Risky Behavior Detection Under Electrical Pole Based on Improved YOLOv8

In order to reduce the occurrence of electric shock accidents caused by risky behaviors under electric poles every year, this paper proposes an improved method based on YOLOv8 to detect these risky behaviors. This method preserves the main network structure of YOLOv8 while adding a small object detection layer and the Focal-EIoU loss function to the model to enhance the detection performance. By incorporating the small object detection layer, the model is able to learn and analyze the features of images, identify small objects, and label their positions and categories. The Focal-EIoU helps adjust the weights of difficult-to-classify samples, thereby improving the focus on boundary ambiguity, small objects, and other challenging samples for classification. The proposed method is tested on a dataset constructed for risky behaviors under electric poles, and the results demonstrate an average precision (MAP) of 85.3%, indicating a 3.9% improvement over the original YOLOv8.