IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 12, No. 3, September 2023, pp. 1007∼1018

ISSN: 2252-8938, DOI: 10.11591/ijai.v12.i3.pp1007-1018 r 1007

Background subtraction challenges in motion detection

using Gaussian mixture model: a survey
Nor Afiqah Mohd Aris, Siti Suhana Jamaian
Department of Mathematics and Statistics, Faculty of Applied Sciences and Technology, Universiti Tun Hussein Onn Malaysia, Pagoh,
Malaysia

Article Info ABSTRACT

Article history: Motion detection is becoming prominent for computer vision applications. The
background subtraction method that uses the Gaussian mixture model (GMM)
Received Jun 25, 2022
is utilized frequently in camera or video settings. However, there is still more
Revised Aug 24, 2022 work that needs to be done to develop a reliable, accurate and high-performing
Accepted Nov 2, 2022 technique due to various challenges. The degree of difficulty for this challenge
is primarily determined by how the object to be detected is defined. It could be
Keywords: influenced by the changes in the object posture or deformations. In this con-
text, we describe and bring together the most significant challenges faced by
Background-foreground detection the background subtraction techniques based on GMM for dealing with a cru-
Background subtraction cial background situation. Therefore, the findings of this study can be used to
Computer vision identify the most appropriate GMM version based on the crucial background
Gaussian mixture model situation.
Motion detection
This is an open access article under the CC BY-SA license.

Corresponding Author:
Siti Suhana Jamaian
Department of Mathematics and Statistics, Faculty of Applied Sciences and Technology
Universiti Tun Hussein Onn Malaysia, Pagoh Education Hub
84600 Pagoh, Johor, Malaysia
Email: [email protected]

1. INTRODUCTION
The study and comprehension of video sequences have become an active area of research in computer
vision over the past few years due to their increasing importance in numerous video analysis applications, such
as video surveillance and multimedia applications that detect motion in the video scene [1], [2]. Identifying
motion in a video sequence is important in target detection [3]–[5] and behaviour interpretation [6], [7]. Thus,
the initial operation is distinguishing between the foreground and background objects. It can be done in various
ways, depending on the available data and whether the object is moving. Any prior information to detect
moving objects from a video sequence is unnecessary. Instead, a series of consecutive frames is required.
Many motion-based video systems still struggle to deal with dynamically changing backgrounds. Dy-
namic or moving background objects can cause massive false detections [8]. It is a significant contributor to
false alarms in event detection. In order to raise the sensitivity to important events of interest, the security indus-
try has a tremendous need to either detect and suppress these false alarms or mitigate the effects of background
changes.
Many new approaches for identifying motion have received attention. In the process of detecting
motion, the optical flow method [9]–[12], the inter-frame difference method [13]–[16], and the background
subtraction method [17]–[20] are the most used approaches. Detecting methods are chosen based on detection
scenarios. The inter-frame difference method extracts a moving target from a continuous video frame image.

Journal homepage: http://ijai.iaescore.com

1008 r ISSN: 2252-8938

It is pretty adaptive. However, this approach causes a cavity effect, reducing target detection accuracy. The
optical flow approach uses the target image’s brightness. This technology is rarely employed because of its
processing complexity and lack of anti-jamming capacity. Meanwhile, background subtraction focuses on
building a stable background model to detect motion.
Subtracting the background from a video sequence is one of the most straightforward approaches to
finding the motion inside the video. It is an essential procedure for the majority of computer vision applica-
tions. In its most basic form, a background subtraction method requires a consistent background, which is
an extremely challenging requirement for applications that run in real time. The video sequence is separated
into its component video frames, and then each video frame is removed from a background model or reference
image. It is assumed that the moving object is represented by the current active structure’s pixels distinct from
the background model. The foreground object goes through additional processing for object localization and
tracking. Because removing the background is the initial step in many applications involving computer vision,
it is essential that the obtained foreground pixels precisely correlate to the moving object of interest.
Background subtraction is a method that uses a fairly straightforward algorithm. However, it is ex-
tremely unstable to shifts in the surrounding environment and has weak anti-interference capabilities. Many
researchers have presented a variety of background subtraction approaches to deal with a variety of problems.
Different approaches to simulating the background can be divided into pixel-based, region-based, and hybrid
approaches. In addition, there is the possibility of classifying modelling techniques for the background as either
parametric or nonparametric. Basic models [21], background estimation [22], background clustering [23], sub-
space learning, kernel density estimation, and the Gaussian mixture model (GMM) are some video background
modelling methodologies that are often utilized in the video. Among these, the GMM technique has gained
much interest due to its ability to easily handle image or video noise, shadow, camouflage, slow-moving ob-
ject, multimodal background and illumination changes, which is particularly noteworthy. However, researchers
are still investigating and making new contributions to the established study to increase the object-detection
performance. The purpose of this work is to provide a summary of researches that has successfully done on
GMM-based object detection in relation to various background environments. The other parts of this paper are
structured as follows. In section 2, the GMM is explained and introduced. In section 3, we comprehensively
summarise the research on the various GMM versions with respect to various background conditions. Section
4 presents our conclusions.

2. GAUSSIAN MIXTURE MODEL

GMM was first introduced by [24], which is a form of modelling technique based on the background
that creates a robust tracking system to handle multiple moving objects, variations in lighting, moving scene
clutter, and other arbitrary changes to the scene. The fundamental concept behind this method is to create a
Gaussian distribution for every pixel contained within a frames-long series. The foreground and background
are each indicated by a different weight, with the foreground having a smaller weight than the background. If
the new pixel fits the parameters of the K Gaussian model, then it will be treated as a background pixel. If there
is no match for the K Gaussian model, the new pixel will be handled as though it were a foreground pixel. The
random variable X is assigned to each pixel and will be modelled as a mixture of K Gaussian distributions, as
shown in (1).
K
X
P (Xt ) = Wk,t · η (Xt , µk , Σk,t ) . (1)
k=1

η(Xt , µk , Σk,t ) where is the weight of k th Gaussian in the mixture of time t, Wk,t is the Gaussian probability
density function with mean µk and covariance matrix Σk,t which is given by (2).
 
−1
1  1 X 
η(Xt , µk , Σk,t ) = n 1 exp − (Xt − µk,t )T (Xt − µk,t ) (2)
(2π) 2 |σk,t | 2  2 
k,t

where n is the pixel intensity dimension. It is assumed that the covariance matrix is
2
Σk,t = σk,t I (3)

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1007–1018

Int J Artif Intell ISSN: 2252-8938 r 1009

for computational reasons and K is also determined by the computational power and the available memory.
Meanwhile, the value of pixels can be modelled by a Gaussian mixture distribution value of K from 3 to 5.
These methods assume that the pixel values with red, green, and blue (RGB) have the same variance and are
independent. This assumption can avoid a costly matrix inversion at some accuracy rate. As a result, the GMM
is applied in this research in order to characterize the arrangement of pixels found in the scene.
The background model estimation process starts when every pixel value is compared to the current
Gaussian distribution value K until a match is found. A pixel value within 2.5 standard deviations of the
distribution is considered a match. If the match distribution found for the new pixel value is one of the back-
ground models, it is regarded as the background. Otherwise, the value of the pixel is the foreground. The B
Gaussian distribution, which is chosen as the background model and exceeds a certain threshold which can be
written as (4)

k=1
!
X
B = argmin Wk > T hreshold . (4)
b

The Gaussian parameters such as the weight, the mean, and the variance must be updated for the
subsequent foreground detection. The weight, the mean, and the variance of the k th Gaussian in the mixture at
the time t are updated as follows

Wk,t = (1 − α)Wk,t−1 + α(Nk,t ) (5)

µt = (1 − ρ)µt−1 + ρ(Xt ) (6)

σt2 = (1 − ρ)µ2t−1 + ρ(Xt − µt )T (Xt − µt ) (7)

where α is the learning rate and ρ is a second learning rate which is equal to αη(Xt , µk , σk ) . If the model
matched, the value of Nk,t is 1 and 0 for the remaining models. If the model is unmatched, the mean and the
variance remain unchanged. After making this first approximation, the weights are subsequently renormalized.
Suppose the current pixel does not match with the k th Gaussian in the mixture. In that case, the distribution
with the most negligible probability is replaced with a distribution that has a current value as its mean value,
high initial variance, and low current weight. The value of ω/σ 2 is used to rank the Gaussians. This value will
increase as the distribution obtains more evidence and the variance decreases.

3. GMM FOR BACKGROUND SUBTRACTION ISSUES

GMM is able to establish multiple distributions rather than clearly modelling the pixel values as
one particular type of distribution [25]. GMM has become a classical parametric model for moving de-
tection and is classified as background and foreground due to its effectiveness and robustness over diverse
distribution [26]. It is possible to have a static or dynamic background. It is straightforward to recognize ob-
jects against a static background. When there is motion in the background, the object detection system that
is needed is quite complicated. The object identification system needs to be able to accommodate dynamic
backgrounds while maintaining satisfactory performance in real time. In the next subsections, we will describe
various different background scenarios and the solutions for these issues using a GMM-based object detection
system.

3.1. Image or video noise

The image size of motion detection can vary due to the effects of camera imaging characteristics,
and the noise can easily influence the findings of motion detection. Because of this, the precision of the
identification of moving objects is easily compromised [27], [28]. The presence of noise in an image will
not only have an impact on how the image appears to the human eye, but it will also have an impact on the
following processing of the image, such as the extraction of image features, the categorization, and recognition
of images, and so on. Therefore, before processing the image, it is required to do denoising processing on the
obtained image. These issues will enhance the image quality and make it easier to process the image once it
has been acquired [26].

Background subtraction challenges in motion detection using ... (Nor Afiqah Mohd Aris)
1010 r ISSN: 2252-8938

Chen and Ellis [29] used a multi-dimensional Gaussian kernel density transform (MDGKT) pre-
processor to reduce noise in the spectral, temporal, and spatial domains. Each spectral component is smoothed
spatially and temporally using a multivariate kernel that can be thought of as the product of two radially sym-
metric kernels. The Euclidean metric enables a single bandwidth setting for each domain [30], [31],
a 2 ! b 2 !
M X X
Kga ,gb (x) = ga k
k (8)
ga gb gb

where the spatial component, denoted by xa , and the temporal component marked by xb , make up the feature
vector x. ga and gb are the bandwidths of the kernel, and M is the constant of associated normalization. This
MDGKT plays a vital role as a pre-processor to enhance the GMM’s reliability. It is possible to control the
size of the kernel with nothing more than a pair of bandwidth parameters (ga , gb ), which in turn determines the
time interval and the resolution of the GMM. This adjustment significantly affects the background subtraction’s
efficiency and precision without compromising spatial flexibility.
Kalti and Mahjoub [32] assumes the incorporation of a fuzzy distance into both the expectation-
maximization (EM) and adaptive distance-based fuzzy-C-means (ADFCM) algorithms. In this research, the
characterization of pixels is based on two features. The first feature explains the inherent attributes of the pixel,
and the second feature defines the entire neighbourhoods of the pixel. After that, the classification is determined
based on adaptive distance, which gives preference to either one of two attributes depending on where the pixel
is located in the picture in terms of its spatial location. The acquired findings have demonstrated that their
method performs considerably better than the conventional fuzzy-C-means (FCM) and EM, particularly with
regard to the toughness of the face-to noise and the precision of the edges between areas.
Zuo et al. [33] improved the conventional GMM for noise interruption by proposing a new method
considering the image block averaging method, wavelet semi-threshold, and adaptive background updating
method. At the step of background modelling, the video frame is blocked to enable computation and enhance
the speed at which modelling is performed. The model of the background is rebuilt by employing the method
of image block averaging. The following step involves utilizing wavelet semi-threshold function denoising in
conjunction with mathematical morphology closed operation. The noise problem is successfully eliminated,
and detection performance is improved in the detection stage of moving targets. During background updating
phase, the method of adaptive background updating is used to bring the background update to produce more
accurate detection results. The new approach in this work is both subjectively and objectively preferable to the
conventional GMM, which validates the efficacy of the system and demonstrates its flexibility.
Wei and Zheng [28] calculate the L2 norm between the GMM corresponding to the two pixels to
determine the degree of similarity between the two sets of pixels. The grayscale of the pixel and the abundance
of features in the local region of the picture are both represented by the GMM of the image pixels in the image
area. Measurements of the pixel grey intensity and the variety of information in the immediate region of the
image can be made with more precision if they are based on the difference between individual pixels. Because
of the similarities, the performance of the model used to denoise the image is improved, and the image’s detailed
information is preserved.
Luo et al. [27] used a method for detecting motion that considers the variation in the spatial image
threshold. The researcher calculates the projected size of motion in the image regions by establishing the
mapping relationship between the geometric features of motion in the image regions and the reasonable level
circumscribed rectangle (BLOB) of motion in the geographic space. This method is able to set an adaptive
threshold for each motion in order to remove unwanted noise during the process of motion detection.

3.2. Sudden illumination change

It is generally accepted that background models can accommodate slow but steady alterations in how
the environment appears. For instance, the amount of light in outdoor environments shifts throughout the day.
Variations in the scene’s illumination can also occur all of a sudden. The sudden turning on or off the light in
an interior space is one example of this shift that can take place. These issues can also occur outside scenarios,
which as a sudden shift from cloudy to sunlight. The amount of illumination significantly impacts how the
background appears and can lead to the detection of false-positive [34].
Chen and Ellis [35] proposed a global illumination of background model adaption and an online
dynamical learning rate in order to deal with this challenge. The researchers devised a revolutionary approach

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1007–1018

Int J Artif Intell ISSN: 2252-8938 r 1011

using a revised adaptive strategy in the iterative Zivkovic-Heijden GMM (ZHGMM) learning procedure. In
this step, they introduce the global illumination change of the median of quotient (MofQ) factor h among the
previously learned background and the binary input picture. The MofQ global illuminating altering factor ic
between the current picture and the reference picture ir is described as follows for all of the pixels that make
up a set S,
 
ic,s
h = medians∈S . (9)
ir,s

the mixture model introduces a counter denoted by the letter c for each Gaussian component. The counter c
keeps track of how many data sets have been sent to the Gaussian estimating algorithms, while the factor h
monitors how the global illumination varies over time. This adjustment makes a significant improvement to the
convergence as well as the background subtraction accuracy while maintaining the same degree of temporal
flexibility. This method is accomplished by recursively adding a modified input adaptive schedule into an
existing filtering system. When applied in sudden illumination changes, the performance is noticeably better
than previous approaches.
Martins et al. [36] developed a novel classification mechanism that blends colour space discriminating
skills with hysteresis and a dynamic learning rate to update the background model with sudden illumination
change. Each channel element L∗ , a∗ , b∗ is studied on its own, and the decisions obtained from each are merged
using the AND rule, which produces good results compared to those obtained from a conclusion based on ma-
jority voting. To prevent noisy pixels, the colour distance of which is very significant to the decision threshold,
from inadvertently altering the classification, a hysteresis method has been implemented. The dynamic back-
ground learning rate denoted by αDBG , depends on the number of Gaussians present in the mixture, marked
by M , such that αDBG = M · αBG , where is a fixed minimum value for the learning rate of the background
if the pixel is categorized as background. This method ensures that the model adaptation is performed more
quickly in dynamic regions and more slowly in static regions. In contrast, if the pixel categorization shifts from
foreground to background, a higher learning rate αU BG is applied. This method encourages rapid adaptation
when the background is shown, which helps prevent the appearance of phantom pictures.
Agrawal and Natu [37] developed GMM with BLOB analysis of the interconnected parts, including
labelling and morphological operations, to increase the accuracy of foreground detection. The suggested model
can be broken down into two stages: the training phase, which is responsible for producing a referential image,
and the testing phase, which is in charge of producing a binary mask. This model determines the difference
between the frame of reference BM G(x, y) and the current, and then it applies a threshold value in order to
extract the region of interest. When using this approach to construct the foreground model, a threshold value
determined using the standard deviation was selected for each pixel. Compared to other methods, the results
demonstrate that the concept of integrating GMM with blob analysis and morphological operation obtained a
lower number of incorrectly classified pixels.
Su [38] presented a GMM with data model optimization to adapt light transitions. Calculating the
gradient picture of the stream is the first step in the approach. The scar operator is used for this calculation.
Then, integrate RGB and gradients and use figurative ways to eliminate noisy movement areas and combine
those which remain. In order to reduce the likelihood of making an incorrect diagnosis, a comparison is made
between the two models’ outputs to arrive at a final make-up area. Ultimately, they conducted the assessment
and comparison using three separate image streams. The results reveal this method increases the accuracy of
the detection process by minimizing the occurrence of erroneous detection areas caused by sudden illumination
changes.

3.3. Shadow
Foreground items frequently have shaded areas due to changing light, which typically affects the
segmentation of foreground items and the execution of subsequent modules of an algorithm that models the
background. More specifically, there is a substantial difference in the lighting, but just a minor variation in
the colour, in a darkened area. A pixel is considered a component of the shadow in the scene if it is a part
of the background model that has been made darker by a shadow produced by another object in the scene.
Therefore, a reliable method should include this technique to eliminate shadows cast by the foreground regions
or disregard shadows that aren’t relevant to the problem at issue [39].

Background subtraction challenges in motion detection using ... (Nor Afiqah Mohd Aris)
1012 r ISSN: 2252-8938

Yadav and Xiaogang [34] presented the innovative hybrid method relying on the GMM, subtraction
of background, hue-saturation-value (HSV) colour model, feature extraction, and neural networks. In order
to provide a clean foreground, the shadows cast by moving objects are identified, and then eliminated using
the HSV colour model, and morphological operations are carried out. After the algorithms indicate that the
detection is finished, the background is modified to conform to the dynamic background. Following the iden-
tification of objects, the shape properties of those objects are retrieved using Hu’s seven-moment invariants of
the training samples of the image data. These recovered shape features are then fed into the back propagation
neural network (BPNN) during the training process. The system could erase the shadow’s influence and ac-
curately detect motion. The findings of the experiments have shown that the suggested approach has strong
resilience as well as real-time performance in realistic environments.
Jin et al. [40] presented an Improved GMM-based automatic segmentation method (IGASM) by im-
proving the approach for updating the background so that it can more effectively segment the floats on the sea
surface. Following the mapping of the GMM’s findings into an HSV colour space, a light-shadow classifier
function is implemented to address the problems associated with shadows. After that, a morphological method
is applied in order to refine the foregrounds that were previously acquired. In the end, the graph cuts technique
is utilized to optimize the segmentation outcomes based on the spatial information in the video images. It is
possible to detect stationary floats by improving the updating approach successfully. In order to solve the is-
sues that are brought about by shadows in the segmentation results, a shadow discriminant function is utilized.
Smoothing out the contour is accomplished through the open operation and the graph cuts algorithm. As a re-
sult, the accuracy of the segmentation results can be increased even more. The results of the experiments reveal
that this method demonstrates a higher efficiency in the surface floats segmentation on the water, particularly
in situations where there is a considerable shift in light. Still, the surface floats on the water do not move.
Lin and Chen [41] provide a method for the detection of moving objects that are based on GMM and
visual saliency maps. This method can eliminate the disruption caused by the shadow situation and accomplish
stable detection of moving objects. The researcher uses the GMM approach to construct models for the video
sequences and then obtains the rough foreground objects. The foreground, however, has a significant amount
of incorrect detection, and as a result, they cannot adequately extract the moving objects. In the second step of
the process, which is refining the crude foreground objects, make use of the vision saliency to achieve reliable
detection results. Following the conversion of each image frame to the L∗ , a∗ , b∗ colour space, the L∗ , a∗ , b∗
channels are smoothed using a Gaussian filter in order to remove small texture features as well as noise. Then,
estimate the saliency maps for each channel and linearly merge those maps to generate the final saliency map.
In the end, the final saliency maps and the foregrounds are combined in order to produce the moving objects.
The shadow issue can be successfully addressed and resolved using this method.

3.4. Multimodal background

A multimodal situation is created whenever there is motion occurring in the background. The scenery
in the background might have some motion, such as a fountain, the movement of clouds, the swaying of
nearby trees, and a wave in the water. This movement can be regular or irregular at various intervals. The
traditional GMM is capable of handling multimodal backgrounds robustly. However, the parameter K is fixed
experimentally, and its value remains unchanged. This method is not ideal regarding the time required for
detection, and computation [24]. An effective GMM improvement should be able to detect regular, or irregular
motion [39].
Ou et al. [42] proposed an adaptive GMM (AGMM) with BPNN to extract the foreground objects
in multimodal background conditions. In most cases, AGMM is employed to accomplish the twin goals of
simplifying the algorithm and enhancing its precision. All image pixels can fit with the hybrid model if more
single-Gaussian models exist. The neural network can figure out the statistical parameters of the image’s noise,
the model can alter the noise very well, and the foreground objects can be preserved entirely. This method
solves the problem of defective foreground objects due to morphological processing. It eliminates the need
for the model structure to make the trade between foreground objects and the noise. Because of this, they are
implementing an adaptive version of the GMM supported by a neural network that can improve the robustness
and performance of the entire system.
Zuo et al. [33] proposed a GMM-based technique for moving picture target recognition that over-
comes multimodal backgrounds. Compared to other methods, this method’s main advantage is that it eliminates
disturbance from dynamic backgrounds and improves algorithm detection. The image sequence is initially used

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1007–1018

Int J Artif Intell ISSN: 2252-8938 r 1013

to divide the image of the video sequence. This method is followed by replacing each pixel value with the over-
all average of each pixel of the image block in the background modelling step. Then, the image block of
the GMM mean approach is applied. The combination of mathematical morphology and a semisoft thresh-
old function is employed in this article to remove noise from an image of the foreground detection during the
process of detecting moving targets. The soft and hard threshold methods have distinct disadvantages. The
researcher presents a revised semisoft threshold function to overcome the limitations of both the hard and soft
threshold functions. The conventional GMM approach cannot update the background in real time, resulting
in a ghosting and poor moving target identification accuracy. The adaptive background update algorithm uses
the existing frame of image detection and the background model to tackle the problem. This quantitative and
qualitative approach is preferable to the method that substantiates the claims made on the system’s efficiency
and adaptability.

3.5. Camouflage
The camouflage means the foreground may contain an element with a colour identical to the back-
ground. It creates confusion during the process of detection and makes it difficult to determine whether some-
thing is in front of the background. Even while the traditional GMM has fine-tuned crucial parameters, it still
tends to create more false alarms.
Zhang et al. [43] use a combination of foreground matching with a measure of short-term stability for
camouflage. The priority is given to matching probable foreground in incoming pixels by foreground models
that have been constructed and updated using the foreground pixels that have been detected. During this time,
the stability at the pixel level is assessed to ensure that an integrated foreground will be identified when a
dynamic foreground procedure is being carried out. Suppose the currently-used pixel value does not match any
pre-existing background models. In that case, the foreground model will be constructed with the currently-used
pixel value acting as the model’s mean and an enormous value working as the model’s variance. Foreground
models are given precedence over background models by this approach. Foreground models always try to
match the pixels that come after them, which helps to lessen the likelihood that foreground pixels will not
accurately represent the background models. Compute the short-term stability by utilizing the previous pixel
only if the existing foreground model does not match the next pixel can be written as (10)

PM P 2
M
M j=1 x2t+j − j=1 xt+j
S= (10)
M (M − 1)

where M are the total frame numbers representing the amount of time that has passed. The GMM’s tolerance to
camouflage at a slow pace is significantly improved by the combination of foreground matching and short-term
stability measures.
Lima et al. [44] proposed a method for estimating a threshold for each region by feedback step.
Spatial analysis is used to determine a threshold, which is then used to classify the pixel. In order to correct
some classification mistakes, the segmentation is filtered before the threshold estimation is performed. This
filtering technique removes disturbances and combines the entire area into one cohesive whole. During the
feedback stage, the fixes are used to estimate the thresholds for the subsequent iteration. Since the filtering
stage only corrects foreground mistakes, the improvement is concentrated on the vehicle areas. This strategy
that has been recommended promotes the segmentation of regions that have already been segmented. As a
result, the estimation of the threshold is comparable to a first-order Markov chain. This technique lowers the
amount of error on subsequent iterations when applied to fixed areas.

3.6. Slow moving object

Motions with unique patterns, such as moving slowly or staying, are significant for security protec-
tion since they make detection exceedingly difficult [45]–[49]. In addition, these kinds of motions make it
more challenging to identify perceived risks. In the actual world, many things move at a slower rate. When
dealing with the kind of objects, the majority of the currently available GMM-based algorithms typically ex-
tract fragmented bodies. This results in a significant reduction in detection precision. The traditional GMM
algorithm is predicated on the idea that the background is more likely to be seen than any particular fore-
ground. Because of this assumption, there is a chance that a long-observed object that moves slowly would be
misidentified as background. This might lead to significant complications. It is not appropriate to include a

Background subtraction challenges in motion detection using ... (Nor Afiqah Mohd Aris)
1014 r ISSN: 2252-8938

slow-moving object in the background until after the object has stopped moving altogether and has been still for
a sufficient time [43].
Zhang et al. [43] use the foreground matching technique to detect slow-moving objects. GMM is
used for the foreground and is continuously updated using newly identified foreground pixels. It takes into
explicit consideration the spatial continuity of the objects that are moving. The foreground model is priori-
tized to fit the pixels, and pixels that provide the foreground models are explicitly designated as foreground.
This is done to prevent the misleading inclusion of long-term foreground into the background. The results
show progress has been made in improving the performance of the identification of slow-moving objects in
complicated circumstances.
Fu et al. [50] studied an initial Gaussian background model (IGBM) based on an extended Kalman
filter to improve GMM performance for the slow-moving object. In order to find a solution to this issue, the
IGBM is not only incorporated with the GMM but also adds an extended Kalman filter [51]–[53] based tracker
into the system whenever the grey value of this static object is comparable to the grey background value.
Doing things this way can keep the moving item in its static condition indefinitely. Meanwhile, the tracker
implemented in the system can reliably determine whether the grey of a stationary object is comparable to the
grey background value. This research assumes a total of K single Gaussian distributions in the GMM. Hence,
they can construct a new background model using these initial Gaussian distributions, which are IGBM. Due
to the fact that it is made up of complete single Gaussian distributions rather than the initial GMM, it can keep
the most abundant information from the model that was used to create the original background. The results
indicate that the EGMM approach can efficiently address the challenge of recognizing moving objects that stop
and start intermittently.
Zhang et al. [54] presented the GMM with confidence measurement (GMMCM) as a potential solution
to the issue that the background subtraction model is susceptible to being easily contaminated by vehicles either
moving slowly or temporarily stopped. The confidence measurement (CM) is then applied to each background
model pixel to quantify each background pixel’s current trust values. In order to avoid contamination of the
background model in complicated urban traffic scenes, a technique for maintaining a balance between the
dynamic changes in the brightness and the background, which makes use of a self-adaptive learning rate to
keep the background model, has been developed. The result reveals that the proposed GMMCM does a great
job of coping with cars that are going slowly or temporarily stopped.

3.7. Bootstrapping
Background initialization for bootstrapping video sequences is a technique that is frequently employed
in intelligent video surveillance systems for the purpose of monitoring crowded public spaces. Background
removal approaches have recently been concentrating on background initialization for bootstrapping video
sequences. This is because a training duration devoid of foreground items is not always accessible in congested
contexts [55], [56]. To put it another way, the definition of background initialization for bootstrapping video
sequences is to estimate a background frame that does not contain any foreground items, given a video sequence
that was captured by a stationary camera and in which the background is occluded by some foreground objects
in each frame of the video sequence [57], [58]. This circumstance occurs when the objects in the foreground
constantly occupy the background. In a situation like this, there are two conceivable outcomes. In the first
scenario, the background is excessively active, and no single frame will be available that is devoid of any
objects. The second scenario has a crowded background, but there are still some beginning frames that do not
include any objects.
Amintoosi et al. [59] proposed the identification of the background through the use of the GMM
with the QR decomposition method in linear algebra. R-values derived via QR decomposition can be used
to deconstruct a given system to reflect the level of relevance possessed by each system’s components that
have been decomposed. After dividing the image into several smaller blocks, the researchers look at each to
determine which makes a minor significant contribution to the overall picture and then choose those blocks.
The simulation findings indicate that the background detection performance is superior to other methods.
Harville et al. [60] used a method for modelling the background that uses per-pixel, time-adaptive
GMM in the combined input space of depth and luminance-invariant colour. In this method, the GMM is
used to model the background. This combination is already novel, but the researchers make it even more so by
introducing the concepts of modulating the background model learning rate based on scene activity and making
colour-based segmentation criteria dependent on depth observations. Both of these ideas serve to improve the

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1007–1018

Int J Artif Intell ISSN: 2252-8938 r 1015

original combination further. These results demonstrate that the method has significantly larger robustness
to troublesome phenomena than the previous state-of-the-art without sacrificing real-time performance. This
makes it well-suited for a wide range of practical applications in video event detection and recognition. Table
1 summarizes all the background challenges in motion detection using GMM approaches.

Table 1. Summary of background subtraction challenges in motion detection using GMM

No. Background issues Modification of GMM Authors

1 Image or video noise 1.MOD-AT based on adaptive threshold. Luo et al. [27]
2.Calculating the L2 Norm. Wei and Zheng [28]
3.Image block averaging method, wavelet semi- Zuo et al. [33]
threshold, and adaptive background updating method.
4.Fuzzy distance in EM and FCM algorithm. Kalti and Mahjoub [32]
5.A multi-dimensional Gaussian Kernal density Chen and Ellis [29]
transform.
2 Sudden illumination change 1.Online dynamic learning. Chen and Ellis [35]
2.Colour space discrimination capabilities with Martins et al. [36]
hysteresis and dynamic learning rate.
3.Blob analysis and morphology. Agrawal and Natu [37]
4.Data model optimization. Su [38]
3 Shadow 1.HSV colour model features extraction and neural Yadav and Xiaogang [34]
network.
2.Background updating strategy. Jin et al. [40]
3.Visual saliency map. Lin and Chen [41]
4 Multimodal background 1.AGMM and BPNN hybrid method. Ou et al. [42]
2.Image block average method, wavelet semi- Zuo et al. [33]
threshold, adaptive background updating method.
5 Camouflage 1.Foreground matching and short-term stability mea- Zhang et al. [43]
sure.
2.The threshold for each region by feedback step. Lima et al. [44]
6 Slow moving object 1.Foreground matching. Zhang et al. [43]
2.Initial GMM and Extend Kalman Filter. Fu et al. [50]
3.Confidence measurement. Zhang et al. [54]
7 Bootstrapping 1.QR decomposition. Amintoosi et al. [59]
2.Learning rate adaptation based on scene activity. Harville et al. [60]

4. CONCLUSION
The goal of this study is to have a comprehensive assessment of the numerous GMM methods that have
been developed to address various background issues. In addition to that, it offers a concise description of GMM
methods. Image or video noise, sudden illumination change, shadow, multimodal background, camouflage,
bootstrapping and slow-moving object are some of the background difficulties that various GMM versions can
address. Image and video noise are the difficulties that receive significant attention from researchers, while
camouflage receives less of their focus. Different variants of GMM are capable of simultaneously addressing
various challenges. This study assists researchers in selecting the correct version of GMM for their applications
based on the study’s findings. Moreover, GMM techniques, including comprehensive bibliography material,
can provide helpful insights into this critical background topic and motivate for future research.

ACKNOWLEDGEMENTS
This research was supported by the Ministry of Higher Education (MOHE) through Fundamental
Research Grant Scheme (FRGS/1/2021/STG06/UTHM/02/1).

REFERENCES
[1] K. Goyal and J. Singhai, ”Review of background subtraction methods using Gaussian mixture model for video surveillance systems,”
Artificial Intelligence Review, vol. 50, no. 2, pp. 241-259, 2018, doi: 10.1007/s10462-017-9542-x.
[2] T. Bouwmans, F. El Baf and B. Vachon, ”Background modeling using mixture of gaussians for foreground detection-a survey,”
Recent patents on computer science, vol. 1, no. 3, pp. 219-237, 2008, doi: 10.2174/1874479610801030219.
[3] T. Liu, X. Cao, and J. Jiang, “Visual object tracking with partition loss schemes,” IEEE Transactions on Intelligent Transportation
Systems, vol. 18, no. 3, pp. 633–642, 2016, doi: 10.1109/TITS.2016.2585663.

Background subtraction challenges in motion detection using ... (Nor Afiqah Mohd Aris)
1016 r ISSN: 2252-8938

[4] T. S. Waykole and Y. K. Jain, “Detecting and tracking of moving objects from video,” International Journal of Computer Applica-
tions, vol. 81, no. 18, pp. 0975–8887, 2013, doi: 10.5120/14224-2410.
[5] Y. Dai, Y. Wu, F. Zhou and K. Barnard, “Attentional local contrast networks for infrared small target detection,” IEEE Transactions
on Geoscience and Remote Sensing, vol. 59, no. 11, pp. 9813–9824, 2021, doi: 10.1109/TGRS.2020.3044958.
[6] M. Devanne, S. Berretti, P. Pala, H. Wannous, M. Daoudi and A. Del Bimbo, “Motion segment decomposition of RGB-D sequences
for human behaviour understanding,” Pattern Recognition, vol. 61, pp. 222–233, 2017, doi: 10.1016/j.patcog.2016.07.041.
[7] X. Y. Zhang, H. Shi, C. Li, K. Zheng, X. Zhu and L. Duan, “Learning transferable self-attentive representations for action recognition
in untrimmed videos with weak supervision,” In Proceedings of the AAAI conference on artificial intelligence, vol. 33, no. 01,
pp. 9227–9234, 2019, doi: 10.1609/aaai.v33i01.33019227.
[8] D. S. Pham, O. Arandjelović and S. Venkatesh, ”Detection of dynamic background due to swaying movements from motion feature,”
IEEE Transactions on Image Processing, vol. 24, no.1, pp. 332-344, 2014, doi: 10.1109/TIP.2014.2378034.
[9] Y. Xin, J. Hou, L. Dong and L. Ding, “A self-adaptive optical flow method for the moving object detection in the video sequences,”
Optik, vol. 125, pp. 5690–5694, 2014, doi: 10.1016/j.ijleo.2014.06.092.
[10] P. Han, J. Du, J. Zhou and S. Zhu, “An object detection method using wavelet optical flow and hybrid linear-nonlinear classifier,”
Mathematical Problems in Engineering, vol. 2013, pp. 1–14, 2013, doi: 10.1155/2013/965419.
[11] R. M. Baby and R. R. Ahamed, “Optical flow motion detection on Raspberry Pi,” In Fourth International Conference on Advances
in Computing and Communications, 2014, pp. 151–152, doi: 10.1109/ICACC.2014.42.
[12] D. Archana and S. Sanjeevani, “Moving object detection using optical flow and HSV,” In Evolution in Signal Processing and
Telecommunication Networks, vol. 839, 2022, pp. 49-55, doi: 10.1007/978-981-16-8554-5 6.
[13] Y. H. Cheng and J. Wang, “A motion image detection method based on the inter-frame difference method,” Applied Mechanics and
Materials, vol. 490–491, pp. 1283–1286, 2014, doi: 110.4028/www.scientific.net/AMM.490-491.1283.
[14] M. Wan, G. Gu, E. Cao, X. Hu, W. Qian, and K. Ren, “In-frame and inter-frame information-based infrared moving
small target detection under complex cloud backgrounds,” Infrared Physics & Technology, vol. 76, pp. 455–467, 2016,
doi: 10.1016/j.infrared.2016.04.003.
[15] J. Guo, J. Wang, R. Bai, Y. Zhang and Y. Li , “A new moving object detection method based on frame-difference and
background subtraction,” in IOP Conference Series: Materials Science and Engineering, 2017, vol. 242, o. 1, p. 012115,
doi: 10.1088/1757-899X/242/1/012115.
[16] Y. Chen and J. Dong, “Target detection based on the interframe difference of block and graph-based,” In Proceedings of the 2016 9th
International Symposium on Computational Intelligence and Design, 2016, vol. 2, pp. 467–470, doi: 10.1109/ISCID.2016.2115.
[17] A. Akula, N. Khanna, R. Ghosh, S. Kumar, A. Das, and H. K. Sardana, “Adaptive contour-based statistical background subtraction
method for moving target detection in infrared video sequences,” Infrared Physics & Technology, vol. 63, pp. 103–109, 2014,
doi: 10.1016/j.infrared.2013.12.012.
[18] M. Kaushaland and B. S. Khehra, “BBBCO and fuzzy entropy-based modified background subtraction algorithm for object detection
in videos,” Applied Intelligence, vol. 47, no. 4, pp. 1008–1021, 2017, doi: 10.1007/s10489-017-0912-5.
[19] R. Kalsotra and S. Arora, “Background subtraction for moving object detection: explorations of recent developments and chal-
lenges,” The Visual Computer,vol.38, pp. 4151-5178, 2022, doi: 10.1007/s00371-021-02286-0.
[20] A. F. Abbas, U. U. Sheikh, F. T. AL-Dhief and M. N. H. Mohd, “A comprehensive review of vehicle detection using com-
puter vision,” TELKOMNIKA(Telecommunication Computing Electronics and Control), vol. 19, no. 3, pp. 838–850, 2021,
doi: 10.12928/TELKOMNIKA.v19i3.12880.
[21] T. Bouwmans, “Recent advanced statistical background modeling for foreground detection-a systematic survey,” Recent Patents on
Computer Science, vol. 4, no. 3, pp. 147–176, 2011, doi: 10.2174/1874479611104030147.
[22] X. Yu, X. Chen and H. Zhang, “Accurate motion detection in dynamic scenes based on ego-motion estimation and op-
tical flow segmentation combined method,” In2011 Symposium on Photonics and Optoelectronics (SOPO), 2011, pp. 1-4,
doi: 10.1109/SOPO.2011.5780637.
[23] M. Shah, J. D. Deng and B. J. Woodford, “Video background modeling: recent approaches, issues and our proposed techniques,”
Machine Vision and Applications, vol. 25, no. 5, pp. 1105–1119, 2014, doi: 10.1007/s00138-013-0552-7.
[24] C. Stauffer and W. E. L. Grimson, “Adaptive background mixture models for real-time tracking,” In Proceedings. 1999 IEEE
Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149), 1999, vol. 2, pp. 246–252,
doi: 10.1109/CVPR.1999.784637.
[25] D. S. Lee, “Effective Gaussian mixture learning for video background subtraction,” IEEE transactions on pattern analysis and
machine intelligence, vol. 27, no. 5, pp. 827–832, 2005, doi: 10.1109/TPAMI.2005.102.
[26] Y. Zhao and Y. Su, “Vehicles detection in complex urban scenes using Gaussian mixture model with FMCW radar,” IEEE Sensors
Journal, vol. 17, no. 18, pp. 5948–5953, 2017, doi: 10.1109/JSEN.2017.2733223.
[27] X. Luo, Y. Wang, B. Cai and Z. Li, “Moving object detection in traffic surveillance video: New MOD-AT method based on adaptive
threshold,” ISPRS International Journal of Geo-Information, vol. 10, no. 11, pp. 742, 2021, doi: 10.3390/ijgi10110742.
[28] H. Wei and W. Zheng, “Image denoising based on improved Gaussian mixture model,” Scientific Programming, vol. 2021, pp. 1–8,
2021, doi: 10.1155/2021/7982645.
[29] Z. Chen and T. Ellis, “A self-adaptive Gaussian mixture model,” Computer Vision and Image Understanding, vol. 122, pp. 35–46,
2014, doi: 10.1016/j.cviu.2014.01.004.
[30] Z. Chen, N. Pears, M. Freeman and J. Austin, “Background subtraction in video using recursive mixture models, spatio-temporal fil-
tering and shadow removal,” in International Symposium on Visual Computing, vol. 5876LNCS, no. PART 2, 2009, pp. 1141–1150,
doi: 10.1007/978-3-642-10520-3-109.
[31] D. Comaniciu and P. Meer, “Mean shift: a robust approach toward feature space analysis,” IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol. 24, no. 5, pp. 603–619, 2002, doi: 10.1109/34.1000236.
[32] K. Kalti and M. A. Mahjoub, “Image segmentation by gaussian mixture models and modified FCM algorithm,” International Arab
Journal of Information Technology, vol. 11, no. 1, pp. 11–18, 2014.
[33] J. Zuo, Z. Jia, J. Yang and N. Kasabov, “Moving target detection based on improved Gaussian mixture background subtraction in
video images,” IEEE Access, vol. 7, pp. 152612-152623, 2019, doi: 10.1109/ACCESS.2019.2946230.

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1007–1018

Int J Artif Intell ISSN: 2252-8938 r 1017

[34] B. K. Yadav and J. Xiaogang, “An Algorithm for intelligent identification of moving objects in natural environment,” International
Journal of Machine Learning and Computing, vol. 10, no. 5, pp. 637–647, 2020, doi: 10.18178/ijmlc.2020.10.5.985.
[35] Z. Chen and T. Ellis, “Self-adaptive Gaussian mixture model for urban traffic monitoring system,” In 2011 IEEE international
conference on computer vision workshops (ICCV Workshops), 2011, pp. 1769-1776, doi: 10.1109/ICCVW.2011.6130463.
[36] I. Martins, P. Carvalho, L. Corte-Real and J. L. Alba-Castro, “BMOG: boosted Gaussian mixture model with
controlled complexity for background subtraction,” Pattern Analysis and Applications, vol. 21, no. 3, pp. 641-654, 2018,
doi: 10.1007/s10044-018-0699-y.
[37] S. Agrawal and P. Natu, “An improved Gaussian mixture method-based background subtraction model for moving object detection
in outdoor scene,” In 2021 4th International Conference on Electrical, Computer and Communication Technologies (ICECCT),
2021, pp. 1–8, doi: 10.1109/ICECCT52121.2021.9616883.
[38] Y. Su, “Target detection algorithm and data model optimization based on improved Gaussian mixture model,” Microprocessors and
Microsystems, vol. 81, pp. 103797, 2021, doi: 10.1016/j.micpro.2020.103797.
[39] Y. Xu, J. Dong, B. Zhang and D. Xu, “Background modelling methods in video analysis: a review and comparative evaluation,”
CAAI Transactions on Intelligence Technology, vol. 1, no. 1, pp. 43–60, 2016, doi: 10.1016/j.trit.2016.03.005.
[40] X. Jin, P. Niu, and L. Liu, “A GMM-based segmentation method for the detection of water surface floats,” IEEE Access, vol. 7,
pp. 119018-119025, 2019, doi: 10.1109/ACCESS.2019.2937129.
[41] L. L. Lin and N. R. Chen, “Moving objects detection based on gaussian mixture model and saliency map,” In Applied Mechanics
and Materials, vol. 63-64, pp. 350–354, 2011, doi: 10.4028/www.scientific.net/AMM.63-64.350.
[42] X. Ou et al., “Adaptive GMM and BP neural network hybrid method for moving objects detection in complex scenes,” . International
Journal of Pattern Recognition and Artificial Intelligence, vol. 33, no. 2, pp. 1950004, 2019, doi: 10.1142/S0218001419500046.
[43] C. Zhang, X. Wu and X. Gao, “An improved Gaussian mixture modelling algorithm combining foreground matching and short-
term stability measure for motion detection,” Multimedia Tools and Applications, vol. 79, no. 11-12, pp. 7049–7071, 2020,
doi: 10.1007/s11042-019-08210-y.
[44] K. A. B. Lima, K. R. T. Aires and F. W. P. D. Reis, “Adaptive method for segmentation of vehicles through local thresh-
old in the gaussian mixture model,” In 2015 Brazilian Conference on Intelligent Systems (BRACIS), 2015, pp. 204–209,
doi: 10.1109/BRACIS.2015.33.
[45] X. Cao, L. Yang and X. Guo, “Total variation regularized RPCA for irregularly moving object detection under dynamic background,”
IEEE Transactions on Cybernetics, vol. 46, no. 4, pp. 1014–1027, 2015, doi: 10.1109/TCYB.2015.2419737.
[46] S. Puhan, D. K. Rout and N. K. Kamila, “Slow and fast moving object detection under illumination variation condition,” American
Journal of Signal Processing, vol. 3, no. 5, pp. 121–131, 2013, doi: 10.5923/j.ajsp.20130305.01.
[47] P. K. Sahoo, P. Kanungo and S. Mishra, “A fast valley-based segmentation for detection of slowly moving objects,” Signal, Image
and Video Processing, vol. 12, no. 7, pp. 1265-1272, 2018, doi: 10.1007/s11760-018-1278-9.
[48] B. N. Subudhi and P. K. Nanda, “Detection of slow-moving video objects using compound Markov random Field model,” In
TENCON 2008-2008 IEEE Region 10 Conference, 2008, pp. 1–6, doi: 10.1109/TENCON.2008.4766385.
[49] A. Utasi and L. Czúni, “Reducing the foreground aperture problem in mixture of Gaussians based motion detection,” In 2007 14th
International Workshop on Systems, Signals, and Image Processing and the 6th EURASIP Conference focused on Speech and Image
Processing, 2007, pp. 157–160, doi: 10.1109/IWSSIP.2007.4381177.
[50] H. Fu, H. Ma and A. Ming, “EGMM: an enhanced Gaussian mixture model for detecting moving objects with intermittent stops,”
In 2011 IEEE International Conference on Multimedia and Expo, 2011, pp. 1–6, doi: 10.1109/ICME.2011.6012011.
[51] X. Kai, C. Wei, and L. Liu, “Robust extended Kalman filtering for nonlinear systems with stochastic uncertainties,”
IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and Humans, vol. 40, no. 2, pp. 399–405, 2009,
doi: 10.1109/TSMCA.2009.2034836.
[52] C. Urrea and R. Agramonte, “Kalman filter: historical overview and review of its use in robotics 60 years after its creation,” Journal
of Sensors, vol. 2021, pp. 1-21, 2021, doi: 10.1155/2021/9674015.
[53] L. Torres, J. Jiménez-Cabas, O. González, L. Molina, and F. R. López-Estrada, “Kalman filters for leak diagnosis in pipelines: brief
history and future research,” Journal of Marine Science and Engineering, vol. 8, no. 3, pp. 173, 2020, doi: 10.3390/jmse8030173.
[54] Y. Zhang, C. Zhao, J. He and A. Chen, “Vehicles detection in complex urban traffic scenes using a nonparametric approach with
confidence measurement,” In 2015 International Conference and Workshop on Computing and Communication (IEMCON) , 2015,
pp. 1-7, doi: 10.1109/IEMCON.2015.7344486.
[55] K. Toyama, J. Krumm, B. Brumitt and B. Meyers, “Wallflower: principles and practice of background maintenance,” In Proceedings
of the seventh IEEE international conference on computer vision, 1999, vol. 1, pp. 255-261, doi: 10.1109/ICCV.1999.791228.
[56] D. Farin, P. H. de With and W. Effelsberg, “Robust background estimation for complex video sequences,” In Proceedings 2003
International Conference on Image Processing, 2003, vol. 1, pp. 1-145, doi: 10.1109/ICIP.2003.1246919.
[57] V. Reddy, C. Sanderson, B. C. Lovell and A. Bigdeli, “An efficient background estimation algorithm for embedded smart
cameras,” In 2009 Third ACM/IEEE International Conference on Distributed Smart Cameras (ICDSC), 2009, pp. 1-7,
doi: 10.1109/ICDSC.2009.5289348.
[58] A. Colombari and A. Fusiello, “Patch-based background initialization in heavily cluttered video,” IEEE Transactions on Image
Processing, vol. 19, no. 4, pp. 926-933, 2009, doi: 10.1109/TIP.2009.2038652.
[59] M. Amintoosi, F. Farbiz, M. Fathy, M. Analoui and N. Mozayani, “QR decomposition-based algorithm for background subtraction,”
In 2007 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP’07), 2007, vol. 1, pp. I-1093-I–1096,
doi: 10.1109/ICASSP.2007.366102.
[60] M. Harville, G. Gordon and J. Woodfill, “Foreground segmentation using adaptive mixture models in color and depth,” In Proceed-
ings IEEE workshop on detection and recognition of events in video, 2001, pp. 3-11, doi: 10.1109/EVENT.2001.938860.

Background subtraction challenges in motion detection using ... (Nor Afiqah Mohd Aris)
1018 r ISSN: 2252-8938

BIOGRAPHIES OF AUTHORS

Nor Afiqah Mohd Aris received her Bachelor of Science (Mathematics Technology)
in 2019 and Master of Science (Mathematics) in 2021 from Universiti Tun Hussein Onn Malaysia,
Malaysia, before pursuing a Doctor of Philosophy of Science (Mathematics) in 2022. She is currently
a Graduate Research Assistant at the Faculty of Applied Sciences and Technology, Universiti Tun
Hussein Onn Malaysia, Malaysia. She runs research in image processing for real-time traffic flow
analysis supported by the Ministry of Higher Education (MOHE) Malaysia through the Fundamental
Research Grant Scheme (FRGS). Her research interests include smart city, pattern recognition and
machine learning. She can be contacted at email: [email protected].

Ts. Dr. Siti Suhana Jamaian received her first degree in Industrial Mathematics from Uni-
versiti Teknologi Malaysia in 2007. She then furthered her study with a Master’s degree in the same
university and graduated in 2009. In 2013, she completed her PhD in Applied and Computational
Mathematics at the School of Mathematics, The University of Edinburgh. She is now a Senior Lec-
turer in the Department of Mathematics and Statistics, Faculty of Applied Sciences and Technology,
Universiti Tun Hussein Onn Malaysia. Her current research interests in applied mathematics include
the chemostat model, nutrient removal, industrial wastewater, parameter estimation and traffic flow.
She can be contacted at email: [email protected].

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1007–1018