Coal mine belt conveyor is the main transportation tool of raw coal, and its normal operation is very important to the safe production of coal mine. In the process of belt transportation, the foreign matter such as large coal gangue bolt is the main cause of belt clamping and deviation tearing, which restricts the safe and efficient operation of belt conveyor.The key factors leading to the low accuracy of intelligent picking and detection of belt foreign objects are complex detection conditions such as the complex environment with high dust concentration and poor lighting in the mine, the fast speed of coal transport belts and the easy obstruction of foreign objects from each other The complex detection environment such as large coal dust underground, poor lighting, and fast coal transport belt speed are the key factors that lead to low detection accuracy of intelligent picking of foreign objects on the belt. Based on the actual monitoring data of an underground belt conveyor in a coal mine in Datong, Shanxi, this paper improves the YOLOv7 detection algorithm; firstly, through the adaptive contrast enhancement algorithm, the contrast of the belt monitoring image is enhanced, which is beneficial to improve the outline definition of the target image; secondly, in the trunk In the extraction network, a multi-scale mixed residual attention mechanism is proposed to enhance the ability of YOLOv7 to extract foreign object features and interfere with the background. At the same time, it introduces full-dimensional dynamic convolution to replace the ordinary convolution in the ELAN module, reducing the redundant data in the convolution. accumulation. Finally, a weighted bidirectional feature pyramid network and 4 detection heads are used to output model prediction results to improve the network's detection efficiency for foreign objects of different sizes. Compared with CPU GPU, TensorRt engine was used to convert the weight of the trained and improved YOLOv7 model and deploy it to the detection platform. Online detection was performed on the belt conveyor surveillance video with a resolution of 1920 1080 in coal mine to realize real-time monitoring of foreign bodies.Through experiments, the improved YOLOv7 model is superior to YOLOv5 and YOLOv7 in the recognition accuracy and speed of foreign objects in the downhole belt, and the recognition accuracy and speed of foreign objects in the downhole belt are 94.3% and 24fps, respectively. Compared with the YOLOv5 model and YOLOv7 model, the average recognition accuracy rate has increased by 4.6% and 3.8% respectively; the average recall rate has increased by 4.5% and 3.8% respectively; the detection time has been improved by 0.007s and 0.003s respectively.