2.2.1

The Power Supply Module.

This module mainly completes the conversion of AC power into a stable 12V DC power supply. The design point is that the module should support three-phase four-wire input.

2.2.2

ARM Processor.

The main point of the ARM processor design is to fully consider the communication method based on USB3.0 with the 5G module, and also to meet the computing power requirements of face recognition and helmet recognition.

2.2.3

Beidou Positioning Module.

This module mainly completes high-precision positioning and is used for “virtual electronic fence” positioning of dangerous areas on power infrastructure sites.

2.2.4

5G Communication Module.

It realizes the remote communication function between the 5G edge IoT proxy communication terminal and the cloud platform, constructs the panoramic multi-dimensional real-time perception of power infrastructure on-site monitoring, and the safety management and control of substation infrastructure on-site mobile operations.

2.2.5

Face Information Acquisition HD Camera Module.

This module collects the face information of the workers who want to enter the power infrastructure construction site, downloads the pre-entered photo information of the construction workers from the cloud platform, and performs real-time comparison on the site. After confirming the illegality, the device keeps the smart lock access control closed, and sends out sound and light alarm signals, and uploads the alarm events to the cloud platform in time through 5G.

2.2.6

Electronic Fence Area Positioning Camera Module.

This module realizes the electronic fence function of the dangerous area by locating and collecting the video stream information of the camera. When there is an illegal entry or when the construction worker is found not wearing a safety helmet, the device will issue an audible and visual alarm signal and report the event to the cloud platform.

2.2.7

Noise Sensor Module.

This module mainly collects the noise of the power infrastructure construction site in real time. When the noise exceeds the relevant set threshold, it will issue sound and light warnings, broadcast prompts to stop related operations, and upload the events of the noise exceeding the threshold to the cloud platform.

2.2.8

Small Weather Station Module.

It mainly monitors the temperature, humidity, wind speed, wind direction, air pressure and rainfall around the power infrastructure construction site in real time. When the relevant parameters exceed the threshold, for example, the wind speed exceeds the set threshold, the dangerous operation will be temporarily stopped through a broadcast notification, monitor the construction site’s various monitoring parameters within the safe range to continue the construction operation to ensure the safety of the power infrastructure.

2.2.9

Sound and Light Alarm Module.

This module mainly includes yellow light, red light and broadcast (horn) and other components, and performs functions such as prompting and alarming of related events.

2.2.10

Smart Lock Access Control Module.

5G edge IoT proxy communication terminal will open the smart lock after successful face recognition. When the face recognition fails for a certain number of times, it will be judged as illegal intrusion, the smart door lock will be kept closed, and the event will be reported to the cloud platform.

2.3.1

Gstreamer Processing Flow.

5G edge IoT proxy communication terminal is connected to the face information acquisition HD camera through Ethernet, the data stream is transmitted through the RTSP protocol, and the Gstreamer open-source framework is used to process the data stream. Gstreamer is a multimedia framework that supports cross-platform. The application program connects the various steps of multimedia processing in series through the pipeline (Pipeline) to achieve the desired effect. Each step is realized by means of plug-ins (plugins) based on the GObject object system through element (Element), which facilitates the expansion of various functions. We utilize Gstreamer to process RTSP data stream from webcam. First, use the rtph264depay plugin to parse out the H264 package. Second, use the decodebin plug-in to select the corresponding plug-in to decode according to the video stream encoding format. Third, use the jpegenc plug-in to encode the pictures in the video stream in jpg format. Finally, the data is output to the filesink terminal and we can get the photo. The Gstreamer processing flow is shown in Figure 2⁷.

Figure 2.

Gstreamer processing flow.

2.3.2

Face Recognition Module.

The face recognition module is based on the Dlib tool, which is a cross-platform general framework based on modern C++, covering machine learning, image processing, numerical algorithms, data compression, etc. The key code of Python is shown in Table 1. The face recognition processing flow is shown in Figure 3⁸, ⁹.

Figure 3.

Face recognition control unlocking flow.

Table 1.

The key code of Python for face recognition.

2.3.3

Face Recognition Control Unlocking and Alarm Reporting Platform.

The control unlocking program is compiled into a dynamic library, which will be called after the face recognition is successful in executing the unlocking function. Finally, the functions of the smart lock can be controlled by the face recognition of the on-site personnel of the power infrastructure. This can improve the work efficiency and prevent the illegal personnel from entering the construction site, and ensure the safety of the construction site of the electric power infrastructure¹⁰.

2.4.1

Prepare the Dataset File.

First, we prepare the dataset that needs to be labeled, create a new folder named VOC2007, and create a folder named JPEGImages in it to store the image files that need to be labeled. Next, we create a label file called Annotations to store the labels; then we use the labeling tool to label the files in the JPEGImages folder, and save the generated label file to the Annotations folder. Finally, we use the complete code to divide the converted data set into training set and validation set to train our own YOLOv5 model. The directory format is shown in Figure 4¹³.

Figure 4.

Prepare the dataset file.

2.4.2

Get Pretrained Weights.

In order to shorten the training time of the network and achieve better accuracy, pre-training weights are generally loaded for network training. YOLOv5 provides several pre-training weights, and we can choose different pre-training weights according to different needs. The larger the pre-training weight, the higher the training accuracy will be, but the slower the detection speed will be. The pre-training weight used for training our own dataset this time is YOLOv5s.pt, as shown in Figure 5¹⁴.

Figure 5.

Pretrained weights.

2.4.3

Modify Data Configuration Files and Model Configuration Files.

Once the pretrained model and dataset are ready, we can start training our own YOLOv5 object detection model. This training target detection model needs to modify the parameters in two yaml files, one is the corresponding yaml file in the data directory, and the other is the corresponding yaml file in the model directory file, as shown in Figures 6 and 7¹⁵.

Figure 6.

Modify data configuration files.

Figure 7.

Modify model configuration files.

2.4.4

Start Training.

The parameters required are modified to train our own model. Weights are the pre-training weight, cfg is the data configuration file, data is the model training file, and epochs are the training round. Here we specify 100 rounds, as shown in Figure 8.

Figure 8.

Start training.

2.4.5

Training Results.

After training, the weights will be saved in weights/best.pt under./runs/exp, and we can see the training results, as shown in Figure 9.

Figure 9.

Training results.