Machine Learning-Based Anomaly Detection in Smart City Traffic:
Performance Comparison and Insights
Mohammad Bawaneh
a
and Vilmos Simon
b
Department of Networked Systems and Services, Faculty of Electrical Engineering and Informatics, Budapest University of
Technology and Economics, M
˝
uegyetem rkp. 3., H-1111 Budapest, Hungary
Keywords:
Anomaly Detection, Machine Learning, Traffic Congestion, Intelligent Transportation Systems, Smart City,
Sustainability.
Abstract:
In recent years, urban roads have suffered from substantial traffic congestion due to the rapidly increasing
number of road users and vehicles. Some traffic congestion patterns on specific roadways, such as the recur-
ring congestion during morning and evening rush hours, can be foreseen. However, unexpected events, such
as incidents, may also cause traffic congestion. Monitoring traffic status poses vital importance for city traffic
operators. They can leverage the monitoring system for resource allocation, traffic lights adjusting, and adapt-
ing the public transport schedules to alleviate traffic congestion. Machine learning-based methods for anomaly
detection are valuable tools for monitoring traffic status and promptly detecting congestion on city roads. In
this paper, we comprehensively study the performance of the common machine learning methods for anomaly
detection in the traffic congestion detection use case. In addition, we provide methods usage insights based on
the study findings by examining the accuracy, detection speed, and computation overhead of the methods to
guide the researchers and city operators toward a suitable method based on their needs.
1 INTRODUCTION
In recent years, metropolitan cities have witnessed a
significant increase in traffic congestion due to the in-
creasing population in urban cities (Gonz
´
alez-Aliste
et al., 2023). Therefore, the burden of traffic conges-
tion has made a direct negative impact on the drivers
and commuters’ daily lives. In addition, traffic con-
gestion significantly slows down the development of
urban cities and weakens the efforts of achieving sus-
tainable smart transport (Nguyen and Jung, 2021). A
potential solution to reduce traffic congestion is to
change the roads’ infrastructure, such as opening new
roads and widening the existing roads that witness
high daily traffic. However, this solution is costly
and sometimes inefficient (Zhu et al., 2021). In ad-
dition, it is not adaptable to different traffic situations
where some roads are susceptible to frequent changes
in traffic status. Therefore, Intelligent Transportation
Systems (ITS) intend to find smart solutions to re-
duce traffic congestion without the need to change
the roads’ infrastructures to avoid the aforementioned
shortcomings (Zhang et al., 2011).
a
https://orcid.org/0000-0003-4402-9254
b
https://orcid.org/0000-0002-7627-3676
ITS integrates information, data analytics, ma-
chine learning, and advanced communications tech-
nologies into one system that can monitor and man-
age the city’s traffic in an automated way (Cheng
et al., 2020). In order to monitor the traffic status and
take actions to prevent or mitigate traffic congestion,
the data-driven system must be capable of analyz-
ing and understanding the underlying traffic patterns.
The traffic data can be provided by vehicles’ trajec-
tories or by measurements collected by roadside sen-
sors. The vehicles’ trajectories can be collected us-
ing the Global Positioning System (GPS) technology,
which collects the movement data of the equipped ve-
hicle, such as the vehicle’s speed and location per
time (Sousa et al., 2020). On the other hand, roadside
sensors, such as inductive loops and cameras, collect
aggregated traffic data such as the traffic flow (ve-
hicles’ count), speed, and occupancy in close vicin-
ity of the sensor’s location (Tasgaonkar et al., 2020).
By collecting this data, various artificial intelligence
(AI) algorithms could be utilized to learn the patterns,
which can help the city operators manage the traffic
and introduce smart transportation services. Some of
the algorithms that have already discovered its path
to ITS are: Convolutional Neural Network (CNN) (Li
Bawaneh, M. and Simon, V.
Machine Learning-Based Anomaly Detection in Smart City Traffic: Performance Comparison and Insights.
DOI: 10.5220/0013141100003941
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 11th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2025), pages 309-318
ISBN: 978-989-758-745-0; ISSN: 2184-495X
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
309

et al., 2017), Graph Convolutional Neural Networks
(GCNs) (Zheng et al., 2023), Long Short-Term Mem-
ory (LSTM) (Zhao et al., 2017), Gated Recurrent
Units (GRUs) (Fu et al., 2016), Bidirectional-Long
Short-Term Memory (Bi-LSTM) (Redhu et al., 2023),
K-Means Clustering (Rao et al., 2019; Pustokhina
et al., 2020), Autoencoders (Li et al., 2022), and Gen-
erative Adversarial Network (GAN) (Shi et al., 2024).
AI algorithms can provide traffic prediction (Nagy
and Simon, 2018; Lee et al., 2021), and traffic conges-
tion detection and prediction services (Djenouri et al.,
2019; Akhtar and Moridpour, 2021). This paper sheds
light on automatizing the process of monitoring traf-
fic status by employing the most common machine
learning-based anomaly detection methods to detect
unusual behavior in road traffic.
Numerous research papers have put forth algo-
rithms for the detection of traffic congestion. The first
category of these research papers relies on traffic data
derived from the trajectories of vehicles, as presented
by references (Pang et al., 2011; Pan et al., 2013;
Kuang et al., 2015; Yu et al., 2017; Kong et al., 2018;
Kohan and Ale, 2020; Zhang et al., 2021; Qin et al.,
2021; He et al., 2023; Makara et al., 2023). However,
it is worth noting that trajectory data, while easier and
cheaper to collect compared to the costs of installing
and operating sensor stations throughout the city net-
work, comes with certain limitations. This data is
typically collected by public transportation vehicles
that have GPS devices, which operate on fixed routes
and schedules, making them less precise in capturing
citywide traffic congestion. Additionally, data from
connected cars could offer more benefits, but the pen-
etration of such vehicles remains low when compared
to the overall vehicle population, and access to these
datasets is often restricted due to their proprietary na-
ture and high market prices. Navigation mobile apps
also collect this kind of data; however, access to such
datasets is restricted.
In the subsequent category of research papers
from the field, aggregated data from roadside sen-
sors is utilized. Various congestion detection defi-
nitions have been developed, including Travel Time
Index (TTI) (Hasanzadeh et al., 2017), Traffic Con-
dition (TC) (Fouladgar et al., 2017), and the Conges-
tion Level (CL) (Chen et al., 2018). These definitions
check and compare the current traffic state (i.e., speed
or travel time) with a free-flow state and trigger a
congestion alert if a significant deviation is observed.
Other statistical-based algorithms have also been pro-
posed, such as the ones mentioned in (Lam et al.,
2017; Nguyen et al., 2016; Kalair and Connaughton,
2021; Bhardwaj et al., 2023; Xu and Li, 2024), which
classify traffic states as congested when they deviate
from a predefined null hypothesis. While statistical-
based algorithms are efficient and straightforward to
implement, they do not acquire correlations between
different traffic states, potentially leading to more
false congestion alerts. Several algorithms have in-
troduced distance-based approaches, such as the K-
Nearest Neighbors Outlier Detection (KNN OD) al-
gorithm (Dang et al., 2015), the Bounded Local Out-
lier Factor (BLOF) algorithm (Tang and Ngan, 2016),
and the Piecewise Switched Linear Traffic-Kalman
Filter based K-Nearest Neighbors (PWSL-KF-based
KNN) algorithm (Harrou et al., 2020). These meth-
ods employ similarity measures to detect dissimilar
traffic states as congestion. However, KNN OD and
PWSL-KF-based KNN may miss certain instances of
congestion when the traffic data exhibits varying dis-
tribution densities. On the other hand, BLOF can
handle traffic data with varying densities, thanks to
its density-based outlier detection nature, but at the
expense of higher computational complexity. Deep
learning techniques have also been applied for traffic
congestion detection, as seen in references (Zhu et al.,
2019; Li et al., 2020; Liu et al., 2023). However, these
algorithms often need more computational complex-
ity, which hinders fast congestion detection, as they
require multiple scans of the traffic data.
Our goal in this paper is to examine the perfor-
mance of the most common machine learning meth-
ods for anomaly detection in the traffic congestion de-
tection use case, as researchers, ITS developers, and
city traffic operators could get easily lost in the abun-
dance of these methods. We could not find a paper
where the most essential methods are compared on
the same dataset, using the same use case scenario
and parameter settings, measuring the comprehensive
performance metrics to show a thorough picture of
the methods’ performance. The papers from the lit-
erature usually validate only a few congestion detec-
tion methods in their performance testing, which are
based on similar principles. Therefore, we have cho-
sen methods that originate from different solution ar-
eas: the Modified Z-Score Method (MZ), the Isola-
tion Forest (IF), the Local Outlier Factor (LOF), the
Support Vector Machine (SVM), and the Long Short
Term Memory (LSTM).
Many aspects are taken into account to provide
a comprehensive overview of each method’s perfor-
mance. The methods are tested for their ability to
provide reliable and accurate congestion detection.
In addition, the speed of detecting the congestion is
measured to verify the methods’ effectiveness for the
traffic management process. In other words, the con-
gestion detection alarm must be triggered in the early
phases of the congestion to be helpful if the traffic
VEHITS 2025 - 11th International Conference on Vehicle Technology and Intelligent Transport Systems
310

management center wants to take actions to prevent
or mitigate the propagation of the detected congestion
to the neighboring roads (Nagy and Simon, 2021).
Moreover, the computation overhead of running the
methods is measured to test their suitability for edge
computing, where the sensors collect the data and
run the machine learning methods simultaneously to
monitor the traffic.
The remainder of this research paper is organized
as follows. Section 2 presents the anomaly detec-
tion methods used in this research, while Section 3
presents the experimental results, performance com-
parison, and insights. Finally, the paper is concluded
in Section 4.
2 ANOMALY DETECTION
METHODS OVERVIEW
This paper follows the second category of research
papers in the literature, which utilizes aggregated traf-
fic data (flow, speed, and occupancy) from roadside
sensors to detect traffic congestion. In particular, this
paper detects congestion by detecting anomalies in
traffic time series data. The speed and occupancy
traffic features are utilized as their behavior during
the congestion period reflects an anomaly compared
to the free-flow period. In the free-flow period (i.e.,
the normal behavior of the traffic), the speed values
will be high. However, when the congestion starts
(i.e., the abnormal traffic behavior), the drivers have
to slow down their vehicles, which will be reflected
in the low-speed values in the speed time series data.
When the congestion ends, the speed values will re-
turn to the high values as the free-flow period. Con-
sequently, abnormal traffic behavior can be found by
finding the anomaly in the speed time series data, and
hence, traffic congestion can be detected. Similarly,
the occupancy values will be low during the free-flow
traffic and high during the congestion period. There-
fore, as the free-flow traffic is the normal behavior
and the congestion is typically the abnormal behav-
ior, detecting the anomaly in the occupancy time se-
ries leads to detecting the congestion. On the con-
trary, this is not the case for the flow traffic feature,
as the anomaly in the flow time series does not ensure
a congestion behavior. We refer the reader to the ar-
ticles (Chakroborty and Das, 2017; Garber and Hoel,
2019; Bawaneh, 2023) for detailed information on the
fundamental behavior of traffic flow data. Based on
the explanation above of the behavior of traffic flow
data, traffic congestion can be detected by studying
the data and detecting the anomalies in the speed and
occupancy time series.
This section provides an overview of the anomaly
detection algorithms used in this paper to detect traf-
fic congestion. They have been chosen in this pa-
per because of their popularity and effectiveness for
anomaly detection problems accompanied by easy
and free access to code implementations, which fa-
cilitates the city operators’ work of automating city
traffic monitoring (Nassif et al., 2021; Scikit-learn,
2024; TensorFlow, 2024). Moreover, they represent
different solution domains (statistical, machine learn-
ing, and deep learning), which leads to a more com-
prehensive study.
2.1 Modified Z-Score Method
The modified Z-score (MZ) method is a simple yet
effective statistical technique for finding outliers or
anomalies within a dataset (Blaise et al., 2020). It rep-
resents a variation of the traditional Z-score method.
The traditional z-score method measures how many
standard deviations a particular data point deviates
from the mean of the dataset. It is computed by sub-
tracting the mean from the data point’s value and di-
viding the result by the value of the standard devi-
ation. The computed z-score is then thresholded to
classify the data point as an outlier or normal value.
Its variation (i.e., the modified z-score method) works
with a similar principle. However, the modified ver-
sion uses the median and the median absolute devia-
tion values instead of the mean and the standard de-
viation. This modification provides the method with
more robustness against extreme values, as the mean
and the standard deviation can be significantly af-
fected by the extreme values. In addition, the mod-
ified z-score method does not require a particular data
distribution, in contrast to the traditional one, which
assumes a normal distribution for the data.
2.2 Isolation Forest
Isolation Forest is a machine learning algorithm used
for anomaly detection (Liu et al., 2008; Togbe et al.,
2020; Chen et al., 2020; Liu et al., 2021). Isola-
tion Forest works by recursively randomly partition-
ing the dataset. It selects a random split point at each
step. This randomness in partitioning provides the al-
gorithm with more effectiveness. The algorithm con-
structs a binary tree collection called Isolation Trees.
Each tree is constructed by recursively partitioning
the data until it isolates individual data points. The
isolation process involves randomly selecting split
points to create splits in the data. Data points that
are isolated quickly (with fewer splits) are likely to
be anomalies, while those that require many splits are
Machine Learning-Based Anomaly Detection in Smart City Traffic: Performance Comparison and Insights
311

considered normal. The anomaly score for each data
point is determined by the path length required to iso-
late that point across all the trees in the forest. Data
points that have shorter-than-average path lengths are
more likely to be anomalies, as they were easier to
isolate. A threshold is set to identify anomalies. Data
points with path lengths less than this threshold are
considered anomalies, while those over the threshold
are considered normal.
2.3 Local Outlier Factor
The Local Outlier Factor (LOF) is a machine learning
algorithm used for identifying local anomalies or out-
liers within a dataset (Breunig et al., 2000; Djenouri
et al., 2018;
ˇ
Sabi
´
c et al., 2021). LOF assesses the
relative density of data points with respect to their lo-
cal neighborhoods to detect regions of the data where
points are significantly less dense than their neigh-
bors. LOF starts by estimating the local density of
each data point within its neighborhood. A neighbor-
hood is defined around each data point by consider-
ing its k nearest neighbors, where k is a pre-defined
parameter. The density of a data point is measured by
considering the average density of its k nearest neigh-
bors. Data points in densely populated regions will
have a higher local density, while those in sparser re-
gions will have a lower local density. For each data
point, LOF calculates its local outlier factor by com-
paring its local density to the local densities of its
neighbors. The LOF of a point is defined as the ratio
of its own local density to the average local density of
its k nearest neighbors. LOF values provide a mea-
sure of how much a data point deviates from its local
neighborhood’s density. Data points with high LOF
values are considered local outliers because they have
significantly lower densities than their neighbors. Set-
ting a threshold can determine which points are con-
sidered outliers based on their LOF values. Points
with LOF values above the threshold are identified as
outliers.
2.4 Support Vector Machine
A One-Class Support Vector Machine (One-Class
SVM) is a machine learning algorithm used for
anomaly detection. It is a special use case of the Sup-
port Vector Machine (Hearst et al., 1998; Pisner and
Schnyer, 2020; Wang et al., 2020; Ioannou and Vas-
siliou, 2021). It is designed to detect anomalies by
learning the characteristics of the majority class (nor-
mal instances) and then flagging any instances that de-
viate significantly from this majority class. One-Class
SVM trains on a dataset that contains mostly normal
instances. The algorithm’s goal is to create a decision
boundary (or hyperplane) that covers the majority of
the normal data points while minimizing the number
of anomalies within this boundary. The hyperplane
is positioned in a way that maximizes the margin be-
tween the hyperplane and the normal data points. This
margin represents the region within which most nor-
mal instances are expected to lie. During training,
the algorithm identifies a subset of data points known
as support vectors. These are the data points closest
to the hyperplane and are essential for defining the
margin. The margin is the distance between the hy-
perplane and the closest support vectors. One-Class
SVM aims to maximize this margin while minimizing
the number of support vectors. Once trained, One-
Class SVM can classify new data points into normal
and anomaly. Normal data points fall within the mar-
gin, while anomaly data points fall outside the margin.
2.5 Long Short Term Memory
Long Short Term Memory (LSTM) (a type of Re-
current Neural Network (RNN)) is a deep learning
method that is well-suited for modeling sequential
data due to its ability to capture long-term depen-
dencies (i.e., avoiding the vanishing gradient prob-
lem), which was a shortcoming in the traditional RNN
(Hochreiter and Schmidhuber, 1997; Tan et al., 2020;
Markovic et al., 2023). The LSTM model works by
integrating units. The unit consists of a cell state, for-
get gate, input gate, and output gate. The cell state
represents the long-term memory, which can be con-
trolled by the forget, input, and output gates. The
forget gate controls the percentage of the memory in
the cell state that should be kept or removed. The
input gate controls the addition or the update of the
cell state based on the new information. The output
gate controls which information should be the out-
put of the unit. The units are connected sequentially,
forming a chain of LSTM units. For each unit, an
input is received from the actual current input value
in addition to the previous output value of the previ-
ous unit. These connections provide an effective pro-
cess of sequential data, where the dependencies are
captured over time. Due to the recurrent nature of
such a structure, the information is propagated across
the network. Consequently, LSTM can learn com-
plex temporal patterns and dependencies. LSTM does
not detect the anomalies in the data directly. It can
be used to train a model to forecast time series data.
Once the LSTM model is trained, it can be used to
make predictions on new unseen data. Anomalies are
typically detected by comparing the model’s predic-
tions to the actual data points. A high deviation be-
VEHITS 2025 - 11th International Conference on Vehicle Technology and Intelligent Transport Systems
312

tween the predicted and the actual values indicates an
anomaly. Therefore, a threshold must be trained to
detect the anomalies points.
Considering the space limitation of the paper, a
brief overview of the anomaly detection methods uti-
lized has been presented. We refer the readers to
the original papers for more details and mathemati-
cal descriptions of the methods. As mentioned pre-
viously, prior studies typically focus on validating a
limited subset of anomaly detection methods, often
based on similar principles (for example, statistical-
based methods). However, in this paper, a diverse ar-
ray of methods spanning different solution domains
has been chosen for this study to assess their effi-
ciency for traffic congestion detection. We have cho-
sen popular statistical-based, machine learning-based,
and deep learning-based methods to cover widespread
solution domains. Despite the spread of such meth-
ods for solving other research problems, existing lit-
erature lacks a comprehensive comparison conducted
under standardized conditions (i.e., using the same
dataset, scenario, and parameter settings) and cover-
ing the overall performance points of view, enabling
a thorough evaluation of their performance for traf-
fic congestion detection. The next section of this re-
search not only offers a comprehensive understand-
ing of their comparative performance but also pro-
vides effective practical recommendations for opti-
mizing congestion detection strategies. This study
thus serves as a pivotal resource for enhancing the
effectiveness of traffic management systems, which
contributes to achieving safer, sustainable, and more
efficient urban mobility.
3 EXPERIMENTAL RESULTS,
PERFORMANCE
COMPARISON, AND INSIGHTS
In the previous section, several methods for anomaly
detection were presented briefly. These methods
have been successfully used for various domains and
have shown superior performance (Blaise et al., 2020;
Chen et al., 2020;
ˇ
Sabi
´
c et al., 2021; Wang et al.,
2020; Markovic et al., 2023). However, their perfor-
mance for traffic congestion detection tasks was not
previously compared with each other. Therefore, tak-
ing into account the diversity in the solution domains,
a comprehensive comparison that shows the entire
performance picture under unified and standardized
conditions is presented in this section.
As mentioned previously, this paper provides an
experimental verification of employing these meth-
ods for the traffic congestion detection use case. The
methods are tested for accuracy, detection speed,
and computation complexity. Consequently, this pa-
per provides the needed comprehensive evaluation to
present the complete picture of the performance of
each method based on the traffic management oper-
ators’ needs. The following subsection presents the
real-life acquired dataset utilized in this paper to con-
duct the experiment.
3.1 The Utilized Dataset
The California Department of Transportation pro-
vides the researchers with the Caltrans Performance
Measurement System (PeMS) real-time collected
traffic dataset (PeMS, 2020). The PeMS dataset also
provides the incident information. This is an advan-
tage of utilizing it to conduct this paper’s experiment
because the incident usually causes congestion.
A separate traffic incident dataset has been ex-
tracted by Nagy et al. (Nagy et al., 2021; Nagy and
Simon, 2020) from the PeMS dataset. The extracted
data covers one year of the traffic dataset (i.e., from
January 1, 2016, to December 31, 2016). It contains
incidents that happened in different locations in Dis-
trict 3. The aggregation period is 5 minutes, which
means that every sample shows the traffic informa-
tion (speed, occupancy, or flow) for 5 minutes. The
total traffic information is collected in a PARQUET
file with a size of 183 MB. The traffic incident dataset
has been extracted to contain only the incidents that
caused congestion on the roads, according to (Nagy
et al., 2021). This makes it suitable to conduct the
experiment on traffic congestion detection.
This paper is experimenting with 100 incidents
from the dataset that occurred in 2016, where 50%
of the dataset was used for training and 50% for
testing. The incidents have occurred in different lo-
cations, which helps to verify the robustness of the
tested anomaly detection methods. We have tested
the methods on 12 hours of traffic for each incident,
where the congestion happened in the middle of this
period. Python programming language has been used
to implement all the methods. As explained in section
2, and based on the fundamental behavior of traffic
congestion, the speed and occupancy traffic features
have been used in this paper. All the methods have
been tested by using both of them in order to evaluate
their performance in both cases.
3.2 Experimental Results
As an initial step, common to all these algorithms, we
perform data cleaning and preprocessing, a necessary
Machine Learning-Based Anomaly Detection in Smart City Traffic: Performance Comparison and Insights
313

step as the data often contains noise and may have
some missing values for certain samples. In the data
cleaning process, we employ the forward-fill method
to fill in any missing sample values. Subsequently, we
apply the exponential moving average technique to
mitigate noise and to smooth the traffic data. To illus-
trate, for a specific traffic feature such as road speed,
flow, or occupancy, we define a time series denoted as
X with n observations in the following manner:
X = (X
1
, X
2
, ..., X
n
) (1)
where the time sequence in which the observations
(speed or occupancy) were captured by a certain sen-
sor is indicated by the indices (1, 2, ..., n). The defini-
tion of the exponential moving average is:
T
t
=

X
1
t = 1
α · X
t
+ (1 − α) · T
t−1
t > 1
(2)
Each of the utilized methods has its own param-
eters that must be configured. A common example
of these parameters between the utilized methods is
the threshold that classifies the sample point as nor-
mal or anomaly point by thresholding the anomaly
score obtained by the anomaly detection method. The
outcome can be significantly affected by the choice
of these parameters. Hence, in order to conduct a
fair comparison, all methods must be tuned to achieve
their best performance, which can be achieved by ap-
plying the method with the best parameter configu-
ration. Consequently, to find the optimal parameters
for each method, the dataset under the study is split
into a training dataset and a testing dataset, where the
training dataset is used to find the optimal parame-
ters that can achieve the highest performance of the
method, which can be measured by the used perfor-
mance metrics in this paper. Grid search has been
used to obtain the optimal parameters for all the uti-
lized methods. The determined optimal parameters
are after applied to the testing dataset to measure and
conclude the method’s performance. This ensures that
all the methods used have been fairly tuned to achieve
the best possible results, which in turn ensures a fair
comparison. In addition, we note that separate learn-
ing for the parameters has been done for each traffic
feature (i.e., speed and flow).
The experiment has been conducted by employ-
ing the testing dataset to evaluate the efficiency of
each anomaly detection method for the traffic con-
gestion detection use case. The standard performance
evaluation metrics in the literature for evaluating traf-
fic congestion detection methods have been used in
this study (AlDhanhani et al., 2019; Kalair and Con-
naughton, 2021). In addition to these metrics, we
have compared the run time of each method to com-
pare their computation complexity and suitability for
edge computing. The used metrics are defined as fol-
lows:
• Detection Rate (DR): The percentage of conges-
tions that have been correctly detected.
• False Alarm Rate (FAR): The percentage of false
congestion alarms to the number of congestion-
free instances.
• Mean Time to Detection (MTTD): The average
time needed to detect the congestions (i.e., the de-
tection speed). It is measured based on the differ-
ence between the timestamp where the congestion
occurred and the timestamp where the congestion
has been detected by the method.
• Runtime: The computation run time needed by
the method to conduct the experiment.
Table 1 shows the experimental results of this
study for all the utilized methods.
Table 1: Performance Comparison: the outcome of the ex-
periments of this study in terms of accuracy and time.
Method Traffic Feature DR FAR
MTTD
(minutes)
Runtime
(seconds)
MZ Speed 94% 6.8% 9.3 1.8
MZ Occupancy 96% 17% 7.5 1
IF Speed 94% 7.4% 5.6 13.9
IF Occupancy 90% 5.5% 4 13.6
LOF Speed 98% 21% 2.8 1.45
LOF Occupancy 88% 10% 8 1.2
SVM Speed 98% 6.5% 5.6 1.5
SVM Occupancy 94% 4% 5.2 5.1
LSTM Speed 96% 11.7% 14 10397
LSTM Occupancy 96% 18% 7.9 11284
3.3 Discussion and Recommendations
Based on Table 1, it can be generalized that high DR
can be achieved with most of the methods. Therefore,
the anomaly detection methods (MZ, IF, LOF, SVM,
and LSTM) have verified their ability to achieve a
reliable system for detecting congestion. Typically,
the speed readings during the congestion period will
reach a significant drop compared to the free flow
readings, which makes it easier for the anomaly detec-
tion methods to capture the congestion. As a conse-
quence, the methods have shown better performance
using the speed traffic feature than the occupancy in
terms of DR. The best performance has been achieved
by LOF and SVM using the speed traffic feature with
a DR of 98%.
In terms of FAR, which reflects the robustness of
the method to the noise and false alarms, SVM and
IF have shown better performance than the rest of the
competing methods.
VEHITS 2025 - 11th International Conference on Vehicle Technology and Intelligent Transport Systems
314

The speed of the detection has been found to vary
between the methods. The fastest method to detect the
congestion was LOF using the speed traffic feature,
achieving only 2.8 minutes of delay. However, a high
delay occurred when LOF used the occupancy traf-
fic feature. Such behavior may happen as each traf-
fic feature explains the traffic from a different point
of view. Therefore, one traffic feature may reflect
the congestion behavior in the data before the other
traffic feature (Bawaneh and Simon, 2023). In gen-
eral, SVM and IF have achieved acceptable MTTD
using both features (the speed and occupancy traffic
features). On the other hand, LSTM has shown the
highest delay among the anomaly detection methods
when using the speed traffic feature.
In this study, we added another important metric
to the comparison: the runtime of the algorithms on
the utilized dataset. This metric measures the com-
putation complexity of the methods and indicates the
computation resources needed to utilize these meth-
ods. It can be noticed in Table 1 that all methods re-
quired roughly low time to conduct the experiment
except for LSTM. LSTM required around 3 hours
(more or less) to run the calculations, while the high-
est runtime needed by the rest of the methods was
13.9 seconds.
Traffic control operators may utilize traffic con-
gestion detection for different needs, such as utiliz-
ing the congestion alarm information to adjust traf-
fic lights to reduce the effect of the congestion on
the neighboring areas (Cao et al., 2016; Tseng and
Ferng, 2021). These needs require different perfor-
mance standards. Thus, we suggest in this paper three
main scenarios to provide guidance for traffic control
operators according to the efficiency of each method:
• Scenario I: Traffic management center operators
intervene to reduce the effect of the congestion by
adjusting traffic lights.
• Scenario II: Traffic management center operators
provide a city mobility application to the city-
dwellers to suggest routes to avoid congestions.
• Scenario III: Traffic management center operators
apply edge computing (i.e., the roadside sensors
collect the data and run the traffic congestion de-
tection locally in parallel).
The traffic congestion detection system can be
considered reliable when it can achieve high DR. This
means that the system is more likely not to miss de-
tecting any congestion. Consequently, traffic manage-
ment centers can safely deploy and rely on such an
automated system. The three scenarios require high
DR to provide system reliability. However, if this
system suffers from a high number of false alarms,
actions that are not needed might be taken with the
intention to mitigate the wrongly detected congestion
effect on the roads network (for example, wrong re-
source allocation). The effect of this is highly depen-
dent on how the traffic congestion detection system is
utilized. For example, suppose the traffic congestion
detection system is needed only for monitoring pur-
poses, such as providing a city mobility application
to provide road users with traffic status information
(Scenario II). In that case, the negative impact of a
high FAR can be dealt with. On the other hand, if the
system is meant to be utilized to intervene to mitigate
the effect of congestion, the cost of a high FAR will
be significant. On one side, it will overload the work
of the traffic management centers with unnecessary
actions, and on the other side, it might lead to unde-
sired events, such as causing real traffic congestion
or resource wasting. An example of that is if wrong
decisions have been made to adjust the traffic lights
(Scenario I) while there was actually no congestion,
congestion might occur as a result of such wrong ad-
justing. Therefore, low FAR is recommended for the
chosen system to prevent such effects.
From another point of view, the detection speed
is a crucial metric for all Scenarios. Fast detection
(i.e., detection with low delay) eases the congestion
effects mitigation work of the traffic operators by pro-
viding them with sufficient time to take proper ac-
tions. These actions may include adapting the traf-
fic lights, sending suggestions for lighter traffic routes
for the drivers, or increasing the operated public trans-
port services.
Last but not least, if edge computing is to be
employed in smart city infrastructure (Scenario III),
the computation complexity of the used method is
a key factor. Even without adopting edge comput-
ing and parallelizing the computations, traffic control
centers in major cities with thousands of measure-
ment points can become overloaded with these cal-
culations. For example, deep learning-based methods
(such as LSTM) are time-consuming, and hence, they
cannot fulfill a low computation complexity to sup-
port edge computing. This can be seen in the perfor-
mance of LSTM, which confirms the time-consuming
limitation of deep learning methods in general, even
if they can achieve powerful results. The recorded
runtime shows that the other methods can support
the new era of Tiny Machine Learning (TinyML),
where machine learning models are deployed on de-
vices with limited computational resources (such as
roadside units). TinyML allows to make intelligent
decisions locally without relying on cloud infrastruc-
ture (or traffic control centers) by running lightweight
and efficient machine learning algorithms directly on
Machine Learning-Based Anomaly Detection in Smart City Traffic: Performance Comparison and Insights
315

the roadside units, which enables better real-time pro-
cessing, lower latency, increased privacy, and reduced
reliance on cloud infrastructure. Table 2 summarizes
the ability of each method to support each Scenario.
Table 2: A summary of the ability of the anomaly detection
methods to support each Scenario.
Method Traffic Feature Scenario I Scenario II Scenario III
MZ Speed ✓ ✓
MZ Occupancy ✓
IF Speed ✓ ✓ ✓
IF Occupancy ✓ ✓ ✓
LOF Speed ✓
LOF Occupancy ✓ ✓
SVM Speed ✓ ✓ ✓
SVM Occupancy ✓ ✓ ✓
LSTM Speed
LSTM Occupancy
4 CONCLUSION
In this research paper, we have studied the perfor-
mance of the most commonly used techniques for
identifying anomalies in the context of traffic conges-
tion detection. Statistical-based, machine-learning-
based, and deep learning-based techniques have been
chosen to cover diverse solution domains. In addi-
tion, various factors have been considered to offer a
comprehensive assessment of each method’s perfor-
mance. Conducted under standardized conditions, the
study outcomes have shown that the methods have
varying efficiency for different needs. Therefore, this
study can help researchers and city operators choose
the optimal method based on their needs, such as the
ability to intervene to reduce the effect of congestion
by adjusting traffic lights and the ability to support
edge computing. Three different scenarios for traffic
management centers have been proposed in this pa-
per. The experimental results have shown that IF and
SVM can support all three scenarios with varying per-
formance. However, the city operators can also use
the results to assess the methods for other needs by
studying their comprehensive performance in terms
of DR, FAR, MTTD, and runtime. This allows for
an assessment from all possible points of view. In
general, the experimental results have shown that all
the utilized methods can achieve relatively high DR.
However, regarding FAR, MTTD, and runtime, vary-
ing efficiency has been found. Hence, the latter met-
rics are key factors in choosing the suitable method
for particular requirements.
REFERENCES
Akhtar, M. and Moridpour, S. (2021). A review of traf-
fic congestion prediction using artificial intelligence.
Journal of Advanced Transportation, 2021:1–18.
AlDhanhani, A., Damiani, E., Mizouni, R., and Wang, D.
(2019). Framework for traffic event detection using
shapelet transform. Engineering Applications of Arti-
ficial Intelligence, 82:226–235.
Bawaneh, M. (2023). Public Service Optimization in Au-
tomated Cities. PhD thesis, Budapest University of
Technology and Economics.
Bawaneh, M. and Simon, V. (2023). Novel traffic conges-
tion detection algorithms for smart city applications.
Concurrency and Computation: Practice and Experi-
ence, 35(5):e7563.
Bhardwaj, A., Iyer, S. R., Ramesh, S., White, J., and Subra-
manian, L. (2023). Understanding sudden traffic jams:
From emergence to impact. Development Engineer-
ing, 8:100105.
Blaise, A., Bouet, M., Conan, V., and Secci, S. (2020).
Detection of zero-day attacks: An unsupervised port-
based approach. Computer Networks, 180:107391.
Breunig, M. M., Kriegel, H.-P., Ng, R. T., and Sander, J.
(2000). Lof: identifying density-based local outliers.
In Proceedings of the 2000 ACM SIGMOD interna-
tional conference on Management of data, pages 93–
104.
Cao, Z., Jiang, S., Zhang, J., and Guo, H. (2016). A unified
framework for vehicle rerouting and traffic light con-
trol to reduce traffic congestion. IEEE transactions on
intelligent transportation systems, 18(7):1958–1973.
Chakroborty, P. and Das, A. (2017). Principles of trans-
portation engineering. PHI Learning Pvt. Ltd.
Chen, H., Ma, H., Chu, X., and Xue, D. (2020). Anomaly
detection and critical attributes identification for prod-
ucts with multiple operating conditions based on iso-
lation forest. Advanced Engineering Informatics,
46:101139.
Chen, M., Yu, X., and Liu, Y. (2018). Pcnn: Deep con-
volutional networks for short-term traffic congestion
prediction. IEEE Transactions on Intelligent Trans-
portation Systems, 19(11):3550–3559.
Cheng, Z., Pang, M.-S., and Pavlou, P. A. (2020). Mitigat-
ing traffic congestion: The role of intelligent trans-
portation systems. Information Systems Research,
31(3):653–674.
Dang, T. T., Ngan, H. Y., and Liu, W. (2015). Distance-
based k-nearest neighbors outlier detection method
in large-scale traffic data. In 2015 IEEE Inter-
national Conference on Digital Signal Processing
(DSP), pages 507–510. IEEE.
Djenouri, Y., Belhadi, A., Lin, J. C.-W., Djenouri, D., and
Cano, A. (2019). A survey on urban traffic anomalies
detection algorithms. IEEE Access, 7:12192–12205.
Djenouri, Y., Zimek, A., and Chiarandini, M. (2018). Out-
lier detection in urban traffic flow distributions. In
2018 IEEE international conference on data mining
(ICDM), pages 935–940. IEEE.
VEHITS 2025 - 11th International Conference on Vehicle Technology and Intelligent Transport Systems
316

Fouladgar, M., Parchami, M., Elmasri, R., and Ghaderi, A.
(2017). Scalable deep traffic flow neural networks for
urban traffic congestion prediction. In 2017 Interna-
tional Joint Conference on Neural Networks (IJCNN),
pages 2251–2258. IEEE.
Fu, R., Zhang, Z., and Li, L. (2016). Using lstm and gru
neural network methods for traffic flow prediction.
In 2016 31st Youth academic annual conference of
Chinese association of automation (YAC), pages 324–
328. IEEE.
Garber, N. J. and Hoel, L. A. (2019). Traffic and highway
engineering. Cengage Learning.
Gonz
´
alez-Aliste, P., Derpich, I., and L
´
opez, M. (2023).
Reducing urban traffic congestion via charging price.
Sustainability, 15(3):2086.
Harrou, F., Zeroual, A., and Sun, Y. (2020). Traffic conges-
tion monitoring using an improved knn strategy. Mea-
surement, 156:107534.
Hasanzadeh, A., Liu, X., Duffield, N., Narayanan, K. R.,
and Chigoy, B. (2017). A graph signal processing ap-
proach for real-time traffic prediction in transportation
networks. arXiv preprint arXiv:1711.06954.
He, Y., Hofer, B., Sheng, Y., Yin, Y., and Lin, H. (2023).
Processes and events in the center: a taxi trajectory-
based approach to detecting traffic congestion and an-
alyzing its causes. International Journal of Digital
Earth, 16(1):509–531.
Hearst, M. A., Dumais, S. T., Osuna, E., Platt, J., and
Scholkopf, B. (1998). Support vector machines. IEEE
Intelligent Systems and their applications, 13(4):18–
28.
Hochreiter, S. and Schmidhuber, J. (1997). Long short-term
memory. Neural computation, 9(8):1735–1780.
Ioannou, C. and Vassiliou, V. (2021). Network attack classi-
fication in iot using support vector machines. Journal
of sensor and actuator networks, 10(3):58.
Kalair, K. and Connaughton, C. (2021). Anomaly detection
and classification in traffic flow data from fluctuations
in the flow–density relationship. Transportation Re-
search Part C: Emerging Technologies, 127:103178.
Kohan, M. and Ale, J. M. (2020). Discovering traffic
congestion through traffic flow patterns generated by
moving object trajectories. Computers, Environment
and Urban Systems, 80:101426.
Kong, X., Song, X., Xia, F., Guo, H., Wang, J., and Tolba,
A. (2018). Lotad: Long-term traffic anomaly detec-
tion based on crowdsourced bus trajectory data. World
Wide Web, 21(3):825–847.
Kuang, W., An, S., and Jiang, H. (2015). Detecting traffic
anomalies in urban areas using taxi gps data. Mathe-
matical Problems in Engineering, 2015.
Lam, P., Wang, L., Ngan, H. Y., Yung, N. H., and Yeh, A. G.
(2017). Outlier detection in large-scale traffic data
by na
¨
ıve bayes method and gaussian mixture model
method. Electronic Imaging, 2017(9):73–78.
Lee, K., Eo, M., Jung, E., Yoon, Y., and Rhee, W. (2021).
Short-term traffic prediction with deep neural net-
works: A survey. IEEE Access, 9:54739–54756.
Li, L., Lin, Y., Du, B., Yang, F., and Ran, B. (2020). Real-
time traffic incident detection based on a hybrid deep
learning model. Transportmetrica A: Transport Sci-
ence, pages 1–21.
Li, L., Lin, Y., Du, B., Yang, F., and Ran, B. (2022). Real-
time traffic incident detection based on a hybrid deep
learning model. Transportmetrica A: transport sci-
ence, 18(1):78–98.
Li, Y., Yu, R., Shahabi, C., and Liu, Y. (2017). Diffusion
convolutional recurrent neural network: Data-driven
traffic forecasting. arXiv preprint arXiv:1707.01926.
Liu, D., Zhen, H., Kong, D., Chen, X., Zhang, L., Yuan,
M., and Wang, H. (2021). Sensors anomaly detection
of industrial internet of things based on isolated forest
algorithm and data compression. Scientific Program-
ming, 2021:1–9.
Liu, F. T., Ting, K. M., and Zhou, Z.-H. (2008). Isolation
forest. In 2008 eighth ieee international conference
on data mining, pages 413–422. IEEE.
Liu, Y., Cai, Z., and Dou, H. (2023). Highway traffic
congestion detection and evaluation based on deep
learning techniques. Soft Computing, 27(17):12249–
12265.
Makara, L. A., Maric, P., and Pekar, A. (2023). Public trans-
port congestion detection using incremental learning.
Pervasive and Mobile Computing, 91:101769.
Markovic, T., Dehlaghi-Ghadim, A., Leon, M., Balador, A.,
and Punnekkat, S. (2023). Time-series anomaly de-
tection and classification with long short-term mem-
ory network on industrial manufacturing systems. In
2023 18th Conference on Computer Science and In-
telligence Systems (FedCSIS), pages 171–181. IEEE.
Nagy, A. M. and Simon, V. (2018). Survey on traffic predic-
tion in smart cities. Pervasive and Mobile Computing,
50:148–163.
Nagy, A. M. and Simon, V. (2020). Traffic incident
dataset. https://gitlab.medianets.hu/anagy/incident
dataset. Accessed: 2024-08-30.
Nagy, A. M. and Simon, V. (2021). A novel congestion
propagation modeling algorithm for smart cities. Per-
vasive and Mobile Computing, 73:101387.
Nagy, A. M., Wiandt, B., and Simon, V. (2021). Transient-
based automatic incident detection method for intelli-
gent transport systems. Infocommunications Journal,
13(3):2–13.
Nassif, A. B., Talib, M. A., Nasir, Q., and Dakalbab, F. M.
(2021). Machine learning for anomaly detection: A
systematic review. Ieee Access, 9:78658–78700.
Nguyen, H., Liu, W., and Chen, F. (2016). Discovering con-
gestion propagation patterns in spatio-temporal traffic
data. IEEE Transactions on Big Data, 3(2):169–180.
Nguyen, T.-H. and Jung, J. J. (2021). Swarm intelligence-
based green optimization framework for sustain-
able transportation. Sustainable Cities and Society,
71:102947.
Pan, B., Zheng, Y., Wilkie, D., and Shahabi, C. (2013).
Crowd sensing of traffic anomalies based on human
mobility and social media. In Proceedings of the
21st ACM SIGSPATIAL International Conference on
Advances in Geographic Information Systems, pages
344–353. ACM.
Machine Learning-Based Anomaly Detection in Smart City Traffic: Performance Comparison and Insights
317

Pang, L. X., Chawla, S., Liu, W., and Zheng, Y. (2011).
On mining anomalous patterns in road traffic streams.
In International conference on advanced data mining
and applications, pages 237–251. Springer.
PeMS (2020). Caltrans Performance Measurement System.
https://pems.dot.ca.gov/. Accessed: 2024-9-15.
Pisner, D. A. and Schnyer, D. M. (2020). Support vector
machine. In Machine learning, pages 101–121. Else-
vier.
Pustokhina, I. V., Pustokhin, D. A., Rodrigues, J. J., Gupta,
D., Khanna, A., Shankar, K., Seo, C., and Joshi, G. P.
(2020). Automatic vehicle license plate recognition
using optimal k-means with convolutional neural net-
work for intelligent transportation systems. Ieee Ac-
cess, 8:92907–92917.
Qin, J., Mei, G., and Xiao, L. (2021). Building the traf-
fic flow network with taxi gps trajectories and its ap-
plication to identify urban congestion areas for traffic
planning. Sustainability, 13(1):266.
Rao, W., Xia, J., Lyu, W., and Lu, Z. (2019). Interval data-
based k-means clustering method for traffic state iden-
tification at urban intersections. IET Intelligent Trans-
port Systems, 13(7):1106–1115.
Redhu, P., Kumar, K., et al. (2023). Short-term traffic flow
prediction based on optimized deep learning neural
network: Pso-bi-lstm. Physica A: Statistical Mechan-
ics and its Applications, 625:129001.
ˇ
Sabi
´
c, E., Keeley, D., Henderson, B., and Nannemann, S.
(2021). Healthcare and anomaly detection: using ma-
chine learning to predict anomalies in heart rate data.
AI & SOCIETY, 36(1):149–158.
Scikit-learn (2024). Scikit-Learn (Machine Learning in
Python). https://scikit-learn.org/. Accessed: 2024-07-
27.
Shi, H., Dong, S., Wu, Y., Nie, Q., Zhou, Y., and Ran, B.
(2024). Generative adversarial network for car fol-
lowing trajectory generation and anomaly detection.
Journal of Intelligent Transportation Systems, pages
1–14.
Sousa, R. S. D., Boukerche, A., and Loureiro, A. A.
(2020). Vehicle trajectory similarity: models, meth-
ods, and applications. ACM Computing Surveys
(CSUR), 53(5):1–32.
Tan, Y., Hu, C., Zhang, K., Zheng, K., Davis, E. A.,
and Park, J. S. (2020). Lstm-based anomaly detec-
tion for non-linear dynamical system. IEEE access,
8:103301–103308.
Tang, J. and Ngan, H. Y. (2016). Traffic outlier detection by
density-based bounded local outlier factors. Informa-
tion Technology in Industry, 4(1):6.
Tasgaonkar, P. P., Garg, R. D., and Garg, P. K. (2020). Vehi-
cle detection and traffic estimation with sensors tech-
nologies for intelligent transportation systems. Sens-
ing and Imaging, 21:1–28.
TensorFlow (2024). TensorFlow: An end-to-end platform
for machine learning. https://www.tensorflow.org/.
Accessed: 2024-07-27.
Togbe, M. U., Barry, M., Boly, A., Chabchoub, Y., Chiky,
R., Montiel, J., and Tran, V.-T. (2020). Anomaly de-
tection for data streams based on isolation forest using
scikit-multiflow. In Computational Science and Its
Applications–ICCSA 2020: 20th International Con-
ference, Cagliari, Italy, July 1–4, 2020, Proceedings,
Part IV 20, pages 15–30. Springer.
Tseng, Y.-T. and Ferng, H.-W. (2021). An improved traffic
rerouting strategy using real-time traffic information
and decisive weights. IEEE Transactions on Vehicular
Technology, 70(10):9741–9751.
Wang, Y., Masoud, N., and Khojandi, A. (2020). Real-time
sensor anomaly detection and recovery in connected
automated vehicle sensors. IEEE transactions on in-
telligent transportation systems, 22(3):1411–1421.
Xu, M. and Li, J. (2024). High-efficiency anomaly detec-
tion of traffic data stream using sequential bi-iteration
svd. IAENG International Journal of Computer Sci-
ence, 51(2).
Yu, Y., Cao, L., Rundensteiner, E. A., and Wang, Q.
(2017). Outlier detection over massive-scale trajec-
tory streams. ACM Trans. Database Syst., 42(2).
Zhang, J., Wang, F.-Y., Wang, K., Lin, W.-H., Xu, X., and
Chen, C. (2011). Data-driven intelligent transporta-
tion systems: A survey. IEEE Transactions on Intelli-
gent Transportation Systems, 12(4):1624–1639.
Zhang, X., Zheng, Y., Zhao, Z., Liu, Y., Blumenstein, M.,
and Li, J. (2021). Deep learning detection of anoma-
lous patterns from bus trajectories for traffic insight
analysis. Knowledge-Based Systems, 217:106833.
Zhao, Z., Chen, W., Wu, X., Chen, P. C., and Liu, J. (2017).
Lstm network: a deep learning approach for short-
term traffic forecast. IET Intelligent Transport Sys-
tems, 11(2):68–75.
Zheng, G., Chai, W. K., Duanmu, J.-L., and Katos, V.
(2023). Hybrid deep learning models for traffic pre-
diction in large-scale road networks. Information Fu-
sion, 92:93–114.
Zhu, L., Krishnan, R., Sivakumar, A., Guo, F., and Polak,
J. W. (2019). Traffic monitoring and anomaly detec-
tion based on simulation of luxembourg road network.
In 2019 IEEE Intelligent Transportation Systems Con-
ference (ITSC), pages 382–387. IEEE.
Zhu, Q., Liu, Y., Liu, M., Zhang, S., Chen, G., and Meng,
H. (2021). Intelligent planning and research on urban
traffic congestion. Future Internet, 13(11):284.
VEHITS 2025 - 11th International Conference on Vehicle Technology and Intelligent Transport Systems
318