Abstract: During electric high-altitude work, it is essential to ensure correct and proper usage of safety belts by workers. This paper presents a method for detecting the wearing of electric high-altitude safety belts based on EPSA-YOLOv5. Initially, EPSANet is employed as the backbone extraction network to reduce parameters and accelerate model recognition speed while maintaining complete feature extraction. Subsequently, adjustments are made to the spatial pyramid pool structure to enhance model recognition accuracy. Finally, the recognition performance of YOLOv5, EPSANet, and EPSA-YOLOv5 target detection models is compared and analyzed using various evaluation metrics. The results demonstrate that the EPSA-YOLOv5 model achieves superior recognition performance with a recall rate of 0.9214 and an accuracy of 0.942