Abstract: To address missed detections, false detections, and bounding box localization deviations in bolt foreign object detection on underground coal mine belt conveyors caused by low illumination, complex backgrounds, and slender target shapes, an improved YOLOv11n-based bolt foreign object detection method for coal mine belt conveyors was proposed. A Self-Calibrated Illumination Network (SCINet) was introduced at the input stage of YOLOv11n to enhance low-light images and improve the clarity of edge and texture details of bolt targets. A Large Selective Kernel Block (LSK Block) was introduced into the Bottleneck branch of the C3k2 module in the backbone network to construct the C3k2_LSK module and replace some traditional convolutions, enhancing the model's representation of the overall structural features of bolt targets and their spatial relationships with the background. The Inner-FocalerIoU loss function was used to optimize bounding box regression and improve localization accuracy for slender bolt targets. The experimental results showed that the precision, recall, mAP@0.5, and mAP@0.5:0.95 of the improved YOLOv11n reached 90.5%, 87.3%, 92.8%, and 62.1%, respectively, which were 1.2%, 2.1%, 0.7%, and 1.7% higher than those of the baseline YOLOv11n, respectively. The model frame rate reached 102.8 frames/s, meeting the real-time requirements of online bolt foreign object detection on underground coal mine belt conveyors. Compared with mainstream object detection models, the improved YOLOv11n model improved detection accuracy while maintaining good real-time performance. In scenarios involving low illumination, bolt inclination, high target-background similarity, and partial occlusion, the improved YOLOv11n model could stably detect bolt targets, with detection boxes closely matching the target regions, showing good adaptability to complex detection scenarios.