The breakthrough progress of deep learning in fields such as computer vision often relies on the support of massive labeled data. However, data annotation is not only time-consuming, but also costly and challenging task. Unsupervised Domain Adaptation (UDA), as a key technical in the field of transfer learning, provides a new research paradigm for solving cross domain generalization problems by constructing a knowledge transfer bridge between source and target domain. Although significant progress has been made in this technology, existing review studies still have shortcomings in terms of systematical and timeliness. To fill this gap, this paper conducts systematic research from three aspects: methodology, dataset, and application practice. Firstly, we conduct a comprehensive and systematic investigation of the existing UDA methods and provide a unified taxonomy framework. Secondly, we systematically reviewed three benchmark datasets and introduced the innovative applications of this technology in cutting-edge fields such as computer vision. Finally,
based on the analysis of existing work, we provide new perspectives and technical paths for future research directions in UDA.