Improved edge lightweight YOLOv4 and its application in on-site power system work

A “cloud-edge-end” collaborative system architecture is adopted for real-time security management of power system on-site work, and mobile edge computing equipment utilizes lightweight intelligent recognition algorithms for on- site risk assessment and alert. Owing to its lightweight and fast speed, YOLOv4-Tiny is often deployed on edge computing equipment for real-time video stream detection; however, its accuracy is relatively low. This study proposes an improved YOLOv4-Tiny algorithm based on attention mechanism and optimized training methods, achieving higher accuracy without compromising the speed. Specifically, a convolution block attention module branch is added to the backbone network to enhance the feature extraction capability and an efficient channel attention mechanism is added in the neck network to improve feature utilization. Moreover, three optimized training methods: transfer learning, mosaic data augmentation, and label smoothing are used to improve the training effect of this improved algorithm. Finally, an edge computing equipment experimental platform equipped with an NVIDIA Jetson Xavier NX chip is established and the newly developed algorithm is tested on it. According to the results, the speed of the improved YOLOv4-Tiny algorithm in detecting on-site dress code compliance datasets is 17.25 FPS, and the mean average precision (mAP) is increased from 70.89% to 85.03%.