The introduction of automated, digitized, and smart factors into parking systems can improve labor costs, user efficiency, and management. Among these improvements, smart factors are the most anticipated innovation. Equipped with network sensing and camera monitoring in the parking space, a smart parking management system (or so-called smart parking system) can transmit the on-site conditions instantly to the central control in order to sense various abnormalities, alleviate real-time problems, and guide vehicles, based on management viewpoints, to assist in the control of parking access ^[1]. Current smart parking management system trends employ innovative technologies to effectively manage and add value to the limited resources. Smart parking is an innovative parking technology that aims to use a smart strategy to deal with these limited resources more efficiently.

The smart parking system includes many basic functionalities: entrance and exit gate passage control, license plate recognition, movement line and parking space guidance, car search query, automatic billing, charging systems, etc. The system also has the ability to cooperate with the cloud and then be integrated as a cloud service ^[2]. All information is assembled on a central management platform, and more automated, digital, and intelligent functions are introduced through cloud-based intelligent computing and analysis services.

The most popular trend in smart parking systems presently is the use of image recognition for license plates. Its main purpose is to allow for control when entering and exiting the parking lot and to allow the driver to choose and pay when paying the toll. This not only makes driving more automated and convenient, but also greatly reduces the time needed for entering and exiting the traditional parking lot. In addition, with the concept of the Artificial Internet of Things (AIoT), the system can provide multiple advanced functionalities, including parking searching, reservation, payment, notifications, statistics, monitoring, and even context awareness ^[3]. Another direction for developing advanced functionalities is via the cloud. Cloud services provide enterprises with more diverse and scalable solutions. For instance, using cloud services for image recognition may allow the heavy usage of local computing resources to be avoided and reduce the need for large image storage.

2. Smart Parking System

The parking area of a large city often occupies 31% of its land use [7]^[4], so the management of parking areas affects the operation of a city enormously every day. If the driver could obtain parking information in real-time, they would be able to adjust their schedules and activities efficiently. Hence, many city officers are eager to introduce smart parking systems, hoping to create more convenient parking services and to help drivers to find parking spaces more effectively [8]^[5]. In general, the parking system has some common business problems that need to be solved; thus, it needs to be upgraded to become more automatic and intelligent. The common business problems are:

• The significant manual workload of parking lot inspection and ticketing;
• The inefficiency of parking space usage;
• Inability to allocate the optimal location and layout of parking space for new incoming cars [9]^[6].

Smart parking systems (SPSs), combined with the artificial intelligence Internet of Things (AIoT), have been widely used in Europe and North America. Smart parking systems (or smart parking management systems) refer to parking management systems that combine cloud computing technology, databases, and mobile devices in a parking environment [9]^[6]. In a smart parking system, a “parking space” is the basic logical unit of all service and management. Knowing the real-time status of all parking spaces can resolve many service problems. In most cases, a well-deployed sensor network can fulfill the requirement of real-time parking space detection. The construction of a sensor network for a parking system involves three different aspects: sensors, computing, and communications. The first is sensors, such as instruments, detectors/actuators, or monitors, that interact with the environment. These multiple sensors respond to changes in the environment and generate signal data. The advantage of using multiple sensors lies in multi-faceted perception; they complement each other to detect blind spots and can be used for sensor fusion to form a multi-dimensional data viewpoint with time series correction, which can be used for back-end spatial–temporal data mining and big data analysis to gain deeper insight. Compared with the complexity of the sensor network, using image recognition to detect the state of a parking space is simpler and more flexible. As long as some cameras are placed in the parking lot, an intelligent image analysis system can replace the role of the parking lot sensor. Like the popular facial recognition technology used by the surveillance system to control human access, license plate recognition is used to control cars’ access to the entrance of the parking lot in order to effectively control the entry and exit of vehicles. It can also be used in public places where trucks often enter and exit in the absence of supervision, such as communities or commercial buildings. Currently, a large number of high-definition cameras, such as 1440 p (4 MP), 1080 p (2 MP), and 720 p cameras, are widely used to fit security needs regarding camera resolution. Therefore, based on the availability of abundant high-quality images, the development of visual recognition technology using deep learning has gradually attracted attention in various fields of application. For instance, parking systems use wide-angle fisheye lenses or catadioptric cameras to realize visual recognition. This method uses a simple camera module to calibrate a space line without knowing its position and direction. It automatically extracts the parking space’s boundary line from the input wide-angle image, and can easily change its direction. The parking space boundary is used to divide the parking grid and to detect empty parking spaces through background subtraction [10,11]^[7][8]. However, visual recognition using a deep learning model with full-field camera identification relies heavily on CPU and GPU for real-time computing, so it is very dependent on hardware and electric power. Many systems use cloud image identification services to relieve the burden of the local computing requirements. However, the customization of cloud services is not easy, and the fees are expensive, which is disadvantageous in scenarios which heavily use computing in smart parking systems. Therefore, some studies have proposed an energy-saving algorithm architecture (such as YOLO), which would use agile object detection algorithms to detect the availability in parking lots using embedded artificial intelligence edge devices (such as TX2) [12]^[9].

3. Edage-Cloud-Dew Computing Architecture

The definition of dew computing and its applications are still under development and need to be further discussed [13]^[10]. Here, the researchers only list the published literature which does not involve business research. Dew computing works together with cloud services and controls edge/fog devices, including sensors, routers, and IoT devices. Initially, the cloud-dew architecture can be viewed as an extension of the existing client–server architecture, which forms the three-level communication link pipe (client–dew server–server) [14,15]^[11][12]. With the mediation of a dew server, a client can still access website services without an Internet connection, using a local form of an actual website. The reason for this is that the related data are stored in the dew server as a local copy, which is synchronized with the master copy in the cloud server. A duplicate (local copy) of the visited website on a user’s local computer is called a “Dew Site” [15]^[12]. Once such a dew site exists in the local machine, the user can browse the web content from the dew site and perform the necessary operations. Moreover, through the dew domain naming system, name mapping between different local dew sites has become possible [15]^[12]. As a result, novel concepts of services can be proposed, such as infrastructure-as-a-dew, a software-as-a-dew service, and software-as-a-dew products [15,16]^[12][13]. Although dew devices are often commodity computers (desktop, laptop, tablet, smartphone, etc.), with appropriate software support, they can provide functionality independently of cloud services and work as mediators between fog and cloud. Dew computing is an additional layer between end-user devices to process and coordinate with the Edge/Fog/Cloud layers, offering autonomy, independence, and collaboration features [17]^[14]. Dew computing, such as fog computing closer to the user, also seeks to optimize resource allocation before processing and before the analysis has been transferred into the cloud, hence reducing the complexity and improving the productivity of distributed computing architecture [18]^[15]. Dew computing provides a solution to organizing non-cloud components when enterprises are required to upgrade IT infrastructure with the cloud. Its main goal is to access a pool of raw data/metadata that can be rapidly created, edited, stored, and deleted offline with minimal internetwork management effort [14]^[11]. Dew computing cam accelerate computing services across devices. Hence, some research has applied dew computing to a 5G IoT coalition formation game [19]^[16]. Dew computing has two key features—independence of external systems and collaboration with cloud servers—and it works in two modes [4,20]^[17][18]:

• Localized mode: where all the services are provided within the internal local network perimeter;
• Global mode: where it functions as an intermediate device in the client–server cloud model.

In addition, the dew computing model is composed of six essential characteristics: rule-based data collection, synchronization, scalability, re-origination, transparency, and anytime/anyhow accessibility [15]^[12]. Cloud-dew computing architecture enables the personal cloud data to be continuously accessible by a new type of application without an Internet connection [14]^[11]. Some of the obvious examples of dew applications are Dropbox, OneDrive, and Google Drive Offline. Users can use these services regardless of their Internet connection, and they can be synchronized with cloud services [18]^[15]. Dew computing accelerates computing services across devices for 5G IoT using a coalition formation game. Dew computing speeds up computing services across devices and allows data output to be generated very quickly. To employ cloud-dew architecture on a local computer, a dew virtual machine is required (DVM). The DVM is an isolated environment for implementing the dew server on a local computer [15]^[12]. While a cloud server is considered as an orating cloud in the sky, a dew server is thought of as dew upon a plant leaf [4]^[17]. The dew server (DS) acts like the cloud service on the local computer, which can work in a closed environment without the use of external server systems [20]^[18]. When Internet is available, it interacts with and periodically synchronizes content with the cloud service. Dew servers store user data/metadata with capacities smaller than those of cloud servers. Generally, a dew server serves only one client with cloud services, and the database of the dew server takes care of the synchronization with the cloud database. Dew servers can be reinstalled with the help of cloud data backup and can be accessed independently of an Internet connection, as they run on local computers [18]^[15]. Based on the application field, Wang [4]^[17] proposed categories of dew computing, including Web in Dew (WiD), Storage in Dew (SiD), Database in Dew (DBiD), Software in Dew (SiD), Platform in Dew (PiD), Infrastructure as Dew (IaD), and Data in Dew (DiD) [4]^[17]. All application fields run on applications in a distributed and hierarchical environment, without requiring continuous intervention from a remote central control server [15]^[12]. In SiD, the client’s local computer has a cloud copy of the software ownership and settings. The existence of the software on the client’s local computer and the setting information recorded in the cloud service both reflect the ownership of the client. The client can re-download the software when necessary [4]^[17]. Apple’s App Store and Google Play are examples of SiD [21]^[19]. However, this use case does not exactly fit our situation. In Ray’s research [15]^[12], he proposed an extended version of the dew cloud definition. The local machine is comprised of four components: (1) the dew server, (2) the dew DBMS, (3) the dew client program, and (4) the dew client service application. Corresponding to the cloud computing service models, he presented the following viewpoints:

• He described the intersection of Software as a Service (SaaS) with the dew computing paradigm as Software-as-a-Dew Service (SaaDS) and Software-as-Dew Product (SaaDP). When one goes offline on SaaS, the other continually holds the business under the SaaS service model. Related data will be stored partly in cloud servers and partly in the local machine. The sew server can access and update the information stored in the local machine and also that synchronized with the cloud server.
• He also described the intersection of Infrastructure-as-a-Service (IaaS) with the dew computing paradigm as Infrastructure-as-a-Dew Service (IaaDS). The dew server ensures that the local infrastructure is safely recovered from the cloud when infrastructure-related data is lost or destroyed. Hence, the dew server must store all settings related to the infrastructure’s data as a backup copy into the cloud.

Like the traditional cognition on dew computing, Ray’s introduction to dew computing [15]^[12] focused on defining dew components in terms of an internal local network of a single machine. However, for some advanced applications, it is impractical and unreliable to use a single-client machine to deploy the entirety of the dew architecture, because enterprise business operations often rely on resources of the local area network (LAN). Specifically, for machine learning techniques, the image recognition service relies heavily on computing resources and data processing space. Hence, the dew architecture was deployed on the LAN-level, and the dew server was a real local server dedicated to performing dew service tasks. Furthermore, the position of the image recognition service here was the cloud API used remotely in smart parking applications. It, like Microsoft’s Cognitive Services APIs, is a cloud platform-based service (PaaS: Platform as a Service) that leverages the latest AI technologies in the cloud. When it is woven into dew computing architecture, it becomes a new type of category: ML API in Dew (MLiD) or Platform-as-a-Dew Service (PaaDS). This is not similar to the “an API for dew computing services” study [22]^[20], which provided services to other users by means of a local device. In the study by Mishra et al. [23]^[21], they proposed a dew-enabled vehicular fog computing model dedicated to dealing with the disaster challenge of latency-sensitive and computing-intensive tasks of vehicular networks. Their work consisted of a practical application, like the research focus, in parking lot management, and dew-enabled AIoT, which adapts Deep Q-Learning-based (FedDQL) techniques for the optimal offloading of tasks in a collaborative computing paradigm, was also facilitated in their architecture. However, due to different application domain, they emphasized fog computing on a vehicular network, which would take place further away from the sensors generating the data. On the other hand, in the application domain, the researchers emphasized the edge computing taking place directly on the AIoT device, which was physically connected to the camera monitoring sensors. We can refer to this as a new category—an ML model in dew (MMiD) or a Model-as-a-Dew Service (MaaDS). To clarify the aforementioned points, these related discussions are summarized in Table 1.

4. AIoT

For parking systems, a changing number of cars competing for a limited number of parking spaces is always a major concern. Through technological innovation, AIot (Artificial Intelligence of Things) presents a viable method for alleviating the inefficient usage of existing parking spaces. AIot, a combination of AI and IoT, helps the system to achieve ambient automation, which gives it self-healing and self-management properties, by leveraging IoT sensing and AI decision-making capabilities. IoT sensors collect spatial–temporal data and response information to the cloud service. In the cloud, the AI deep learning pipeline incrementally trains the model and uses prescriptive and predictive analytics for advanced applications. For instance, a smart pest monitoring system uses the environment and wireless image sensors installed inside greenhouses to monitor the population density of pests in hotspots [24]^[23]. Deep learning-based techniques for object segmentation, detection, and classification are used to analyze pest populations.