Abstract: Existing vision-based drill rod counting methods for underground coal mines mostly rely on single or local visual features. When state transitions such as drilling advance, drilling retreat, and short pauses occur during drilling operations, accompanied by common underground interferences including complex illumination, target occlusion, and target jitter, the counting features may exhibit abnormal fluctuations, leading to miscounting or missed counts. To address these issues, a drill rod counting method based on underground coal mine drilling rig operation state identification was proposed. First, a lightweight object detection model was constructed by introducing a Large Separable Kernel Attention (LSKA) module, a Generalized Efficient Layer Aggregation Network (GELAN), and a DynamicHead into the YOLOv11 model to stably detect key components of the drilling rig (the chuck and the gripper). Then, the DeepSORT algorithm was used to continuously track the chuck and the gripper, obtain the relative distance variation curve between them, and smooth the curve using Kalman filtering to suppress noise. Finally, trough positions were extracted from the smoothed curve; drilling advance and drilling retreat states were identified according to the spacing between adjacent troughs, and troughs under the drilling advance state were accumulated to achieve drill rod counting. Experimental results showed that the improved YOLOv11 model achieved an mAP@0.5 of 0.966 and accurately detected targets under complex conditions such as complex illumination, target occlusion, and target jitter. The proposed method achieved an average accuracy of 97.97%, meeting the accuracy requirements for automatic drill rod counting in underground environments.