With the progress of science and technology, the UAV market is expanding rapidly, but many UAV are “flying in the dark” and “flying indiscriminately”, which seriously affects the social security, so the management of UAV should not be delayed. At present, UAV identification and tracking have the problems of inaccurate target identification and insufficient extraction of UAV attitude features in the air. In this thesis, the YOLOv8 network is trained using Det-Fly UAV dataset, and the precision (B) value reaches 0.93 and the recall rate reaches 0.74 on the validation set, mAP@0.5 is 0.838, which indicates that the YOLOv8 model can accurately and efficiently recognize the UAV, and transfer the information such as coordinates to the DeepSort tracking network. DeepSort accepts the target information to track the target stably. However, the YOLOv8 training results show the problem of inaccurate capture of UAV image features leading to a fallback in the evaluation index and discontinuous tracking effect. For this problem the method of inserting the STN module into the YOLOv8 network architecture before the affine transformation of the UAV spatial changes is used, so that the subsequent YOLOv8 network architecture can learn the UAV features better, and the experimental results on the test set have a good enhancement to provide a good technical framework for the anti-UAV system.