In the context of power construction, safety accidents resulting from unsafe behaviors among staff are common occurrences. In complex construction site scenarios, hidden safety hazards may arise when workers neglect to wear safety equipment, which poses a serious threat to their safety and lives [1]. Compliance with safety regulations, including the proper usage of safety helmets, safety belts, work clothes, and safety gloves, is of paramount importance in safeguarding the well-being of individuals engaged in construction work. Despite the existence of safety regulations, enforcement among workers often falls short, which poses challenges in ensuring safety at construction sites. Consequently, power grid enterprises face difficulties in monitoring and verifying workers’ compliance with these regulations. Identifying and warning workers about non-compliance is crucial to maintaining safety on construction sites.

In recent years, video surveillance cameras have been widely adopted in substation construction safety management systems [2]. However, traditional detection methods that rely on manual labor exhibit several drawbacks, such as high rates of missed detection, low efficiency, labor-intensiveness, and low reliability [3]. These drawbacks are largely attributed to the complex backgrounds and terrains of Yunnan province, where substation construction takes place. To address the challenges of slow transmission speed and lack of timely warning for violations, real-time detection of compliance wearing is necessary to prevent accidents. Deep learning-based object detection algorithms have been increasingly applied in recent years to replace manual monitoring of on-site safety regulations. These algorithms offer a more reliable and efficient way to ensure the safe operation of the power grid [4]. Furthermore, recent advances in drone and camera phone technology have made it possible to capture video sequences with moving cameras for detecting and tracking of moving objects [5]. These object detection algorithms rely on cloud platforms with substantial computing power or high-performance GPU computer clusters to achieve superior detection rates. However, GPU chips have limitations in terms of server power consumption, size, and portability, making them unsuitable for front-end scenarios. Therefore, low-cost edge devices offer a viable solution by simplifying hardware complexity and providing efficient and effective power construction safety detection in real-world settings.

At a construction site that requires significant power, video acquisition equipment is connected to the substation’s intelligent management system via a high-speed 4G or 5G network. This advanced technology allows the substation’s remote safety construction management and control system center to capture live video feeds and closely monitor every aspect of the construction process. However, environmental constraints such as incomplete 4G or 5G signal coverage and signal delays in certain field construction environments must be taken into consideration. For instance, Yunnan Province exhibits a complex topography, with mountains and plateaus comprising over 90% of its landscape. As a result, signal delays can easily trigger electric shock hazards and high-altitude accidents. Furthermore, traditional safety measures primarily focus on post-incident actions rather than proactive prevention.

Object detection is a crucial task in computer vision, and numerous methods and algorithms have been proposed to address this issue. However, most of these methods are designed for detecting and analyzing a single wearable object, which limits their effectiveness in the multi-specification wearable application environment of power systems. To improve the detection accuracy of targets in power systems and to more effectively locate illegal wearing, advanced target detection algorithms are often combined with traditional or binary classification algorithms. However, these models tend to be relatively complex and struggle to achieve a balance between speed and accuracy. Additionally, there has been a lack of emphasis on the development of real-time detectors for low-cost edge hardware, such as the Jetson Xavier NX.

2. Real-Time Intelligent Detection System for Illegal Wearing of On-Site Power Construction Worker Based on Edge-YOLO and Low-Cost Edge Devices

Wearing compliance on power construction sites is an important measure to ensure the safety of power grid construction. Previously, substation supervision mainly relied on manual monitoring. However, with the development of surveillance technology, human supervision is gradually being replaced by video surveillance. Nevertheless, video monitoring tasks still require manual observation of the camera system, leading to challenges in achieving real-time and accurate feedback due to the extensive time required for monitoring multiple surveillance cameras. In recent years, intelligent video analysis has emerged as a promising approach to replace traditional methods to support electric power enterprises in behavior recognition and prediction, employee safety, perimeter intrusion detection, and vandalism deterrence. To leverage the benefits of video monitoring systems for the aforementioned purposes, a computer vision-based automatic solution is necessary to enable real-time detection of the substantial volume of unstructured image data collected. Traditional object detection algorithms, such as ViolaJones detector (VJ.Det) ^[6][7], histogram of oriented gradient (HOG.Det) ^[7][8], and the deformable part-based mode ^[8][9], adopt window sliding and manual feature extraction methods. However, these methods have limitations in adapting to multiscale features and tend to exhibit low detection efficiency and accuracy. In recent years, deep learning-based object detection methods have been proposed and continuously applied, taking advantage of the ongoing development of deep learning. Therefore, many object detection algorithms based on deep learning have been introduced, including AlexNet ^[9][10] and residual learning ^[10][11]. Based on their approaches, current object detection methods using deep learning can be categorized into two-stage and one-stage object detection. The former includes faster R-CNN ^[11][12], FPN ^[12][13], Certernet ^[13][14], etc. While achieving high accuracy, the previously mentioned two-stage object detectors often exhibit low efficiency, making real-time detection challenging. In contrast, one-stage target detectors, such as the YOLO (You Only Look Once) series ^[14][15], RetinaNet ^[15][16], EfficientNet ^[16][17], etc., strike a balance between real-time detection and accuracy. Additionally, lightweight models are designed for mobile devices and other computing resource-constrained environments, such as drones and edge devices ^[17][18][18,19]. Among these models, SSDLite ^[19][20] is a typical architecture used in lightweight object detection. The MobileNet series ^[20][21] gradually improve the performance of the constructed model with depthwise separable convolutions, while Xception ^[21][22] is designed to enhance network performance without increasing network complexity. Moreover, several breakthroughs have been achieved in the field of object detection, particularly with the YOLO family of detectors. Many researchers have succeeded in reducing the size of YOLO models, enabling real-time detection. YOLO-LITE ^[22][23] offers a more efficient model for mobile devices, while YOLObile ^[23][24] introduces block-punched pruning and mobile acceleration using a collaborative scheme between mobile GPUs and CPUs. However, research on real-time object detection for safety construction monitoring is still in its early stages. YOLOv5s is an improvement of the YOLO series for lightweight networks, which provides a trade-off between detection speed and precision ^[24][25]. Yan et al. ^[25][26] proposed a system that combines remote substation construction management and artificial intelligence object detection techniques during construction in real-time based on YOLOv5s. Several researchers have conducted studies on related variants of YOLOv5 to improve the model’s performance. For example, Liao et al. ^[26][27] built the device components using a simple online and real-time tracking and counting algorithm on pruned YOLOv5. Liu et al. ^[27][28] improved YOLOv5n by optimizing the configuration of the target detector head and the network structure, solving the problems of low efficiency and redundant parameters in feature extraction in the model. Additionally, Xu et al. ^[28][29] proposed a target detection algorithm based on the YOLOv5 algorithm to address the issues of low accuracy and strong interference in existing safety helmet wearing detection algorithms and successfully improves the detection accuracy of safety helmets. The algorithm effectively demonstrated that by adding the SE (squeeze-and-excitation) block module to the YOLOv5 model, it is not only possible to obtain the weights of image channels but also to accurately separate the foreground and background of the image.