Journal of Intelligent Learning Systems and Applications, 2010, 2: 19-27
doi:10.4236/jilsa.2010.21003 Published Online February 2010 (
Copyright © 2010 SciRes JILSA
Continuous Arabic Sign Language Recognition in
User Dependent Mode
K. Assaleh1, T. Shanableh2, M. Fanaswala1, F. Amin1, H. Bajaj1
1Department of Electrical Engineering, American University of Sharjah, Sharjah, UAE; 2Department of Computer Science and En-
gineering; American University of Sharjah, Sharjah, UAE.
Received October 4, 2009; accepted January 13, 2010.
Arabic Sign Language recognition is an emerging field of research. Previous attempts at automatic vision-based recog-
nition of Arabic Sign Language mainly focused on finger spelling and recognizing isolated gestures. In this paper we
report the first continuous Arabic Sign Language by building on existing research in feature extraction and pattern
recognition. The development of the presented work required collecting a continuous Arabic Sign Language database
which we designed and recorded in cooperation with a sign language expert. We intend to make the collected database
available for the research community. Our system which we based on spatio-temporal feature extraction and hidden
Markov models has resulted in an average word recognition rate of 94%, keeping in the mind the use of a high perplex-
ity vocabulary and unrestrictive grammar. We compare our proposed work against existing sign language techniques
based on accumulated image difference and motion estimation. The experimental results section shows that the pro-
posed work outperforms existing solutions in terms of recognition accuracy.
Keywords: Pattern Recognition, Motion Analysis, Image/ Video Processing and Sign Language
1. Introduction
The growing popularity of vision-based systems has led
to a revolution in gesture recognition technology. Vi-
sion-based gesture recognition systems are primed for
applications such as virtual reality, multimedia gaming
and hands-free interaction with computers. Another
popular application is sign language recognition, which
is the focus of this paper.
There are two main directions in sign language recog-
nition. Glove-based systems use motion sensors to cap-
ture gesture data [1–3]. While this data is more attractive
to work with, the gloves are expensive and cumbersome
devices which detract from the naturalness of the hu-
man-computer interaction. Vision-based systems, on the
other hand, provide a more natural environment within
which to capture the gesture data. The flipside of this
method is that working with images requires intelligent
feature extraction techniques in addition to image proc-
essing techniques like segmentation which may add to
the computational complexity of the system.
Note that while respectable results have been obtained
in the domains of isolated gesture recognition and finger
spelling, research on continuous Arabic sign language
recognition is non-existent.
The work in [4] developed a recognition system for
ArSL alphabets using a collection of Adaptive Neuro-
Fuzzy Inference Systems (ANFIS), a form of supervised
learning. They used images of bare hands instead of col-
ored gloves to allow the user to interact with the system
conveniently. The used feature set comprised lengths of
vectors that were selected to span the fingertips’ region
and training was accomplished by use of a hybrid learning
algorithm, resulting in a recognition accuracy of 93.55%.
Likewise [5] reported classification results of Arabic sign
language alphabets using Polynomial classifiers. The
work reported superior results when compared with their
previous work using ANFIS on the same dataset and fea-
ture extraction techniques. Marked advantages of poly-
nomial classifiers include their computational scalability,
less storage requirements and absence of the need for it-
erative training. This work required the participants to
wear gloves with colored tips while performing the ges-
tures to simplify the image segmentation stage. They ex-
tracted 30 features involving the relative position and
orientation of the fingertips with respect to the wrist and
each other. The resulting system achieved 98.4% recogni-
tion accuracy on training data and 93.41% on test data.
Continuous Arabic Sign Language Recognition in User Dependent Mode
Sign language recognition of words/gestures as op-
posed to alphabets depends on analyzing a sequence of
still images with temporal dependencies. Hence HMMs
are a natural choice for model training and recognition as
reported in [6]. Nonetheless, the work in [7] presented an
alternative technique for feature extraction of sequential
data. Working with isolated ArSL gestures, they elimi-
nate the temporal dependency of data by accumulating
successive prediction errors into one image that repre-
sents the motion information. This removal of temporal
dependency allows for simple classification methods,
with less computational and storage requirements. Ex-
perimental results using k-Nearest Neighbors and Bayes-
ian classifiers resulted in 97 to 100% isolated gesture
recognition. Variations of the work in [7] include the use
of block-based motion estimation in the feature extrac-
tion process. The resultant motion vectors are used to
represent the intensity and directionally of the gestures’
motion as reported in [8].
Other sign languages such as American Sign Language
have been researched and documented more thoroughly.
A common approach in ASLR (American Sign Language
Recognition) of continuous gestures is to use Hidden
Markov Models as classifier models. Hidden Markov
Models are an ideal choice because they allow modeling
of the temporal evolution of the gesture. In part, the suc-
cess of HMMs in speech recognition has made it an ob-
vious choice for gesture recognition. Research by Starner
and Pentland [9] uses HMMs to recognize continuous
sentences in American Sign Language, achieving a word
accuracy of 99.2%. Users were required to wear colored
gloves and an 8-element feature set, comprising hands’
positions, angle of axis of least inertia, and eccentricity
of bounding ellipse, was extracted. Lastly, linguistic
rules and grammar were used to reduce the number of
Another research study by Starner and Pentland [10]
dealt with developing a Real-time ASLR system using a
camera to detect bare hands and recognize continuous
sentence-level sign language. Experimentation involved
two systems: first, using a desk mounted camera to ac-
quire video, that attained 92% recognition and second,
mounting the camera in the user’s cap, which achieved
an accuracy of 98%. This work was based on limited
vocabulary data, employing a 40-word lexicon. The au-
thors do not present sentence recognition rates for com-
parison. Only word recognition and accuracy rates are
This paper is organized as follows. Section 2 describes
the Arabic sign language database constructed and used
in the work. The methodology followed is enumerated in
Section 3. The classification tool used is discussed in
Section 4. The results are discussed in Section 5. Con-
cluding remarks along with a primer on future work in
presented in Section 6.
2. The Dataset
Arabic Sign Language is the language of choice amongst
the hearing and speech impaired community in the Mid-
dle East and most Arabic speaking countries. This work
involves two different databases; one for isolated gesture
recognition and another for continuous sentence recogni-
tion. Both datasets are collected in collaboration with
Sharjah City for Humanitarian Services (SCHS) [11], no
restriction on clothing or background was imposed. The
first database was compiled for isolated gesture recogni-
tion as reported in [7]. The dataset consists of 3 signers
acting 23 gestures. Each signer was asked to repeat each
gesture a total of 50 times over 3 different sessions re-
sulting in a total of 150 repetitions of each gesture. The
gestures are chosen from the greeting section of the Ara-
bic sign language.
The second database is of a relatively high perplexity
consisting of an 80-word lexicon from which 40 sen-
tences were created. No restrictions are imposed on
grammar or sentence length. The sentences and words
pertain to common situations in which handicapped peo-
ple might find themselves in. The dataset itself consists
of 19 repetitions of each of the 40 sentences performed
by only one user. The frame rate was set to 25Hz with a
spatial resolution of 720×528. The list of sentences is
given in Table 1. Note that this database is the first fully
labeled and segmented dataset for continuous Arabic
Sign Language. The entire database can be made avail-
able on request.
A required step in all supervised learning problems is
the labeling stage where the classes are explicitly marked
for the classifier training stage. For continuous sentence
recognition, not only do the sentences have to be labeled
but the individual boundaries of the gestures that make
up that sentence have to be explicitly demarcated. This is
a time-consuming and repetitive task. Conventionally, a
portion of the data is labeled and used as ‘bootstrap’ data
for the classifier which can then learn the remaining
boundaries. For the purposes of creating a usable data-
base, a segmented and fully labeled dataset was created
in the Georgia Tech Gesture Recognition toolbox (GT2K)
format [12]. The output of this stage is a single master
label file (MLF) that can be used with the GT2K and
HTK Toolkits.
3. Feature Extraction
In this section we introduce a feature extraction tech-
nique suitable for continuous signing. We also examine
some of the existing techniques and adapt them to our
application for comparison reasons.
Copyright © 2010 SciRes JILSA
Continuous Arabic Sign Language Recognition in User Dependent Mode21
Table 1. Table type styles list of Arabic sentences with Eng-
lish translation used in the recognition system
NO. Arabic Sentence & English translation
1 مﺪﻘﻟا ةﺮآ يدﺎﻧ ﻰﻟا ﺖﺒهذ
I went to the soccer club
2 تارﺎﻴﺴﻟا قﺎﺒﺳ ﺐﺣا ﺎﻧا
I love car racing
3 ﺔﻨﻴﻤﺛ ةﺮآ ﺖﻳﺮﺘﺷا
I bought an expensive ball
4 مﺪﻗ ةﺮآ ةارﺎﺒﻣ يﺪﻨﻋ ﺖﺒﺴﻟا مﻮﻳ
On Saturday I have a soccer match
5 مﺪ ةﺮآ ﺐﻌﻠﻣ يدﺎﻨﻟا ﻲﻓ
There is a soccer field in the club
6 تﺎﺟارد قﺎﺒﺳ كﺎﻨه نﻮﻜﻴﺳ اﺪﻏ
There will be a bike racing tomorrow
7 ﺐﻌﻠﻤﻟا ﻲﻓ ةﺪﻳﺪﺟ ةﺮآ تﺪﺟو
I found a new ball in the field
8 ؟ﻚﻴﺧا ﺮﻤﻋ ﻢآ
How old is your brother?
9 ﺎﺘﻨﺑ ﻲﻣا تﺪﻟو مﻮﻴﻟا
My mom had a baby girl today
10 ﺎﻌﻴﺿر لاﺰﻳ ﻲﺧا
My brother is still breast feeding
11 ﺎﻨﺘﻴﺑ ﻲﻓ يﺪﺟ نا
My grandfather is at our home
12 ﺔﺼﻴﺧر ةﺮآ ﻲﻨﺑا ىﺮﺘﺷا
My kid bought an inexpensive ball
13 ﺎﺑﺎﺘآ ﻲﺘﺧا تأﺮﻗ
My sister read a book
14 حﺎﺒﺼﻟا ﻲﻓ قﻮﺴﻟا ﻰﻟا ﻲﻣا ﺖﺒهذ
My mother went to the market this morning
15 ؟ﺖﻴﺒﻟا ﻲﻓ كﻮﺧا ﻞه
Is your brother home?
16 ﺮﻴﺒآ ﻲﻤﻋ ﺖﻴﺑ
My brother’s house is big
17 ﺮﻬ ﺪﻌﺑ ﻲﺧا جوﺰﺘﻴﺳ
In one month my brother will get married
18 ﻦﻳﺮﻬﺷ ﺪﻌﺑ ﻲﺧا ﻖﻠﻄﻴﺳ
In two months my brother will get divorced
19 ؟ﻚﻘﻳﺪﺻ ﻞﻤﻌ ﻦﻳا
Where does your friend work?
20 ﺔﻠﺳ ةﺮآ ﺐﻌﻠﻳ ﻲﺧا
My brother plays basketball
21 ﻦﻳﻮﺧأ يﺪﻨﻋ
I have two brothers
22 ؟ﻚﻴﺑا ﻢﺳا ﺎﻣ
What is your father’s name?
23 ﺲﻣﻻا ﻲﻓ ﺎﻀﻳﺮﻣ يﺪﺟ نﺎآ
Yesterday my grandfather was sick
24 ﺲﻣﻻا ﻲﻓ ﻲﺑا تﺎﻣ
Yesterday my father died
25 رﺔﻠﻴﻤﺟ ﺎﺘﻨﺑ ﺖﻳأ
I saw a beautiful girl
26 ﻞﻳﻮ ﻲﻘﻳﺪﺻ
My friend is tall
27 مﻮﻨﻟا ﻞﺒﻗ ﻞآّا ﺎﻧا
I do not eat close to bedtime
28 ﻢﻌﻄﻤﻟا ﻲﻓ اﺬﻳﺬﻟ ﺎﻣﺎﻌﻃ ﺖﻠآا
I ate delicious food at the restaurant
29 ءﺎﻤﻟا بﺮﺷ ﺐﺣا ﺎﻧا
I like drinking water
30 بﺮ ﺐﺣا ﺎﻧاءﺎﺴﻤﻟا ﻲﻓ ﺐﻴﻠﺤﻟا
I like drinking milk in the evening
31 جﺎﺟﺪﻟا ﻦﻣ ﺮﺜآا ﻢﺤﻠﻟا ﻞآا ﺐﺣا ﺎﻧا
I like eating meat more than chicken
32 ﺮﻴﺼﻋ ﻊﻣ ﻨﺒﺟ ﺖﻠآا
I ate cheese and drank juice
Next Sunday the price of milk will go up
34 ﺲﻣﻻا حﺎﺒ ﺎﻧﻮﺘﻳز ﺖﻠآأ
Yesterday morning I ate olives
35 ﺮﻬﺷ ﺪﻌﺑ ةﺪﻳﺪﺟ ةرﺎﻴﺳ يﺮﺘﺷﺎﺳ
I will buy a new car in a month
36 ﺢﺒﺼﻟا ﻲﻠﺼﻴﻟ ﺄﺿﻮﺗ ﻮه
He washed for morning prayer
37 ةﺮﺷﺎﻌﻟا ﺔﻋﺎﺴﻟا ﺪﻨﻋ ﺔﻌﻤﺠﻟا ةﻼﺻ ﻰﻟا ﺖﺒهذ
I went to Friday prayer at 10:00 o’clock
38 ﺎﺘﻴﺑ تﺪهﺎﺷ اﺮﻴﺒآ زﺎﻔﻠﺘﻟﺎﺑ
I saw a big house on TV
39 ةﺮﺷﺎﻌﻟا ﺔﻋﺎﺴﻟا ﺪﻨﻋ ﺖﻤﻧ ﺲﻣﻻا ﻲﻓ
Yesterday I went to sleep at 10:00 o’clock
40 ﻲﺗرﺎﻴﺴﺑ حﺎﺒﺼﻟا ﻲﻓ ﻞﻤﻌﻟا ﻰﻟا ﺖﺒهذ
I went to work this morning in my car
3.1 Proposed Feature Extraction
The most crucial stage of any recognition task is the se-
lection of good features. Features that are representative
of the individual gestures are desired. Shanableh et al. [7]
demonstrated in their earlier work on isolated gesture
recognition that the two-tier spatial-temporal feature ex-
traction scheme results in a high word recognition rate
close to 98%. Similar extraction techniques are used in
our continuous recognition solution.
First, to represent the motion that takes place as the
expert signs a given sentence, pixel-based differences of
successive images are computed.
It can be justified that the difference between two im-
ages of similar background results in an image that only
preserves the motion between the two images. These
image differences are then converted into binary images
by applying an appropriate threshold. A threshold value
of µ+xσ is used where µ is the mean pixel intensity of the
image difference, σ is the corresponding standard devia-
tion and x is a weighting parameter which was empiri-
cally determined based on subjective evaluation whose
criteria was to retain enough motion information and
discarding the noisy data.
Figure 1 shows an example sentence with thresholded
image differences. Notice that the example sentence is
temporally downsampled for illustration purposes.
Next, a frequency domain transformation such as the
Discrete Cosine Transform (DCT) is performed on the
binary image differences.
The 2-D Discrete Cosine Transformation (DCT) given
by [13]:
where NxM are the dimensions of the input image ‘f’ and
F(u,v) is the DCT coefficient at row u and column v of
the DCT matrix. C(u) is a normalization factor equal to
1for u=0 and 1 otherwise.
ﺐﻴﻠﺤﻟا ﺮﻌﺳ ﻊﻔﺗﺮﻴﺳ مدﺎﻘﻟا ﺪﺣﻻا مﻮﻳ
Copyright © 2010 SciRes JILSA
Continuous Arabic Sign Language Recognition in User Dependent Mode
(a) An image sequence denoting the sentence ‘I do not eat close to bed
(b) Thresholded image differences of the image sequence in part (a)
Figure 1. An example sentence and its motion representa-
Figure 2. Discrete cosine transform coefficients of a thresh-
olded image difference
In Figure 2, it is apparent that the DCT transformation
of a thresholded image difference results in energy com-
paction where most of the image information is repre-
sented in the top left corner of the transformed image.
Subsequently, zig-zag scanning is used to select only a
required number of frequency coefficients. This process
is also known as zonal coding. The number of coeffi-
cients to retain or the DCT cutoff is elaborated upon in
the experimental results section.
These coefficients obtained in a zig-zag manner make
up the feature vector that is used in training the classifier.
3.2 Adapted Feature Extraction Solutions
For completeness, we compare our feature extraction
solution to existing work on Arabic sign language recog-
nition. Noteworthy are the Accumulated Differences
(ADs) and Motion Estimation (ME) approaches to fea-
ture extraction as reported in [8,16]. In this section we
provide a brief review of each of mentioned solutions
and explain who it can be adapted to our problem of con-
tinuous Arabic sign language recognition
3.2.1 Accumulated Differences Solution
The motion information of an isolated sign gesture can
be computed from the temporal domain of its image se-
quence through successive image differencing. Let
denote image index j of the ith repetition of a gesture at
index g. The Accumulated Differences (ADs) image can
be computed by:
jgjjg IIAD (2)
where n is the total number of images in the ith repetition
of a sign at index g. j
is a binary threshold function of
the jth frame.
Note that the ADs solution cannot be directly applied
to continuous sentences (as opposed to isolated sign ges-
tures). This is so because the gesture boundaries in a
sentence are unknown, thus one solution is to use an
overlapping sliding window approach in which a given
number of video frame differences are accumulated into
one image regardless of gesture boundaries. The window
Copyright © 2010 SciRes JILSA
Continuous Arabic Sign Language Recognition in User Dependent Mode23
is shift by one video frame at a time. In the experimental
results section we experiment with various window sizes.
Examples of such accumulated differences are shown in
Figure 3 with a window size of 8 video frames. Notice
that the ADs capture the frame difference between the
current and previous video frames and it also accumulates
the frame differences from the current window as well.
Once the ADs image is computed it is then trans-
formed into the DCT domain as described previously.
The DCT coefficients are Zonal coded to generate the
feature vector.
3.2.2 Motion Estimation solution
The motion of a video-based sign gesture can also be
tracked by means of Motion Estimation (ME). One well
known example is the block-based ME in which the input
video frames are divided into non-overlapping blocks of
pixels. For each block of pixels, the motion estimation
process will search through the previous video frame for
the “best match” area within a given search range. The
displacement, in terms of pixels, between the current
block and its best match area in the previous video frame
is represented by a motion vector
Formally, let C denote a block in the current video
frame with bxb pixels at coordinates (m,n). Assuming
that the maximum motion displacement is w pixel per
video frame then the the motion estimation process will
find the best match area P within the (b+2 w)(b+2w) dis-
tinct overlapping bxb blocks of the previous video frame.
An area in the previous video frame that minimizes a
certain distortion measure is selected as the best match
area. A common distortion measure is the mean absolute
difference given by:
 
nynxmnm wyxwPC
11 ,,
where refer to the spatial displacement in be-
tween the pixel coordinates of C and the matching area in
the previous image. Other distortion measures can be
used such as mean squared error, cross correlation func-
tions and so forth. Further details on motion estimation
can be found in [17] and references within.
yx  ,
The motion vectors can then be used to represent the
motion the occurred between two video frames. These
vectors are used instead of the thresholded frame differ-
ences. In [8] it was proposed to rearrange the x and y
components of the motion vectors into two intensity im-
ages. The two images are then concatenated to generate
one representation of the motion that occurred between
two video frames. Again, once the concatenated image is
computed it is then transformed into the DCT domain as
described previously. The DCT coefficients are Zonal
coded to generate the feature vector.
(a) Part of the gesture “Club”
(b) Part of the gesture “Soccer”
Figure 3. Example accumulated differences images using an
overlapping sliding window of size 8
4. Classification
For conventional data, naïve Bayes classification pro-
vides the upper bound for the best classification rates.
Since sign language varies in both spatial and temporal
domains, the extracted feature vectors are sequential in
nature and hence simple classifiers might not suffice.
There are two main approaches to dealing with sequen-
tial data. The first method aims to combine the sequential
feature vectors using a suitable operation into a single
feature vector. A detailed account of such procedures is
outlined in [14]. One such method involves concatenat-
ing sequential feature vectors using a sliding window of
optimal length to create a single feature vector. Subse-
quently, classical supervised learning techniques such as
maximum-likelihood estimation (MLE), linear discrimi-
nants or neural networks can be used. The second ap-
proach makes explicit use of classifiers that can deal with
sequential data without concatenation or accumulation,
such an approach is used in this paper. While the field of
gesture recognition is relatively young, the related field
of speech recognition is well established and documented.
Hidden Markov Models are the classifier of choice for
continuous speech recognition and lend themselves
suitably for continuous sign language recognition too. As
mentioned in [15], a HMM is a finite-state automaton
characterized by stochastic transitions in which the se-
quence of states is a Markov chain. Each output of an
HMM corresponds to a probability density function.
Such a generative model can be used to represent sign
language units (words, sub-words etc).
The implementation of an HMM framework was car-
ried out using the Georgia Tech Gesture Recognition
Toolkit (GT2K) which serves as a wrapper for the more
general Hidden Markov Model Toolkit (HTK). The
GT2K version used was a UNIX based package. HTK is
the de-facto standard in speech recognition application
Copyright © 2010 SciRes JILSA
Continuous Arabic Sign Language Recognition in User Dependent Mode
using HMM’s.
5. Experimental Results
A logical step in proceeding from isolated gesture recog-
nition would be connected gesture recognition. This can
be simulated by concatenating individual gestures into
artificial sentences. Intuitively, one would expect better
results for connected gesture recognition as opposed to
continuous gesturing. This is because concatenated ges-
tures do not suffer from the altered spatial gesturing that
occurs as gestures are signed continuously without
pauses. The first database consisting of isolated gestures
was used to create concatenated sentences of varying
length. These sentences were created without any con-
sideration of whether the constructed sentence held any
meaning or grammatical structure. This concatenated
data was divided into a training set and a testing set
comprised of 70% and 30% of the total data respectively.
The GT2K Toolkit was then used to perform recognition
based on individual words as the basic unit of Arabic
sign language. While concatenation is not the aim of this
work, the results obtained provide a valuable benchmark
for subsequent experiments with continuous sentence
signing. An average of 96% sentence recognition and
98% word recognition was obtained on the concatenated
testing dataset. The word recognition rate is comparable
to previous work in ArSL [7] using similar feature ex-
traction schemes. It would be prudent to note that due to
the nature of concatenation, the boundary between ges-
tures is prominent and this might account for the high
sentence recognition rate.
The second database was then used to perform con-
tinuous sentence recognition. This was also performed
with the help of GT2K. This data is also divided into a
training (70%) and testing set (30%). An average of 75%
sentence recognition and 94% word recognition was ob-
tained on the testing set. A detailed analysis of the vari-
ous associated parameters is given in the following dis-
There are several parameters that affect the recognition
rates in continuous sign language recognition. Namely,
the sections below discuss the effect of varying the
number of hidden states, number of guassian mixtures,
length of feature vectors and the threshold used for bi-
narizing the image differences. Unless otherwise stated,
the length of the feature vectors used throughout the ex-
periments is 100 DCT coefficients.
The following results are based on the word and sen-
tence recognition rates. The former is computed through
the following equation:
 1 (4)
Figure 4. The effect of number of states on the sentence
recognition rate (3 Gaussian mixtures are used)
Where D is the number of deletions, S is the number of
substitutions, I is the number of insertions, and N is the
total number of words. On the other hand, sentence rec-
ognition rate is the ratio of the correctly recognized sen-
tences to the total number of sentences. Correctness in
this case entails correct recognition of all the words con-
stituting the sentence without any insertions, substitu-
tions, or deletions.
5.1 Number of Hidden States
In Figure 4, the effect of increasing the number of hidden
states in the HMM topology on sentence and word rec-
ognition rates is examined.
An increasing trend with the recognition rates is ob-
served as the number of states is increased to a certain
number and then the classification rates saturates with a
subsequent drop. As the number of states in the Hidden
Markov Model is increased, we are in effect increasing
the degrees of freedom allowed in modeling the data.
Working with video data sampled at 25 frames per sec-
ond, the classification rate increased to a maximum at
nine states.
The saturation in recognition accuracy is attributed to
the fact that certain gestures do not extend for a long time
duration and are only represented by few frame differ-
ences. The increase in number of states only serves to
increase computation time while adding redundant data
that does not contribute to the classification rate.
5.2 Length of the Feature Vector
Figure 5 shows the recognition rates for increasing the
number of DCT coefficients within the feature vector.
It is expected that the increase in feature vector size be
accompanied by a corresponding increase in recognition
rates. This is due to the fact that each DCT coefficient is
Copyright © 2010 SciRes JILSA
Continuous Arabic Sign Language Recognition in User Dependent Mode25
uncorrelated with other coefficients and hence no redun-
dant information is present in increasing coefficients.
Experimental results shown in Figure 5 show a general
increase in recognition rates as the number of DCT coef-
ficients is increased.
The trend shows that any increase in recognition ac-
curacy beyond 100 coefficients is only slightly signifi-
cant. The increase in computation time is however a lim-
iting factor in increasing the length of the feature vector
5.3 Number of Gaussian Mixtures
The effect of increasing the number of Gaussian mixtures
is shown in Figures 6 and 7.
Gaussian mixtures are used to model the emission
probability densities of each state of a continuous Hidden
Markov Model.
For multi-dimensional data like the long feature vec-
tors used in this work it is desirable to have a number of
Gaussian mixtures so that any emission pdf can be effec-
tively fit. Increasing the number of Gaussian mixture
shows substantial improvement in recognition rates. The
results depict a general increase in recognition rates as
the mixtures are increased. However, the authors feel that
the limitation in collecting large amounts of data does not
allow the use of more mixtures.
5.4 Choice of Threshold
In the feature extraction process, the image differences
are thresholded into binary images based on a threshold
of µ+xσ, where µ is the mean pixel intensity of the image
difference, σ is the corresponding standard deviation and
x is a weight parameter. Results are shown in Figure 8
and 9 for different values of the weight parameter.
The recognition accuracy peaks at a weighting pa-
Figure 5. The effect of length of the feature vector on sen-
tence recognition rates (3 Gaussian mixtures are used with
a HMM topology using 9 states)
Figure 6. The effect of the number of gaussian mixtures on
the sentence recognition rate (HMM topology using 9
Figure 7. The effect of the number of gaussian mixtures on
the word recognition rate (HMM topology using 9 states)
rameter value between 1 and 1.25. A subjective com-
parison of the thresholded image differences shows that
these parameter values retain most of the motion infor-
mation whilst discarding spurious information such as
small stray shifts in clothing, illumination and the like.
Lastly as mentioned in Section 3.2 above, we compare
our solution against existing work on Arabic sign lan-
guage recognition. Namely we consider both the Ads and
ME approaches to feature extraction. Table 2 summaries
the sentence and word recognition rates using various
feature extraction solutions. The experimental parameters
are similar to those used in Figure 9.
The recognition results presented in the table indicate
that the proposed solution provides the highest sentence
Copyright © 2010 SciRes JILSA
Continuous Arabic Sign Language Recognition in User Dependent Mode
Figure 8. The effect of the weighting factor on the sentence
recognition rate (3 Gaussian mixtures are used with a
HMM topology using 9 states)
Figure 9. Effect of the weighting factor on the word recog-
nition rate (3 Gaussian mixtures are used with a HMM to-
pology using 9 states)
and word recognition rates. The ADs with the overlap-
ping sliding window approach was not advantageous.
Intuitively the ADs image puts the difference between 2
video frames into context by accumulating future frame
differences to it. However in HMMs temporal informa-
tion is preserved and therefore extracting feature vectors
from video frame differences without accumulating them
will suffice. It is also worth mentioning that increasing
the window size beyond 10 frames did not further en-
hance the recognition rate.
In the ME approach, the image block size and the
search range are set to 8x8 pixels which is a typicalset-
ting in video processing. The resultant recognition rates
are comparable to the ADs approach. Note that ME tech-
-niques do not entirely capture the true motion of a video
sequence. For instance with block-based search tech-
Table 2. Comparisons with existing feature extraction solu-
Feature extraction approach
recognition rate
ADs with an overlapping
Sliding window of size 4. 64.1% 91.0%
ADs with an overlapping
Sliding window of size 7. 65.2% 90.6%
ADs with an overlapping
Sliding window of size 10. 68% 93.71%
Motion estimation 67.9% 92.9%
Proposed solution 73.3% 94.39%
niques object rotations are not captured as good as trans-
lational motion. Therefore the recognition results are
inferior to the proposed solution.
6. Conclusions and Future Work
The work outlined in this paper is an important step in
this domain as it represents the first attempt to recognize
continuous Arabic sign language. The work entailed
compiling the first fully labeled and segmented dataset
for continuous Arabic Sign Language which we intend to
make public for the research community. The average
sentence recognition rate of 75% and word recognition-
rate of 94% are obtained using a natural vision-based
system with no restrictions on signing such as the use of
gloves. Furthermore, no grammar is imposed on the sen-
tence structure which makes the recognition task more
challenging. The use of grammatical structure can sig-
nificantly improve the recognition rate by alleviating
some types of substitution and insertion errors. In the
course of training, the dataset was plagued by an unusu-
ally large occurrence of insertion errors. This problem
was mitigated by applying a detrimental weight for every
insertion error which was incorporated into the training
stage. As a final comment, the perplexity of the dataset is
large compared to other work in related fields. Future
work in this area aiming to secure higher recognition
rates might require a sub-gesture based recognition sys-
tem. Such a system would also serve to alleviate the mo-
tion-epenthesis effect which is similar to the co-articu-
lation effect in speech recognition.
Finally, the frequency domain transform coefficients
used as features perform well in concatenated gesture
recognition. The average word recognition rate is also
sufficiently high with an average of 94%. The authors
feel that geometric features might be used in addition to
the existing feature to create an optimum feature set.
7. Acknowledgements
The authors acknowledge Emirates Foundation for their
support. They also acknowledge the Sharjah City for
Humanitarian Services for availing Mr. Salah Odeh to
Copyright © 2010 SciRes JILSA
Continuous Arabic Sign Language Recognition in User Dependent Mode
Copyright © 2010 SciRes JILSA
help in the construction of the databases.
[1] J. S. Kim, W. Jang, and Z. Bien, “A dynamic gesture rec-
ognition system for the Korean sign language (KSL),”
IEEE Transations on Systerms, Man, Cybern. Part B, Vol.
26, pp. 354–359, April 1996.
[2] H. Matsuo, S. Igi, S. Lu, Y. Nagashima, Y. Takata, and T.
Teshima, “The recognition algorithm with noncontact for
Japanese sign language using morphological analysis,”
Proceedings of Gesture Workshop, pp. 273–284, 1997.
[3] S. S. Fels and G. E. Hinton, “Glove-talk: A neural network
interface between a data-glove and a speech synthesizer,”
IEEE Transactions on Neural Networks, Vol. 4, pp. 2–8,
January 1993.
[4] O. Al-Jarrah and A. Halawani, “Recognition of gestures in
Arabic sign language using neuro-fuzzy systems,” Artifi-
cial Intelligence, Vol. 133, No. 1–2, pp. 117–138, De-
cember, 2001.
[5] K. Assaleh and M. Al-Rousan, “Recognition of Arabic
Sign language alphabet using polynomial classifiers,”
EURASIP Journal on Applied Signal Processing, Vol.
2005, No. 13, pp. 2136–2146, 2005.
[6] M. AL-Rousan, K. Assaleh, and A. Tala'a, “Video-based
signer-independent Arabic sign language recognition us-
ing hidden Markov models,” Applied Soft Computing,
Vol. 9, No. 3, June 2009.
[7] T. Shanableh, K. Assaleh, and M. Al-Rousan, “Spa-
tio-temporal feature-extraction techniques for isolated
gesture recognition in Arabic sign language,” IEEE
Transactions on Systems, Man and Cybernetics Part B,
Vol. 37, No. 3, pp. 641–650, 2007.
[8] T. Shanableh and K. Assaleh, “Telescopic vector compo-
sition and polar accumulated motion residuals for feature
extraction in Arabic sign language recognition,” EURA-
SIP Journal on Image and Video Processing, vol. 2007, Ar-
ticle ID 87929, 10 pages, 2007. doi:10.1155/2007/ 87929.
[9] T. Starner, J. Weaver, and A. Pentland. “Real-time
American sign language recognition using desk and
wearable computer based video,” IEEE Transations on
Pattern Analysis and Machine Intelligence, Vol. 20, No.
12, pp. 1371–1375, 1998.
[10] T. Starner, J. Weaver and A. Pentland. “A wearable com-
puter-based American sign language recogniser,” in per-
sonal and ubiquitous computing, 1997.
[11] Sharjah City for Humanitarian Services (SCHS), website:
[12] T. Westeyn, H. Brashear and T. Starner, “Georgia tech
gesture toolkit: Supporting experiments in gesture recog-
nition,” Proceedings of International Conference on Per-
ceptive and Multimodal User Interfaces, Vancouver, B.C.,
November 2003.
[13] K. R. Rao and P. Yip, “Discrete cosine transform: Algo-
rithms, advantages, applications,” Academic Press, ISBN
012580203X, August 1990.
[14] T. Dietterich, “Machine learning for sequential data: A
review,” In T. Caelli (Ed.) Structural, Syntactic, and Sta-
tistical Pattern Recognition; Lecture Notes in Computer
Science, Vol. 2396, pp. 15–30, Springer-Verlag, 2002.
[15] L. Rabiner, B. Juang, “Fundamentals of speech recogni-
tion,” New Jersey: Prentice-Hall Inc., 1993.
[16] F. S. Chen, C. M. Fu, and C. L. Huang, “Hand gesture
recognition using a real-time tracking method and hidden
Markov models,” Image and Vision Computing, Vol. 21,
No. 8, pp. 745–758, 2003.
[17] M. Ghanbari, “Video coding: An introduction to standard
codecs,” Second Edition, The Institution of Engineering
and Technology, 2003.