The computer (Figure 2) is essential to the functioning of the system, accelerating data processing. It is equipped with Wi-Fi and Ethernet connectivity. The system is built with a sensor that collects information from the external environment in order to detect movement and capture images. The captured image is stored in the mini-computer and, depending on the device, will either be sent to the cloud to be processed with an artificial intelligence model for facial recognition in order to identify the person or processed locally with artificial intelligence models. Table 1 shows the characteristics of the mini-PC.

Figure 2. Lenovo ThinkCentre M700 Intel mini-PC.

Table 1. Specifications of the Lenovo ThinkCentre M700 Intel Mini PC.

In this project, we use an Arduino Uno board that is interconnected with components (Figure 3). It can be programmed for various tasks; in our case, to control three LEDs and a sensor. Table 2 summarizes all the connections established between the Arduino board, the proximity sensor, the PC and the LEDs.

Figure 3. Input and output connections to the Arduino board.

Table 2. Addressing of the electronic board.

The display interface (Figure 4) is a tool that displays data or information in a way that is understandable to the user. It is developed using the Python tkinter library and can be accessed either locally or remotely on the same network, depending on the prototype. The information displayed includes the surname, first name, date, arrival and departure times, and status, which manages the agent’s attendance, absence, and lateness times.

The first visualization interface stores data in an AWS cloud database, using a Python script developed with Tinker. This data is then displayed on the interface. The second interface stores data in a local WampServer database, which can be used to visualize data either through SQL queries or via a Tinker Python script, either locally or remotely on the same network. As for working hours, we have set a limit of 8 a.m. to 8:30 a.m.; beyond this range, the agent is considered late, and from 12 p.m. onwards, they are considered absent. Each interface has a menu with several options, such as attendance management, working hours, and daily, weekly, monthly, and overall reports (Figure 5), not to mention holiday management, all in real time for better organization.

Figure 4. Visualization interface.

Figure 5. Weekly report.

The first device has one camera, while the second has two extendable cameras. They are used to take photos if movement is detected in order to identify the person.

2.5.1. Photoelectric Sensor

This measures the distance between different objects, with a detection range of 5 to 30 cm. It is set to detect whether there is a person in the device.

2.5.2. Infrared Motion Detector

The heat emitted by an individual produces infrared rays. It is thanks to this infrared radiation that the sensor can detect the presence of intruders.

2.5.3. LED

There are three LEDs: red, green and yellow. Red indicates that the device is out of service, yellow that it is processing information and green that it is waiting for information.

In order to enhance the educational environment, an intelligent real-time monitoring system linked to cameras has been set up. The device uses two USB cameras, a Tinker Edge T, and algorithms. The Tinker Edge is equipped with Google Edge TPU, a machine learning (ML) accelerator that speeds up processing efficiency, reduces energy requirements, and facilitates the creation of connected devices and intelligent applications. Thanks to this integrated ML accelerator, the Tinker Edge T is capable of performing 4 tera-operations per second (TOPS) using only 0.5 watts per computing unit. It is also optimized for TensorFlow Lite models, which facilitates the compilation and execution of common ML models [19]. The object detection model used is the TensorFlow Lite API developed by Google AI Edge.

The device detects and identifies objects such as phones and computers, which are not allowed during exams, and sends an alert message with a screenshot as proof. It is also capable of counting the number of people in a room. At this stage, the system is being tested, and several improvements are planned for future versions.

The scoring device is a box divided into two compartments (Figure 6). The first

Figure 6. Overview of the scoring system.

part of the device is dedicated to taking photos. It has an illuminated opening that receives the photo captured by the camera. The second part, for control and command, is equipped with an Arduino board that manages Boolean inputs and outputs, indicating the activation of the system (photo capture, processing, and shutdown).

The first scoring system (Figure 7) was developed using Python scripts to manage the connection between the Arduino board, the photoelectric sensor, the three LEDs (green, yellow, and red), and the camera, and to send the data to the cloud server.

Figure 7. Architecture of the scoring system.

When a person is detected by the photoelectric sensor, the green LED turns yellow to indicate that the image capture program is active, then turns green again after capture. The image is saved with the date, hour, minutes, and seconds. The image name is saved in a file called enregistrement.txt, then the connection is checked. When the connection is established, the image is transmitted to the cloud for identity verification; if this fails, the image name is saved in the file erreur.txt.

For this system, the person must stand in front of the prototype and face the camera at a distance of 25 cm to ensure proper identification. The position of the device is important, as lighting is a key factor in obtaining a better photo. The captured image is then sent to the cloud for processing using AWS Rekognition, a deep learning-based image recognition service.

The cloud is used for data processing by AWS artificial intelligence. It enables savings to be made by pooling services across a large number of customers. According to a study by the American think tank Brookings Institution, governments could achieve savings of 20% to 25% on their IT budgets if they migrated to cloud computing [20]. The use of cloud computing also relieves the company of certain tasks: maintenance, security, and service upgrades are the sole responsibility of the service provider. These are generally performed better and faster than when they are the responsibility of the customer, especially when the customer is not an IT organization [21][22].

However, cloud computing also has its limitations, as the data collected is sent to other countries for storage and processing, which can lead to a loss of control over the information. Data access and management depend on the provider’s policies and tools, with a risk of dependence on the provider. There are also security issues, such as data confidentiality, online threats, and privacy violations, if the provider does not comply with strict standards. Finally, connection problems can make it difficult to send data to the cloud for processing. Control over personal data, therefore, remains a major challenge for privacy protection. Figure 8 shows the device as it was built, and Figure 9 shows an example of processing performed by the system.

Figure 8. First prototype.

Figure 9. Processing performed by the system.

After the first prototype based on a cloud-connected scoring system for identity verification, the second prototype (Figure 10) takes a local approach with two cameras positioned differently to improve the probability of individual recognition. The two images are verified to ensure they have been captured correctly, resized to the same height, combined side by side, and then saved with the date, hour, minutes, and seconds. The name of the image is saved in a file called enregistrement.txt. The officer concerned does not need to be present in front of the device to trigger the prototype, as the cameras are automatically activated as soon as they enter the motion detector’s field of action.

Figure 10. Scoring system architecture.

To process the collected data, a pre-trained minimum norm vector recognition model (dlib 128D descriptor) is used, which is open source locally. This model is built using dlib’s advanced facial recognition technology with deep learning. It achieves 99.38% accuracy on the Labeled Faces in the Wild benchmark [23]. Dlib is a modern C++ toolkit containing machine learning algorithms and tools for creating complex C++ software to solve real-world problems [24]. It stands out for its easy-to-use Python interfaces, which make it particularly accessible for integration into various projects. The model performs facial recognition by projecting each face into a 128-dimensional feature space. It then compares the representations obtained by calculating the Euclidean distance between them: if this distance is sufficiently small, the system concludes that the faces match.

All processing is performed locally from a single reference image. The position of the two cameras used to take the facial image is crucial, because if neither camera captures the face correctly, identification will be impossible. This is because the system relies entirely on capturing the face: if it is not photographed correctly or if the person is not within the cameras’ field of view, identification cannot be performed.

After the second prototype, which uses two cameras and local processing to enhance recognition reliability, the third device (Figure 11) simplifies the architecture by using only one camera while retaining the same processing model. The image is saved with the date, hour, minutes, and seconds. The image name is saved in a recording.txt file. The agent concerned must be present in front of the device to trigger the prototype and ensure identification. For data processing, the model applied in the second device was used. The information is recorded in a database that is entirely local with WampServer.

Figure 11. Scoring system architecture.

Most do not use any system or use a manual system. Among the six components, only 16.7% have a formal system that uses logbooks to record the surname, first name, arrival, and departure times for each employee. This indicates a low level of automation or digitization of the time and attendance system. All entities want a modern, biometric, and automatic solution that integrates an entry/exit or multi-clocking system. This reflects a significant need for reliability, strict monitoring, and security. Figure 12 illustrates the result obtained.

Figure 12. Types of clocking systems used.

The system is intended solely for managing administrative staff. Almost all structures have fewer than 50 employees; only the rectorate has more than 50 and fewer than 500 employees.

The data collected is personal information based on the photos, names, and identities of employees. Most UIDT organizations wish to maintain internal control over their data, but some are considering a hybrid approach, possibly for security or connectivity reasons. Two-thirds of entities (66.7%) prefer local data storage, while one-third (33.3%) opt for a hybrid solution combining local and cloud storage. Figure 13 shows these results.

We also compared cloud and local processing times on the university’s Wi-Fi network for 30 processes to obtain the average processing time for each environment. We noticed greater consistency in processing times with the local environment, which had a peak but did not exceed 50 seconds. Overall, local processing time ranged from 26 to 50 seconds. Cloud processing, on the other hand, varied much more, ranging from 7 to 70 seconds. This result can be explained by heavy traffic on the university’s Wi-Fi network during the test. The results are illustrated in Figure 14.

Figure 13. Preferred types of storage.

Figure 14. Local and cloud processing.