocks (MBs) or pictures to minimize the effect
of error propagation, while redundant picture coding methods
allocate redundant pictures to increase the delivered video
quality. However, these techniques introduce certain reductions
in coding efficiency. On the other hand, the coding efficiency
penalty of video coding with RPS is much lower and results
in [5] show that the RPS algorithms provide better error
resilience than conventional intra refresh techniques. However,
RPS requires extra overhead to transmit control messages
between the encoder and the decoder, which may lead to
higher network traffic load and increased delay.
This research has been supported by the UC Discovery Grant Program and
Applied Signal Technology, Inc.
In our previous work [6], we proposed to use the standard
ad-hoc routing messages to estimate the possible packet losses
in the networks and select reference frames accordingly to
alleviate error propagation caused by packet losses. It has been
shown to be an effective method to enhance error resilience
of video without introducing extra network overhead or delay.
In [6], we use a simple method to utilize the routing messages,
that is, every time a route error (RERR) message is received
by the source node, we assume that the previously transmitted
packet is lost. However, due to the transmission delay of the
video packets and routing messages, a RERR may indicate
possible losses of several previously transmitted video packets.
In this paper, we establish a new model to estimate the
packet loss probability of preceding packets transmitted from
the source node before a RERR is received. This model uti-
lizes information from MAC access mechanism and network
parameters for the packet loss estimation. We further estimate
the corrupted frames using the estimation model and routing
messages received from the network layer. We run experiments
using the Qualnet simulator to test the effectiveness of our
design and we show that our estimation method can effectively
detect the frame corruption in the network while maintaining a
low probability of false alarm. We further demonstrate that our
estimation model works well under various network settings.
The estimated results can be adapted to the reference
frame selection approach for multiple description coding with
multipath transport [7] (referred to as RA-MDC) and help to
alleviate the effect of error propagation. In this paper, we show
that the RA-MDC achieves PSNR gains of up to 2.16 dB
under different packet loss rates. In [7], we discuss the RA-
MDC method in more detail and provide more comprehensive
results on delivered video quality for multiple users.
II. FRAME CORRUPTION ESTIMATION
In this section, we present our frame corruption estimation
method. First, we build a packet loss probability model for
the preceding packets sent from the source node when a route
error message is received at the source node. Based on the
model and the routing messages, we then estimate the packet
loss probability of each transmitted packet at the source node
and determine whether the corresponding frame is corrupted
or not.

Page 2

V5
V6
V7
V8
V9
V1
Transmitter
Intermediate
nodes
RERR
Receiver
RERR
RREP
RREQ
RREQ
RREP
...
...
...
Tretrans
TRERR
Tdata
V2
V3
V4
V5
V6
V7
V10
V1
V2
V3
V4
V10
V10
...
V1
V2
V3
V5
V7
Routing messages
Packets in GOOD state
Packets in FAIL state
Packets in BAD state
Packets discarded
Fig. 1.An example to illustrate the packet losses in the network and the corresponding routing messages
A. Packet Loss Probability Model
For most on-demand routing protocols over ad-hoc net-
works, a route error (RERR) message is initiated when the
MAC layer fails all retransmission attempts to transmit a
packet to the next hop destination. This RERR indicates that
a link becomes unreliable and packets transmitted through
this link suffer a high packet loss rate. Before the source
node receives the RERR, video packets sent from the source
node are still transmitted through this error-prone link and are
susceptible to losses. When the source receives the RERR, it
either reconstructs the route from the route cache or initiates
a route recovery process to find a new route. A new route
is established when the route reply (RREP) message reaches
the source node. Packets scheduled to be transmitted in the
broken route during the route recovery process are discarded
and marked as lost. According to the routing mechanism, when
the source node receives a RERR message, it indicates a link
becomes unreliable and packets previously sent through this
link suffer packet losses. We derive a model to estimate the
packet loss probability of the preceding packets sent through
this unreliable link.
Figure 1 illustrates how the RERR message correlates to the
packet losses in the networks. As shown in Fig. 1, a RERR
is initiated at the intermediate node when video packet v4
exhausts all retransmission attempts and still fails to transmit
to the next hop destination. We define the retransmission
delay of this packet as Tretrans. After time TRERR, the source
node receives the RERR and stops transmitting video packets
through the unreliable link. We see that packets v5, v6, v7
sent during time period Tretrans+TRERRare still transmitted
through the unreliable link and are very susceptible to packet
loss.
We assume that anytime the source receives a RERR, the
preceding video packets sent from the source node follow
the same packet loss distribution under the same network
conditions. Therefore, we denote Pr(n) as the packet loss
probability of the nthpreceding packet sent from the source
node before the source node receives a RERR. Our main goal
is to model Pr(n) and utilize it to determine the potential
corrupted frames. Due to the random delay between link
failure and RERR reception at the source, the nthpreceding
packet before RERR can be sent at a time before, right at or
after the link failure happens. We use three states to represent
these three cases: GOOD means the packet is sent before
the link failure, FAIL means the packet fails to transmit and
triggers RERR, and BAD means the packet is sent after the
link failure. According to our above analysis, we define Pr(n)
as
Pr(n) = λg·pg(n) + λf·pf(n) + λb·pb(n)
where λg, λf, and λbrepresents the packet loss probability in
GOOD, FAIL, or BAD state respectively, and pg(n), pf(n),
and pb(n) denotes the probability of the nthpreceding packet
in these three states, respectively. In the following, we estimate
the state probability distribution and packet loss probabilities
in these three states.
1) Estimation of State Probability Distribution: The state
of a video packet depends on the delay of the link failure
feedback and the transmission interval of video packets. For
example, as shown in Fig. 1, v4 is the packet that triggers
RERR and hence is in FAIL state. The packets sent before
v4 (e.g. v3) are in GOOD state while the packets sent after
v4 are in BAD state. Therefore, we can compare the video
packet transmission interval Tdata and the delay of the link
failure feedback Tdelay to determine the state probability of
the packets sent before receiving the RERR by


We can calculate the video packet interval Tdataby
(1)

pg(n) = p(Tdelay≤ (n − 1)Tdata)
pf(n) = p(nTdata≥ Tdelay> (n − 1)Tdata)
pb(n) = p(Tdelay> nTdata)
(2)
Tdata= L/Rt
(3)
where Rtis the transmission bitrate of the video sequence and
L is the payload size. Then in order to calculate Eq. (2), we
need to estimate the probability distribution of Tdelay.
As shown in Fig. 1, we see that Tdelay consists of two
parts: the retransmission delay of a packet that fails all
retransmission attempts (denoted as Tretrans) and the time
period to transmit the RERR to the source (denoted as TRERR).
So we have
Tdelay= Tretrans+ TRERR
(4)
The values of both Tretransand TRERRdepend on the MAC
layer access mechanism. In this paper, our estimation is based
on the 802.11 DCF basic access mechanism [8].

Page 3

Data
SIFS+ACK
ACK Timeout
Slot time
CW=CWmin
Successful
transmission
CW=2CWmin+1
First
retransmission
Backoff
BackoffUnsuccessful transmission
Backoff
CW=4CWmin+3
Defer
...
Second
retransmission
Defer
DIFS
Fig. 2.Packet retransmission procedure based on basic access mechanism
a) TRERR: We first estimate the time period to transmit
the RERR to the source TRERRby
TRERR= nhop·TC
(5)
where nhopis the average number of hops to transmit RERR
to the source, TC is the transmission time for a successful
RERR transmission defined in [8].
b) Tretrans: We then estimate the transmission delay
Tretrans of a packet that fails to transmit from the current
station to the next hop destination after exhausting all re-
transmission attempts. As shown in Fig. 2, each transmission
period consists of a defer access and a backoff process.
The transmission procedure starts when the station senses an
idle channel and invokes a backoff procedure. The backoff
time is uniformly chosen in the range of [0,CW], where
CW is the current contention window (CW) size. Then the
station sends out the video packet. If the transmitting station
does not receive the ACK, the station concludes that the
transmission has failed and invokes a retransmission process
until the retransmission limit is reached. Note that CW takes
an initial value of CWmin and exponentially increases after
each unsuccessful transmission, until it reaches the maximum
CW size of CWmax.
Based on the above analysis, the transmission delay of a
packet that fails all retransmission attempts is
Tretrans= mTD+ Tbackoff
(6)
where m is the retransmission limit, TDis the time period of a
defer access defined in [8], and Tbackoffis the overall backoff
time.
The overall backoff time is a random variable that is the
sum of a series of independent random variables uniformly
distributed in the range of [0,Wi]·Tslotand Wiis the CW size
in the ithretransmission defined in [8].
We define TBias the backoff time in the ithretransmission,
then we have TBi∼ U(0,Wi·Tslot), where U(0,Wi·Tslot)
represents an uniform distribution in the range [0,Wi]·Tslot.
Thus the overall backoff time is
Tbackoff=
m−1
?
i=0
TBi∼ Us(0,
m−1
?
i=0
Wi·Tslot)
(7)
where Us(·) represents the probability distribution of the
overall backoff time Tbackoff, which is the sum of m uniform
random variables. We use Ps(t) to represent the CDF of
Tbackoff, i.e. the probability that Tbackoff is shorter than time
t is represented by Ps(t).
Finally, based on Eqs. (2)-(7), we have the state probability
distribution by


2) Estimation of Packet Loss Probability λg, λf, and λb: λg
refers to the packet loss rate of a good link, in which the ACK
is received to indicate a successful transmission. Therefore,
we assume λg = 0. λb is defined as the packet loss rate
of an unreliable link, which is the probability that the video
packet does not reach the next hop destination successfully.
λf is the packet loss rate for the video packet that fails all
transmission attempts and triggers the RERR. Based on the
MAC layer mechanism, we know that each time a video packet
fails a transmission, it means either the video packet fails to
transmit to the next hop destination or the ACK message is not
received by the transmitter. Thus, λfis the conditional packet
loss probability for the video packet that fails all transmission
attempts.
Let A0denote the event that the video packet is lost and A1
denote the event that the video packet fails all transmission
attempts. We assume that each transmission is independent
and the loss probability of a video packet and an ACK for
an unreliable link are pdataand pACKrespectively. Then we
have p(A0) = pm
Finally, λf and λbare represented by
p(A0|A1) =p(A1|A0) · p(A0)

pg(i) = Ps(∆T)
pf(i) = Ps(Tdata+ ∆T) − Ps(∆T)
pb(i) = 1 − Ps(Tdata+ ∆T)
where ∆T = (n−1)Tdata−TRERR−mTD.
(8)
dataand p(A1) = [pdata+(1−pdata)·pACK]m.
λf
=
p(A1)
=
pm
data
[pdata+ (1 − pdata) · pACK]m
p(A0) = pm
data
(9)
λb
=
(10)
By Eqs. (9) and (10), we have
λb
λf
= [pdata+ (1 − pdata)·pACK]m≤ 1
(11)
i.e. λf is generally larger than λb.
B. Frame Corruption Estimation
Section II-A presents a packet loss probability model for the
preceding packets sent from the source node when a RERR
message is received. Every time the source node receives a
RERR message, we estimate the packet loss probability of the
preceding packets using the proposed estimation model. For
the packets that are dropped during the route recovery process,

Page 4

TABLE I
SIMULATION PARAMETERS
Region
Number of nodes
500 m× 500 m
50
Random waypoint model:
node speed 0 ∼ 10 m/s,
pause time 120 s
5.5 Mbps
15 dBm
802.11b CSMA/CA
350 ms
Mobility model
PHY data rate
Transmission Power
MAC layer protocol
Playout deadline
we set their packet loss probabilities to 1. Given the estimated
packet loss probability of each video packet, we can estimate
the frame corruption probability by
?
where p(vi) is the packet loss probability of packet vi, p(fk) is
the frame corruption probability of frame fk, and {vi|vi∈ fk}
is the set of packets that contain information of frame fk.
We define a threshold pthresto determine whether a frame is
corrupted or not, i.e. if p(fk) ≥ pthres, we consider frame fkas
corrupted. We propose to use the frame corruption estimation
results to assist the reference frame selection in video coding.
That is, we do not use any estimated corrupted frame as a
reference frame in the motion compensation process in order
to mitigate error propagation.
p(fk) = 1 −
{vi|vi∈fk}
(1 − p(vi))
(12)
III. EXPERIMENTAL RESULTS
To investigate the effectiveness and robustness of our esti-
mation model, we simulated a two-path transport system over
a mobile ad-hoc network using the Qualnet simulator. The
network parameters chosen are shown in Table I. In this ad-
hoc network, 50 nodes are uniformly placed in a 500m×500m
region, where the connectivity of any two nodes is determined
by the network topology and the transmission power. The
movement of each node is characterized by a random waypoint
model [9] with parameters shown in Table I. We use IEEE
802.11b with 5.5 Mbps PHY transmission rate and CSMA/CA
basic access protocol. The values of IEEE 802.11b parameters
used for the packet loss probability model can be found in [8].
In the network layer, we implement the split multipath routing
(SMR) protocol [10] to generate two routes for multipath
transport. A pair of source and destination nodes is randomly
chosen to transmit video packets and packets are dropped if
they do not reach the destination by the playout deadline of
350 ms.
We use two error probabilities to measure the accuracy
of our estimation. We define a false alarm probability PFA
as the probability of detecting a corrupted frame when the
frame is actually correctly received and define a miss-detection
probability PMISS as the probability of detecting a correctly
received frame when the frame is actually corrupted.
We plot a receiver operating curve (ROC) [11] to represent
the possible values of PFA and PMISS under various pthres
in the range of [0, 1] in Fig. 3. This figure is generated
00.10.20.30.40.5
PFA
0.60.7 0.80.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
PMISS


Random Guess
Our Estimation
pthres=0.6
pthres=0.1
pthres=0
pthres=1
pthres=0.8
Worse
Better
Perfect Estimation
Fig. 3.
between PFAand PMISS.
ROC curve for pthres∈ [0,1]. Varying pthresprovides a trade-off
TABLE II
PFA, PMISSUNDER DIFFERENT TRANSMISSION POWERS WITH
pthres= 0.5
Transmission Power
15 dBm
14 dBm
13 dBm
12 dBm
11 dBm
PFA
0.008
0.010
0.012
0.014
0.016
PMISS
0.155
0.195
0.229
0.267
0.315
using the default network settings. In this ROC space, the
(0, 0) point represents perfect estimation and the diagonal
line denotes a completely random guess. The overall accuracy
of the estimation depends on how close the point is to the
lower left corner. In Fig. 3, we see that the ROC curve of
our estimation method is close to the lower left corner for the
pthres values in the range of [0.1, 0.6], where we achieve a
very low probability of a false classification while maintaining
a fairly low probability of missing a corrupted frame. Thus, by
choosing a pthresvalue in that range, our estimation method
yields fairly good performance.
Next, we examine the robustness of our estimation method
under different network settings. Here we choose pthres =
0.5. Table II presents PFAand PMISSvalues for transmission
power varied in the range of 11∼15 dBm. We see that the
false alarm probability is constantly low under these network
settings, which leads to negligible unnecessary reduction in
coding efficiency. Meanwhile, the miss-detection probability is
in the range of 0.16∼0.32, which indicates that our proposed
estimation method can detect most of the corrupted frames. We
notice that the performance of the estimation becomes worse
as the transmission power is reduced. This is because when
the transmission power decreases, the network connectivity
becomes worse. The loss probability of routing messages may
increase, which leads to more miss-detections.
Table III shows the impact of the number of nodes on
our estimation accuracy. As shown in Table III, the es-
timation accuracy does not have distinguishable difference

Page 5

TABLE III
PFAAND PMISSUNDER DIFFERENT NUMBER OF NODES WITH
pthres= 0.5
Number of Nodes
20
40
50
60
80
PFA
0.008
0.008
0.008
0.008
0.008
PMISS
0.144
0.164
0.155
0.169
0.164
under a different number of nodes, which means that the
estimation accuracy is insensitive to the number of nodes in
the network. Similarly, we study estimation accuracy under
different network sizes, transmission bitrates, and number of
retransmissions. In general, the false-alarm probability is lower
than 1% while the miss-detection probability is lower than
32%. Therefore, our estimation model is effective and robust
under various network settings.
Finally, we evaluate the delivered video quality while our
frame corruption estimation is used for video coding. We
incorporate our frame corruption estimation with the refer-
ence frame selection technique for multiple description video
coding with multipath transport (MPT) [7]. We refer to this
method as routing-aware multiple description coding (RA-
MDC). We compare our proposed method to single description
coding (SDC) and multiple description coding (MDC) with
MPT. We implement multiple state video coding (MSVC) [12]
as the MDC method and apply refined error concealment
method as proposed in [13]. For the three methods, we use
the same MPT strategy such that even and odd frames are
transported through two routes respectively.
In the simulation shown here, we use foreman sequence
(CIF) at 15 fps with 150 frames and the bitrate is 400 kbps.
We vary the transmission power from 15 dBm to 11 dBm to
achieve packet loss rates in the range of 2.2% ∼ 8.8%. For
each packet loss rate, we simulate 500 realizations and plot
the average PSNR in Fig. 4. We see that our RA-MDC method
achieves up to 1.44 dB gain in PSNR compared to MDC and
up to 2.16 dB gain compared to SDC, which shows that our
proposed RA-MDC can effectively mitigate error propagation
and improve the delivered video quality. More details on the
RA-MDC method can be found in [7].
IV. CONCLUSIONS
In this paper, we propose a frame corruption estimation
method for video coding over mobile ad hoc networks. We
notice that the routing mechanisms in most of the routing
protocols provides feedback concerning the network condi-
tions. This inspires us to utilize these feedback messages to
estimate the packet losses in the network and to adapt the video
coding accordingly. We build a packet loss probability model
from the network parameters and estimate the frame corruption
probability based on the estimation model and the routing
messages received at the network layer. We demonstrate that
our estimation method is effective and robust under various
network settings. In addition, we use the estimated results to
guide the reference frame selection for multiple description
23456789
30.5
31
31.5
32
32.5
33
33.5
34
34.5
Packet Loss Rate (%)
Average PSNR (dB)


SDC
MDC
RA−MDC
Fig. 4.
(CIF, 15fps) at 400 kbps. Transmission power varies from 15 dBm to 11 dBm
to achieve packet loss rate from 2.2% to 8.8%.
Performance under different packet loss rates for Foreman sequence
video coding and show that our method achieves up to 2.16
dB gains in PSNR.
REFERENCES
[1] R. Zhang, S. L. Regunathan, and K. Rose, “Video coding with optimal
inter/intra-mode switching for packet loss resilience,” IEEE Journal on
Selected Areas in Communications, vol. 18, no. 6, pp. 966 –976, Jun
2000.
[2] T. Stockhammer, D. Kontopodis, and T. Wieg, “Rate-distortion opti-
mization for JVT/H.26L video coding in packet loss environment,” in
International Packet Video Workshop, 2002.
[3] C. Zhu, Y.-K. Wang, M. M. Hannuksela, and H. Li, “Error resilient
video coding using redundant pictures,” IEEE Transactions on Circuits
and Systems for Video Technology, vol. 19, no. 1, pp. 3–14, Jan. 2009.
[4] Y.-K. Wang, C. Zhu, and H. Li, “Error resilient video coding using
flexible reference frames,” in Proceedings of SPIE, vol. 5960, no. 1.
SPIE, 2005.
[5] Y. Wang, A. R. Reibman, and S. Lin, “Multiple description coding for
video delivery,” Proceedings of the IEEE, vol. 93, no. 1, pp. 57–70, Jan.
2005.
[6] Y. Liao and J. D. Gibson, “Routing-aware multiple description video
coding over wireless ad-hoc networks using multiple paths,” in Interna-
tional Conference on Image Processing (ICIP), Sep. 2010.
[7] ——, “Routing-aware multiple description coding with multipath trans-
port for video delivered over mobile ad-hoc networks,” accepted by IEEE
Workshop on Multimedia Communications & Services, 2010.
[8] “IEEE Standard Part 11: Wireless LAN Medium Access Control (MAC)
and Physical Layer (PHY) Specifications,” IEEE Std 802.11-2007 (Re-
vision of IEEE Std 802.11-1999), pp. C1–1184, 2007.
[9] J. Broch, D. A. Maltz, D. B. Johnson, Y. C. Hu, and J. Jetcheva, “A
performance comparison of multi-hop wireless ad hoc network routing
protocols,” in Proceedings of the 4th annual ACM/IEEE international
conference on Mobile computing and networking, 1998, pp. 85–97.
[10] S. J. Lee and M. Gerla, “Split multipath routing with maximally
disjoint paths in ad hocnetworks,” in IEEE International Conference
on Communications, vol. 10, 2001, pp. 3201–3205.
[11] R. Yates and D. Goodman, Probability and Stochastic Processes: A
Friendly Introduction for Electrical and Computer Engineers.
Wiley & Sons, 2004.
[12] J. G. Apostolopoulos, “Reliable video communication over lossy packet
networks using multiple state encoding and path diversity,” in SPIE
Proceedings on Visual Communications and Image Processing, vol.
4310, no. 1, 2001, pp. 392–409.
[13] Y. Liao and J. D. Gibson, “Refined error concealment for multiple
state video coding over ad hoc networks,” in Proceedings of the 42nd
Asilomar Conference on Signals, Systems and Computers, Oct. 2008,
pp. 2243 –2247.
John