Abstract: With the rapid development of the autonomous driving and the virtual reality technologies, visual simultaneous localization and mapping (SLAM) has become a research hotspot in recent years. Three main problems of loop-closure detection of visual SLAM in complex environments are surveyed, i.e. place description, decision model, and evaluation of loop-closure detection. Firstly, the place description methods are introduced based on classical image features, deep learning, depth information and time-varying map, and the advantages and disadvantages of different methods are analyzed in detail. Secondly, some decision models are summarized which are commonly used in the process of loop recognition based on place description, especially for the probability model and the sequence matching. Thirdly, the performance evaluation method of loop-closure detection is explained, and its connection with the backend optimization is analyzed. Finally, the future directions that contribute to the development of loop-closure detection are discussed, focusing on several key points, such as deep learning, backend optimization and fusion of multiple descriptors.