Abstract: In order to locate underground operators and various types of underground equipment in real time, find abnormal situations in time and take measures to prevent safety accidents, in view of the complex underground scenes of coal mines, the existence of a large number of small targets and the large-scale changes of tracking targets, an algorithm for tracking moving objects based on deep learning is proposed. Based on the Siamese tracking algorithms, a local-global matching network is proposed to improve the tracking accuracy of the algorithm in different scenarios, and an anchor-free classification-regression network is designed to reduce the computational complexity of the model. In order to verify the performance of the model, 40 video sequences were collected for testing. The test results show that the proposed algorithm can effectively improve the tracking performance of small targets, targets with large scale changes and targets in complex backgrounds, which is better than commonly used object tracking algorithms and achieves a real-time running speed of 57fps.