Abstract: Conveyor belt is one of the most important transportation equipment in underground coal mines. Anchor rod, channel steel, large gangue, and other foreign objects that are mixed with the conveyor belt during conveying operations cause the belt to tear, clog the coal drop opening, and cause other serious safety incidents. They seriously affect the efficiency of transportation and even threaten the lives of workers. Aiming at the existing conveyor belt foreign object detection algorithm’s problems such as weak characterization of slender objects and poor weak semantic feature extraction, a foreign object detection algorithm based on low-level feature enhancement with Transformer is designed, notated as FET–YOLO. Firstly, to address the problem that existing detection networks have difficulty in extracting features of elongated objects, deformable convolution is introduced to enhance the network’s adaptability to the shape characteristics of elongated foreign objects, and the MobileViT module is used to increase the differentiation between the foreign objects and the background in the image, in order to extract features that are more consistent with the diversity of the elongated foreign objects and to weaken the interference of background noise.Secondly, constructing a low-level feature enhancement module LFEM, to improve the representation of weak semantic features of foreign objects in the detection network in order to reduce the probability of wrong detection.Finally, the introduction of gsconv reduces the information loss due to changes in the size of the feature map and ensures that the network extracts features efficiently while reducing the number of model parameters. The training set and validation set are produced by using the video of conveyor belt work in an underground coal mine. The proposed algorithm is compared to three other conveyor belt detection algorithms, and the experimental results show that the proposed algorithm can better solve the problems of poor detection of elongated objects and difficulty in weak semantic feature extraction in conveyor belt foreign object target detection, with higher detection accuracy. For images with a resolution size of 640×640, the performance metrics mAP@0.5 can be up to 0.875, mAP@0.5:0.95 can be up to 0.543, and the detection speed is 75 fps.