| Summary: | Accidents at substation sites have occurred frequently in recent years due to workers violating power safety regulations by not wearing hardhats. Therefore, it is necessary to provide real-time warnings when detecting workers without hardhats. Nevertheless, the deployment of deep learning-based algorithms necessitates the utilization of a multitude of parameters and computations, which consequently engenders an augmented expenditure on hardware. Therefore, using a lightweight backbone can address this issue well. This paper explored methods, such as deep learning, power Internet of Things (PIoT), and edge computing and proposed a lightweight and effective method called hardhat-YOLO for hardhat-wearing detection. First, the MobileNetv3-small backbone replaced the backbone of You Only Look Once (YOLO) v5s to reduce parameters and increase detection speed. In addition, the Convolutional Block Attention Module (CBAM) was effectively integrated into the network to improve detection precision. Finally, the hardhat-YOLO model, trained with a customized dataset, was transmitted to edge computing terminals in substations through PIoT for hardhat-wearing detection. Compared to the YOLOv5s model, the Parameters and Giga Floating Point Operations (GFLOPs) of the proposed model decreased by about 35.5% and 54.4%, respectively, and Frame per Second (FPS) increased by 17.3% approximately. The experimental results indicated that the hardhat-YOLO model achieved a Mean Average Precision of 83.3% at 50% intersection over union (mAP50), correctly and effectively conducting hardhat-wearing detection tasks. © (2024), Science and Information Organization. All Rights Reserved.
|
|---|