Abstract: The control of gas in coal mining is a key factor in ensuring the safety of miners. Therefore, achieving automatic counting of drill rods during the drilling process is crucial for the intelligent construction of coal mines. To meet the demand for real-time monitoring of drill rod depth during drilling, a large-scale image dataset for coal mining drill bit identification and counting, CMDRC_OBB, was created, which collects monitoring videos from various underground coal mine scenarios. Addressing issues such as lighting changes, strong light interference, angle variation, as well as drill bit jamming and irregular drill rig movement, a real-time drill rod counting method, DPC_OBB_SPC, which integrates rotated object detection and saliency peak inference algorithms, is proposed. The method includes two stages: target detection recognition and drill rod counting inference. In the target detection recognition stage, data augmentation techniques are employed to enhance the model's generalization ability under lighting changes and varying viewpoints. The C2f-DG module is introduced to improve feature extraction efficiency, reduce parameters and computational load, and achieve model lightweighting. In the drill rod counting inference stage, the movement patterns of the drill rig tail are tracked by the target detection model from the first stage. Combined with the change in saliency peak values of key points and a state transition mechanism, the periodic statistics of the drill rig movement are used to achieve precise counting. Experimental results show that, compared to the original model, the improved target detection algorithm YOLOV8_OBB_DG achieved a 2.52% improvement in mAP@0.5, and a 4.75% improvement in mAP@0.5:0.95. Meanwhile, the number of parameters (Params) reduced from 3.1×10⁶ to 2.3×10⁶, and the computation load (FLOPs) decreased from 8.3×10⁹ to 7.0×10⁹. Finally, the proposed counting method DPC_OBB_SPC was tested on eight videos containing both drill-in and drill-out scenarios and compared with a simple peak counting method and a weighted peak counting method based on computer vision. The experimental results show that the DPC_OBB_SPC method achieved an average prediction accuracy (MNA) of 97.95%, significantly outperforming the other two algorithms with accuracies of 86.61% and 90.97%, and it maintained an average FPS of over 76 frames, fully meeting real-time requirements. The experiments fully validate the adaptability, accuracy, and robustness of the proposed drill rod counting method DPC_OBB_SPC in underground environments.