Wiley

Journal of Advanced Transportation

Volume 2024, Article ID 8474973, 16 pages
https://doi.org/10.1155/2024/8474973

Research Article
Machine Learning-Based Prediction of Parking Space
Availability in IoT-Enabled Smart Parking Management Systems

Anchal Dahiya ,1 Pooja Mittal ,1 Yogesh Kumar Sharma ,2 Umesh Kumar Lilhore ,3,4
Sarita Simaiya ,4 Ehab Ghith ,5 and Mehdi Tlija 6
1
Department of Computer Science and Applications, MDU, Rohtak 124001, Haryana, India
2
Department of Computer Science and Engineering, Koneru Lakshmaiah Education Foundation, Green Field, Vaddeswaram,
Guntur, Andhra Pradesh, India
3
Arba Minch University, Arba Minch, Ethiopia
4
School of Computing Science and Engineering, Galgotias University, Greater Noida 201310, Uttar Pradesh, India
5
Department of Mechatronics, Faculty of Engineering, Ain Shams University, Cairo 11566, Egypt
6
Department of Industrial Engineering, College of Engineering, King Saud University, P.O. Box 800, Riyadh 11421, Saudi Arabia

Correspondence should be addressed to Umesh Kumar Lilhore; [email protected]

Received 7 May 2024; Revised 23 June 2024; Accepted 24 July 2024

Academic Editor: Qi-zhou Hu

Copyright © 2024 Anchal Dahiya et al. Tis is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Parking space management has become a critical challenge in urban areas due to increasing vehicle numbers and limited parking
infrastructure. Tis paper presents a comprehensive study of machine learning (ML) models in IoT-enabled environments
focusing on proposing an ML-based model for predicting available parking space. Te study evaluates the performance of various
models including K-nearest neighbors (KNNs), support vector machines (SVMs), random forest (RF), decision tree (DT), logistic
regression (LR), and Naı̈ve Bayes (NB) based on “precision, recall, accuracy, and F1-score performance metrics”. Te results
obtained by implementing ML models on the data with 65% and 85% threshold values are compared to draw meaningful
conclusions regarding their performance in predicting parking space availability. Among the evaluated models, random forest
(RF) demonstrates superior performance with high precision, recall, accuracy, and F1-score values. It showcases its efectiveness in
accurately predicting parking space availability in the IoT-enabled environment. On the other hand, models such as K-nearest
neighbors (KNNs), decision tree (DT), logistic regression (LR), and Naı̈ve Bayes (NB) show relatively lower performance in
complex parking scenarios. Te paper concludes that the use of advanced predictive models, particularly random forest, sig-
nifcantly enhances the accuracy and reliability of IoT-enabled parking management systems and also reduces the waiting time of
the vehicles, leading to more efcient resource utilization, reduced trafc congestion in real-time scenarios, and better user
satisfaction in the IoT-enabled environment.

1. Introduction Te advent of the Internet of Tings (IoT) has brought

transformative changes to various domains, including smart
As more people move to cities and their lifestyles change, the parking systems, by enabling the seamless integration of
number of vehicles on the roads has increased signifcantly. physical infrastructure with digital connectivity, data col-
However, the number of parking spaces has not kept pace lection, and advanced analytics [1].
with this growth, causing trafc jams and making parking Smart parking management systems majorly use dif-
difcult to manage. Many current technologies used to fnd ferent devices such as sensor communication channels to
parking spaces are slow and depend too much on human gather and store real-time parking data by using IoT
intervention, exacerbating the problem. technologies.
2 Journal of Advanced Transportation

Birmingham, the second largest city in the United Kingdom Furthermore, this research study explores the potential
witnessing signifcant urban development, is facing trafc implications of the fndings for enhancing parking space
congestion problem; therefore, there is a need for efcient city prediction systems and improving the overall urban mobility
planning including parking and transportation. Tough experience in the IoT environment. By identifying the most
a number of models are developed using diferent technologies, efective prediction models and highlighting areas for im-
results were not appropriate, leading to trafc congestion provement, this research aims to contribute to the devel-
problems creating chaos among people [2]. opment of more reliable and efcient parking management
In smart parking systems, real-time data are collected solutions for IoT-enabled parking systems. Te main ob-
continuously by using diferent sensors such as ultrasonic, jective of the smart parking system is to optimize the uti-
infrared, and magnetic sensors, which will be used for lization of available parking space. Implementation of ML
further data analysis for free parking space prediction [3]. models in the IoT environment can efciently organize and
Further, low power wide area networks (LPWAN) and allocate parking spaces, particularly in high-demand areas
cellular networks are used for efective data transmission during peak times.
with low latency, providing timely parking information to
the users [4].
Tough IoT has played a signifcant role in parking 1.1. Organization. Te structure of this paper is organized as
management systems, it faces challenges due to the large follows. Section 2 provides a related work that highlights
data collected by IoTdevices. To analyze such enormous data relevant works in the feld. In Section 3, the research
is itself a big challenge. To handle this challenge, machine methodology is discussed, along with the introduction of the
learning techniques can be integrated with IoT devices. proposed model. Section 4 covers the dataset used for
Machine learning techniques such as random forest (RF), K- analysis. In Section 5, the results are evaluated based on four
nearest neighbors support vector machines (SVMs), Naı̈ve parameters: accuracy, F1-score, recall, and precision. A
Bayes, logistic regression, and decision tree are among most comparison with the previous work is also performed in this
popularly used machine learning models used for the section. Finally, in Section 6, the paper concludes with
decision-making process [5]. a summary of the fndings.
Further cloud computing can be integrated with IoT for
enhancing the system performance [6]. Cloud-based plat- 2. Related Work
forms provide efcient management and analysis of large
volumes of sensor data and ensure data security and op- Te literature survey conducted revealed several relevant
erational reliability in IoT-enabled parking systems. Tese studies and research papers that have contributed to the
components jointly contribute to more efective parking understanding and development of parking management
space utilization, reduced congestion, and improved user systems. In this paper, the main aim is to build upon the
experiences in metropolitan environments. existing knowledge and address the limitations of previous
Our work focuses on implementing the proposed approaches by utilizing IoT-enabled ML models. Our focus is
framework for predicting parking space availability. Tis on comparing diferent ML models, such as “random forest,
paper presents a comparative analysis of various ML models, logistic regression, decision tree, support vector machine
with the goal of recommending an ML-based model by (SVM), Naı̈ve Bayes, and K-nearest neighbors (KNNs),” to
assessing how well each model predicts the availability of predict the availability of parking spaces accurately.
parking spaces in the IoT environment. Te sensor-collected IoT technologies have revalorized almost all possible
IoT dataset is preprocessed, and then, we hypertuned it for domains such as healthcare, smart homes, smart cities, and
evaluation using diferent threshold values. In this paper, we agriculture.
evaluate the performance of various models, including K- Te Internet of Tings (IoT) refers to a network of
nearest neighbors (KNNs), support vector machines physical objects fxed with sensors, software, and other
(SVMs), random forest (RF), decision tree (DT), logistic technologies to connect and exchange data with other de-
regression (LR), and Naı̈ve Bayes (NB), based on the per- vices and systems over the Internet [3]. It includes a wide
formance metrics such as precision, recall, accuracy, and F1- range of applications, from smart home devices to industrial
score. After performing the analysis, an ML model is pro- automation and healthcare monitoring systems.
posed to enhance prediction accuracy and reduce waiting IoT-enabled smart parking systems optimize space avail-
time when the number of vehicles increases in an IoT- ability and improve urban mobility [3]. However, challenges
enabled environment. By evaluating existing models using such as data security, privacy, and interoperability persist.
the performance metrics, this study seeks to provide insights IoT technology is inherently innovative, creating inter-
into the strengths and limitations of diferent prediction connected ecosystems that leverage edge computing, AI and
models in the IoT-enabled environment. Understanding the ML, 5G connectivity, and blockchain for enhanced ef-
comparative performance of these models can support ciency, advanced analytics, and secure data transactions.
decision-making processes for city planners, parking facility Tese innovations reduce latency, improve decision-making,
operators, and developers of parking-related applications. and address security issues in IoT networks.
Te study focuses on evaluating the predictive accuracy, Keote [7] explores the development and deployment of
strength, and scalability of diferent models across diferent an IoT-enabled smart parking system designed to optimize
metropolitan environments. parking space utilization, reduce trafc congestion, and
Journal of Advanced Transportation 3

minimize environmental impact. Te system integrates Yan et al. [12] introduced a real-time parking-sharing
various IoT components to manage parking spaces in real program where private parking owners use a mobile app to
time and employs a mathematical model to dynamically ofer their spaces. It addresses the challenges of matching
allocate spots based on sensor data. Machine learning is used uncertain supply and demand and scheduling to avoid
to enhance predictive capabilities over time. Real-world tests conficts. Te proposed approach employs a rolling-horizon
demonstrate signifcant improvements in parking efciency framework and a mixed integer programming model,
and urban sustainability. However, limitations include supported by a two-stage heuristic for near-optimal solu-
challenges with scalability, dependency on data quality, high tions. While demonstrating computational efciency, po-
implementation costs, privacy and security concerns, en- tential limitations include scalability, user adoption, and
vironmental impact of IoT devices, and adaptability to system reliability in practical urban deployments, necessi-
diferent urban layouts. tating further research for efective implementation.
Raj and Shetty [8] discuss the necessity of smart parking Sarangi et al. [13] presented an IoT-based smart parking
systems due to urban population growth and the resultant system that ofers several functionalities. Te authors used
increase in private vehicle usage, which leads to trafc three distinct techniques: “computer vision techniques, wire-
congestion and inefcient parking. It examines smart less sensor networks, and an Android mobile platform.” By
parking solutions from a technical perspective, focusing on using all these approaches, the system aims to enhance parking
advancements in the Internet of Tings (IoT) and machine management and improve the overall user experience. Users
learning. Tis paper includes a comprehensive survey of the can utilize an Android-based application to quickly locate
latest developments in smart parking systems that integrate vacant parking spaces. In addition, the application allows users
these technologies. In addition, it explores the role of smart to make payments for the parking spaces they choose. Tis
parking in smart city environments and the benefts of convenient feature enables users to efciently fnd and secure
adopting Parking 4.0. Tis paper also addresses current parking spaces through their mobile devices.
challenges in smart parking systems, proposes possible so- Kumar et al. [14] employed a blend of radar and ultra-
lutions, and outlines future implementation prospects. sonic detection technologies to develop a smart parking
Kokate et al. [9] propose a machine learning-based smart system as part of their construction of smart cities. Te
parking system to address the challenges of parking man- primary objective of this model is to establish a self-regulating
agement caused by the exponential growth of vehicles over parking system that mitigates trafc congestion, reduces air
the past two decades. To simplify parking operations, the pollution, and minimizes user waiting time. By leveraging
system makes use of cloud computing and a cyber framework these technologies, the system enables users to quickly locate
rather than sensor technology. Utilizing HTML, CSS, and available parking slots, even during peak hours. Tis con-
JavaScript, the system is implemented as a web application tributes to a more efcient and user-friendly parking expe-
that ofers real-time information on parking slot availability, rience in urban areas. Te proposed parking system solution
efectively manages both reserved and unreserved slots, aims to minimize human involvement by utilizing a Wi-Fi
identifes anomalies, and enables intelligent trafc manage- system to transmit all data to the cloud for storage. In peak
ment. Te proposed model provides an efcient cost-efective hours, when the number of vehicles exceeds a certain limit, an
time-saving solution for parking management systems. alarm is triggered for further necessary action.
Grbić and Koch [10] proposed an automatic parking slot After conducting the extensive literature survey, we
detection and occupancy classifcation (APSD-OC) algo- found number of gaps in the existing research that motivated
rithm considering diferent key points such as weather us to propose an efcient parking space prediction model
conditions, obstacles, and diferent view positions for integrating IoT with machine learning techniques with
detecting parking spaces. Furthermore, the ResNet34 clas- minimum human intervention. To achieve this objective,
sifer is used to fnd out the occupancy status. Te research a novel hybrid model is proposed for efcient parking space
was carried out on PKLot and CNRPark + EXT datasets management. Firstly, preprocessing is done to extract the
producing robust results. However, the quality of the results most relevant data, and then, a comparative analysis of
was largely dependent on camera positions and weather diferent machine learning techniques is carried out to select
conditions, afecting system performance. the most promising technique for data analysis. Finally, the
Piccialli et al. [11] introduced the integration of smart derived results are verifed and validated to ensure the
parking systems in smart cities, aiming to improve mobility system performance.
and reduce pollution using IoT sensors and mobile appli- By shedding light on this topic, the paper aims to drive
cations. It presents a new deep learning-based ensemble further research and innovation in the feld of innovative
technique for predicting parking space occupancy, sup- parking systems, paving the way for more intelligent, more
ported by a genetic algorithm for parameter optimization. sustainable, and user-friendly cities. Table 1 presents the
Despite showing superior predictive performance over fndings of the literature survey regarding the techniques used,
conventional methods on extensive IoTdata, challenges such their corresponding fndings, and the potential future scope.
as scalability and reliability of IoT sensor networks, along
with ongoing operational and maintenance issues in urban 3. Research Methodology
settings, compromise accuracy. Addressing these challenges
is essential for sustaining the efectiveness of smart parking Figure 1 illustrates the block diagram of the proposed smart
solutions in urban environments. parking prediction model. Te current research aims to
 4

Table 1: Key fndings from literature survey.

Authors Advantages Disadvantages Technique used Findings Future scope
IoT components for real-time
Optimizes parking space management and a mathematical Signifcant improvements in Improves capabilities using
Challenges with scalability.
Keote [7] utilization. Reduces trafc model for dynamic allocation. parking efciency. Enhances advanced machine learning
Dependency on data quality.
congestion. Machine learning for predictive urban sustainability. techniques.
capabilities.
Addresses urban population Comprehensive survey of latest
Raj and Challenges in integration. IoT and machine learning for Proposes solutions only for
growth. Reduces trafc developments. Highlights benefts
Shetty [8] Potential high costs. smart parking solutions. current challenges.
congestion. of Parking 4.0.
ML-based system without Machine learning, cloud Efcient slot management and
Relies on accurate data inputs. Integrates IoTfor better anomaly
Kokate et al. sensors or IoT. computing. HTML, CSS, anomaly detection. Saves time,
Scalability and cybersecurity detection. Adapt to diverse
[9] Real-time parking slot JavaScript for web application. cost, and energy. Minimal human
concerns. parking environments.
information. Cyber-physical framework. intervention.
Efciently detects and classifes Focuses on improving
parking slots using adaptability to environmental
Dependency on consistent Utilizes vision-based approaches, Demonstrates robust performance
vision-based solutions and changes, enhancing system
camera angles and the potential vehicle detection, clustering, and across varied conditions such as
Grbić et al. vehicle detection with reliability through predictive
need for recalibration due to deep learning (ResNet34) for slot weather and occlusions, validated
[10] clustering. Achieves high maintenance, and optimizing
environmental changes may detection and occupancy on datasets (PKLot and
accuracy in occupancy scalability and efciency,
afect reliability over time. classifcation. CNRPark + EXT).
classifcation with a ResNet34 possibly integrating AI for
deep classifer. resilience.
Outperforms traditional IoT scalability and reliability. Deep learning-based ensemble Integrates smart parking in smart
Piccialli and Enhances predictive accuracy
methods. Uses IoT for real-time Operational issues afect predicts parking occupancy, cities for mobility and pollution
Koch [11] and operational reliability.
data. accuracy. optimized by genetic algorithms. reduction using IoT and apps.
Utilizes computer vision,
Accurate identifcation of vehicles,
wireless sensor networks, and Integration of IoT technologies
Sarangi Requires user adoption of the Computer vision and wireless verifcation of vacant spaces, and
Android mobile platform for for real-time monitoring and
et al. [13] Android-based application. sensor networks. easy user interface for locating
enhanced parking management of parking spaces.
parking spots.
management.
Journal of Advanced Transportation
Journal of Advanced Transportation 5

Data Pre-Processing

Remove
Parking Outliers
Dataset Pre-Processed
Parking Dataset
Normalize
Dataset

Data Splitting

Training Dataset
Testing Dataset
(Treshold 65%,85%)

Predictive Models

Performance
Logistic
Evaluation
Regression
SVM
Hyper Parameter Naive Bayes
Tuning Random Forest
KNN Proposed RF
Decision Tree

Predicted Output

Figure 1: Proposed RF model for parking space prediction.

selectively extract and optimize results using machine 3.2. Data Splitting. Once the dataset is preprocessed, it is
learning by collecting relevant parking data from the vast divided into two subsets: a training set and a testing set.
volume of sensor-generated data. “Te primary objective of Partitioning the dataset into two parts is important for
the smart parking system is to optimize the utilization of assessing the performance of our proposed model. Te
available parking spaces.” Te methodology of the proposed stratifed splitting technique is used for splitting the parking
model is described as follows. dataset that ensures that each subset maintains the equal
distribution of classes as the original parking dataset. For
splitting the parking dataset, two diferent threshold values
3.1. Data Collection and Preprocessing. In this initial phase, are used, 65% and 85%, popularly used by many researchers
we collect the smart parking dataset from sensors embedded in existing prediction models [15]. Tese threshold values
in smart parking, ensuring it encompasses all necessary ensure a sufcient amount of data for learning as well as
variables for our analysis; the detailed data description is evaluation. With a 65% threshold value, 65% of the pre-
presented in Section 4. Following this, the dataset undergoes processed data are used as training set, keeping 35% for
preprocessing steps to ensure its quality and suitability for testing purpose. Tis division will explore larger data for
modeling. Outlier detection techniques are employed to training purpose for better learning of the proposed model.
identify and remove any anomalous data points that could Similarly, with an 85% threshold value, 85% of the pre-
skew our analysis. In addition, normalization techniques are possessed data are selected for training and 15% for testing
applied to standardize the range of features within the dataset, purposes, respectively. Tis division will explore larger data for
facilitating fair comparisons between diferent variables. training purpose for better learning of the proposed model.
6 Journal of Advanced Transportation

3.3. Model Training. For training, our proposed model Table 2: Comparative analysis of cross-validation results for each
dataset is divided into two diferent parts: training dataset ML model with threshold 65%.
and testing dataset. Diferent models are tested, such as Technique Precision Recall Accuracy F1-score
random forest, KNN, Naı̈ve Bayes, decision tree, logistic KNN 64.64 51.08 71.72 56.57
regression, and support vector machine. Each model is SVM 72.03 66.35 77.72 69.13
trained to learn the patterns and relationships as described in RF 92.13 90.39 93.16 91.70
Sections 5.9.1 and 5.9.2 in Tables 2 and 3, respectively, DT 85.91 79.12 85.49 82.36
between features and the target variable that is parking space LR 63.62 50.04 71.68 56.20
detection. NB 64.64 51.08 71.72 56.57

Table 3: Comparative analysis of cross-validation results for each

3.4. Hyperparameter Tuning. Predictive models are fnely
ML model with threshold 85%.
tuned with the hyperparameters that will accurately predict
smart parking scenarios. Hyperparameters are used in such Technique Precision Recall Accuracy F1-score
a manner that our proposed model will learn from the KNN 63.40 51.41 71.10 56.05
historical dataset, and setting the suitable confguration can SVM 72.80 65.05 77.02 69.87
signifcantly enhance the performance of the proposed RF 91.53 90.89 93.96 91.06
model. Te grid search technique is used for systematically DT 85.70 78.54 84.17 82.43
exploring diferent combinations that will enhance the LR 62.22 51.32 71.08 56.10
predictive accuracy of the proposed model. Te hyper- NB 64.03 51.38 71.02 56.07
parameters are tuned in Section 5.7 in Table 4. Te main goal
of this tuning is to maximize the accuracy of the proposed Table 4: Tuning of hyperparameters used for evaluation.
model that will fnely work in real-time parking systems.
Algorithm Parameter Value
n_estimators 100
3.5. Model Evaluation and Prediction of Parking Slot. Te Random forest max_depth 10
min_samples_split 2
proposed model is evaluated with diferent testing datasets
that are used for parking space prediction. Diferent eval- C 0.01
Logistic regression
penalty 12
uation metrics are used such as accuracy, precision, recall,
and F1-score. Metrics are compared against predefned max_depth 5
Decision tree
thresholds for gaining valuable observation of each model to min_samples_leaf 1
predict parking spaces in the IoT-enabled environment. kernel RBF
SVM
Tables 2 and 3 highlight the comparative analysis using C 1.0
diferent threshold values. We observe that the random Naı̈ve Bayes — —
forest (RF) model consistently outperforms the other models n_neighbors 5
KNN
across all evaluation metrics, reafrming its efectiveness in weights Uniform
accurately predicting parking instances within a 10-minute
timeframe. By conducting this thorough evaluation process, (3) Time: Te timestamp indicates the moment at which
we can identify the most promising models for real-world the data are collected.
deployment in IoT-enabled parking systems. Figure 1
(4) Start time and fnish time: Te specifc times at which
presents a structure of the proposed model for parking
a parking slot is assigned or becomes available.
space prediction.
(5) Status: Te status of the parking area, indicating
whether it is free or allocated. Tis information is
4. Description about Datasets updated using a status queue.
(6) By organizing the data in this way, it becomes easier
Te initial phase involves gathering data from sensors in-
to analyze and predict the availability of parking
stalled across diferent IoT-enabled parking areas in smart
spaces within the specifed time intervals. Table 5
cities. In our study, we collected data from Birmingham,
represents the attribute of the dataset used.
a city in the United Kingdom, specifcally from various car
park slots. Te data were collected over three months and
stored as a CSV fle. Te primary goal is to predict the 5. Result Evaluation
availability of parking spaces within a time interval of
10 minutes [16]. Te data are structured and organized in the Tis section focuses on data analysis and result evaluation
following manner: for the proposed model. First of all, the six machine learning
techniques are defned, and then, their comparative analysis
(1) Parking unique ID: Each parking slot is assigned is performed. To enhance the result, machine learning
a unique ID for identifcation purposes. techniques are tuned with hyperparameter described in
(2) Duration: Te total amount of time that a parking Section 5.7. Accuracy, F1-score, recall, and precision are the
slot remains occupied or vacant. performance evaluators described in Section 5.8. Finally,
Journal of Advanced Transportation 7

Table 5: Attributes of the dataset. 1

􏼒P􏼒Availability � 􏼓􏼓, (2)
Feature Range x
Unique ID Assigned ID where x � (X1 , X2 , . . . , Xp ) represents the vector of input
Day 1–7 features (e.g., time of day and location).
Start hour 0–23
Finish hour 0–23 1 1
P􏼒Availability � 􏼓 � , (3)
Start minute 0–59 x 1 + exp(−z)
Finish minute 0–59
Status Free or engaged where z is the linear combination of the input features
weighted by their respective coefcients:
z � β0 + β1 ∗ X1 + β2 ∗ X2 + · · · + βp ∗ Xp . (4)
a comparison between the proposed model and existing
models is carried out to verify and validate the result in Te LR model is used to predict new data after it has been
Section 5.9.3. trained. Te model produces result on the probability that
a parking space will be available after processing the input
variables from a parking dataset. Te parking space can be
5.1. Random Forest (RF). Random forest is a powerful ML classifed as available or unavailable by the model by setting
model. It combines multiple decision trees for making a threshold on the projected probabilities. LR can ofer
predictions [17]. In this paper, RF is utilized to predict the insights on the variables afecting parking spot availability
availability of parking spaces based on various features and [18]. We can use the performance metric to validate the
historical data collected from the sensor’s technology. RF result of predicted parking space.
model is trained using the collected parking datasets, such as
parking occupancy, time of day, day of the week, and other
signifcant factors, that can accurately predict the availability 5.3. Decision Tree (DT). A well-known machine learning
of parking spaces in real-time scenarios. RF efciently works model for classifcation and regression applications is the
on the parking prediction model that accurately predicts the decision tree. Tey ofer a simple and understandable
parking spaces using large datasets with numerous features method for making choices depending on a collection of
and handles missing or noisy data efectively. input attributes. DT is also used efectively for predicting
A comparative analysis is performed for the prediction parking space in a smart environment depending on a va-
of available parking space. In this paper, we compare the riety of input features. Te DT algorithm builds a model that
accuracy, strength, and predictive power of the model with resembles a tree, with each leaf node representing a pre-
other ML techniques for parking space prediction. Mathe- dicted value and each inside node representing a choice
matical formula used for parking space prediction is given in based on a particular attribute [11]. Te following equation
the following equation: explains how we use it for predicting parking space:
􏽢 � f(x; θ),
y (5)
1 T
􏽢 � 􏽘t�1 ft (x),
y (1)
T where y 􏽢 is the predicted availability of parking space based
where y􏽢 is the predicted availability of a parking space based on input feature x (e.g., time of day and location), the
on input features x, T is the number of decision trees in function f segments the feature space into regions based on
random forest, and ft (x) represents the prediction of the t- the values of input features and makes predictions for each
th decision tree. segment, and θ is used as a parameter, for example, splitting
criteria and leaf node values learned during training.
Figure 2 represents the basic structure of the decision tree.
Te top-down methodology is used by the DT algorithm.
5.2. Logistic Regression (LR). Logistic regression is a super- A decision node determines whether the number of occu-
vised learning model that is used for binary classifcation. pied parking spaces is greater than a predetermined
Working of LR is to predict the probability of an instance threshold value in order to predict parking space. It split the
belonging to a particular class [18]. LR is efciently used in data into various branches, and the process continues re-
parking space prediction. Te LR model is trained on peatedly until a stopping criterion is met [11]. Its efcacy in
a historical parking dataset, where the input variables in- predicting parking spaces can be assessed and contrasted
clude factors such as time of day, day of the week, location, with other methods.
and any other relevant features. Te output variable would
indicate whether a parking space is available (1) or not (0).
Te logistic function maps the input variables to a proba- 5.4. Support Vector Machine (SVM). Another powerful
bility value between 0 and 1. Te model parameters are machine learning algorithm is support vector machine
learned through an optimization process, such as maximum (SVM). In order to divide the data points into distinct
likelihood estimation or gradient descent, to fnd the best ft classes, SVM fnds the best hyperplane [19]. Te objective is
for the training data. Mathematical term used for parking to maximize the margin or the separation between the
space prediction is given in equations (2)–(4), respectively: closest data points from each class and the hyperplane. Based
8 Journal of Advanced Transportation

Root Node

Sub-Tree

Decision Node Decision Node

Decision Node
Leaf Node Leaf Node Leaf Node

Leaf Node Leaf Node

Figure 2: Generic structure of decision tree.

on the provided input data, such as time, duration, vehicle In the above equation, w stands for weight vectors, b is
type, and other relevant parameters, SVM can be used to used for bias, xi represents input vector, yi is the corre-
identify parking spaces as either vacant or occupied. Te sponding class label (−1 or 1), and C is the regularization
SVM model can accurately predict the availability of parking parameter, ξ. i are the slack variables. Te main aim is to fnd
spaces. the optimal values for w and b that minimize the objective
Te mathematical formula for SVM involves fnding the function while satisfying the constraints. Once the model is
optimal hyperplane that maximizes the margin while en- trained, it is used for predicting parking space.
suring the correct classifcation of data points. It can be
expressed in the following equations:
5.5. Naı̈ve Bayes. Naı̈ve Bayes is a classifer popularly used
1 for predicting parking space availability based on the input
minimize: ∗ ‖w‖2
2 provided. Many researchers had used time, day, weather
(6)
conditions, past experience, and many other signifcant
T i
+ C ∗ 􏽘(max􏼐0, 1 − y ∗ 􏼐w ∗ x − b􏼑􏼑, features as input for predicting parking space. Naı̈ve Bayes is
a statistical method that follows Bayes’ theorem and assumes
yi ∗ 􏼐wT ∗ xi − b􏼑 > � − _ci . (7) that features are independent from each other [20]. Parking
space availability is predicted by calculating posterior
probability in terms of class information and prior proba-
bility by using the following equation:

1 (P(X | Available � 1) ∗ P(Available � 1))

P􏼠Available � � 􏼡. (8)
X P(X)

5.6. KNN. Another popular algorithm that can be used for (1) Calculate the distance between the target location
predicting parking space is k-nearest neighbor. It is and all known parking spots using a distance metric
a simple yet efective algorithm. For predicting parking, this such as Euclidean distance
algorithm is used to fnd the K-nearest parking spots to (2) Select the nearest parking spots in K based on the
a given location and use their occupancy status for in- calculated distances
dicating the available parking space [21]. By considering the
characteristics of nearby parking spots, the model can make (3) Determine the majority class (occupied or vacant)
predictions about the availability of parking spaces in real among the K-nearest parking spots
time. Te KNN algorithm operates based on the following (4) Assign the majority class as the predicted class for the
steps: target location
Journal of Advanced Transportation 9

Equations (9)–(12) explain how we use it for parking variance. Te standardized weight function is chosen to treat
space prediction: all neighbors equally and avoid bias towards any particular
x � x1 , x2 , . . . , xM 􏼁, (9) subset of neighbors.

where x represents the features of the parking space in-

stances (e.g. time of day and location): 5.8. Parameters Used for Evaluation. Hyperparameters are
􏽳�������������� used for evaluating our proposed model. Hyperparameters
M 2 are chosen to maximize the performance of each algorithm
d x, xi 􏼁 � 􏼐x − xi,j 􏼑 , (10)
j�1 j for predicting smart parking space. To evaluate the strength
and potential overftting of the proposed model, we
where xi is the feature vector of the i-th training instances. employed the k-fold cross-validation method. Tis tech-
Identify the K training instances with the smallest nique divides our dataset into k folds. After that, iteratively
train and evaluate the model k times, each time using
d x, xi 􏼁, (11) a diferent fold as the validation set and the remaining folds
as the training set. By averaging the performance metrics
􏽢 � arg max
y 􏽘 π yi � c􏼁, across these iterations, we gain a more robust estimate of the
c∈ (12)
{0,1} i∈Nk (x) model’s simplifcation ability and its ability to handle unseen
data. Te k-fold cross-validation helps us validate the
where Nk (x) is the set of K-nearest neighbors and π (.) is the model’s performance and ensure that it is not excessively
indicator function. tailored to the training data, justifying the risk of overftting.
KNN ofers the advantage of simplicity and fexibility, as
it can handle both numerical and categorical features. It can (i) Accuracy: Out of all the samples provided, it cal-
also adapt to changing patterns in the data and make real- culates the portion of correctly predicted samples
time predictions. However, it is essential to choose an ap- [22]. It is computed by dividing the total number of
propriate value for K, as a low value may lead to overftting samples by the number of correctly predicted
and a high value may lead to underftting. samples. Improved model performance in precisely
predicting the target variable is indicated by higher
accuracy:
5.7. Tuning Hyperparameters for Machine Learning Models. accurate prediction
Te chosen hyperparameters for each algorithm were de- accuracy � . (13)
total samples
termined based on careful experimentation and consider-
ations specifc to our study. Table 4 represents the tuned (ii) F1-score: A metric called the F1-score aggregates
hyperparameters used for the prediction of parking space. recall and precision into a single number [22]. Better
For the random forest algorithm, we set the number of outcomes are indicated by a higher F1-score, which
decision trees (n_estimators) to 100 to achieve a good takes into account recall (the capacity to locate all
balance between model complexity and computational ef- positive instances) as well as precision (the accuracy
fciency. Te maximum depth (max_depth) of 10 was se- of positive predictions):
lected to prevent overftting and promote the generalization
of the model. Te minimum number of samples required to 2 ∗ (recall ∗ precision)
F1score � . (14)
split an internal node (min_samples_split) was set to 2 to recall + precision
ensure that the algorithm considers a sufcient number of
samples for meaningful splits. In LR, we selected a small (iii) Recall: Te ratio of true positive predictions to the
regularization parameter (C � 0.01) to prevent overftting total number of actual positive instances is known as
and encourage simpler models. In order to enhance model recall [22]. It measures how well the model can
stability and account for multicollinearity, the L2 (ridge) distinguish true positive predictions from the total
regularization (penalty) is selected. A maximum depth of 5 is set of actual positive instances. A higher recall in-
selected for the decision tree algorithm in order to control dicates a higher level of completeness in capturing
the tree’s complexity and avoid overftting. positive instances:
To enable fne-grained splits and capture more focused TP
patterns in the data, the minimum number of samples re- recall � . (15)
TP + FN
quired at a leaf node (min samples leaf ) is set to 1. For SVM,
the radial basis function (RBF) kernel, which is well-known (iv) Precision: Te ratio of true positive predictions to
for performing well with a variety of data types, is used. Te the total number of positive predictions made by the
regularization parameter (C) is set to 1.0 to balance the model is known as precision [22]. From the set of all
trade-of between model complexity and training error. predicted positive instances, it measures how well
Since NB is a probabilistic algorithm, it does not require the model can predict positive instances:
specifc hyperparameters to be tuned. For KNN, we chose 5
TP
neighbors (n_neighbors) based on experimental in- precision � . (16)
terpretation, as it strikes a good balance between bias and Tp + FP
10 Journal of Advanced Transportation

5.9. Performance Evaluation. In this section, our main focus interesting observations can be made. Random forest (RF)
is the performance evaluation of the proposed model. To continues to demonstrate a high level of performance across
achieve this objective, frst of all, a comparative analysis of all metrics. With a precision of 91.53%, recall of 90.89%,
following six machine learning techniques is carried out: accuracy of 93.96%, and F1-score of 91.06%, RF maintains its
random forest, decision tree, support vector machine, lo- position as the top-performing model. Support vector
gistic regression, Naı̈ve Bayes, and KNN. Ten, the evalu- machine (SVM) also remains consistent in its performance.
ation is done based on the results of each cross-validation It achieves a precision of 72.80%, recall of 65.05%, accuracy
iteration by using performance evaluators explained in of 77.02%, and F1-score of 69.87%. While SVM performs
Section 5.8. We conducted a comparison specifcally to well, it does not surpass RF in terms of overall accuracy.
predict available parking spaces within 10 minutes with Te decision tree (DT) model exhibits a precision of
diferent threshold values [23]. Te threshold is used as 85.70%, recall of 78.54%, accuracy of 84.17%, and F1-score of
a criterion to determine the predicted parking space avail- 82.43%. Although DT shows a similar performance to the
ability. Te threshold represents a minimum assurance level previous evaluation, it falls slightly behind RF in terms of
or probability required for a prediction to be measured as accuracy. Logistic regression (LR) and Naı̈ve Bayes (NB)
a positive outcome (i.e., available parking space). By models once again demonstrate relatively lower perfor-
establishing a threshold, we create a standard for decision- mance compared to the other methods. LR achieves a pre-
making based on the scores or prediction probabilities cision of 62.22%, recall of 51.32%, accuracy of 71.08%, and
created by the models [24]. F1-score of 56.10%. NB achieves a precision of 64.03%, recall
We utilized a training dataset with a 65% threshold for of 51.38%, accuracy of 71.02%, and F1-score of 56.07%.
the 10-minute prediction and later applied an 85% threshold Tese results are consistent with the previous evaluation,
for comparison presented in Section 5.9.1 and 5.9.2, re- highlighting the limitations of LR and NB in accurately
spectively. Our goal is to gain deeper insights into each predicting parking space availability. Te results from Ta-
model performance and signifcance for the parking space ble 3 are graphically represented in Figure 5.
prediction task by analyzing the results and considering Figure 5 proves that random forest outperforms all other
various prediction timeframes. machine learning models by achieving 93.96% accuracy and
91.53% precision even with a higher threshold value of 85%.
Ten, decision tree achieves 84.17% accuracy, SVM with 77.02%,
5.9.1. 10-Minute Prediction (Treshold 65%). To select the
and KNN with 71.10%, and accuracy followed by logistic re-
best machine learning technique, a comparative analysis of
gression and Naı̈ve Bayes with relatively lower performance.
all six approaches under consideration is performed. Te
Furthermore, to verify the applicability of the random
results of the comparative analysis with a threshold value of
forest technique with a higher threshold value, it is imple-
65% for the 10-minute prediction are shown in Table 2.
mented on the preprocess dataset fne-tuned with selected
Based on the above results, random forest (RF) is se-
hyperparameters for predicting parking space occupancy.
lected as the best technique with 93.16% accuracy, 90.39%
Figure 6 represents the comparison between actual and
recall, 92.13% precision, and 91.70% F1-score. Te results
predicted results for predicting parking occupancy with an
from Table 2 are graphically represented in Figure 3.
85% threshold value.
From Figure 3, it can be clearly witnessed that random
It is clearly evident from Figures 4 and 6 that the random
forest is the best performer followed by decision tree with
forest technique is achieving higher accuracy in predicting
85.49% accuracy and 85.91% precision value. SVM achieves
parking occupancy with diferent threshold values in the
77.72% accuracy and precision of 72.03%. Whereas,
IoT-enabled environment.
remaining approaches including KNN, logistic regression,
By comparing the results obtained using both the 65%
and Naı̈ve Bayes exhibit poor performance.
and 85% threshold values, we can draw several conclusions
Finally, the random forest technique is implemented on
about the performance of the techniques in predicting
the preprocessed dataset tuned with hyperparameters for
parking space availability.
predicting parking occupancy with a 65% threshold value. A
graphical comparison of actual and predicted results is (1) First, we observe that for most models, there is
shown in Figure 4. a slight decrease in precision, recall, accuracy, and
Te parking occupancy results are compared in terms of F1-score when transitioning from the 65% threshold
diferent timing of a day as shown in Figure 4. It clearly to the 85% threshold. Tis implies that raising the
proves the efectiveness of the random forest technique in threshold encourages a more careful approach and,
predicting parking space availability in smart parking areas. as a result, fewer positive predictions (available
parking spaces) are generated. Because there is
a greater likelihood of false negatives—erroneously
5.9.2. 10-Minute Prediction (Treshold 85%). For 10-minute
predicting parking spaces as unavailable—precision
prediction with an 85% threshold, a comparative analysis
consequently declines.
was conducted based on the performance metrics of dif-
ferent models. Table 3 showcases the precision, recall, ac- (2) Second, the random forest (RF) model outperformed
curacy, and F1-score of each model, allowing us to assess all other techniques by achieving higher accuracy,
their efectiveness in predicting available parking spaces. recall, and precision values in predicting parking
Comparing the results to the previous evaluation, some space availability efciently.
Journal of Advanced Transportation 11

Performance Metrics

80

60
Percentage

40

20

0
Precision Recall Accuracy F1-Score
Metrics

SVM Random Forest

KNN Logistic Regression
Decision Tree Naive Bayes
Figure 3: Comparative analysis of diferent machine learning techniques with 65% threshold in terms of diferent performance metrics.

Actual vs Predicted Occupancy with 65% Threshold

120

100

80
Parking Occupancy

60

40

20

–20
08:00
08:10
08:39
09:08

09:52
10:06
10:35
11:04

11:48
12:02
12:31
13:00
13:10

13:54
14:08
14:37
15:06

15:50
16:04
16:33
17:02

17:46
18:00
18:10
18:39
19:08

19:52
20:06

Time of Day
Actual
Predicted
Figure 4: Monitoring parking occupancy trends: actual vs. predicted with 65% threshold.

(3) Furthermore, it can be concluded that the logistic parking space in smaller as well as larger smart
regression (LR) technique, with low precision and parking areas due to its limitations in capturing and
accuracy values, is not a good choice for predicting analyzing complex features of parking areas.
12 Journal of Advanced Transportation

Performance Metrics

80

60
Percentage

40

20

0
Precision Recall Accuracy F1-Score
Metrics

SVM Random Forest

KNN Logistic Regression
Decision Tree Naive Bayes
Figure 5: Comparative analysis of diferent machine learning techniques with 85% threshold in terms of diferent performance metrics.

Actual vs Predicted Occupancy with 85% Threshold

120

100

80
Parking Occupancy

60

40

20

–20
07:40
08:00

08:38
09:07

09:51
10:05
10:34

11:03

11:47
12:01
12:30
12:40
13:09

13:53
14:07
14:36
15:05

15:49
16:03
16:32
17:01

17:45
17:55
18:09
18:38
19:07

19:51
20:05

Time of Day
Actual
Predicted
Figure 6: Monitoring parking occupancy trends: actual vs. predicted with 85% threshold.

(4) Finally, it can be concluded that random forest is the and identifying impending areas for further improvement
most signifcant machine learning technique that can and exploration.
be deployed for predicting the availability of parking
spaces with varied threshold values.
5.9.3. Comparison of the Proposed RF Model with Existing
Table 6 presents the detailed summary of strengths, Models. In a comparative analysis of our proposed RF
shortcomings, and fndings of diferent existing models used model with diferent existing machine learning models for
for parking space prediction for an in-depth understanding IoT-enabled parking space prediction, we applied diferent
of all signifcant factors. It serves as a valuable resource for threshold values to evaluate performance. Te random forest
understanding the strengths and limitations of each model (RF) model consistently outperformed other models across
 Journal of Advanced Transportation

Table 6: Key fndings of comparative analysis.

Model Advantages Disadvantages Findings
Computationally expensive during prediction, which may Showed moderate performance in terms of precision,
KNN Simple and easy to implement.
limit real-time applicability in IoT environments. recall, accuracy, and F1-score.
Demonstrated good performance in precision, recall,
SVM Efective in handling high-dimensional data. Sensitive to the choice of kernel and parameters.
accuracy, and F1-score.
Ability to handle large datasets and complex May overft noisy data, especially with a large number of Achieved high accuracy in predicting parking space
Random forest
feature interactions. trees. availability.
Prone to overftting with complex datasets, potentially Proven efective in predicting parking availability using
Decision tree Easy to understand and interpret.
impacting generalization in varying IoT parking scenarios. simple decision rules.
Assumes linear relationship between features and
Logistic Interpretable and provides insights into feature Showed limitations in complex parking scenarios with
outcome, limiting its performance in capturing nonlinear
regression importance. lower precision, recall, accuracy, and F1-score.
relationships in IoT parking data.
Assumes independence among features, which may not Demonstrated moderate performance in precision, recall,
Naı̈ve Bayes Fast and simple algorithm.
hold true in IoT parking data, afecting its performance. accuracy, and F1-score.
13
14 Journal of Advanced Transportation

Comparative Analysis of Waiting Time between Existing Models and Proposed RF Parking Prediction Model

25.0

22.5

20.0
Waiting Time (minutes)

17.5

15.0

12.5

10.0

7.5

5.0
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Number of Vehicles

SVM Model LR Model

KNN Model DT Model
Naive Bayes Proposed RF Model
Figure 7: Comparative analysis of the proposed RF model with existing models in terms of waiting time and number of vehicles.

Table 7: Comparison of proposed random forest model with existing models in terms of performance evaluators.
S. No Model Accuracy (%) Precision (%) Recall (%) F1-score (%)
1 SVM (Zhu et al. [26]) 86 — — —
2 KNN (Suheryadi et al. [27]) 77 87 85 85
3 Decision tree (Feng et al. [28]) 72% — — —
4 Naı̈ve Bayes (Filali et al. [29]) 84 — — —
5 Logistic regression (Errousso et al. [30]) 92.52 91.01 90.99 90.55
6 Proposed RF model (threshold 65%) 93. 6 92. 3 90.39 9 .70
7 Proposed RF model (threshold 85%) 93.96 9 .53 90.89 9 .06
Te bold values emphasize the superior performance of the proposed RF models, particularly at diferent threshold levels, compared to existing models.

performance metrics. Given its superior performance, ro- (1) Utilization factor (ρ):
bustness, and accuracy, we proposed the RF model as the c
optimal solution for predicting parking space availability in ρ� . (17)
cμ
IoT-enabled systems.
Several researchers have focused on calculating waiting (2) Probability of zero vehicles in the system ( P0 ):
times in IoT-enabled parking systems, but their proposed
c−1 n c −1
models often fall short in efciently reducing these times (c/μ) (c/μ) 1 ⎤⎦
[25]. In contrast, we addressed this critical issue and pro- P0 � ⎡⎣ 􏽘 + · . (18)
n�0 n! c! 1−ρ
posed a model that signifcantly mitigates waiting time.
Mathematical equations used for calculating waiting time in
(3) Average no. of vehicles in the queue (Lq ):
the IoT-enabled smart environment are given in equations
(17)–(21). (c/μ)c · c · μ
Te waiting time W can be calculated using principles Lq � · Po . (19)
c
from queuing theory. For the system with multiple parking
spaces, the M/M/c queue model is used: (i) arrival rate c: the (4) Average waiting time in queue ( Wq ):
rate at which vehicles arrive; (ii) service rate μ: the rate at
Lq
which parking space becomes available; and (iii) no. of Wq � . (20)
servers (c): no. of parking spaces. c
Journal of Advanced Transportation 15

(5) Total average waiting time (W): 6. Conclusion

1 With fewer parking spots and more vehicles on the road,
W � Wq + . (21)
μ efective parking systems in cities are important. Drivers
often deal with trafc jams, wasted time, and frustration
Figure 7 presents a comparative analysis on the basis of when they cannot fnd parking space using traditional
the waiting time between various existing models and our methods. Tis paper addresses these limitations and pro-
proposed IoT-enabled parking model. Te number of ve- poses a model to solve the problem faced by the users. In this
hicles is shown on the x-axis, and the waiting time is shown study, the random forest model is proposed for parking
on the y-axis in minutes. Te proposed model is represented space prediction based on a comprehensive analysis of
in red on the graph, with multiple lines indicating each diferent ML models using the IoT-enabled dataset of Bir-
existing model. mingham and also compares the proposed model with
Results from Figure 7 clearly prove that our proposed existing models. Diferent evaluation parameters are used for
model is consistently outperforming all other existing validating the performance of the proposed model such as
models irrespective of all other deciding factors. For our precision, recall, F1-score, and accuracy.
proposed model, with an increase in the number of vehicles, Results prove that the proposed random forest model
there is a very slow rise in waiting time, compared to other achieves highest accuracy with minimum waiting time in
existing models. predicting parking space availability in IoT-enabled areas,
It can be concluded that our proposed model is a better compared to other existing models. To ensure the appli-
choice for all parking management systems whether for cability of the proposed model in diferent scenarios, its
smaller or larger areas. It can provide a signifcant solution to performance is evaluated with diferent threshold values.
all problems related to parking systems such as trafc Te proposed model achieves 93.16% accuracy for 65%
congestion, air pollution due to more emission of harmful threshold value and 93.96% accuracy for 85% threshold
gases, and wastage of time and other resources. Our pro- value. Furthermore, random forest has shown tremendous
posed model can predict the available parking space more potential for further scalability and extension.
fast and accurately, guiding the users to assist them. To Tis work has laid a foundation for future research, in
validate and verify the performance of our proposed model, this important area, to improve the living quality standards
a comparative analysis is performed between our proposed of all. Tis model can be extended to ensemble approaches
model and the existing state of the arts in terms of diferent for enhancing the prediction results for sizeable parking
performance evaluators, presented in Table 7. areas. Finally, it can be concluded that this work has reduced
From Table 7, it can be clearly seen that our proposed the waiting time for searching for parking slots, leading to
model achieves the highest accuracy and precision value pollution reduction and reduced trafc congestion and
compared to all other existing models. Moreover, it is leading to a happy place to live in, by efciently predicting
performing consistently better even with diferent the parking space availability.
threshold values. Tus, it can be considered as the best
prediction model among all existing models for predicting Data Availability
parking space.
Te dataset is available from the corresponding author.

5.9.4. Recommendations. Tough our proposed random Conflicts of Interest

forest model improves the parking space availability many
folds for IoT-enabled parking areas, its performance can be Te authors declare that they have no conficts of interest
enhanced by implementing the following suggestions. Te that could have appeared to infuence the work reported in
prediction results must be transmitted to the drivers timely this paper.
to provide them better insights about parking space avail-
ability. It may improve user satisfaction by reducing search
time, pollution, and wastage of resources. It also ensures Acknowledgments
maximum utilization of parking facilities by occupying Te authors extend their appreciation to King Saud Uni-
overall free areas, which will further enhance the complete versity for funding this work through the Researchers
city planning and transportation systems. In addition, the Supporting Project number (RSPD2024R685), King Saud
deployment of appropriate IoT sensors will make the University, Riyadh, Saudi Arabia.
parking areas safer and maintained. Te proposed model
may also suggest appropriate numbers and locations for
electric vehicle charging stations, supporting a green and References
sustainable environment. It is also important to ensure the [1] F. Al-Turjman and A. Malekloo, “Smart parking in IoT-
adaptability and scalability of the proposed model along with enabled cities: a survey,” Sustainable Cities and Society,
IoT infrastructure. Tese recommendations will surely im- vol. 49, Article ID 101608, 2019.
prove the overall quality of life in urban areas by optimizing [2] M. Aljohani, S. Olariu, A. Alali, and S. Jain, “A survey of
parking management systems. parking solutions for smart cities,” IEEE Transactions on
16 Journal of Advanced Transportation

Intelligent Transportation Systems, vol. 23, no. 8, pp. 10012– [18] W. He, G. Yan, and L. Da Xu, “Developing vehicular data
10029, 2022. cloud services in the IoT environment,” IEEE Transactions on
[3] R. Gupta, N. Budhiraja, S. Mago, and S. Mathur, “An IoT- Industrial Informatics, vol. 10, no. 2, pp. 1587–1595, 2014.
based smart parking framework for smart cities,” Data [19] J. Shen, Y. Xia, H. Ding, and W. Cabrel, “Smart parking locks
Management, Analytics and Innovation, vol. 1, pp. 19–32, based on extended UNET-GWO-SVM algorithm,” Sensors,
2020. vol. 23, no. 20, p. 8572, 2023.
[4] Y. Agarwal, P. Ratnani, U. Shah, and P. Jain, “IoT based smart [20] A. Khanna and R. Anand, “IoT based smart parking system,”
parking system,” in 2021 5th International Conference on in 2016 International Conference on Internet of Tings and
Intelligent Computing and Control Systems (ICICCS), Applications (IOTA), pp. 266–270, IEEE, Pune, India, January
pp. 464–470, IEEE, Pune, India, January 2021. 2016.
[5] S. U. Raj, M. V. Manikanta, P. S. S. Harsitha, and M. J. Leo, [21] D. Susandi, W. Nugraha, and S. F. Rodiyansyah, “Per-
“Vacant parking lot detection system using random forest ancangan smart parking system pada prototype smart ofce
classifcation,” in 2019 3rd International Conference on berbasis internet of things,” Pros. Semnastek, 2017.
Computing Methodologies and Communication (ICCMC), [22] F. M. Awan, Y. Saleem, R. Minerva, and N. Crespi, “A
pp. 454–458, IEEE, Erode, India, March 2019. comparative analysis of machine/deep learning models for
[6] T. N. Pham, M.-F. Tsai, D. B. Nguyen, C.-R. Dow, and parking space availability prediction,” Sensors (Switzerland),
D.-J. Deng, “A cloud-based smart-parking system based on vol. 20, no. 1, p. 322, 2020.
Internet-of-Tings technologies,” IEEE Access, vol. 3, [23] Y. Zheng, S. Rajasegarar, and C. Leckie, “Parking availability
pp. 1581–1591, 2015. prediction for sensor-enabled car parks in smart cities,” in
[7] M. L. Keote, “Design of mathematical model and imple- 2015 IEEE 10th Int. Conf. Intell. Sensors, Sens. Networks Inf.
mentation of IoT enabled smart secure parking system,” Process. ISSNIP 2015, Singapore, April 2015.
Communications on Applied Nonlinear Analysis, vol. 31, [24] Z. C. Lipton, C. Elkan, and B. Naryanaswamy, “Optimal
no. 2s, pp. 373–387, 2024. thresholding of classifers to maximize F1 measure,” Machine
[8] A. Raj and S. D. Shetty, “Smart parking systems technologies, Learning and Knowledge Discovery in Databases, vol. 8725,
tools, and challenges for implementing in a smart city en- pp. 225–239, 2014.
vironment: a survey based on IoT and ML perspective,” In- [25] H. Brožová and M. Růžička, “Te prediction of parking space
ternational Journal of Machine Learning and Cybernetics, availability,” Transport, vol. 35, no. 5, pp. 462–473, 2020.
vol. 15, no. 7, pp. 2673–2694, 2024. [26] D. Zhu, S. Song, H. Zhang, Z. Shi, W. Zheng, and H. Liu,
[9] R. Kokate, D. Kohad, A. Hiwarkar, P. Godbole, and “Prediction of parking spaces and recommendation of
R. Bhasarkar, “Real time parking system using ML,” 2024, parking area in urban complex,” in Proceedings of the 2020 4th
https://www.researchgate.net/publication/377459619_Real_ International Conference on Big Data and Internet of Tings,
Time_Parking_System_using_ML. pp. 59–63, New York, NY, USA, October 2020.
[10] R. Grbić and B. Koch, “Automatic vision-based parking slot [27] A. Suheryadi, W. P. Putra, M. A. Al Hilmi, and
detection and occupancy classifcation,” Expert Systems with K. A. Cahyanto, “Vehicles position tracking in parking lots
Applications, vol. 225, Article ID 120147, 2023. using K-nearest neighbor and fngerprinting based on RSSI
[11] F. Piccialli, F. Giampaolo, E. Prezioso, D. Crisci, and bluetooth,” in 2021 Sixth International Conference on In-
S. Cuomo, “Predictive analytics for smart parking: a deep formatics and Computing (ICIC), pp. 1–6, IEEE, Jakarta,
learning approach in forecasting of iot data,” ACM Trans- Indonesia, November 2021.
actions on Internet Technology, vol. 21, no. 3, pp. 1–21, 2021. [28] N. Feng, F. Zhang, J. Lin, J. Zhai, and X. Du, “Statistical
[12] P. Yan, X. Cai, D. Ni, F. Chu, and H. He, “Two-stage analysis and prediction of parking behavior,” in Network and
matching-and-scheduling algorithm for real-time private Parallel Computing: 16th IFIP WG 10.3 International Con-
parking-sharing programs,” Computers and Operations Re- ference, Hohhot, China, August 2019.
search, vol. 125, Article ID 105083, 2021. [29] Y. Filali, H. Errousso, N. Aghbalou, E. A. A. Alaoui, and
[13] M. Sarangi, S. Mohapatra, S. V. Tirunagiri, S. K. Das, and M. A. Sabri, “Real-time parking availability classifcation on
K. S. Babu, “IoT aware automatic smart parking system for a large-area scale,” in Te Proceedings of the International
smart city,” in Cognitive Informatics and Soft Computing: Conference on Smart City Applications, pp. 215–228, Springer,
Proceeding of CISC 2019, pp. 469–481, Springer, Cham, Cham, Switzerland, February 2023.
[30] H. Errousso, E. A. A. Alaoui, S. Benhadou, and A. Nayyar,
Switzerland, January 2020.
“Intelligent parking space management: a binary classifcation
[14] I. Kumar, P. Manuja, Y. Soni, and N. S. Yadav, “An integrated
approach for detecting vacant spots,” Multimedia Tools and
approach toward smart parking implementation for smart
Applications, pp. 1–41, 2024.
cities in India,” in Advances in Data and Information Sciences:
Proceedings of ICDIS 2019, pp. 343–349, Springer, Cham,
Switzerland, January 2020.
[15] S. Mudaliar, S. Agali, S. Mudhol, and C. Jambotkar, “IoT
based smart car parking system,” Int J Sci Adv Res Technol,
vol. 5, no. 1, pp. 270–272, 2019.
[16] E. A. Abdellaoui Alaoui and S. C. Koumetio Tekouabou,
“Parking availability prediction in smart city,” in Innovations
in Smart Cities Applications Volume 4: Te Proceedings of the
5th International Conference on Smart City Applications,
pp. 450–462, Springer, Cham, Switzerland, February 2021.
[17] E. Scornet, “Random forests and kernel methods,” IEEE
Transactions on Information Teory, vol. 62, no. 3,
pp. 1485–1500, 2016.