A Literature Review of Automated Roadside Parking Monitoring Using Artificial Intelligence Algorithms

Abstract

The issue of parking has been a major concern in urban centers, primarily due to the increasing demand and daily traffic congestion. This paper endeavors to explore, process, and evaluate the existing literature on parking space detection methodologies, integrating photogrammetric techniques with deep learning models. Towards that end, various existing systems, applications, and models that have been studied were evaluated, and their impact on different test cases was assessed. The literature review was based on the guidelines of the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA). Results indicated that smart parking systems significantly enhance dynamic parking management by leveraging deep learning techniques, particularly convolutional neural networks (CNNs). These systems process visual data from monitoring sources to generate statistics, diagrams, and maps that highlight occupied and available parking spaces, allowing for more efficient parking management and improved traffic flow. These methods contributed to improved urban mobility by providing real-time information to drivers about parking conditions along their routes. This not only enhanced convenience but also supported the development of smarter and more sustainable urban transportation solutions.

1. Introduction

In recent years, rapid population growth combined with escalating urbanization has created significant challenges in urban areas, particularly regarding traffic congestion and parking [1,2]. Traffic congestion remains one of the most pressing issues in metropolitan areas worldwide, contributing significantly to time loss, economic costs, and environmental degradation. One of the most pressing issues arising from these trends is parking, which requires immediate and effective solutions. The demand for parking spaces in central and peripheral areas varies widely on a daily basis, making vehicle movement a time-consuming process that negatively affects traffic flow and the quality of the environment.

Research has shown that, in highly congested countries, drivers spend an average of up to 150 h per year searching for an available parking space [3]. In London, this figure reaches 67 h per year [4], while in Paris, drivers spend approximately 15 min per trip searching for parking [5]. Despite the availability of cost-effective alternatives, such as shared mobility services and public transportation, car ownership continues to rise, with the EU and US accounting for roughly one car for every two people [6]. The inability to reserve a parking space in advance, before starting the journey to it, is observed to hamper urban mobility and negatively affect the driver’s psychological state [7]. Additionally, parking demand peaks during weekends and public holidays, particularly in commercial areas, leading to longer search times and higher environmental impact [8]. For instance, studies estimate that parking search leads to the burning of 8.37 million liters of gasoline, which in turn causes the emission of 29,000 tons of CO₂ per year [9], while the vehicle trips of the UK population for 2012 exceeded 80% [8].

Effectively tackling these complex challenges necessitates solutions that surpass traditional methods. In the past, transportation planning and management largely depended on deterministic models and heuristic approaches. While these methods proved useful in certain situations, they are often inadequate for addressing the growing complexity and dynamic behavior of contemporary mobility systems. The rise of digital transformation, however, has introduced new possibilities, particularly through the adoption of advanced technological tools. Notably, artificial intelligence (AI), machine learning (ML), and deep learning (DL) have emerged as key drivers of innovation in transportation [10]. These technologies facilitate data-driven decision-making, enabling real-time processing of large datasets, optimization of operations, trend forecasting, and enhanced system resilience.

In response, smart mobility and parking systems (SPSs) have been developed to facilitate more efficient parking management [11]. These systems leverage technologies such as the Internet of Things (IoT), artificial intelligence (AI), big data analytics, sensor networks, and unmanned aerial vehicles (UAVs) to monitor parking occupancy and predict availability in real time [12]. Emerging approaches increasingly exploit UAVs to capture high-resolution images of parking areas, which can then be processed through photogrammetric techniques. Access to real-time information significantly improves drivers’ efficiency, with studies reporting up to a 45% reduction in the time spent searching for parking, thereby decreasing unnecessary vehicle emissions [13].

Many current smart parking frameworks rely on sensors to monitor parking occupancy and communicate this information to users via Bluetooth or WiFi. More recent approaches integrate computer vision and deep learning techniques, such as convolutional neural networks (CNNs), to process real-time images and provide immediate updates on space availability [12]. However, these methods still face challenges, including high computational requirements, increased system complexity, and limited scalability, particularly in resource-constrained environments.

Various types of parking environments have been considered in the literature, including urban streets, private lots, university campuses, and commercial centers, each presenting unique challenges for data collection, system scalability, and user interaction. Despite these advances, current systems continue to face challenges related to architecture, connectivity, and scalability, underscoring the need for more efficient, user-friendly solutions, particularly in constrained environments such as university campuses [14]. These gaps highlight the need for more efficient, user-friendly, scalable solutions and justify a comprehensive review of existing methodologies, algorithms, and technologies in the field of parking management.

Within this context, the aim of this paper was to critically review the methods and algorithms that have been developed to study the inventory of available parking spaces. To achieve this objective, a thorough literature review was conducted, in which the various methodologies, applications, and technologies that have been implemented and used were explored. Particular emphasis was placed on examining the capabilities and results offered by the tools for recording the area of interest, as well as the algorithms that have been modeled for the detection of available parking spaces.

This paper is structured as follows: The first part presents an overview of the parking issue, challenges, and possible solutions. The theoretical background is then discussed, which includes the interdisciplinary approaches to parking management and the methods used to implement an effective management plan. This is followed by the section, which focuses on the methodological framework adopted in each research and addresses the advanced methods and algorithms considered to create an automated smart parking management methodology. This is followed by a discussion, which aims to highlight the main conclusions about parking management. Finally, the main results are summarized, and recommendations for possible ways of implementation are made.

2. Background

In recent years, the management of urban parking spaces has evolved rapidly as researchers have sought to overcome the limitations of traditional detection and monitoring systems. Technological innovation has enabled the shift toward more adaptive and automated approaches to parking management. Modern detection algorithms commonly employ image-processing techniques supported by motion or ultrasonic sensors, fixed cameras, and, more recently, UAVs.

Although sensor-based systems are widely used in parking management, they often involve considerable installation and maintenance costs. On the other hand, the use of cameras is in the same cost range, but there is a high risk of interference from environmental and weather parameters. For example, light-colored vehicles exposed to intense sunlight can mislead detection software, leading to the incorrect identification of an occupied space as vacant. Similarly, a shaded area can be misinterpreted as a dark-colored vehicle, resulting in an empty space being considered occupied. In addition, sudden changes in lighting due to cloud movement can affect detection accuracy [15].

Due to these limitations, UAVs have emerged as a promising alternative [6]. They provide a significant advantage through a bird’s-eye view, facilitating coverage of a wider area that varies depending on flight altitude [16]. The drone industry has experienced significant growth and continues to be applied in various research contexts [17]. UAV imagery provides high visual resolution and a broad field of view compared to closed-circuit cameras, enabling data collection from multiple directions simultaneously [18]. However, automatic interpretation of UAV data is computationally demanding, requiring integration with computer vision techniques to extract accurate and high-quality information.

AI tools have further enabled the development of scalable data pipelines that combine distributed computing, spatial data systems, and geospatial machine learning methods for large-scale transportation analysis [10]. Cloud-based frameworks enable real-time processing, while AI at the edge reduces latency and communication overhead. Collectively, these technologies enable intelligent transportation models to learn from complex, heterogeneous datasets while maintaining spatial and contextual integrity [19].

The analyses and interpretations required for data from any type of equipment, especially given the specificity of study cases and environmental conditions, are demanding over time. As a result, computer vision has become integral to these technologies [20]. Deep convolutional neural networks (CNNs) describe image content optimally, and research has shown that pre-trained CNNs achieve excellent performance in various image recognition and object detection tasks [21]. However, due to the increased particle collision rate at higher altitudes, CNNs deployed on UAVs must be designed to be resilient to minor errors [22].

Through applications combining image analysis, object recognition, deep learning algorithms, and real-time control systems, efficient allocation of parking spaces under varying lighting and occupancy conditions has been observed [23]. Intelligent system design provides drivers with a comprehensive overview of the current parking situation, including the number of parked vehicles and the location of vacant spaces [24]. Automated parking search systems employ methods such as user interface design, classification of vacant and occupied spaces, space marking, and infrastructure analysis [25]. The performance of these systems varies with environmental conditions, creating the need for continuous testing.

Recent developments in autonomous vehicles have also influenced parking management research. Pheromone-based route planning algorithms allow autonomous vehicles to navigate parking areas efficiently, avoiding conflicts between multiple vehicles and improving both safety and traffic flow [13].

With the integration of these intelligent systems, users are expected to experience improved mobility [26], while parking becomes more convenient, sustainable, and efficient for all road network users [27]. The combination of UAV monitoring, AI-based analytics, and autonomous vehicle technologies points toward a future of adaptive and fully automated urban parking management, capable of responding to dynamic traffic demands and complex urban environments.

3. Methodology

3.1. Study Selection Process and Criteria

This paper explores the data collection methods and methodologies related to parking space detection. To achieve this purpose, the literature search was performed using interdisciplinary sources, such as ScienceDirect, Scopus, ΙΕΕ, Google Scholar, with the aim of identifying studies investigating parking availability detection and smart parking management, covering approaches from classical photogrammetry to artificial intelligence and deep learning techniques. Search terms were combined using Boolean operators (and, or) to ensure comprehensive coverage.

The selection process initially included reviewing the titles, abstracts, and methodologies of the studies. Studies published from 2000 onwards were considered, focusing on research that investigated parking space detection, intelligent monitoring, and related algorithms, while non-English articles, conference abstracts without full text, and studies unrelated to parking management were excluded. From the initial set of records identified through the databases, duplicates were removed, and the remaining studies were screened, resulting in 68 papers included for in-depth analysis. It is worth noting that interest in the study of parking availability has increased significantly since 2015, with a greater emphasis on methods based on algorithms and Deep Learning models compared to older approaches.

For each included study, data were extracted systematically, including author(s), year, country, type of data collection (e.g., cameras, UAVs, sensors, multi-tool systems), detection algorithm or method (e.g., Deep Learning models, YOLO, photogrammetric techniques), accuracy and performance metrics, and the type of parking environment (e.g., on-street, off-street, university campus, or truck parking). Additionally, a risk-of-bias assessment was conducted considering methodological robustness, sample size, dataset coverage, validation procedures, reproducibility, and potential conflicts of interest.

Table 1. Search terms utilized for the systematic literature review.

Database	Key Search Phrase	Search String (Boolean Operators)	Identification Papers	Included Papers
ScienceDirect	parking space detection	“parking space detection” OR “smart parking” AND “drones” OR “photogrammetric methods” OR “Artificial Intelligence” OR “Machine Learning” OR “Deep Learning”	264	68
Scopus	parking monitoring		48
IEE	photogrammetric methods for parking		182
Google Scholar	parking management		183

presents the key search phrases, search terms, and the number of identified and included papers used in the systematic literature review. It summarizes for each of the 68 selected studies the authors, year of publication, country, data collection method (e.g., cameras, UAVs, sensors), algorithmic approach (e.g., deep learning models, photogrammetry), performance indicators, and parking environment type.

3.2. The PRISMA Procedure

The literature review adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework for searching and selecting scientific studies and was conducted in February 2025. The initial process involved identifying relevant studies, screening them for appropriateness, selecting suitable research, and integrating it into the main theme of the article. The assessment and extraction of key elements from each study were the most time-consuming stages, with the initial search beginning in October 2024. Figure 1 presents the PRISMA flowchart illustrating the systematic process of identifying, screening, selecting, and including studies in the literature review.

4. Related Works

Τhe increasing demand for free parking spaces and congestion in urban areas worsen traffic conditions. While organized parking facilities offer a potential solution, their high cost limits their widespread use. Therefore, an effective parking management tool is essential to promote a more user-friendly and sustainable urban environment. In this context, this paper presents an extensive literature review of studies, focusing on parking detection and monitoring through different technologies and algorithms, evaluating their effectiveness and comparing the results depending on the equipment and methods used.

Figure 2 illustrates the percentage distribution of studies based on the equipment used for data collection, highlighting the varying methodologies employed in parking space detection research. It reveals that 38% of the studies considered had developed relevant methodologies based on camera data, while 19% used motion or sound sensors, 16% used combined equipment, and 16% and 11% used UAV and AVM systems, respectively.

4.1. Camera

One of the most widespread systems that have been used to record the area of interest in order to study parking availability is cameras. In general, several photogrammetric methods have been analyzed, and in recent years, there has been the development of methodologies based on deep learning techniques, through which more direct results are obtained. To better understand the evolution of smart parking detection systems, several studies have investigated both traditional and advanced approaches, ranging from early computer vision methods to modern deep learning architectures.

Through recent studies, a literature review of smart parking methods has been carried out. To begin with, Hadi [28] searched, presented, and compared smart systems implemented based on computer vision, while Liu [23] examined cost-effective parking management techniques and the vehicle guidance system based on Dijkstra’s algorithm for predicting the parking space status, with remarkable results.

Chandrasekaran [29] developed a web-based application based on an algorithm written in Python 3.0. By monitoring pixel intensity variations, the system visually indicates available (green) and occupied (red) spaces. To distinguish the accuracy of the system, they applied the methodology on 2 different parking spaces and obtained remarkable detection results. Similarly, Naraharisetti [24] created a web-based application to support a university community with visual entity, feature recognition, and illegal parking detection. Using the database data, they evaluated the collected data with high accuracy for different parking lots and environmental conditions.

Tatulea [30] studied the availability of parking spaces detection using the MOG2 (Mixture of Gaussians) algorithm to remove the background and separate moving objects (vehicles) from the fixed background. In addition, they smoothed the edges of entities and removed noise. The accuracy rates are obtained through 4 basic indicators, namely false positive (FP), false negative (FN), true positive (TP), and true negative (TN), and in the tests performed, they were quite remarkable.

Mehta [31] proposed a way to separate available and occupied parking spaces using the SURF match and the RANSAC algorithm, converting RGB to HSV to image sets, while through the application of transformation and normalization techniques, the areas of interest appear in different colors depending on their status: red (crowded) or green (empty).

Postigo [32] detected parking spaces using video images, background removal, and transient. The proposed system determined its reliability through simulations in urban areas under different conditions. Another study on the analysis of camera data was carried out by Jung [33] through stereo classification of pixel structures based on their characteristics. They proposed an algorithm using Hough transform and histograms of the depth of the obstacles and to create a panoramic view of the marking of the boundaries of each parking space, in order to locate the parking spaces.

While these traditional computer vision techniques achieved promising results, their performance was often affected by lighting conditions, perspective, and computational limitations. Consequently, research attention gradually shifted toward deep learning-based models capable of extracting higher-level visual features and performing more robust classification.

Amarasooriya [34] applied the YOLO n3 model on PKLot database images. The model was trained on different images from the same database. They used an algorithm to compare the coordinates of the parking space’s boundary frames with those of the cars, in real time with high accuracy. Garcia [35] studied the analysis of images obtained from a camera mounted in zenithal position using a CNN to detect and monitor vacant parking spaces. On a set of images, they applied photogrammetric techniques (filtering, thresholding, etc.) to create pre-scan polygons and check them using their real coordinates. Through CNN techniques, the system maps and classifies parking spaces according to their status while simultaneously evaluating performance time and accuracy.

Taylor [36] created two models: one using a Haar Cascade Classifier, which showed 95% accuracy, and another one using a pre-trained Mask-R-CNN algorithm, which created segmentation masks, separated objects from the background, and determined their exact boundaries. Similarly, Li [37] developed a methodology by reconstructing the parking location in a coordinate system, using photogrammetric procedures such as checkerboard, least squares, and triangulation techniques. By comparing mask R-CNN, YOLOn3, and RetineNet algorithms, the accuracy of identifying feature points was measured, which gave the coordinates of the center of the parking space, and then reconstructed a 3D coordinate system.

A different approach with remarkable results was applied by Zambanini [38]. They studied vehicle detection using stereo satellite images and the Faster-R-CNN algorithm. They isolated regions of interest through DSM and then performed classification of the objects using the geometric and qualitative information. Amato [39] created a new dataset of images taken at different times of the year and implemented a CNN-based algorithm, training it on PKLot, which already exists in the literature, and CNRPark EXT (proposed set), and the results were quite remarkable.

Building on these CNN-based approaches, recent studies have focused on optimizing YOLO architectures and integrating them with edge computing, cloud platforms, and Internet of Things (IoT) frameworks to enable real-time, cost-effective, and highly accurate smart parking systems.

Lu [40] developed a comprehensive intelligent parking space prediction and management system based on the PkLot dataset. They applied preprocessing techniques including resizing, brightness enhancement, background removal using a mixture of Gaussians, and grayscale conversion. Feature extraction was performed with an improved pre-trained U-Net, and feature selection used the modified chaotic BAT algorithm. Classification was conducted via deep cascaded fine-tuned active learning combined with Inception V3, while digital twin technology enabled dynamic simulation and real-time parking management. The system achieved exceptionally high performance, with 99.72% accuracy, 99.87% sensitivity, 98.68% precision, and 98.85% F1-score. Morris [41] implemented a system specifically for truck parking compliance using a multi-camera array and structure-from-motion techniques to reconstruct a 3D representation of the parking area. Predefined 3D volumes corresponding to each truck parking space were analyzed, and the presence of vehicles was determined based on point cloud density. This approach achieved high detection accuracy per parking space.

Aswath [14] developed a lightweight smart parking system designed for university campuses, integrating YOLOn8 with TensorFlow Lite (TFLite) for efficient real-time operation on a Raspberry Pi edge device. The system enabled license plate detection and recognition while supporting online booking functionality. Similarly, Luong [42] proposed a low-cost, cloud-based real-time parking system that uses a YOLO-based algorithm for vehicle recognition, with data processed on the AWS Cloud and made accessible to users through a Progressive Web Application (PWA). Their approach incorporated automatic data labeling using Grounding DINO and the Segment Anything Model (SAM) before YOLO training, achieving over 98% accuracy in clear-view areas, with minor errors observed in distant or partially occluded regions.

Ranjan [43] developed an optimized YOLOn5-based system by incorporating multi-scale feature pyramid networks and attention modules. Their methodology aimed to accurately detect both parked vehicles and empty spaces under diverse conditions, including varying lighting and vehicle types. The researchers also created a new comprehensive dataset from CCTV images, and their experimental results demonstrated high accuracy and real-time effectiveness, outperforming existing methods. Similarly, Ramesh [44] designed a YOLOn8 and customized a convolutional neural network (CNN) system for parking space recognition and classification. They created a dataset of 13,786 images labeled as “empty” or “occupied” under various environmental and lighting conditions. Using data augmentation techniques such as inversion, rotation, and scaling, their system achieved 99% accuracy on the training set and 97% on the validation set, demonstrating robust generalization and efficiency.

Santosh and Kumar [45] proposed the SmartPark Visionaire system, which leveraged YOLOn8 combined with extensive preprocessing and data augmentation. Their approach achieved very high mean average precision (mAP) for both empty and occupied spaces while reducing latency compared to previous YOLO versions. Morell [1] addressed on-street parking detection using non-static public traffic cameras. Their method combined pre-trained YOLOn8 and Detectron2 models with image processing techniques like perspective transformation and CLAHE (contrast-limited adaptive histogram equalization). Although detection was affected by low image quality and frequent camera movement, this approach provided a low-cost solution by utilizing existing CCTV infrastructure instead of installing new sensors.

In parallel with these vision-based detection frameworks, several researchers have extended their work toward integrated management systems combining detection, recognition, billing, and navigation functionalities.

License plate recognition (LPR) and automated management of entry, exit, and billing have also been widely explored. Manivannan [46] designed an automated parking management system that combined Deep Learning and OCR technology to read and process license plate information for entry-exit monitoring, permit validation, and billing. The system maintained high accuracy, achieving over 94% in detection and counting even under challenging conditions. Pranaav [47] expanded on this approach by integrating ANPR with a special space allocation algorithm considering factors such as floor preference, distance from destination, and special zones. They also incorporated an augmented reality (AR) navigation system to guide users to their assigned spots and applied dynamic pricing to optimize occupancy during peak hours, all managed through an IoT-driven engine. Sarker [48] developed an intelligent parking system that combines vehicle detection and license plate recognition for real-time monitoring and automatic assignment of free spaces, improving parking space utilization and overall system efficiency.

Beyond detection and management, some studies have also explored intelligent monitoring and behavioral analysis within parking environments. Micheloni [49] developed a monitoring model, which operates based on the classification models of the behavior of the events that occur. Practically, it can recognize its property and conclude whether this movement leads to a normal movement or to a suspicious or dangerous movement.

Overall, these studies demonstrate the continuous evolution from traditional image-based systems toward fully automated, AI-driven smart parking frameworks capable of real-time decision-making and dynamic resource optimization. Comparing the reviewed studies, it becomes evident that early vision-based systems laid the foundation for parking detection, while modern YOLO and CNN-based approaches significantly enhanced accuracy and scalability. Furthermore, the integration of IoT, AR, and cloud computing has transformed these systems from simple detection tools into comprehensive smart mobility solutions.

4.2. AVM System—Fisheye Images

Beyond traditional static cameras, there are a number of studies that have tested AVM systems. This is a special camera system placed on each side of the vehicle that provides a 360° panoramic view of the environment. These systems are usually used in automatic vehicles so that they can, with the development of special algorithms, guide them in movement along a specific path.

Huang [50] combined panoramic imaging with deep learning techniques, testing two different models based on FASTER-R-CNN for parking detection. They applied techniques such as noise reduction, binarization, and Hough transformation to obtain the coordinates of the vertices of the parking spaces and calculated the degree of sensitivity. Hsu and Chen [51] implemented a fisheye image-based detection-edge algorithm. The algorithm detects the boundaries of the spaces, determines the type (parallel, perpendicular, and angled) according to the width of the boundaries, and uses the RANSAC algorithm to classify the status of the spaces with green and red boxes.

Zhang [52] applied PSDL code to large-scale image data. This new approach consists of “T” or “L” shaped marking patterns, while detecting marking points and extracting valid parking spaces from the detected marking patterns. The accuracy rates showed high variation in relation to the detection success.

Li [53] tested the BEV classification method, which operates with top-down scene views to better recognize the parking space status. The proposed method can correctly recognize three types of marking points under various lighting conditions and angle types. After training the method, they performed tests and statistical analysis with the MV classifier, and the results of the precision and recall indices were very high.

Zong and Chen [54] investigated the line segment detector (LSD) method for detecting parking lines and the Kalman Filter algorithm for monitoring the parking space status. The research emphasizes that the proposed system is robust to shadow and scene diversity, while its comparative analysis with the corresponding ultrasonic sensor-based and pillar-based methods revealed its superior performance.

Lee and Seo [55] used Bayesian networks and a support vector machine (SVM) classifier for space classification, proving that they are more economical and reliable. Its tests were compared with other similar methods on different sets of images regarding lighting and obstacles, and the results showed the highest success rate or very little variation compared to existing methodologies.

Finally, Wang [56] used the Radon transform for edge detection and developed a dual circular path tracking algorithm. The proposed algorithm detects the lines by forming sinograms, and in the tests carried out by the researchers, it was compared with the Hough classifier based on the precision and recall values. To examine the proposed vehicle trajectory, a simulation was performed in MATLAB and compared with the traditional Pure PID method.

Collectively, these AVM-based approaches demonstrate the transition from static, infrastructure-dependent vision systems to onboard, autonomous solutions capable of understanding the parking environment in real time. By integrating multi-camera panoramic imaging with advanced algorithms such as CNNs, BEV models, and probabilistic classifiers, these studies highlight the growing convergence between intelligent vehicle systems and smart parking technologies.

4.3. Unmanned Aerial Vehicles (UAVs)

In several recent studies, unmanned aerial vehicles (UAVs) combined with deep learning techniques offer innovative findings for parking management applications. These approaches aim to overcome the spatial limitations of fixed cameras and provide a broader, real-time view of parking environments.

Firstly, Lai [57] attempted to identify occupied and vacant parking spaces using aerial images (geo-tagged). Object detection was performed using YOLOn8, which distinguishes between three categories: empty spaces, occupied spaces, and license plates. EasyOCR was used for license plate recognition, with preprocessing techniques to improve accuracy. The results showed high performance of YOLOn8 and a significant improvement in OCR accuracy after preprocessing.

Margaris [58] studied parking detection through predefined routes within organized parking lots. Detection was based on the YOLOn3-tiny algorithm, which was trained on images from the PkLot and CARPK databases to recognize empty and occupied parking spaces in real time. The results achieved a mAP of 69% (at a threshold of 0.45). However, the main weakness was found in the recognition of free spaces, with the AP falling below 50%. To address this, the researchers suggest enriching the dataset through image augmentation.

Kim [6] proposed an automated parking recording system based on aerial images from drones. In the city of Pully, Switzerland, they took multiple shots, recording different parking spaces, and applied the Oriented R-CNN model, which incorporates the new boosted pseudo-labeling technique. This model was used to study the stability of parked vehicles in successive frames, improving detection accuracy without the need for manual annotations. The methodology allowed for the recording of occupancy rates, the investigation of usage patterns in on-street and off-street spaces, as well as the turnover rate per space. The results showed that the proposed approach significantly improved detection accuracy, especially in difficult cases with partially hidden vehicles.

Ammour [59] created a vehicle detection algorithm using a VGG16 CNN model. More specifically, after acquiring the images, they were processed using image segmentation, feature extraction through convolutional layers, and classification with an SVM to detect the presence or absence of vehicles. The proposed system, after being trained, was tested in a university campus parking lot by analyzing the sensitivity of the results, changing the image segmentation parameters, and comparing them with the corresponding results of the normalized correlation similarity measure.

Beyond conventional detection methods, several researchers have explored hybrid or optimization-based frameworks to enhance accuracy and adaptability under real-world conditions. Li [60] implemented a parking status detection based on a generative adversarial network (GAN) and Voronoi partition for entity separation. They tested it on different datasets (normal or rotated images) and under different conditions, and the success results were very high. Jausevac [7] trained a CNN model combined with the traveling salesman problem (shortest route search) and vehicle routing problem (capacity, time windows) algorithms in different environments, and the results of the system’s accuracy were quite high; however, they emphasized the importance of taking more images in real environments in future studies.

D’Aloia [61] developed a method for detecting empty parking spaces by printed ground markers in UAV images. The method is based on comparing the profiles of parking spaces with a reference marker in place. The photogrammetric image processing procedures applied included grayscale application, edge detection, Sobel operator application, normalized cross-correlation, labeling, and distortion techniques.

Some researchers focused on supporting functionalities such as vehicle identification, traffic management, and space utilization analysis. Dasilva [62] created a license plate recognition application by using Open Alpr algorithm, which recognizes license plates in order to notify the illegal or non-illegal presence of vehicles in a space on a university campus.

Kujawski and Nürnberg [63] tried to integrate UAV images into the analysis and management of parking and traffic in urban areas, using photogrammetric techniques (Gaussian blur and threshold removal) and the SURF algorithm. The first scenario they tested was about monitoring the impact of delivery truck lane occupation on vehicle speed, and the second was about parking space occupancy and car rotation throughout the day. Smith [64] developed a Python-based system applying the YOLO v8 algorithm to count vehicles, and they managed to achieve high-quality results in the university campus.

Finally, Srivastava [22] reviewed deep learning methods through UAV images, analyzing accuracy, computational cost, their similarities and differences, laying the foundation for future challenges in this field.

Overall, UAV-based approaches demonstrate a clear shift from static, ground-level monitoring systems to dynamic, aerial intelligence frameworks. These methods not only improve detection coverage and accuracy but also enable large-scale, flexible, and adaptive parking management solutions that can be integrated with IoT and smart city infrastructures.

4.4. Sensors

The use of sensors in the detection of vehicles and free parking spaces has been extensively studied, showing promising findings. The review analyzed and presented the existing smart parking systems, for which artificial intelligence technologies have been developed, as well as the difficulties and possibilities they offer, and gave a series of suggestions on how future proposals can be improved.

A similar analysis was carried out by Meshrama [65] and Fahim [66], who reviewed smart parking management systems, classifying them based on equipment, algorithms, and data security concerns. Furthermore, Chandrahasan [67] presented the smart parking systems that have been proposed so far, through the analysis of the advantages of each method, their technologies, and architecture. Given disadvantages such as requiring a large number of sensors, application exclusively at short distances, and problems associated with user identification, the authors emphasize the need to eliminate these and improve the methods.

In more practical applications, Ram [68] implemented an availability control based on an ultrasonic sensor through the Arduino microcontroller in real time. The sensors and the control system were appropriately configured, and a mock-up was created by the researcher; the results were positive, suggesting that the proposal would be very useful in parking management in an environmentally friendly way. Similarly, Cynthia [69] created an Arduino-based sensor network application for monitoring parking availability through sensor networks for managing reservations.

Expanding the focus from simple sensor-based detection to intelligent prediction, Awan [70] evaluated five different parking availability detection models based on artificial intelligence and deep learning technologies, using the Santander database. The results showed that the decision tree model presented the best results, while a less computationally complex algorithm, such as KNN, had better results than the more computationally complex algorithms.

Arjona [71] tested LSTM and GRU methods for studying parking availability in different areas. The results showed that the GRU architecture performed better, while errors were reduced through the proposed exogenous variables. Additionally, weather and calendar changes contributed to better fluctuation modeling.

Chatzigiannakis [72] emphasized the importance of hiding drivers’ personal data in a smart parking system, proposed a new application that provides privacy to user information using elliptic curve cryptography (ECC) and zero-knowledge proofs (ZKPs). For its testing, they used a large number of sensors and information displays to examine the performance of the algorithms on different IoT devices.

In a more specialized context, Barone [7] created SPARK, a wireless sensor-based parking system for parking management. They performed a statistical comparison with existing parking systems through different scenarios and indicators, and it showed promising results despite being in its early stages. Groh [73] developed an algorithm using a particle filter to detect static and semi-static objects in indoor parking lots. An autonomous vehicle with a sensor and a laser scanner identifies parking spaces, visualizes detected points, and classifies elements into a mapped representation. Similarly, Zhou [74] used embedded laser scanners for bumper recognition, clustering scanned data to detect parking spaces. The algorithm was tested in various environments and parking conditions, showing high accuracy.

Yang and Riesgo [75] created a real-time parking space detection system based on wireless sensor networks, while Park [76] proposed a parking space detection method using an ultrasonic sensor and an algorithm based on edge detection through the multi-echo function. The test concerned the comparative analysis of single and multi-echo in real time on different vehicles and with different echo operating values, demonstrating its effectiveness in various parking conditions.

Overall, sensor-based parking systems demonstrate a broad evolution—from basic Arduino and ultrasonic prototypes to AI-enhanced, laser-guided, and privacy-aware architectures. Their integration with artificial intelligence and IoT infrastructures allows for real-time, accurate, and secure parking management, although challenges remain regarding sensor deployment costs, data processing efficiency, and user privacy.

4.5. Combination of Equipment

Of particular interest are studies in which researchers have used multiple tools to capture images and, more generally, the data obtained to record the area. These approaches aim to enhance detection accuracy, optimize space allocation, and provide comprehensive coverage through the combination of cameras, sensors, and advanced algorithms.

Sree [77] conducted a literature review, comparing smart parking lot management by comparing their effectiveness and analyzing their advantages and disadvantages. Similarly, Ma [78] conducted a literature review about visual data collection techniques in parking detection, covering methods such as obstacle distance measurement, 3D mapping reconstruction, and straight-line identification. In the same context, Al-Turjman and Malekloo [79] classified smart parking systems based on their technology, equipment, coverage, cost, and reservation offers to users.

Elfaki [80] proposed an algorithm to optimize space allocation. More specifically, using cameras and motion sensor data, they created a system that initially detects the vehicle, recognizes its license plate number, and communicates with a central server to verify legality, proving effective in optimizing space allocation.

Suhr and Jung [81] proposed a parking detection system using cameras, motion, and ultrasonic wave sensors, and applied the Sobel operator, RANSAC algorithm, and dividing line calculation method. Differences in lighting were identified as the parameter that most affected the variation in accuracy. Bura [82] examined the detection of parking lots using a network of cameras and a lidar sensor for range measurement. Initially, they created an open access image database and a model based on edge calculation, achieving high accuracy.

Deka [83] presented a 3D mapping method for parking space detection, which is based on near-field terrestrial photogrammetry techniques. Practically, through multiple different tests, they aimed to achieve the selection of the most ideal camera calibration parameters for the best possible recording of points of interest (parking location).

Suhr [84] proposed stereoscopic 3D motion reconstruction. More specifically, using a fisheye camera at the rear of a vehicle, they detected points of interest and tried to render them in 3D, through the methods of denotation and similarity, which were found to be quite high-performing, especially when it was supported by ground measurement data from a laser scanner sensor. Moreover, Jung [85] proposed a parking detection method using a camera and a laser scanner, based on scanning angle and the area value calculations. The results showed that the algorithm is capable of detecting the points of interest; however, the cost of implementation and maintenance is quite high.

Xu [86] designed an intelligent system for detecting empty parking spaces and generating a route to them through the RCE neural system. Ιmages were scanned to isolate contour points, and by using the least square method, the contour lines and equations in the image plane were obtained. Detected elements were converted to a 2D coordinate system, checked for empty spaces, and mapped to ground coordinates.

Canli and Toklu [87] created a methodological model for mobile device parking search based on long short-term memory (LSTM) technology. The system calculates the probability of finding parking through an open database. They combined with parameters such as capacity, density, time, day, and holidays; the accuracy rate reached 99%.

5. Discussion

Detecting free parking spaces and, more generally, estimating the availability of parking spots at specific locations remains a significant challenge for research. Early computer vision approaches, such as traditional image processing methods applied to camera data [30,32,48,61], achieved high accuracy but were highly sensitive to lighting conditions, camera perspective, and background complexity. Similarly, automated number plate recognition (ANPR) systems [47,48,82] offered effective monitoring in controlled environments but faced limitations in large-scale urban settings.

Recent advances in deep learning have significantly improved detection accuracy and real-time performance. CNN-based architectures [34,35,44] and YOLO variants [14,42,43] have demonstrated high performance across diverse datasets, achieving over 95% accuracy in many cases. When combined with preprocessing and feature extraction techniques from high-resolution images, these models enable reliable classification of parking spaces under varying environmental conditions.

A comparative analysis of the reviewed studies revealed that performance differences between methods, such as YOLO and Mask R-CNN, are closely tied to their underlying architectures and deployment contexts. Single-shot detectors like YOLO are optimized for speed and perform well in real-time monitoring scenarios with moderate occlusion and clearly defined parking boundaries, making them suitable for camera and UAV-based applications. In contrast, two-stage models such as Mask R-CNN provide more precise segmentation and boundary delineation, which is advantageous in densely parked or overlapping vehicle scenarios but requires higher computational resources and longer inference times. Similarly, sensor-based approaches (e.g., ultrasonic, LiDAR, motion sensors) tend to outperform camera-based systems in adverse weather or low-visibility environments, as they are less affected by illumination changes or occlusions. However, cameras and deep learning models excel in scalability, spatial coverage, and integration with user-facing applications. These performance trade-offs highlight that the effectiveness of each method is not universal but depends on environmental conditions, hardware constraints, and the need for spatial granularity versus processing efficiency.

Beyond static cameras, UAV-based methods [6,57,58] offer the advantages of broader coverage and flexibility, enabling real-time monitoring of on- and off-street parking spaces. Similarly, advanced vehicle-mounted (AVM) systems [50,52,55] provide 360° environmental awareness, improving detection in complex scenarios, although implementation costs and integration overhead remain higher than traditional setups.

Hybrid approaches that combine cameras, sensors, and AI algorithms [77,81,82] further enhance detection reliability and allow more comprehensive parking management, integrating occupancy measurements, turnover, and demand data. Collectively, these developments illustrate a clear shift from simple detection systems toward intelligent AI-driven frameworks capable of real-time, scalable, and context-aware parking management.

Despite numerous studies integrating image feature extraction, classification technologies, and sensor-based applications (e.g., motion, ultrasound), a methodology for detecting parking spaces in urban areas using orthophotos and photogrammetric processing remains underexplored.

This study proposes a comprehensive and innovative methodology for monitoring and managing parking spaces in both urban and non-urban areas. The approach leverages camera data with photogrammetry and AI algorithms, combining 2D and 3D models with real textures and digital surface models (DSMs) through techniques such as structure-from-motion (SfM) and stereoscopic multi-view system (MVS). The methodology can also incorporate critical traffic indicators, including demand, turnover, and occupancy rates.

The results contribute to a detailed understanding of parking dynamics and their impact on vehicle traffic, providing data that can inform future policies and support parking management strategies aimed at road safety and enhancing the driver experience. As previous studies have shown, integrating deep learning and AI algorithms improves detection results while delivering real-time information to users. The novelty of this study lies in training models directly on high-resolution orthophotos rather than typical aerial or ground images, enabling learning from geometrically corrected and georeferenced data, improving accuracy in complex urban environments, and facilitating integration with 2D and 3D spatial analyses.

Evaluation includes detecting vacant spaces through feature classification based on pixel intensity, identifying space boundaries, and measuring distances between points of interest. The primary applications focus on organized parking spaces and small-scale areas, with factors such as parking conditions, weather, environment, and lighting affecting accuracy, while economic considerations influence sustainability assessment.

To better illustrate these methodological gaps, the following table summarizes the main technological categories used in the reviewed studies, including ground-level cameras, UAVs, AVM systems, sensor-based approaches, and combined frameworks, and evaluates them according to key criteria such as accuracy, cost, coverage, and weather sensitivity. This synthesis provides a clear basis for understanding the trade-offs among technologies and highlights where empirical testing and cross-validation remain limited. The qualitative scale (+++, ++, + corresponds to high, medium, or lowlevels, respectively.

Table 2. Overview of the main types of equipment used for parking detection.

Type of Equipment	Accuracy	Cost	Coverage	Weather Sensitivity
Camera (Ground-level)	+++	+	++	++
AVM (360° Vehicle Cameras)	++	++	+	++
UAV	+++	++	+++	+++
Sensors (Ultrasonic, Lidar, Motion)	++	++	+	+
Combination	+++	+++	+++	++

+++ high, ++ medium, + low.

presents an overview of the main types of equipment used for parking detection.

While camera-based methods provide cost-effective real-time monitoring at small scales, UAV-based systems extend coverage but require complex image processing and regulatory considerations. AVM solutions enable onboard detection with panoramic awareness, though they are primarily applicable to autonomous vehicles and involve higher integration overhead. Sensor-based systems (e.g., ultrasonic or LiDAR) are robust under controlled conditions but have limited spatial reach and higher deployment costs. Finally, hybrid systems that combine multiple modalities achieve superior accuracy and spatial awareness, yet remain technically complex and economically demanding.

At the same time, many previous studies have focused on isolated types of environments, such as university campuses [14,48,58] or organized parking facilities [40,44], without performing comparative analyses across different contexts. Urban and rural areas, and mixed-use parking facilities have rarely been evaluated in parallel, limiting the ability to generalize results and highlighting the need for cross-comparisons in heterogeneous environments.

Statistical analyses of key traffic metrics, such as parking demand, occupancy rates, and turnover, have often been overlooked, reducing the ability to assess real-world applicability. Most studies report algorithmic performance using metrics like accuracy, repeatability, and recall, but few examine false positives, false negatives, or environmental dependencies (e.g., lighting, weather, obstruction). These omissions make it difficult to determine which detection approach performs best under varying operational conditions.

Sensor-based methods (ultrasonics, lasers, motion) and ANPR systems [7,47,82] offer reliable detection in controlled environments but face challenges regarding scalability, cost, and privacy risks. Camera-based methods, including traditional image processing [30,32] and deep learning approaches (YOLO, CNN) [34,35,43], generally achieve high accuracy but remain sensitive to environmental factors such as lighting, perspective, and weather conditions. UAV-based systems [6,57,58] provide high-resolution aerial coverage and real-time monitoring, yet few studies integrate UAV data with 3D reconstruction or realistic surface digital models (DSMs), limiting the potential for spatially integrated traffic analysis.

Comparative evaluations across the reviewed studies indicate that YOLO-based models frequently outperform two-stage detectors such as Faster R-CNN in parking space detection due to their real-time inference capabilities and lower computational cost. The single-shot architecture of YOLO enables rapid detection across large monitoring areas, making it advantageous in dynamic environments such as on-street surveillance, UAV-based imaging, and smart parking systems requiring immediate feedback. However, its limitations are also evident in specific edge cases: motorcycles, bicycles, and compact vehicles are more challenging to detect due to smaller object size and reduced feature resolution. Similarly, snow-covered cars, backlit vehicles, and shadowed regions introduce ambiguity that can increase false negatives or reduce localization accuracy. In contrast, two-stage models like Faster R-CNN may provide more precise bounding in controlled or high-resolution scenarios, yet their slower processing time and higher computational demands restrict real-time deployment.

Advanced vehicle-mounted (AVM) systems [50,52,55] provide 360° environmental awareness, but their economic feasibility and integration complexity are significantly higher than camera- or UAV-based alternatives. Hybrid approaches that combine cameras, sensors, and artificial intelligence [77,81,82] enhance detection reliability and allow multi-parameter analysis.

In addition to performance considerations, the operational deployment of UAV-based parking monitoring requires a clear Concept of Operations (CONOPS) [88]. This involves defining flight cadence in relation to parking area size and revisit frequency, estimating recharge or battery—swap intervals, and determining mission—planning parameters such as fleet size, coverage routes, and coordination with ground infrastructure. Scenario-based assessments also help evaluate service levels by accounting for connectivity stability, bandwidth availability for real-time data transmission, and weather-related constraints that may limit flight endurance or sensor performance [89].

Although deep learning and hybrid frameworks improve detection performance, most studies still underrepresent challenging conditions such as night-time scenes, backlit environments, and weather-induced occlusions (e.g., rain, fog, snow). Techniques such as multispectral or thermal sensors, temporal fusion, and targeted data augmentation remain underutilized.

The above comments are summarized in detail for each study in the following table (

Table 3. Overview of the technologies used for parking assessments, along with the advantages and disadvantages per study.

Author	Year	Equipment Type	Technical Performance and Efficiency				Operational Trade-Offs
			Detection Algorithms	Accuracy	Dataset	Environment	Cost	Integration Overhead	Privacy Risk	Weather Condition Dependence
[14]	2025	Camera	YOLOn8/Tesseract (OCR)	>75%	Custom	University Campus	Low	Low	High	High
[40]	2025	Camera	Several models (KNN, MLP, DT, etc.)	>95%	●	Parking Lot	High	High	Low	●
[42]	2025	Camera	YOLO-based	87.39% and 98.13%	Custom	University Campus	Low	Medium	Low	High
[46]	2025	Camera	Several yolo version/OCR	>95%	Real time	●	Low	Medium	High	Medium
[58]	2025	UAV	YOLOn3/YOLOn3-tiny	69%	CARPK, PKLot	University Campus	Low	High	Low	Medium
[1]	2025	Camera	YOLOn8/Detectron2	●	Open Data Images	On-street parking space	Low	Low	Medium	Low
[43]	2025	Camera	YOLOn5	95%	Custom	Parking Lot	Low	Low	Low	Medium
[44]	2025	Camera	YOLOn8	>95%	Custom	Parking Lot	Low	Low	Low	Medium
[47]	2025	Camera	ANPR, OCR	>95%	Custom	Parking Lot	●	Medium	High	●
[45]	2025	Camera	YOLOn8	●	●	Parking Lot	●	Medium	Medium	High
[48]	2025	Camera	Traditional Image Processing	>95%	●	Parking Lot	●	Medium	High	●
[28]	2024	Camera	●	●	●	●	●	●	●	●
[57]	2024	UAV	YOLOn8/EasyOCR	>95%	Custom	Parking Lot	Medium	High	High	High
[41]	2024	Camera	Multi-view Structure and Motion	>90%	Region-wide Deployment	Truck Parking	Medium	High	Low	Medium
[64]	2024	UAV	YOLOn8	100%	Real time	On-street parking space	High	High	Low	High
[6]	2024	UAV	YOLOn5x/Oriented R-CNN	90.20%	COCO/DOTA 1.0	On/Off-street parking space	Medium	High	Low	High
[23]	2023	Camera	Computer Vision	>90%	PKLot	University Campus	Low	High (Computing power)	Low	High
[63]	2023	UAV	●	●	●	●	●	●	●	High
[17]	2023	UAV	TSP/VRP	>95%	Custom	On-street parking space	●	High	●	High
[77]	2023	Camera, sensors	●	●	●	●	●	●	●	●
[80]	2023	Camera, motion sensor, and range-finder sensor	SSD MobileNet V2/EasyOCR	>85% and >70%	●	University Campus	Medium	Medium	High	Medium
[65]	2023	Sensor	●	●	●	●	●	●	●	●
[34]	2023	Camera	YOLOn3	>80%	PKLot	University Campus	Low	●	Low	Medium
[29]	2022	Camera	Django	●	Real time	Parking Lot	High	Medium	Medium	High
[24]	2022	Camera	ANPR	●	Real time	University Campus	High	High	High	High
[35]	2022	Camera	Traditional Image Processing	>90%	Real time	Parking Lot	●	High	Low	●
[50]	2022	AVM	faster R-CNN	>80%	●	●	●	High	●	Low
[83]	2022	Camera, total station	3d Model/DSM	●	●	On-street parking space	High	High	Low	Medium
[68]	2022	Sensor	Arduino	●	●	●	Low	High	●	●
[36]	2021	Camera	Mask R-CNN	>75%	COCO	On/Off-street parking space	●	High	Low	●
[37]	2021	Camera	RetineNet/Mask RCNN/YOLOn3	●	MC COCO	Indoor Parking Lot	●	High	Low	●
[22]	2021	UAV	YOLO/FRCNN	>95%	Custom	Parking Lot	Medium	High	Low	Medium
[78]	2021	Camera, sensors	●	●	●	●	●	●	●	●
[66]	2021	Sensor	●	●	●	●	●	●	●	●
[87]	2021	●	LSTM	>95%	ISTPark dataset	Parking Lot	Medium	Low	Medium	High
[38]	2020	Camera	SIFT, SURF, ORB	>80%	Stereo satellite images	Parking Lot	Medium	High	Low	High
[70]	2020	Sensor	Multilayer Perceptron, K-Nearest Neighbors, Decision	70–95%	Santander’s parking dataset	●	●	High	Low	Low
[71]	2020	Sensor	Tree, Random Forest, and Voting Classifier	>95%	Real time	Parking lot	●	High	Low	Low
[30]	2019	Camera	Traditional Image Processing	>90%	Custom	Outdoor Parking Lot	Low	Low	Low	Medium
[51]	2019	AVM	Traditional Image Processing	●	Custom	Indoor/Outdoor Parking Lot	Medium	Medium	Low	Medium
[79]	2019	Camera, sensors	●	●	●	●	●	●	●	●
[52]	2018	AVM	Vision-Based PSD L	>95%	Custom	Parking Lot	Medium	Medium	Low	Medium
[53]	2018	AVM	Line segment detection (LSD), line clustering method, BEV, and MV classifier	>95%	Custom	Parking Lot	Low	Medium	Low	High
[54]	2018	AVM	Fusion method (Ultrasonic + Pillar-based fusion)	>75%	Real time	Parking Lot	Medium	Medium	Low	High
[81]	2018	AVM, sensors	RANSAC, Sobel edge operator	>90%	Real time	Parking Lot	Low	Low	Low	High
[82]	2018	Camera, sensors	Ting YOLO, AlexNet	>90%	PKLot, CNRPark, Custom	Parking Lot	Low	●	High	●
[69]	2018	Sensor	Arduino IDE	●	Real time	Parking Lot	●	●	●	●
[39]	2017	Camera	AlexNet	>95%	CNRPark-EXT/PKLot	Parking lot	●	●	Low	Medium
[31]	2017	Camera	SURF, RANSAC	●	Custom	Parking Lot	●	High	Medium	Medium
[60]	2017	UAV	GAN	●	PKLot	University Campus	High	High	●	High
[62]	2017	UAV	Bayesian network, SVM classifier	Custom	Real time	Parking Lot	High	●	Medium	●
[59]	2017	UAV	SVM	>70%	Custom	University Campus	High	High	High	●
[55]	2016	AVM	Slot Context Analysis (SVM/HOG)	●	Custom	Parking Lot	Meduim	Medium	Medium	●
[72]	2016	Sensor	Elliptic Curve Cryptography—ECC, Zero-Knowledge Proofs—ZKP	●	●	●	High	High	Low	●
[67]	2016	Sensor	●	●	●	●	●	●	●	●
[32]	2015	Camera	Traditional Image Processing	●	Custom	Parking Lot	Low	●	Low	High
[61]	2015	UAV	Traditional Image Processing	>95%	Custom	Parking Lot	Low	●	Low	●
[56]	2014	AVM	Levenberg–Marquardt, Bilinear interpolation algorithm, Radon Transform, K-means	>95%	Custom	Parking Lot	Low	Medium	Low	Low
[7]	2014	Sensor	IPA system	●	●	On-street parking space	Low	High	Low	●
[73]	2014	Sensor	Particle filter algorithm, KLD algorithm	●	Real time	Indoor Parking Lot	High	High	Low	●
[74]	2012	Sensor	AdaBoost	>80%	●	On/Off-street parking space	High	High	Low	●
[75]	2012	Sensor	Multiple echo function	●	Real time	On/Off-street parking space	High	High	Low	Low
[84]	2010	AVM, laser scanner	RANSAC	Medium	Real time	Parking Lot	High	High	Low	Medium
[85]	2008	Camera, scanning laser radar	Scanning laser radar	●	Real time	Parking Lot	High	Medium	Low	●
[76]	2008	Sensor	Ultrasonic Sensor	●	Real time	●	●	●	●	●
[33]	2006	Camera	3D Stereo Vision/Template Matching	●	Real time	Parking lot	High	High	Low	Low
[49]	2006	Camera	Visual Surveillance/Behavior Analysis (Conceptual)	●	●	Parking lot	●	High	High	●
[86]	2000	AVM, laser scanner	Vision-Guided (Onboard)	●	Real time	University Campus	●	●	Low	●

(●) indicate that the corresponding element is not explicitly mentioned in the study or has not been experimentally evaluated.

), which provides a comparative evaluation of the proposed methods based on multiple indicators, such as cost, diversity of datasets, real-time performance, and weather dependence. More specifically, for each work, the type of equipment and algorithms used, the type of database (either custom or publicly available), the reported accuracy, and the testing environment, as well as operational indicators such as cost, integration complexity, privacy risk, and weather dependence, are presented.

Finally, a practical gap remains in the use of orthophotos and photogrammetric data as inputs for deep learning models. While traditional aerial and ground images have been widely used, geometrically corrected and georeferenced orthophotos could improve detection accuracy in complex urban environments, facilitate 3D modeling, and enable integration with traffic demand and turnover measurements, areas that remain largely unexplored.

Overall, parking space detection intersects with several fundamental challenges in computer vision, including occlusion, lighting variability, and perspective distortion. Vehicles partially hidden by other objects, shadows, or structural elements reduce detection robustness, while strong illumination contrasts, nighttime visibility, and weather-related interference further degrade model performance. Perspective distortions from elevated cameras, UAV imagery, or angled street views complicate geometric interpretation and weaken segmentation accuracy. Addressing these challenges is essential for integrating parking monitoring into broader smart city strategies, where interoperability with traffic management systems, IoT infrastructures, and urban planning policies is required. As cities adopt intelligent transport and mobility frameworks, parking detection technologies can contribute to demand-responsive pricing, congestion mitigation, multimodal connectivity, and data-driven land-use decision-making.

5.1. Limitations

Through the review and analysis of the existing literature, several limitations have been identified that should be considered when interpreting the results. Firstly, numerous studies have explored camera-based, UAV-based, sensor-based, and hybrid systems; however, a lack of consistent comparative evaluation across these technologies remains evident. Differences in spatial scale, environmental conditions, and data acquisition settings make it difficult to generalize results and determine the most efficient approach for real-world applications.

In addition to the above limitations, current approaches remain vulnerable to visual degradation caused by night-time conditions, backlighting, rain, fog, or snow. Only a few studies have explored targeted strategies such as multispectral or thermal imaging, de-weathering and contrast-enhancement techniques, or temporal consistency across video frames to mitigate occlusions and illumination variability. Furthermore, data augmentation methods addressing adverse weather and low-light scenarios are rarely implemented, despite their proven value in improving model generalization. Future research should integrate these techniques to enhance robustness and reliability under real-world operational conditions.

5.2. Future Research Directions

To address the limitations discussed above regarding methodologies based on Artificial Intelligence algorithms for parking monitoring, particularly those utilizing UAV imagery, several avenues for future research emerge. UAV-based monitoring has demonstrated high spatial resolution and improved accuracy in detecting available parking spaces and violations. However, factors such as lighting variations, feature similarities, and environmental conditions can significantly affect the performance of deep learning models. This underscores the need for enhanced training datasets and the exploration of alternative image processing techniques to improve model robustness.

Moreover, the inclusion of diverse data types is crucial for evaluating and improving the performance of vehicle monitoring and detection methodologies. Collecting data across multiple parking environments, times of day, geographical areas, traffic levels, demand, and occupancy conditions enables a more thorough assessment of algorithm performance and identifies situations in which models may fail. Integrating real-time contextual information, including driving conditions, traffic levels, and temporal patterns, can further enhance model training and operational performance.

The creation of 2D and 3D urban environment models through structure-from-motion (SfM) and multi-view stereo (MVS) techniques allows for more sophisticated analyses and innovative approaches to parking management. Transportation planning can benefit from combining accurate infrastructure mapping with spatial-temporal traffic and demand modeling. Future research should also consider demographic and environmental parameters to analyze varied urban needs and behaviors. Overall, research directions should aim at improving the accuracy, generalization, and practical applicability of AI algorithms by integrating advanced analytical techniques, diverse datasets, and social and traffic parameters to support efficient parking management and a user-friendly urban environment.

A critical technical insight concerns the choice of deep learning algorithms for parking detection. Single-shot detectors, such as YOLO, often outperform two-stage detectors like Faster R-CNN in parking applications because they process entire images in a single pass, enabling real-time detection with lower computational costs [35,43]. Nevertheless, YOLO exhibits limitations in edge cases: small objects like motorcycles or scooters may be missed; overlapping vehicles can cause false detections; and extreme conditions, including snow-covered cars, shadows, or backlighting, can reduce accuracy. Addressing these issues may require preprocessing techniques, multispectral or thermal imaging, and training datasets that incorporate diverse edge scenarios to improve robustness.

Despite the potential of UAV imagery, real-time occupancy prediction remains underexplored. Technical challenges include limited UAV endurance, high-resolution image transmission bandwidth and latency, and the computational demands of onboard or edge-based AI inference. Synchronizing UAV data with ground-based sensors or IoT networks adds complexity but offers the opportunity to create dynamic, large-scale monitoring systems for parking availability. Overcoming these limitations would enable richer datasets, facilitate real-time decision-making, and enhance the scalability of AI-based parking monitoring.

Fundamental computer vision challenges, such as occlusion, lighting variance, and perspective distortion, continue to affect detection performance. Vehicles partially blocked by other objects, diverse illumination conditions, and variable camera perspectives can lead to misdetections. Integrating UAV- and AI-based approaches with urban planning frameworks and smart city infrastructures can optimize parking allocation, improve traffic flow, and enhance driver convenience.

Additionally, autonomous vehicles (AVs) introduce a critical, yet overlooked, direction for future research regarding urban parking management and land use, which will significantly complement these technologies. AVs drastically reduce central urban parking demand by enabling passengers to be dropped off while the vehicle self-parks in remote, low-cost peripheral lots. Furthermore, they allow for tighter, more efficient parking designs [90], potentially reducing the urban parking footprint and freeing up central urban land currently reserved for parking. The authors of [91] further emphasize the need for research into the local policy implications required to facilitate these shifts in AV-related areas. Integrating AV operations with UAV monitoring, AI inference, IoT networks, and 3D urban modeling offers the most comprehensive approach to sustainable and adaptive urban parking management.

Overall, future research should focus on developing robust, context-aware AI models, integrating diverse data sources (UAVs, ground sensors, IoT), and addressing practical constraints such as computational costs, bandwidth, and environmental variability. Emphasis on edge cases, real-time performance, and urban integration will enhance both the generalization and operational utility of AI-based parking monitoring systems, driving smarter, more resilient, and user-centric urban mobility solutions.

6. Conclusions

Parking is an issue that has been a concern for many countries for many years and remains one of the most difficult issues to resolve to this day. It is a major factor contributing to traffic congestion and, in many cases, is the cause of road crashes.

The aim of this research was to evaluate existing studies that have led to the development of methodologies and applications for the identification of empty parking spaces and monitoring of their variation. This objective was achieved through the analysis of the literature review, highlighting the positive effects they bring to the quality of life of citizens and the environment. Through this research, a thorough classification of the equipment used for data collection, with regard to parking spaces, analysis of the technologies and methods that have been used to date, including different photogrammetric techniques and deep learning techniques, was performed.

In addition, the main goal was to identify areas that have not been examined until now according to the current literature by utilizing new methods and techniques, as well as the use of different equipment and application cases. Through a review of the existing literature, an attempt was made to understand the different methods used both in photogrammetric techniques and in the creation of deep learning models. More specifically, of utmost importance for the research was the study of the results obtained in each case of application of entity classifiers—image features and models, the analysis of the advantages and disadvantages in each situation, and the parameters that affect accuracy.

Research showed that the results of applying such methodologies offer proper parking management, organization, and optimization of traffic, while reducing congestion and fuel consumption. The use of smart systems allows for dynamic allocation, reducing driver search for spaces and increasing parking efficiency while enhancing urban sustainability.

The integration of smart systems technologies offers more direct and easier movement of citizens in cities while playing an important role in minimizing delays. It offers the monitoring and simultaneous analysis of traffic flow in real time, reducing environmental pollution and facilitating the development of a friendly urban environment. By highlighting the systems and methods used, the possibility of studying different parameters in various conditions is given while distinguishing the needs of users. Through continuous coordination of all developing systems, the sustainability and long-term effectiveness of the solutions are ensured.