Aiming at the issues of low statistical accuracy and performance degradation of existing coal mine underground drill pipe counting methods in low-light environments, this paper proposes a low-light image drill pipe counting method for coal mine underground based on improved YOLOv8. This method calculates the number of drill pipes by detecting the center point coordinates of the two prediction boxes of the drill chuck and the holder, drawing the spacing curve, and counting the peaks. Firstly, the SCI module is used for pre-processing low-light images to address issues such as uneven illumination and low contrast, ensuring that the subsequent model can extract more effective feature information. Secondly, the EMA attention mechanism is integrated into the C2f module in the backbone network to retain information from each channel and establish long- and short-term context dependencies, enhancing the focus on targets in low-light and complex backgrounds. Additionally, the BiFPN structure is used as the feature fusion method in the neck network to reduce feature information loss and enhance the network′s feature fusion capability, improving the detection accuracy of the model in low-light scenarios. Finally, the Inner-CIoU loss function is designed, based on auxiliary bounding box regression of different sizes, to enhance the model′s adaptability to low-light and noise. Experimental results show that the improved YOLOv8-GC algorithm achieves a 5.7% increase in mAP@0.5, with a detection speed of 151 fps; the drill pipe counting accuracy in low-light environments reaches 97.2%, fully demonstrating the effectiveness and application potential of the proposed improved algorithm.