Abstract: To address the issues of personnel trespassing into hazardous zones and the need for safety helmet supervision at blasting sites, this paper proposed a real-time supervision system based on a deep learning algorithm. This system incorporated the convolutional block attention module （CBAM） mechanism and SPPFCSPC structure into YOLOv5 and introduced an MPDIoU loss function to optimize target box regression, thereby improving small-object detection accuracy, reducing computational cost, and enhancing target localization performance in complex scenes. Concurrently, the system integrated digital surface models with buffer zones to achieve precise detection of electronic fence breaches using the "ray-casting method". Based on PyQt5, it constructed a graphical user interface （GUI）, providing visual alerts that display real-time helmet detection results and boundary violation alarms. Experimental results show that compared with the basic YOLOv5 model, the improved algorithm increases detection precision to 93.6 %; the response time of the proposed model has been shown to outperform that of mainstream models such as YOLOv7, YOLOv8, and DETR. The efficacy and real-time performance of the system have been demonstrated in mining operations, satisfying the real-time supervision requirements for safety supervision at blasting sites.