Abstract: To address the problems of low efficiency and high error rate of manual counting during drill pipe disassembly in underground coal mines, as well as the insufficient robustness of existing visual counting methods in complex environments such as low illumination and dust interference, an intelligent detection and counting method for coal mine drill pipes based on improved YOLO11, named YOLO11-DP (YOLO11-Drill Pipe), is proposed. In the detection stage, the lightweight multi-scale feature enhancement is carried out on the YOLO11 network structure. The lightweight EfficientNetV2 backbone is introduced, integrating the Mobile Inverted Bottleneck Convolution (MBConv) and Fused Mobile Inverted Bottleneck Convolution (Fused-MBConv) modules. A hierarchical attention mechanism is adopted: the efficient channel attention (ECA) is introduced in shallow layers to strengthen edge feature response, and the convolutional block attention module (CBAM) is applied in middle and high layers to enhance the focusing ability of target regions, so as to improve multi-scale feature extraction and detail capture capability. In the feature fusion stage, an improved path aggregation network-feature pyramid network (PAN-FPN) structure is used, in which depthwise separable convolution (DWConv) and residual modules replace standard convolution, effectively improving the model’s feature discrimination ability for densely arranged and occluded drill pipes. A rotated bounding box annotation strategy is designed to accurately describe drill pipe targets at arbitrary angles, and the angle-sensitive Kullback-Leibler Divergence (KLD) loss function is introduced to improve the precision of rotated object detection. K-means clustering is used to analyze the aspect ratio distribution of the dataset and optimize anchor box configuration, making the model more suitable for the slender geometric characteristics of drill pipes. In the counting stage, combined with the DeepSORT object tracking algorithm, a multi-condition joint decision mechanism is constructed based on three features: displacement, angle and velocity, which effectively suppresses counting errors caused by occlusion, trajectory interruption, motion blur and similar action interference. Field test results show that compared with the original YOLO11 model, the YOLO11-DP model can more accurately cover inclined drill pipes. After introducing all the improved strategies, the mAP@0.5 of the model in the drill pipe detection task is increased from 76.8% to 88.2%, and the counting error rate is reduced to 6.4%.