Abstract: The underground environment of coal mines is special, and the working scene of drilling rigs is complex and variable. There are often factors such as strong light, water vapor, and obstruction that interfere with the recognition effect of drilling rigs, which can easily lead to false or missed detections; At the same time, abnormal behaviors such as lagging and ineffective movement often occur at the tail of the drilling rig during operation, which affects the accuracy of drill Pipe Counting. A real-time counting method for coal mine drill pipes based on deep learning technology is proposed to address the above issues. It consists of two parts: the drill recognition model Drill-YOLOv8 optimized based on AM-NT and the drill pipe counting inference algorithm Pipe Count based on two-level judgment regions. Firstly, collect the working video of the drilling rig from the real monitoring footage of the coal mine underground, annotate and construct the Coal Mine drill Pipe Counting image dataset CMDPC; Then, an improved drilling rig recognition model Drill-YOLOv8 is constructed, which enhances the feature extraction ability of the backbone network for drilling rig targets by introducing triplet attention. The Slim-Neck network structure is improved to balance the accuracy and complexity of the model, and the DDA Head module is designed to replace the detection head structure to enhance the attention of the model head to the multi-dimensional information of drilling rig targets; Finally, a pipe count inference method called Pipe Count is designed based on the principle of calculating the effective number of movements at the tail of the drilling rig using a two-level polygon judgment area. Based on the results of the Drill-YOLOv8 detector, a two-level judgment area is automatically generated, and ByteTrack method is used to track the target at the tail of the drilling rig. By updating the motion status of the tail of the drilling rig through the judgment area, real-time counting of the drilling rig is indirectly achieved. Experiments on the CMDPC dataset show that the improved Drill-YOLOv8 model performs well in mAP@0.5 The mAP @ 0.5:0.95 index has increased by 3.0% and 2.7% respectively, effectively solving the problem of false detection and missed detection of drilling head and drill pipe targets under strong light, water vapor, and occlusion environments, and the detection speed has reached 86 frames per second; At the same time, the weighted average error rate of the counting inference algorithm Pipe Count is 2%, showing good robustness against multi scene data, and the processing speed reaches 40 frames per second, meeting real-time counting requirements.