International Journal of Machine Learning and Computing, Vol. 9, No.

6, December 2019

Hand Gesture Detection Using Neural Networks

Algorithms
N. Alnaim and M. Abbod

detection, and tracking.

Abstract—Human gesture is a form of body language usually Several gesture-based techniques [1]-[3] have been
used as a mean of communication and is very critical in developed to support human computer interaction. According
human-robot interactions. Vision-based gesture recognition to Pradipa and Kavitha [1], the main aim of gesture
methods to detect hand motion are vital to support such
recognition is developing a system that can detect human
interactions. Hand gesture recognition enables a convenient
and usable interface between devices and users. In this paper, actions to be used for extracting meaningful information for
an approach is presented for hand gesture recognition based on device control.
image processing methods, namely Wavelets Transform (WT), Hand motion can be detected using any type of camera
Empirical Mode Decomposition (EMD), besides Artificial supporting reasonable image quality. 2-D cameras such as
Intelligence classifier which is Artificial Neural Networks (ANN) Microsoft's Kinect, Intel's RealSense Technology and
and Convolutional Neural Network (CNN). The methods are
Apple's iPhone high quality camera can easily be used in
evaluated based on many factors including execution time,
accuracy, sensitivity, specificity, positive and negative detecting most hand motions on a constant surface. Video
predictive value, likelihood, receiver operating characteristic content (composed of several images) is processed in several
(ROC), area under roc curve (AUC) and root mean Square. phases including data input, pre-processing, image
Results indicate that WT have less execution time than EMD segmentation, feature extraction, and classification.
and CNN. In addition, CNN is more effective in extracting The objective of this study is to investigate the best method
distinct features and classifying data accurately compared to
available to extract features. Deep learning techniques are
EMD and WT.
evaluated including WT, EMD, and CNN comparing their
Index Terms—Artificial neural networks, convolutional classification accuracy. A hand gesture recognition system
neural network, empirical mode decomposition, hand gesture was developed based on various image processing methods.
recognition, wavelet transform. The performance of hand gesture recognition methods was
evaluated using various metrics including: execution time,
accuracy, sensitivity, specificity, positive predictive value,
I. INTRODUCTION negative predictive value, positive likelihood, negative
Currently, direct contact is the dominant form of likelihood, receiver operating characteristic, area under roc
interaction between the user and the machine. The interacting curve and root mean square. ANN was used to classify the
channel is based on devices such as the mouse, the keyboard, gestures using the features extracted as inputs. Multiple
the remote control, touch screen, and other direct contact training sessions were performed applying filters. A CNN
methods. Human to human interaction is achieved through deep learning tool was used to minimize previous stages.
more natural and intuitive noncontact methods, such as sound The rest of the paper is structured as follows. A literature
and physical movements. The flexibility and efficiency of review of hand gesture recognition methods and techniques is
noncontact interaction methods has led many researchers to provided in Section II. Related theory to gesture extraction is
consider exploiting them to support the human computer presented in Section III. Details on the proposed system's
interaction. Gesture is one of the most important noncontact implementation is provided in Section IV followed in Section
human interaction methods and forms a substantial part of the V by a presentation and discussion on the results obtained.
human language. Historically, wearable data gloves were Finally, conclusion and future work are discussed in Section
usually used to obtain the angles and positions of each joint in VI.
the user's gesture. The inconvenience and cost of a wearable
sensor have limited the widespread use of such method.
Gesture recognition methods based on noncontact visual II. LITERATURE REVIEW
inspection are currently popular due to their low cost and Image processing is central to hand gesture recognition.
convenience to the user. Hand gesture is an expressive Digital image processing is most relevant to our work where
interaction method used in healthcare, education and the useful information related to hand gesture and movement
entertainment industry, in addition to supporting users with need to be extracted from digital images. Segmentation is
special needs and the elderly. Hand tracking is essential to crucial in gesture recognition [4], [5]. Hand detection and
hand gesture recognition and involves undertaking various background removal are vital to the success of the gesture
computer vision operations including hand segmentation, recognition algorithm. In previous work, a monocular camera
was used in gesture recognition algorithms to filter out the
Manuscript received August 20, 2019; revised October 17, 2019. background, which can be inconvenient in a real-world
The authors are with the Department of Electronic and Computer
Engineering, Brunel University London, Uxbridge, UK, UB8 3PH (e-mail:
setting. Most methods used in hand detection are based on
[email protected], [email protected]). Harr features, colour, context information, or even shape.

doi: 10.18178/ijmlc.2019.9.6.873 782

 International Journal of Machine Learning and Computing, Vol. 9, No. 6, December 2019

Such methods can provide accurate performance given the recognition technique based on deep transfer learning. Their
successful identification of the background and the hand in work involved recognizing six static and eight dynamic hand
the image. However, there are limitations, e.g. a hand gestures under various light conditions and backgrounds.
detection method relying on the skin colour will fail if the Pang et al. [18] presented a technique based on deep learning,
person is wearing a glove [6]. heuristic, and transfer learning. It uses a convolutional neural
Feature detection is a crucial part of 2-D and 3-D image network tool, which can arrange a set of areas based on how
processing [7]. Before any feature extraction technique is to fit the target. 36 different algorithms on 50 videos were
applied, the image data is pre-processed, and different evaluated.
pre-processing techniques are applied to it including In this work, a hand gesture recognition system was
thresholding, binarization, and normalization. Features are developed to investigate and compare various methods
then extracted and used for classification purposes. The including WT, EMD and CNN. Their classification accuracy
behaviour of an image is captured based on its features. A is compared based on measuring various performance
good feature set contains attributes with high information metrics.
gain and can be used to effectively classify images into
different groups. A method that utilizes the symmetric
properties of visual data to detect spare and stable image III. THEORY
features was presented by Huebner and Zhang [8]. Regional
features were formed by using Qualitative Symmetry Wavelets Transforms
operator together with quantitative symmetry range Wavelet transforms support image processing by
information. performing signal analysis where signal frequency differs at
Extraction and classification of local image structures are the end of time [19]. Wavelet transforms analysis offers
crucial to gesture recognition. Gevers et al. [9] proposed a accurate information on signal data in comparison to other
method to classify the physical nature of local image analysis techniques. Daubechies orthogonal wavelet is used
structure using the geometrical and photometrical in this study. It is the known as dbN wavelets where N is the
information. number of fading moments. The Daubechies wavelets are
To make the hand gesture recognition more accurate and defined as follows:
thus ensuring a more natural user experience interacting with
the machine interface, Bouchrika et al. [10], [11] applied a
∫ 𝑥𝑛 𝜓(𝑥)𝑑𝑥 = 0, 𝑛 = 0,1, … , 𝐾 (1)
Wavelet Network Classifier (WNC) in a remote computer The equation has a combination of scaling functions used
ordering application using hand gestures to place orders. to represent numerical approximations on a secured scale.
Hands detection, tracking and gesture recognition techniques The value of K is directly proportional to the orthogonality
were applied. WNC was used for its effective classification condition.
results. An approach also proposed by Bouchrika et al. [12]
made amendments to the Wavelet Network classification
phase by making separated Wavelet Networks discriminating IV. EMPIRICAL MODE DECOMPOSITION
classes (n − 1) with the purpose of training each image. This EMD is an innovative technology used in both
resulted in less time required to complete the testing phase. non-stationary and non-linear data [20]. It is based on
The proposed Wavelet Network architecture enables quick decomposing a signal into Intrinsic Mode Functions (IMF)
learning and recognition of actions by avoiding unnecessary with respect to the time domain [20], [21]. The EMD method
hand movements. Another hand gesture recognition approach can be compared to other analysis techniques such as
[13] was based on wavelet enhanced image pre-processing Wavelet Transforms and Fourier Transforms [21].
and supervised Artificial Neural Networks (ANN). Neuroscience experiments, seismic readings,
Contour segmentation was supported in the pre-processing. gastro-electrograms, electrocardiograms, and sea-surface
Reference points were used to provide 2-D hand gesture height readings are some of the data type to which the EMD
contour images to 1D signal conversion. Wavelet technology might be applied [21]. EMD is defined as
decomposition was used for 1D signals. Four statistical follows:
features were extracted from the wavelet coefficients. Six
hand gestures were tested. An accuracy of 97% was achieved 𝑥(𝑡) = ∑𝑁
𝑛=1 𝑐𝑛 (𝑡) + 𝑟𝑛 (𝑡) (2)
with fast feature extraction and computation. Murthy and where rn is the mean trend of x(t), the value of cn is the of
Jadon [14] proposed a method for hand gesture recognition amplitude and frequency modulated output set. The
using Neural Networks. It is based on supervised frequency decreases as the value of cn increases.
feed-forward neural network net-based training and
back-propagation technique to classify hand gesture in ten
various categories including hand pointing up, down, left, V. ARTIFICIAL NEURAL NETWORKS
right, front, etc. Convolution Neural Networks (CNNs) were
used [15] to evaluate hand gesture recognition, where ANN is a simple electronic model similar to the neural
depth-based hand data was employed with CNN to obtain structure of the human brain. The brain functions by learning
successful training and testing results. Another CNN method from human experiences [22], [23]. ANN is a system that
was proposed [16] that uses a skin model, hand position processes information in a similar manner to the biological
calibration and orientation to train and test the CNN. nervous system. The system is composed of enormous
Hussain and Saxena [17] presented a hand gesture number of unified processing elements working together to

783
 International Journal of Machine Learning and Computing, Vol. 9, No. 6, December 2019

solve certain issues [22], [23]. Specifically, it is structured for 5. Apply the appcoef2 function which is used to compute
data classification or pattern recognition via learning an approximation coefficient of 2-D signals.
processes [22], [23]. Self-learning and the ability to handle 6. Extract each level using the wrcoef function to
large data are some of the many benefits associated with reconstruct the coefficients of each level in the video.
using ANN. A trained neural network is regarded as an expert 7. The execution time of WT is estimated only once.
in the set of information which it has been given to analyse. 8. The image data is trained and tested using a Neural
Network system. The NN has 20 hidden neurons in a
single hidden layer to train data and it stops when the
VI. CONVOLUTIONAL NEURAL NETWORKS error is reached in 20 epochs. When more hidden layers
CNN is a multi-layer neural network with a special are added the depth of the neural network is increased,
architecture used for deep learning [18]. A CNN architecture the neural network model becomes a deep learning
is composed of three layers: convolutional layer, Pooling model. Thus, A single layer is selected for this
layer, and fully-connected layer. CNN is frequently used in experiment.
recognizing scenes, objects, and carrying out image detection, 9. The execution times of image data training and testing
extraction and segmentation. CNN has been significantly are also calculated.
used in the last few years due to the following three factors:
(1) it removes the necessity for feature extraction by using
image processing tools and can directly learn the image data,
(2) it is exceptionally good for recognition results and can be
easily retrained for new recognition missions, and (3) it can
be built on the pre-existing network [18].
a) Sweep motion b) Shrink motion

VII. IMPLEMENTATION

A. Hand Gestures Input

Hand gestures represent the input to different gesture
detection methods evaluated in this study. Fig. 1 illustrates
ten 2-D and 3-D hand gestures with plain backgrounds. They
are recorded within long distances and used in the study's c) Circular motion d) Squeeze motion
experimental work.
The implementation framework illustrating the extraction
and the classification steps is shown in Fig. 2. Using an
iPhone 6 Plus camera with resolution 4k at 30 fps, the hand
motions shown in Fig. 1 are recorded. Each recording lasts 10
seconds and the resolution of the recorded video is 3840 ×
2160. The first system is created using optical flow object by
e) 2 Fingers Shrink f) Back/Forth
estimating and displaying the optical flow of objects in the
video. The length of videos is between 15 to 65 frames. Each
video has a different number of frames, which depends on the
first section of motion.
B. Computing Platform Specification
The experiment was performed using a Dell laptop XPS 15
9550 with 6th processor Generation Intel Quad Core i7, g) Rub motion h) Click motion
memory type DDR4 16 GB, speed 2133 MHz, 512 GB
storage hard drive, 15.6-inch Ultra-HD 15.6" IPS 1920×1080
RGB Optional 3840×2160 IGZO IPS display w/Adobe RGB
colour space and touch. Windows 10 (64 bits) operating
system was used and the system is implemented using
MATLAB R2017bV language.
C. Implementing Wavelet Transform with ANN i) Dance motion j) Pinch motion
The system is implemented using the db8 WT tool Fig. 1. Hand gestures used in the study.
following the steps outlined below:
1. Read each video using a video reader function. D. Empirical Mode Decomposition with Neural Network
2. Create an optical flow object that spreads the object Implementation
velocities in an image. The implementation of EMD is similar to WT with the
3. Estimate and display the optical flow of objects in the addition of reshape function that returns the M-by-N matrix
video. whose elements takes column-wise from X. The function
4. Divide a video into certain frames; each frame contains 8 used is ceemd representing a noise improved data analysis
IMFs. algorithm.

784
 International Journal of Machine Learning and Computing, Vol. 9, No. 6, December 2019

3, there is a substantial fluctuation in the blueline in the IMF

graphs for different hand gestures. In addition, there is
notable variation in the red and light blue signals in all graphs.
Other signals exhibit steadier paths.
Fig. 2. The implementation framework.

E. Convolutional Neural Network Implementation

CNN forms an integral part of deep learning, as it is used to
train data without using any image processing tool. In our
experiment, a new directory is created for each video. Ten
images are generated to transfer the image frame RGB to
grey and resize it to 48×27 from the original image size . All
videos have 70 frames. The image’s data is split into training a) Sweep motion b) Shrink motion
and testing datasets. The CNN topology is created in seven
layers with each layer having the following functionality and
size: ImageInputLayer Input size [48,27,1],
Convolution2-DLayer Filter size [5,5], ReLULayer
(Rectified Linear Unit), MaxPooling2-DLayer Pool size [2,2],
FullyConnectedLayer Input size [auto] and Output size [10],
SoftmaxLayer and ClassificationOutputLayer Output size
[auto]. The hyperparameters of the CNN are generated inside
c) Squeeze motion d) Circular motion
the training options function. The value of max epochs
parameter is set to 200 epochs.
F. Parameters for Comparison
The performance of WT, EMD and CCN algorithms was
compared using number of parameters. This includes
execution time, that is the duration taken by the software to
process the given task. Sensitivity measures the percentage of
positives which are properly identified. Specificity is a e) 2 Fingers Shrink f) Back and Forth
measure of the false positive rate. The PPV and NPV are the
percentages of positive and negative results in diagnostic and
statistics tests which also describe the true positive and true
negative results. The LR+ and LR− are known measures in
diagnostic accuracy. Area under ROC curve (AUC) is the
typical technique to measure the accuracy of predictive
models.
g) Rub motion h) Click motion

VIII. RESULTS
The experiments were executed ten times to obtain the
mean of ten-hand motions. Two different training and testing
were presented and compared to find the best mini gestures
detection tool. Training accuracy is achieved by
implementing a model on the training data and determining
the accuracy of the algorithm.
i) Pinch motion j) Dance motion
Fig. 3 and Fig. 4 show the signal extracted features using Fig. 3. IMF for 10 different motions using WT.
IMF method for 10 different gestures in WT and EMD
techniques respectively. The IMF function is applied under For the IMF graphs for EMD shown in Fig. 4, signals for
two conditions. The first condition involves the entire data, different gestures are generally steadier compared to WT
where the number of extrema and the number of zero IMF graphs shown in Fig. 3. Slight fluctuations can be
crossings are equal or vary at most by one. The second noticed in the blue signal, especially for the back and forth
condition is that the mean of the envelope explained by the hand gesture. Minimum variation can be seen in the path of
local maxima or the envelope clarified by the local minima the red signal for all hand gestures. All other signals show
has value zero [24]. The extracted features are each assigned steady lines for different gestures.
a class and fed to ANN for training. A summary of the values acquired for various parameters
In the IMF graphs shown in Fig. 3 and Fig. 4, the X axis in training mode is listed in Table I. It can be noticed that the
represents time (in microseconds) and the Y axis represents execution time of WT is less than that of EMD and CNN. The
frequency with 8 signals of IMF (levels). The speed of accuracy result of CNN is better than WT and EMD. The
motion starts from 0 microsecond till the end of time with a value of sensitivity in CNN is higher than WT and EMD.
stable signal rate. Examining the IMFs for WT shown in Fig. Specificity in CNN is the highest followed by EMD and the

785
 International Journal of Machine Learning and Computing, Vol. 9, No. 6, December 2019

lowest result was recorded for WT. The PPV and NPV of WT higher PPV and NPV values than WT. For LR+ and LR−,
is lower than EMD and CNN. The best value for LR+ and CNN values are higher than WT and EMD. The value of
LR− are recorded for CNN. For RMS, the value of EMD and RMS for EMD is the lowest while WT has the highest value.
CNN are lower than WT. Finally, The AUC is 0.90 for WT, Lastly, The AUC is 0.93 for WT, 1 for EMD, and 1 for CNN.
99 for EMD, and 1 for CNN. As in the training phase, the duration of CNN execution
took similar time i.e. 636 minutes, an impractical timing
given that only ten images were tested. It is notable that CNN
has a significantly low value of 1 for the Positive Likelihood
(LR+) compared to WT (19.29) and EMD (20.81).
TABLE I: COMPARISON BETWEEN WT, EMD AND CNN IN TRAINING MODE
WL EMD CNN
Exe Time ±SD 636.433±
6.0755 ±1.243 9.171 ±2.329
(sec) 113.922
a) Sweep motion b) Shrink motion Accuracy ±SD 0.440 ±0.068 0.909±0.051 1±0
Sensitivity ±SD 0.433 ±0.167 0.803 ±0.151 1±0
Specificity ±SD 0.968±0.0126 0.996 ±0.004 1±0
Positive Predictive
0.490 ±0.158 0.968±0.041 1±0
Value (PPV)
Negative Predictive
0.958 ±0.013 0.983 ±0.010 1±0
Value (NPV)
Positive Likelihood 16.783
129.406±158.592 1±0
(LR+) ±13.312
Negative Likelihood
0.599 ±0.196 0.237 ±0.155 1±0
c) Squeeze motion d) 2 Fingers Shrink (LR−)
RMS ±SD 1.223±0.129 0.371±0.053 1±0
AUC ±SD 0.901±0.038 0.999±0.001 1±0

TABLE II: COMPARISON BETWEEN WT, EMD AND CNN IN TESTING MODE
WL EMD CNN
636.433
Exe Time ±SD (sec) 0.158 ±0.027 0.195 ±0.053
±113.922
Accuracy ±SD 0.437±0.079 0.915 ±0.049 1±0
Sensitivity ±SD 0.389±0.262 0.769±0.211 1±0
e) Rub motion f) Click motion
Specificity ±SD 0.970±0.0162 0.995±0.007 1±0
Positive Predictive
0.496 ±0.254 0.979±0.044 1±0
Value (PPV)
Negative Predictive
0.955±0.012 0.981±0.014 1±0
Value (NPV)
Positive Likelihood
19.292±18.316 20.81 ±44.378 1±0
(LR+)
Negative Likelihood
0.622±0.197 0.254±0.215 1±0
(LR−)
g) Pinch motion h) Back and Forth RMS ±SD 1.204 ±0.122 0.354±0.073 1±0
AUC ±SD 0.937±0.030 1.00 ±0 1±0

i) Circular motion j) Dance motion

Fig. 4. IMF for 10 different motions using EMD.

The parameter values of CNN are constant for all

categories. Its execution time is approximate 636 minutes, a
substantially long and unacceptable duration to train the
system using only ten hand movement pictures (which are
used in the experiment). Positive Likelihood (LR+) of EMD
is higher indicating more accuracy compared to WT and
CNN. Overall, CNN has the best values in most parameters Fig. 5. ROC for 10 different classes in WT.
when training was performed except for execution time.
Comparative performance values for the three methods are Fig. 5 and Fig. 6 show the Receiver Operating
listed in Table II. CNN's execution time is higher than WT Characteristic ROC curve which is applied in binary
and EMD. For accuracy, WT achieved a lower value classification to learn the output of a classifier. There are two
compared to EMD and CNN. Accuracy results of CNN strategies of ROC to be drawn for multiclass curve: One VS.
outperformed WT and EMD. Similarly, CNN has a higher One and One VS. Multi, with the latter being used in this
sensitivity value compared to WT and EMD. Specificity in study. According to the WT and EMD graphs, the 10 classes
WT is lower than EMD and CNN. EMD and CNN have had 10 ROC curves reached the upper left corner which are

786
 International Journal of Machine Learning and Computing, Vol. 9, No. 6, December 2019

100% True Positive Rate (Sensitivity) and 100% False Conference on Microelectronics, Computing and Communications,
Durgapur, India, 2016.
Positive Rate (1−Specificity). The ROC curve of EMD is [7] R. Szeliski, Computer Vision: Algorithms and Applications, Springer,
extremely near to the upper left corner compared to WT. 2010.
CNN provides a better accuracy when compared with WT [8] K. Huebner and J. Zhang, “Stable symmetric feature detection and
and EMD. However, CNN's duration of execution is classification in panoramic robot vision systems,” in Proc. IEEE/RSJ
International Conference on Intelligent Robots and Systems, 2006.
substantially high. WT and CNN memory usage is lower than [9] T. Gevers, S. Voortman, and F. Aldershoff, “Color feature detection
EMD. and classification by learning,” in Proc. IEEE International
Conference on Image Processing, 2005.
[10] T. Bouchrika, M. Zaied, O. Jemai and C. B. Amar, “Ordering
computers by hand gestures recognition based on wavelet networks” in
Proc. 2nd International Conference on Communications, Computing
and Control Applications, Merseilles, France, 2012.
[11] T. Bouchrika, M. Zaied, O. Jemai, and C. Ben Amar, “Neural solutions
to interact with computers by hand gesture recognition,” Multimedia
Tools and Applications, vol. 72, no. 3, pp. 2949-2975, 2013.
[12] T. Bouchrika, O. Jemai, M. Zaied, and C. Amar, “Rapid and efficient
hand gestures recognizer based on classes discriminator wavelet
networks,” Multimedia Tools and Applications, vol. 77, no. 5, pp.
5995-6016, 2017.
[13] X. Fu, J. Lu, T. Zhang, C. Bonair, and M. L. Coats, “Wavelet enhanced
image preprocessing and neural networks for hand gesture recognition,”
in Proc. IEEE International Conference on Smart
City/SocialCom/SustainCom (SmartCity), Chengdu, China, 2015, pp.
838-843.
[14] G. R. Murthy and R. S. Jadon, “A review of vision based hand gestures
recognition,” International Journal of Information Technology and
Knowledge Management, vol. 2, pp. 405-410, 2009.
Fig. 6. ROC for 10 different classes in EMD. [15] J. Pyo, S. Ji, S. You, and T. Kuc, “Depth-based hand recognition
using convolutional neural ne tworks,” in Proc. 13th International
Conference on Ubiquitous Robots and Ambient Intelligence, Xi'an,
China, 2016, pp. 225-227.
IX. CONCLUSIONS AND FUTURE WORK [16] H. Lin, M. Hsu and W. Chen, “Human hand gesture recognition using a
convolution neural network,” IEEE.
Hand gesture recognition is essential to support a natural [17] S. Hussain, R. Saxena, J. A. Khan, and H. Shin, “Hand gesture
HCI experience. The most important aspects of gesture recognition using deep learning,” in Proc. International SoC Design
recognition are segmentation, detection, and tracking. In this Conference (ISOCC), Seoul, South Korea, 2017, pp. 48-49.
[18] S. Pang, J. del Coz, Z. Yu, O. Luaces, and J. Díez, “Deep learning to
study, a system has been created for hand motion detection frame objects for visual target tracking,” Engineering Applications of
using WT and EMD for features extraction. Classification is Artificial Intelligence, vol 65, pp. 406-420, 2017.
supported using ANN and CNN. Ten 2-D and 3-D motion [19] I. Daubechies, Ten Lectures on Wavelets, Philadelphia, PA, SIAM,
1992.
images with plain backgrounds and recorded within long [20] N. Huang, Z. Shen, S. Long, M. Wu, H. Shih, Q. Zheng, N. Yen, C.
distances were used. Experiments were performed to Tung, and H. Liu, “The empirical mode decomposition and the hilbert
compare the performance of various methods using number spectrum for nonlinear and non-stationary time series analysis,”
Proceedings of the Royal Society A: Mathematical, Physical and
of measures. Results showed that CNN provides better Engineering Sciences, vol. 454, pp. 903-995, 1998.
accuracy compared to WT and EMD. However, its [21] M. Lambert, A. Engroff, M. Dyer and B. Byer. Empirical Mode
computational requirements are relatively high. Memory Decomposition. [Online]. Available:
usage of WT and CNN was lower than EMD. In future work, https://www.clear.rice.edu/elec301/Projects02/empiricalMode
[22] I. Aleksander and J. Taylor, Artificial Neural Networks, vol. 2,
the number of motions will be extended using a 3-D Amsterdam, North-Holland, 1992.
Holoscopic imaging system. [23] B. Fritzke, “Growing cell structures — A self-organizing network in
k-dimensions,” Artificial Neural Networks, pp. 1051–1056, 1992.
ACKNOWLEDGEMENTS [24] C. Junsheng, Y. Dejie, and Y. Yu, “Research on the Intrinsic Mode
Function (IMF) criterion in EMD method,” Mechanical Systems and
The first author would like to show his appreciation in Signal Processing, vol. 20, pp. 817-824, 2016.
terms of the financial support received from the Ministry of
Higher Education in Saudi Arabia. Norah M. Alnaim is a faculty at the Computer Science
Department in Imam Abdulrahman Bin Faisal University
REFERENCES in Saudi Arabia. She had a master of science in computer
information systems in 2012 from St. Mary’s University,
[1] R. Pradipa and S. Kavitha, “Hand gesture recognition-analysis of San Antonio, TX, USA. She is currently a PhD student in
various techniques, methods and their algorithms,” in Proc. the Department of Electronic and Computer Engineering,
International Conference on Innovations in Engineering and Brunel London University, London, UK. She is working
Technology, Tamil Nadu, India, 2014, pp. 2003-2010. on gesture recognition and artificial intelligence field.
[2] A. Yilmaz, O. Javed, and M. Shah, “Object tracking: A survey,” ACM
Computing Surveys, vol. 38, no. 4, 2006.
[3] F. Mahmoudi and M. Parviz, “Visual hand tracking algorithms,” IEEE Maysam Abbod is a reader in intelligent system in the
Gemometric Modeling and Imaging - New Trends, London, UK, 2006. Department of Electronic and Computer Engineering at
[4] N. M. Zaitoun and M. J. Aqel, “Survey on image segmentation Brunel University London. Dr Abbod received BSc in
techniques,” Procedia Computer Science, vol. 65, pp. 797–806, 2015. electronic engineering in 1987, PhD in control
[5] N. Dhanachandra and Y. J. Chanu, “A Survey on image segmentation engineering from Sheffield University in 1992. He has
methods using clustering techniques,” European Journal of published 117 refereed journal papers, 16 chapters in
Engineering Research and Science, vol. 2 no. 15, 2017. edited books and 122 refereed conference papers. His
[6] M. Rao, P. Kumar, and A. Prasad, “Implementation of real time image main research interests are in intelligent systems for
processing system with FPGA and DSP,” in Proc. IEEE International modelling, control and optimisation

787