Abstract: The existing methods for detecting foreign objects in underground coal mine conveyor belts have poor adaptability to complex scenarios, cannot meet real-time and lightweight requirements, and perform poorly when dealing with foreign objects with large size differences. In order to solve the above problems, a Stair-YOLOv7-tiny model is proposed based on the lightweight YOLOv7-tiny model for improvement, and applied to the detection of foreign objects in coal mine underground conveyor belts. This model adds feature concatenation units to the efficient layer aggregation network (ELAN) module to form a Stair-ELAN module. The model fuses low dimensional features from different levels with high-dimensional features, strengthens the direct connection between feature levels, enhances information capture capabilities, and strengthens the model's adaptability to objects of different scales and complex scenes. The introduction of Stair-head feature fusion (Stair-fusion) for detection heads forms a Stair-head module. The model enhances the feature expression capability of medium and low resolution detection heads by fusing detection head features of different resolutions layer by layer, achieving complementary feature information. The experimental results show that the Stair-YOLOv7 tiny model has better detection performance than CBAM-YOLOv5, YOLOv7 tiny, and its lightweight model on the open-source dataset CUMT BelT for conveyor belt foreign objects. The accuracy, average precision, recall, and precision are 98.5%, 81.0%, 82.2%, and 88.4%, respectively, and the detection speed is 192.3 frames per second. In the video analysis of conveyor belt monitoring in a certain mine, the Stair-YOLOv7-tiny model does not have any missed or false detection, achieving accurate detection of foreign objects in the conveyor belt.