Abstract: Mine video surveillance plays a key role in ensuring the production safety of coal mining enterprises. At present, mine intelligent monitoring technology mainly processes monitoring data in the cloud, which has problems such as network congestion and high computing requirements. In response to this problem, the overall architecture of edge-cloud collaborative mine video monitoring was studied, and an adaptive video frame offloading strategy based on task offloading was proposed. The advantages of edge detection speed, high accuracy, and strong real-time performance were used to optimize and update the model. Realize the continuous improvement of edge-cloud collaboration architecture. In view of the low detection accuracy of the YOLOv5 model, and the deep network structure is prone to gradient disappearance and over-fitting, Transformer is applied to the visual field and faces the problem of multi-scale unrecognizability of the same target and the sequence of high-resolution images is too long and requires a large amount of calculation. Due to the problem of insufficient video memory resources, a target detection model based on Swin Transformer-YOLOv5 was constructed. Experimental results show that the target detection model based on ST-YOLOv5 improves the average detection accuracy and is suitable for the deployment of edge devices on mine intelligent working surfaces.