Transformer-based moving target tracking method for Unmanned Aerial Vehicle

Unmanned Aerial Vehicle (UAV) moving target tracking is one of the fundamental implementations in remote sensing and has been widely applied in monitoring, search and rescue, pursuit-escapes, and other fields. Currently, most UAV tracking algorithms merely establish the local relationship between the template and search region without fully using the global context information, leading to problems such as target loss and misclassification, and imprecise bounding boxes. This paper innovatively proposes a UAV tracker, TransUAV, overcoming the above challenge by a feature correlation network based on the self-attention mechanism. The method efficiently combines global features between the search region and the template to reduce the influence of external interference, enhancing the precision and robustness of the tracking algorithm. Moreover, the global spatio-temporal features are acquired by learning query embedding and temporal update strategies to make predictions, enhancing the adaptability to rapid changes in the appearance of target object. There is no proposal or predetermined anchor in this method to satisfy the requirements of onboard operational speed, therefore, no post-processing procedure is required, and the entire approach is end-to-end. The superiority of the proposed TransUAV is verified by an exhaustive evaluation of six challenging target tracking video datasets benchmarks, and the accuracy and robustness of the proposed TransUAV are compared with state-of-the-art methods.