Abstract: The coal flow of the belt conveyor will be mixed with anchor rod, angle iron, wood, gangue, and lump coal. This will easily lead to the tearing of the conveyor belt, the blockage of the transition and even the breakage of the belt. It is difficult for the inspection robot of the belt conveyor to efficiently and accurately recognize foreign objects in the environment of uneven lighting and high-speed running of the belt conveyor. The model deployment is inconvenient. The YOLOv7 model has a high capability to extract target features, but its recognition speed is slow. In order to solve the above problems, a foreign object recognition method of belt conveyor in coal mine based on improved YOLOv7 is proposed. The method of adaptive histogram equalization with limited contrast is used to enhance the collected monitoring image of the belt conveyor to improve the clarity of object contour in the image. The YOLOv7 model is improved by introducing a simple and parameter-free attention module into the backbone extraction network. The improved model can improve the model's anti-interference capability against the complex background of the image and the capability to extract foreign object features. The depthwise separable convolution is introduced to replace the ordinary convolution in the backbone feature extraction network to improve the speed of foreign object recognition. TensorRT engine is used to convert the improved YOLOv7 model after training and deploy it on NVIDIA Jetson Xavier NX, realizing the acceleration of the model. The video of the belt conveyor with the resolution of 1 920 × 1 080 in the underground coal mine is recognized. The experimental results show that the recognition effect of improved YOLOv7 is better than YOLOv5L and YOLOv7. The recognition accuracy rate is 92.8%, and the recognition speed is 25.64 frames/s, meeting the requirements of accurate and efficient recognition of foreign objects in the belt conveyor.