Abstract: To address the issues of poor generalization in downhole drilling image recognition and inaccurate drill pipe counting in coal mines, this study collected and annotated a dedicated dataset CMDPC_OBB, and proposed a method for drill pipe counting based on YOLOv11_OBB. The method consisted of two components: the YOLOv11_OBB-based drilling image recognition model and a scene-adaptive drill pipe counting algorithm. YOLOv11_OBB used rotated bounding boxes to accurately capture drilling images with inclined angles. By applying L2 regularization to improve the neck network of YOLOv11, weight fluctuation interference in feature fusion was reduced, ensuring more stable model training. The scene-adaptive counting algorithm tracked the motion trajectory between target drill pipes and key points at the drill rig's tail while employing multi-condition peak judgment to achieve automatic counting, thereby minimizing the impact of irrelevant pipes on accuracy. In addition, the image features learned by YOLOv11_OBB served as latent knowledge to guide the counting logic. Experimental results on the CMDPC_OBB dataset show that YOLOv11_OBB achieves an image recognition accuracy of 93.5%, outperforming YOLOv5L_OBB and YOLOv8L_OBB by 8.2% and 3.0%, respectively. The counting algorithm achieves an accuracy of 97.96%, with a model recognition speed of 34 FPS and a counting speed of 79 FPS, meeting real-time requirements.