Object Detection Related to Irregular Behaviors of Substation Personnel Based on Improved YOLOv4

The accurate and timely detection of irregular behavior substation personnel plays an important role in maintaining personal safety preventing power outage accidents. This paper proposes a method for behaviors (IBD) based on improved YOLOv4 which uses MobileNetV3 to replace the CSPDarkNet53 feature extraction network, depthwise separable convolution efficient channel attention (ECA) optimize SPP PANet networks, four scales maps fuse improve accuracy. First, image dataset was constructed using video data still photographs preprocessed by gamma correction method. Then, model trained combining Mosaic enhancement, cosine annealing, label smoothing skills. Several cases were carried out, experimental results showed that proposed has high accuracy, with mean average precision (mAP) 83.51%, as well fast speed, frames per second (FPS) 38.06 pictures/s. represents better performance than other object methods, including Faster RCNN, SSD, YOLOv3, YOLOv4. study offers reference IBD provides automated intelligent monitoring