Introduction

Information and communication technology integration is a fundamental element of the smart city concept, which aims to improve quality of life, encourage sustainable development, and boost economic growth1,2. To achieve the goals of smart cities, it is necessary to develop smart transit, implement efficient traffic management systems that try to mitigate the effects of congestion, and improve trip planning3.

Smart parking systems have emerged as essential solutions to these issues, addressing everyday urban challenges and enabling the development of smart, sustainable cities. Cruising for parking not only consumes time and fuel but also contributes to environmental and energy inefficiencies. Analytical models show that fully distributed edge/fog architectures can use 14–25% less total energy (including IT, networking, and cooling) than centralised clouds4. Drivers will have to search for parking manually, which might be expensive and time-consuming5. Additionally, in densely populated areas, cars increase greenhouse gas emissions and waste fuel6. One solution is to implement smart parking systems that allow cars to locate and reserve parking spaces ahead of time. Big cities throughout the world hold major events like conferences, religious gatherings, parades, and festivals, which makes this kind of infrastructure very important. These policies ease transportation, safeguard cities and the environment, and lessen pollution by encouraging the effective use of automobiles5.

Nowadays, traffic management strategies must include smart parking systems since they facilitate parking by enabling more efficient car parking, which shortens arrival times and eases traffic. Smart parking systems have drawn interest from both academic and commercial researchers because of their advantages in terms of cost, environmental impact, and aesthetic appeal. The IoT is a key technology in this area since it enables real-time monitoring and control of smart items as well as communication between them. Looking ahead, emerging sixth-generation (6G) networks promise ultra-low latency, enhanced reliability, precise localisation, and native edge-computing support–capabilities that will further empower IoT-driven intelligent transportation and smart parking solutions7. The IoT improves the effectiveness of smart parking systems by gathering real-time sensor data and merging it with cutting-edge technologies such as computational techniques, machine learning, and sensor fusion8,9,10.

Due to the growing number of cars and parking spots, parking has become a significant issue in both urban and corporate environments. Due to a lack of demand and accessible space, parking has become a challenging and time-consuming task for drivers. Finding parking frequently becomes increasingly difficult, time-consuming, and frustrating for drivers as cities and infrastructure expand. Ineffective parking management also affects the environment by increasing emissions and fuel consumption from parked cars as they look for a spot11. Improvements in transportation are required to shorten routes, lessen time limitations, and make better use of available space due to urbanisation and traffic congestion12.

Conventional parking systems are wasteful and unproductive because they require drivers to manually look for parking spots. In this situation, cars spend a lot of time looking for their location without receiving precise information, which can lead to traffic bottlenecks and congestion in already congested locations13. This lack of direction leads to congested parking lots and irritated drivers, particularly during events or peak hours. This leads to overcrowded parking lots, wasted time, and frustration for drivers. The lack of real-time information about parking availability results in increased travel time, traffic congestion, and unnecessary air pollution. According to estimates, cars looking for parking in crowded metropolitan areas contribute for up to 30% of traffic congestion, which not only wastes time but also raises pollution levels in the area14. In many cases, vehicles circle around parking areas looking for a spot, creating bottlenecks and contributing to urban chaos. Moreover, poor parking management systems often fail to effectively utilise available parking spaces, leaving some spots unused while others remain overly crowded. Several commercial parking systems offer some features such as remote parking-spot booking, fee payment, interactive parking maps, and others. However, their main drawback is the high cost and the limited development capability for public developers since they are not open source. Furthermore, the practical value of many of these commercial solutions for larger, dynamic contexts is limited since they are not scalable to meet the changing demands of smart cities or to adjust to the particular requirements of various institutional and urban settings. Additionally, the existing systems are unable to fulfil the demands of the evolving smart city as well as the parking requirements of today15.

A smart parking system can address these issues by providing real-time information to drivers and optimising the use of parking spaces. Smart parking systems improve visibility into parking availability and provide a user-friendly experience that increases efficiency and lessens congestion by combining digital infrastructure with data-driven insights. A key technology in this solution is fog computing, which processes data closer to where it is generated, offering faster response times than traditional cloud computing. Fog computing allows data from parking sensors to be processed locally, reducing latency and enabling drivers to receive immediate updates on available parking spots. Recent research in environmental monitoring systems has similarly highlighted the need for in-situ data processing near sensor nodes to maintain continuity, reduce latency, and ensure resilience, particularly in energy-constrained or hard-to-reach locations. Such findings further validate the advantages of fog computing for real-time, distributed applications across various domains, including smart parking infrastructures16. This real-time data helps drivers make informed decisions, reducing the time spent searching for a parking space and alleviating traffic congestion17.

Cloud computing, while popular for centralised data storage and processing, presents challenges in high-demand environments like parking systems. Cloud-based systems often experience delays due to the time required to send data to and from centralised servers18,19,20,21. This delay creates a significant disadvantage in applications like parking management, where seconds matter, as drivers depend on immediate information to make decisions quickly. In the context of smart parking, where real-time information is critical, these delays can negatively impact the user experience. Fog computing, on the other hand, provides faster, decentralised data processing by distributing computing tasks across local nodes near parking areas. This allows for real-time updates and more reliable service, especially in areas with fluctuating demand.

In this paper, we propose a four-layer smart parking system architecture that leverages fog computing as an intermediary layer. The critical task of processing parking lot images and displaying the status of parking spots on LED displays is handled by strategically positioned fog nodes within the fog layer. If a specific lot is fully occupied, the LED display provides real-time information about the nearest available parking spot. This dynamic and immediate update minimises traffic congestion and reduces vehicle emissions by significantly cutting down the time drivers spend searching for parking. Parking management is a persistent challenge for institutions like SDU University, which experiences high volumes of vehicular traffic on campus. Drivers often face long queues and stress while searching for parking, leading to traffic congestion and suboptimal utilisation of limited parking spaces. These challenges underscore the pressing need for sustainable and efficient parking solutions. Fog computing-based smart parking systems offer a promising approach by processing parking data locally and providing real-time updates, thereby streamlining operations, reducing search times, and improving user experience.

In the proposed fog-based smart parking system, sensors installed across parking lots detect available spots and relay this information to nearby fog nodes for rapid processing22,23. This architecture not only optimises parking space utilisation but also eliminates the need for manual spot searches. By distributing data processing across multiple fog nodes, the system enhances resilience to network outages and improves traffic management within parking facilities.

This paper aims to design and implement a fog computing-based smart parking system to address parking management challenges, alleviate traffic congestion, and enhance the overall driver experience. The primary contributions of this work are as follows:stop

  1. 1.

    Multi-layer Design: We propose a multi-layer architecture combining fog and cloud computing to improve data processing efficiency and reduce the time required to access parking information.

  2. 2.

    Real-time Monitoring: We introduce a distributed sensor-based real-time monitoring system that provides immediate updates on parking availability, enabling drivers to locate vacant spots more efficiently.

  3. 3.

    Predictive Analytics Model: A predictive analytics framework is presented, leveraging historical parking data to forecast availability, facilitating better decision-making for drivers and improving overall parking space utilisation.

  4. 4.

    Experimental Validation: Deployment at SDU University demonstrates significant reductions in search times and traffic congestion, validating the effectiveness of the fog computing approach compared to traditional cloud-based systems.

Related work

The evolution of smart parking systems has been propelled by rapid advances in IoT, cloud, and more recently, fog and edge computing technologies24. These technological innovations have not only enhanced the capacity to process vast amounts of data in real-time but have also enabled the development of sophisticated, multi-tier architectures that can address the dynamic nature of urban mobility. Early contributions to the field laid the foundation for such architectures, with researchers such as Badr et al.25, Buldakov et al.26, and Zhang et al.27 exploring pricing structures and contract models to optimise resource allocation and user engagement. Meanwhile, other studies by Karare et al.28, Aravinthkumar et al.29, and Singh et al.30 delved into the usability challenges and application-specific requirements, ensuring that these systems could be effectively integrated into diverse urban settings. In addition to addressing user-centric challenges, research by Karare et al.28, Buldakov et al.26, and Zhang et al.27 also focused on developing robust methods for forecasting parking availability and optimally allocating resources, which are critical for managing traffic flow and reducing congestion. Collectively, this body of work has established the theoretical underpinnings necessary for understanding and implementing complex system integrations in urban environments, paving the way for more advanced, real-time smart parking solutions.

A research by Rathore et al.31 emphasised the broader applications of IoT in intelligent transportation, including dynamic ride-sharing and enhanced data security through edge/fog computing. Although Rathore et al. offered significant insights into system optimisation for large urban networks, their focus on large-scale environments left a gap in addressing the challenges of smaller, controlled contexts.

In a recent study by Wang et al.32, a loosely coupled middleware approach was proposed to integrate automatic valet parking (AVP) and smart parking systems, enabling the seamless management of both autonomous and manual vehicles within a unified parking area. This middleware acts as an intermediary layer, ensuring that all vehicle types can enter the parking system regardless of their operational mode. The study employs a Markov Chain model to evaluate improvements in system usability, and comparative assessments reveal that the proposed approach outperforms traditional tightly coupled systems. Notably, the middleware achieves an enhancement in space availability and reduction in traffic congestion, alongside improved cost efficiency.

Building on this foundation, researchers such as Awaisi et al.33 and Thangam et al.34 advanced the field by incorporating image processing and sensor-based approaches into their system designs. Awaisi et al. proposed a three-tier system featuring LED displays and multi-area fog nodes, while Thangam et al. developed a five-tier fog computing model that demonstrated superior performance in terms of latency, energy consumption, and overall execution time compared to traditional cloud-based approaches. Similarly, Hadi Ghahremani et al.35, introduce a multi-criteria IoT-based smart parking framework that prioritises candidate lots based on distance to the lot, vehicle distance, free-space count, previous failed attempts, and visit frequency. Their Arena-based simulation shows up to a 20% reduction in driver waiting time over existing methods.

Subsequent studies shifted focus towards integrating real-time tracking and dynamic pricing strategies. Singh et al.36 introduced a multi-layered intelligent parking system that leveraged resilient edge and cloud computing alongside low-power wireless communications and sensor nodes. Their work, although primarily targeted at urban settings, provided valuable insights into continuous vehicle tracking and the optimisation of parking space allocation through dynamic pricing.

Balfaqih et al.37 demonstrated a practical implementation of an IoT-based smart parking system in Makkah, employing a multi-layer approach that included sensor nodes, fog computing, and a reservation system. Similarly, Renuka et al.38 developed a camera-based real-time image processing system aimed at reducing search times for available parking spaces in congested areas. Both implementations highlight tangible benefits in reducing travel time and improving revenue, yet, they predominantly address urban scenarios.

Recent work has explored the use of intelligent scheduling techniques to optimise fog computing performance, particularly with respect to latency and energy constraints. For example, FoRLess applies deep reinforcement learning for Function-as-a-Service (FaaS) placement, achieving significant improvements in response time and energy efficiency across fog layers39. While our architecture does not yet incorporate reinforcement learning, the latency-aware, energy-conscious design aligns with similar motivations.

The literature reveals a progressive shift from broad conceptual models and pricing strategies to more integrated, technology-driven solutions that emphasise real-time data processing and predictive analytics. However, most existing studies focus on urban settings, often neglecting the unique challenges presented by localised environments such as university campuses.

Our study addresses this gap by implementing a fog-enabled IoT smart parking system specifically tailored to the context of SDU University. Recognising that university campuses face irregular parking demands influenced by academic schedules, peak periods, and special events, our research focuses on developing a system that minimises reliance on centralised cloud servers. By prioritising local, real-time decision-making through fog computing, we aim to optimise network usage, reduce energy consumption, and improve overall system responsiveness. The system has been evaluated using real-time data from SDU University, demonstrating significant improvements in latency reduction, efficient energy distribution between cloud and fog layers, and a marked decrease in overall execution time. All crucial factors for an environment with limited data centre capacity and variable visitor patterns.

Table 1 summarises the key studies discussed above, highlighting the proposed architectures, key technologies employed, and the limitations or gaps identified in each study.

Table 1 Summary of key studies in smart parking systems.
Full size table

Proposed system architecture

The proposed architecture leverages a multi-layered design to provide an efficient and scalable solution for smart parking management. By integrating edge processing, fog computing, and cloud computing, the system ensures real-time data processing, reduced latency, and optimised bandwidth usage. Through localised data gathering and processing, this architecture improves network efficiency in addition to addressing the issues with high-latency cloud systems. As seen in Fig. 1 , the architecture is organised into four separate levels, each of which is in charge of carrying out particular duties.

Figure 1
figure 1

A four-tiered smart parking system with fog functionality.

Full size image

Together, these layers provide smooth data transfer, real-time slot availability updates, and efficient parking resource management, enabling the system to adjust to a variety of parking situations and requirements.

System workflow

Drivers no longer need to use smartphone apps to verify the availability of parking spaces according to the planned system. Instead, cars can effectively find available spots due to real-time information shown on LED panels placed around the parking area. This strategy minimises the amount of time spent looking for parking spots, which substantially decreases fuel consumption and eases traffic congestion in the parking area.

The system uses a number of carefully positioned cameras to cover every parking lane in order to provide thorough surveillance. These cameras transmit data to microcontroller-equipped edge processing devices, which act as intermediary. Before interacting with fog nodes, which manage additional processing and aggregation, the edge units locally process the raw data. Fog nodes optimise network capacity and lower latency by periodically sending important data to the cloud. Furthermore, fog nodes placed throughout various parking lots work together to provide information about available slots. If a specific parking area is full, the fog node retrieves information from adjacent nodes and updates the LED displays accordingly.

By integrating edge processing and fog computing layers, the system mitigates the limitations associated with cloud computing, such as high latency and excessive bandwidth usage. Localised data processing improves responsiveness and network efficiency, providing timely updates to drivers while minimising environmental impact.

The simulation workflow, as illustrated in Fig. 2, begins with the initialisation of iFogSim2 and the configuration of fog devices, including cloud servers, proxy servers, fog nodes, and cameras.

Figure 2
figure 2

Workflow of the proposed architecture.

Full size image

Each parking area is equipped with fog nodes that connect to a central cloud server via a proxy. The application model, defining modules for detecting and tracking parking slots, is then mapped to these devices. Simulations are conducted to measure performance metrics such as latency, allowing evaluation of the system under varying configurations. This workflow enables flexible testing scenarios, including adjustments in the number of cameras and fog nodes.

To clarify the end-to-end processing pipeline from sensing through analytics, we summarise the key functional stages as follows:

  1. 1.

    Image Capture & Edge Processing: Overhead cameras capture a frame every 5 s. Edge units (500 MIPS, 1000 MB RAM) locally filter and detect slot occupancy, sending only occupancy metadata onward.

  2. 2.

    Fog Aggregation & Display Update: Fog nodes (2800 MIPS, 4000 MB RAM) receive edge metadata, merge data from multiple cameras, and update nearby LED panels in real time. When a lot is full, peer fog nodes exchange availability to guide drivers to adjacent lots.

  3. 3.

    Proxy Routing: A proxy server (2800 MIPS, 4000 MB RAM) securely routes summaries between fog clusters and the cloud, performing load balancing, protocol translation, and request filtering.

  4. 4.

    Cloud Storage & Analytics: The cloud layer (44,800 MIPS, 40,000 MB RAM) ingests only aggregated summaries for historical storage and trend analysis, avoiding high-frequency data flow.

The following sub-sections provide detailed descriptions of each architectural layer, including their primary functions and the hardware profiles used in our iFogSim2 simulations.

Layer 1: data acquisition and edge processing

There are two primary ways that the suggested fog-based vehicle parking system may be used: for a single parking space (as seen in Fig. 3) and for numerous parking spaces (as illustrated in Fig. 4). Cameras positioned above each parking slot capture a still image every 5 seconds. In our simulation model, we assume that basic occupancy detection is performed locally at the edge using lightweight filtering logic, and only occupancy metadata is forwarded to fog nodes. The actual image files are not transmitted; instead, image capture is modelled in terms of data size and transmission impact on network usage.

An edge processing unit with micro-controllers that determine slot availability and filter the input handles the first image processing locally. By doing this, less data is sent to the fog node, saving bandwidth on the network and cutting down on latency. One fog node manages the data in a single parking lot (Fig. 3) and sends it to the cloud server for long-term storage. There is a single fog node connected to a central cloud server for every parking space (Fig. 4). The amount of time and resources needed to upload and retrieve data from the centralised server rises as additional parking spaces connect to it, even while fog nodes maintain constant latency and network utilisation.

Figure 3
figure 3

System architecture for a single parking space.

Full size image

Layer 2: fog computing

Fog nodes comprise the second layer, which acts as a bridge between the proxy server and the edge processing units. In addition to doing more complicated calculations and updating local displays, such LED panels that highlight parking availability, these fog nodes also collect data from many edge units. Fog nodes further enhance system responsiveness by lowering latency by controlling frequent data updates locally and only sending vital information to the cloud at predetermined intervals.

Figure 4
figure 4

Architecture of the multi-space parking system.

Full size image

Layer 3: proxy server

The architecture relies heavily on the proxy server to improve security and speed. By load balancing, it equally distributes traffic among several backend servers, routes requests to them, and caches frequently requested data to lessen server strain and streamline response times. The proxy server enhances security further by screening out malicious requests, converting communication protocols between various system levels, and concealing the IP addresses of back-end servers. Strong security, optimal system performance, and effective data transfer are all ensured by these several roles working together.

Layer 4: cloud computing

The final tier is the cloud server, which manages storage, advanced analytics, and long-term data management. This layer makes it possible to analyse past data and optimise the system. It retains processed data for later use after receiving it from fog nodes. The cloud server ensures that comprehensive parking data is continuously updated, enabling trend analysis and possible system enhancements.

System architecture and component specifications

The proposed smart parking architecture is a hierarchical, four-layer system that integrates edge processing, fog computing, a proxy layer, and cloud computing. Each layer has a well-defined role in achieving low-latency, real-time parking availability detection and display. Figure 1 illustrates the complete architectural layout.

To enhance replicability and transparency, this subsection provides additional detail on how the proposed architecture was instantiated and simulated in the iFogSim2 environment. Each layer described earlier–edge, fog, proxy, and cloud–was mapped to specific virtual entities with realistic hardware profiles and network constraints.

Smart cameras at the edge act as both sensors and image sources, capturing real-time visuals of parking lanes at 5 s. intervals. These images are preprocessed by microcontroller-equipped edge units to minimise upstream data transmission. The processed results are then transmitted to fog nodes for aggregation and analysis. LED display units are updated locally based on this data, guiding drivers in real time.

In the simulation environment, fog nodes are modelled as medium-capacity devices (2800 MIPS, 4000 MB RAM), suitable for embedded servers or smart gateways. The proxy server plays a key intermediary role by handling routing, load balancing, and protocol conversion, while the cloud layer serves for historical analytics and storage.

Configuration parameters for each simulation component, including processing capacity, bandwidth, power usage, and latency, are summarised in Table 2. These values reflect the design choices for replicable, real-world system performance. Figure 2 shows the operational workflow, while Fig. 5 illustrates the topology of the simulated environment, including fog nodes, sensors, and connectivity paths.

This configuration ensures that each component’s capabilities are realistically constrained, enabling evaluation of latency, execution time, network usage, and energy distribution under scalable conditions, as later discussed in Sect. Implementation and Results.

Entity Interactions: Smart cameras operate as both sensors and image capture modules. Once images are processed at the edge, results are forwarded to the fog node, which in turn communicates with the LED display unit and peer fog nodes. When a specific area is full, fog nodes retrieve slot information from neighbouring lots and update the displays accordingly. This cooperative model ensures distributed decision-making and system scalability.

System Performance Flow: A request-response loop begins with the camera capturing occupancy data and ends with a visual LED signal to drivers. This loop is optimally partitioned: lightweight image processing happens at the edge, inference and aggregation at the fog, routing and security at the proxy, and archival and analytics at the cloud. This layered distribution enables latency-sensitive operations to remain local while offloading computationally expensive tasks to the cloud.

Scalability and Real-Time Responsiveness: The architecture is modular, allowing the addition of fog nodes and cameras without overloading the system. As the number of monitored parking lots increases, fog nodes continue to operate with stable latency (see Table 4), while the cloud remains unburdened by high-frequency transmissions. This ensures real-time performance and energy efficiency, key to practical deployment in smart city environments.

Simulation Realism: All components in the architecture are instantiated using the iFogSim2 simulator. The simulator is configured to replicate the communication latency, processing time, and energy consumption of each device in realistic urban deployment conditions. Parameters for each component–including processing rates, memory, bandwidth, power usage (idle and active), and uplink latency–are detailed in Table 2.

Together, these architectural decisions create a highly responsive, scalable, and energy-efficient smart parking system tailored for real-time urban mobility applications.

Implementation and results

This section describes the experimental setting used to assess the suggested parking management system, including how its components were put into practice and how performance measures were examined. A user interface for real-time feedback, processing nodes for analysis, and cameras for data acquisition are just a few of the modules that are integrated into the system’s architecture. A simulation tool for fog computing settings, iFogSim240, was utilised to simulate and evaluate our architecture’s performance. In order to guarantee efficient communication and peak performance in both fog and cloud computing environments, these modules have to be configured during installation. We thoroughly examined the system’s performance after deployment using a number of important measures, such as latency, execution time, cost of execution, and network use. According to the results, fog computing is more efficient than cloud computing and offers notable advantages in terms of reduced latency and expenses. These results show that fog architectures can handle the increasing demand of real-time applications, highlighting their potential.

Experimental setup

This section describes how we put our parking space tracking and detection technology into practice. We describe the setup of the cameras and fog nodes, among other system components. We also explain how we gradually increased the number of cameras and nodes to simulate different situations and evaluate their effects on system performance metrics like processing efficiency and latency. An extra node for the LED display was incorporated into each arrangement to guarantee real-time updates on open parking spaces.

Latency reduction is essential in real-time, high-performance settings. In order to respond to client devices promptly, fog computing processes data at the network edge and reduces reliance on the cloud. Images of parking slots are sent to fog nodes, which process them locally and speed up reaction times. Each fog node has enough processing power to update slot availability on LEDs with minimal latency since it covers a defined area. The latency is computed using Eq. (1).

$$\begin{aligned} L = \alpha + \mu + \varphi \end{aligned}$$
(1)

where \(\alpha\) is the multiple CPU execution delay for taking photographs, \(\mu\) is the time it takes to upload images to the fog node for processing and storage. Lastly, \(\varphi\) is the amount of time it takes for the LED to display the data following processing at the fog node.

We simulated three parking areas, with the first two initially equipped with 2 and 3 cameras, respectively, while the third area featured a display screen. Each parking area included a fog node that connected to a central cloud device via a proxy server. The smart Wi-Fi-enabled cameras functioned as sensors according to established policies. To evaluate different scenarios, we increased the number of cameras and analysed how this affected latency when processing data using either fog or cloud computing.

Figure 5 illustrates the iFogSim2 topology, which consists of three fog nodes, each connected to 2 or 3 cameras, with the final fog node designated for system monitoring. The cameras are equipped with a picture snippet module that captures images of parking slots every 5 seconds.

Figure 5
figure 5

iFogSim topology of 5 cameras and 1 display connected to 3 fog nodes and a cloud server via a proxy server.

Full size image

Figure 6 illustrates the application model, which is designed to efficiently detect and track available parking slots through a series of interconnected modules. This model serves as the backbone of the parking management system, facilitating real-time data processing and enhancing user experience by providing timely information on parking availability. Each module within the system plays a crucial role in ensuring seamless communication and functionality, allowing for quick detection and display of available slots. The model is structured as follows:

  1. 1.

    Camera: The camera records live video of the parking lot while looking for both occupied and available spots. The Slot Finder, the following module, receives this video data.

  2. 2.

    Slot Finder: The raw video data from the camera is processed by this module. Using the video stream, it locates vacant parking spaces and sends the information to the Slot Detector for additional examination.

  3. 3.

    Slot Detector: The Slot Finder processes the data and sends it to the Slot Detector. It identifies the precise position of open parking spaces and transmits this data to the Slot Tracker and User Interface. For users to be guided, the identified slot information is essential.

  4. 4.

    User Interface: The User Interface acts as the cloud-based component of the system. It displays the location of available parking slots to end users, providing real-time information about slot availability. The interface can show a map or a list of open slots.

  5. 5.

    Slot Tracker: This module tracks the status of parking slots in real-time, based on the input from the Slot Detector. It maintains up-to-date information about which slots are occupied or available and sends this data to the LED display.

  6. 6.

    LED Display: The LED Display visually indicates available parking slots. The Slot Tracker sends updated slot information to the LED display, which guides drivers to the available parking spots through visual indicators.

Figure 6
figure 6

Application model.

Full size image

The configuration parameters for the cloud server, proxy server, and fog server used in the fog-based scenario are presented in Table 2. These parameters encompass processing capability (measured in MIPS), RAM, uplink and downlink bandwidth, level, processing cost, as well as power consumption during busy and idle states.

Table 2 Configuration parameters for simulation entities.
Full size table

Figure 7 depicts the parking areas that we have considered in our research. The areas are numbered and blue square on the main gate where the LCD display screen is located to indicate the available spot.

Figure 7
figure 7

Overview of the parking area in SDU University.

Full size image
Figure 8
figure 8

Area 2.

Full size image
Figure 9
figure 9

Area 3.

Full size image
Figure 10
figure 10

Area 4.

Full size image
Figure 11
figure 11

Area 5.

Full size image
Figure 12
figure 12

Area 6.

Full size image

We explored various cases by systematically increasing the number of cameras and nodes deployed in the parking areas. Each case was designed to assess the impact of scaling on system performance, specifically focusing on latency and processing efficiency. Table 3 shows that the number of cameras varied from 7 to 42, while the number of nodes varied from 1 to 6. It is significant that there was always an extra node devoted to the LED display, which was essential in providing consumers with information about the available parking slots. We were able to examine how the overall efficacy of the parking spot recognition and tracking system was impacted by the setup’s growing complexity due to this structured methodology. Additionally, a number of performance indicators were calculated using the data from these simulations, offering insights into how modifications to the system design impact metrics like latency, execution time, and network utilisation.

Table 3 Simulation cases.
Full size table

Results and discussion

This section provides a thorough examination of the performance measures that contrast cloud computing with fog infrastructures in a number of scenarios. Table 4 summarises the findings, which show notable variations in latency, execution time, cost of execution, and network use. These measurements are essential for assessing each architecture’s viability and efficiency in real-time applications. Figures 8 and 9 show the parking lots in the back of SDU University, Figs. 10, 11 show the lots in front of the university, and Fig. 12 shows the lot behind the university residence hall. All figures show the available parking spots as green rectangles, emphasising the stationed camera systems’ monitoring coverage. To ensure effective and well-organised parking for students, staff, and guests, LED screens are also positioned thoughtfully across these spaces to show real-time parking information. The analysis underscores the advantages of fog computing, particularly in terms of lower latency and costs, even as the system scales. Specifically, the data illustrates that fog computing not only minimises end-to-end latency but also offers a more cost-effective solution when compared to its cloud counterpart. The following subsections delve deeper into each metric, providing mathematical formulations and visual representations to enhance understanding and facilitate comparison.

The fog-based architecture enhances scalability and network efficiency by processing data locally at fog nodes located near parking areas. This localised processing reduces the volume of data sent to the cloud, thereby minimising network congestion and latency. As more cameras and sensors are added, the distributed nature of fog computing allows the system to scale effectively by adding additional fog nodes, preventing bottlenecks associated with centralised cloud processing. Our results demonstrate that this approach maintains low latency and network usage even as the system expands, ensuring efficient real-time updates and improved overall performance.

Table 4 Summary of results for fog and cloud architectures.
Full size table

End-to-end Latency quantifies the total delay experienced in a system, which is essential for assessing performance in real-time applications. Latency is particularly critical in surveillance systems, where timely data processing and response are necessary for effective monitoring. As illustrated in Fig. 13, fog-based latency starts at a low value of 14.65 ms and shows a gradual increase to 367.34 ms under higher loads. In contrast, cloud-based latency begins at 209.25 ms, rapidly escalating to 2213.10 ms as the operational demands increase.

Figure 13
figure 13

End-to-end latency.

Full size image

This comparison underscores the significant advantage of fog computing, which demonstrates a more stable and lower latency profile compared to cloud solutions, especially as the system scales.

To compute end-to-end latency L we calculate the sum of delays at various stages, including sensor-to-tracker \(D_{st}\) and tracker-to-display \(D_{td}\) delays, as in Eq. (2):

$$\begin{aligned} L = D_{st} + D_{td} \end{aligned}$$
(2)

Cost of Execution evaluates the total expenses incurred during system operation, which is crucial for determining the economic feasibility of different architectures. As shown in Fig. 14, fog-based execution costs are significantly lower compared to cloud-based costs. For example, with one node, the fog-based cost is 36 thousand, while the cloud-based cost is much higher at about 1.27 million. As the number of nodes increases, fog costs rise steadily, reaching 131.8 thousand for six nodes, whereas cloud costs escalate significantly, peaking at around 3.86 million. This comparison highlights the economic advantage of fog computing over cloud solutions, especially as the system scales.

Figure 14
figure 14

Cost of execution.

Full size image

The total cost of execution C can be calculated using the following Eq. (3):

$$\begin{aligned} C = \sum (C_{n} + C_{d} + C_{p}) \end{aligned}$$
(3)

The cost of computational resources for each node is indicated by \(C_{n}\) in this equation, the cost of data transmission between nodes by \(C_{d}\), and the cost of data processing by \(C_{p}\). These factors together make up the overall cost structure, showing that whereas fog computing keeps prices steady and cheaper, cloud computing costs rise dramatically as the system expands because of increased demands on data processing and transport.

As a critical performance metric for evaluating real-time responsiveness, execution time calculates the overall amount of time needed to finish activities in the system. Figure 15 illustrates that execution speeds based on fog technology are consistently faster than those based on cloud computing. For instance, with one node, the fog-based execution time is 1242 ms, while the cloud-based execution time is 1633 ms. As the number of nodes increases, the gap widens, with fog times reaching 7859 ms for six nodes, whereas cloud times escalate to 10989 ms. This clearly demonstrates the superior efficiency of fog computing in reducing execution time, especially as the system scales.

Figure 15
figure 15

Total execution time.

Full size image

The total execution time T can be calculated using the Eq. (4):

$$\begin{aligned} T = \sum (T_{n} + T_{d} + T_{p}) \end{aligned}$$
(4)

In this equation, \(T_{n}\) represents the time required to perform the actual task at each node, \(T_{d}\) refers to the time taken to transfer data between nodes, and \(T_{p}\) accounts for the time spent processing the data. This equation helps in understanding the breakdown of execution times. While fog computing consistently maintains shorter times due to more localised processing, cloud computing faces higher delays, especially as more nodes are added to the system.

Network usage measures the total amount of data transmitted within the system, which is critical for evaluating the load on the network in different architectures. As shown in Fig. 16, fog-based network usage is considerably lower than cloud-based usage.

Figure 16
figure 16

Network usage comparison between fog-based and cloud-based systems.

Full size image

For instance, in the case of one area, the fog-based usage is around 22.5 thousand MB, while the cloud-based usage is significantly higher at approximately 307.6 thousand MB. As the number of areas increases, fog usage rises to 100.6 thousand MB for six areas, while cloud usage grows dramatically to 1.16 million MB. This highlights how fog computing can alleviate network congestion in large-scale systems.

The total network usage N can be calculated using the following Eq. (5):

$$\begin{aligned} N = \sum ((N_{ds} + N_{dr})\times \frac{\alpha }{B}). \end{aligned}$$
(5)

In this equation, N represents the total network usage, where \(N_{ds}\) is the data sent by the node or camera in area, \(N_{dr}\) is the data received by the processing unit, \(\alpha\) is the number of cameras in area, and B is the constant bandwidth set at 10,000. This formula captures the relationship between the number of cameras, bandwidth, and network usage. The number of cameras, increasing from 7 to 42 as shown on the secondary axis in Fig. 4, contributes to the rise in network usage in both fog and cloud systems. However, cloud systems create a much heavier burden on network resources, making them prone to congestion as the system scales.

Energy distribution quantifies the energy consumption across different components of the system, providing insights into the efficiency of fog and cloud architectures. As illustrated in Fig. 17 for fog execution and Fig. 18 for cloud execution, the energy consumption for fog systems is primarily concentrated in the camera and router components.

Figure 17
figure 17

Energy consumption distribution in fog-based smart parking system.

Full size image
Figure 18
figure 18

Energy consumption distribution in cloud-based smart parking system.

Full size image

For instance, with one node, the energy consumed by the camera is approximately 846,302 J, while the router consumes about 22,503.76 J, and the data centre accounts for 1,750 J. As the number of nodes increases to six, the energy consumed by the camera remains constant, while the router’s energy consumption rises to around 100,584.75 J, and the data centre’s consumption increases to approximately 15,025.5 J. Summary of obtained results shown in Table 5.

Table 5 Summary of energy distribution for fog and cloud architectures.
Full size table

In contrast, the cloud architecture exhibits a markedly different distribution of energy usage. With one node, the camera energy consumption is again 846,302 J, while the router does not consume any energy, and the data centre requires 307,638.06 J. As the system scales to six nodes, the camera’s energy consumption remains stable, but the data centre’s energy demand rises significantly to about 1,164,546.92 J. This stark difference in energy distribution emphasises that while fog computing maintains lower energy consumption in routers and utilises energy more efficiently, the cloud architecture relies heavily on data centres, leading to higher overall energy demands as the system scales.

The findings presented in this section underscore the significant advantages of fog computing relative to cloud computing across several key performance metrics. As summarised in Table 4, fog-based architectures consistently exhibit lower latency, reduced execution costs, shorter execution times, and decreased network usage across all evaluated scenarios. Notably, the end-to-end latency of fog systems remains markedly lower than that of cloud systems, thereby affirming their suitability for applications that require real-time data processing. Furthermore, the cost analysis demonstrates that fog computing is considerably more cost-effective than cloud computing, particularly as system scale increases. These results suggest that fog architectures can effectively alleviate the resource constraints commonly associated with cloud-based systems while better accommodating the demands of real-time applications. Overall, this analysis highlights the critical importance of selecting computing architectures that align with specific application requirements, positioning fog computing as a robust and economically viable alternative for enhanced performance and efficiency.

Conclusion and future directions

The comparative evaluation of cloud-based and fog-based smart parking systems has revealed that fog architectures offer significant advantages in scalability and network efficiency. The analysis of the fog computing-based smart parking system deployed at SDU University indicates that, compared to traditional cloud-based solutions, the fog approach achieves enhanced responsiveness and efficiency. By positioning fog nodes in close proximity to data sources, real-time processing is accelerated and network resources are managed more effectively, particularly as the number of surveillance cameras increases. This results in lower latency, reduced network usage, and an overall improved parking management experience, making fog computing especially well-suited for environments characterised by extensive sensor deployments, such as urban centres and academic campuses. Moreover, evaluations across diverse parking lot configurations and camera setups further substantiate the scalability and adaptability of the proposed system, underscoring the potential of fog computing for real-time IoT applications.

This study primarily focuses on the design and implementation of a fog-based architecture to enable efficient real-time data processing. As a natural progression, future work will explore the integration of machine learning techniques to improve the detection and prediction of available parking slots, enabling more accurate, real-time updates and personalised driver guidance. Expanding the system to cover larger metropolitan areas and incorporating a broader range of IoT devices, such as occupancy sensors and smart meters, will further enhance data collection and parking management efficiency. To improve the system’s environmental sustainability, future deployments should consider powering fog nodes and sensors with renewable energy sources, thereby reducing operational costs and carbon footprint. Additionally, while this work simulates the communication and performance impact of image-based data, it does not implement actual image recognition algorithms. Integrating computer vision models in future versions could enable the evaluation of processing overhead and detection accuracy, in tandem with latency and bandwidth metrics. Collectively, these advancements will further strengthen the role of fog computing as a secure, scalable, and energy-efficient infrastructure for real-time IoT applications in smart cities.